opensource/coder
3
0
Fork 0
mirror of https://github.com/coder/coder.git synced 2026-10-28 02:42:54 +08:00
Code Issues Packages Projects Releases Wiki Activity
14,055 Commits 479 Branches 412 Tags
d8a030bb35bae317b8db682c12a44fa42d716623
Commit Graph
100 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Thomas Kosiewski c3794d54ac fix: avoid PTY for ssh command mode (#24862) 2026-05-01 15:02:05 +02:00
Thomas Kosiewski 0b9b94ae4e feat(site/src): add advisor admin settings UI (#24624) 
## Summary

Add the admin-only Advisor settings form on the Agents Behavior page, wired to the PR 3 HTTP API via TanStack Query. Admins can enable/disable the advisor, tune per-run caps and max output tokens, pick a reasoning effort, and optionally select a dedicated advisor model.

## Motivation

PR 3 introduced the configuration storage and API but no surface to change it. This PR puts the knobs in the existing admin Behavior UI so operators can turn the feature on/off and tune it without touching the database.

## Changes

- `site/src/pages/AgentsPage/components/AdvisorSettings.tsx`: the admin-only settings panel.
  - TanStack Query hooks (`chatAdvisorConfig`, `updateChatAdvisorConfig`) with optimistic cache invalidation on save.
  - Inputs for `Enabled`, `MaxUsesPerRun`, `MaxOutputTokens`, `ReasoningEffort`, `ModelConfigID`.
  - Model-override dropdown renders "Use chat model" when unset.
- `ZERO_UUID` / `isUnsetModelConfigId` helpers normalize the model-override payload in both directions: the backend serializes the unset sentinel as `"00000000-0000-0000-0000-000000000000"`, and the frontend must both recognize that sentinel on read and never send `""` on write (which would trigger a backend `invalid UUID length: 0`).
- Behavior page wiring so the panel appears in the existing settings flow alongside the rest of the agent behavior toggles.

## Stack context

This is **PR 6 of 6** in the advisor feature stack. It is stacked directly on **PR 3** (the config API) and is a sibling of PRs 4/5, so it can be reviewed independently of the chat-rendering work.

## Scope / non-goals

- Does not change advisor runtime behavior; that is PR 4's responsibility.
- No new audit entries for this settings change; it rides on the same storage mechanism as the existing site-configs pattern.

## Validation

- `pnpm --filter site lint:types`
- Manual dogfood against `./scripts/develop.sh`: toggle Enabled on/off, save, confirm `GET /api/experimental/chats/config/advisor` reflects changes, confirm non-admin users cannot reach the form.

---

<details>
<summary>📋 Implementation Plan (shared across the advisor stack)</summary>

# Plan: Add a Mux-style advisor tool to coder agents/chatd

## Outcome

Add a first-class `advisor` tool to agent chats in `coderd/x/chatd` that feels native to Coder:

- it is a built-in server-side tool, not an MCP/dynamic-tool workaround;
- it performs a nested **tool-less** model call for strategic advice;
- it is exposed only when eligible, and the prompt mentions it only when it is actually available;
- it is treated as a **planning-only** tool so it does not run alongside action tools in the same batch;
- it tracks usage/cost separately enough for operators to reason about it;
- it has a minimally polished UI in the Agents page;
- and it ships with explicit dogfooding evidence, including screenshots and repro videos.

## Design decisions to lock before coding

1. **Primary architecture:** native built-in tool in `chattool/`, backed by a small `chatadvisor` package.
2. **Nested model execution:** reuse chatd's existing model/provider stack for a one-step, tool-less advisor call rather than inventing a new provider pathway.
3. **Execution policy:** treat `advisor` as an exclusive/planning-only tool; mixed batches must return structured policy errors and force the model to retry cleanly.
4. **Availability:** initial rollout is for root agent chats only; disable for child/sub-agent chats until recursion/cost policy is proven.
5. **Prompt sync:** use one eligibility boolean to drive both tool registration and advisor guidance injection.
6. **Persistence/cost split:** MVP should keep advisor usage visible in result metadata and server metrics; only add DB schema if product/billing explicitly needs queryable advisor-specific cost.
7. **UI scope:** generic tool rendering is an acceptable temporary milestone during backend bring-up, but the release candidate should include a dedicated lightweight advisor renderer.

## Delivery model

The work should be executed as coordinated workstreams with one integration owner and parallel contributors for low-conflict areas. The integration owner should own `coderd/x/chatd/chatd.go` because prompt assembly, tool registration, and model resolution all converge there.

## Detailed workstreams

### Repo evidence used for this plan

<details>
<summary>Mux reference and current chatd seams</summary>

**Mux reference implementation**

- `src/node/services/tools/advisor.ts` — native advisor tool implementation.
- `src/common/constants/advisor.ts` — advisor prompt/constants and truncation policy.
- `src/common/utils/tools/tools.ts` — conditional tool registration.
- `src/node/services/streamContextBuilder.ts` — injects advisor guidance only when the tool is available.

**Current chatd seams**

- `coderd/x/chatd/chatd.go`
  - `processChat()` — tool assembly, prompt assembly, and chatloop invocation.
  - `resolveChatModel()` — current model/provider/key resolution seam.
  - `type Config struct` — server-level chatd configuration surface.
- `coderd/x/chatd/chatloop/chatloop.go`
  - `Run()` — main streaming/model loop.
  - `executeTools()` — built-in tool execution/batching seam.
- `coderd/x/chatd/chattool/` — built-in tool implementations.
- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx` — tool renderer dispatch.
- `site/src/pages/AgentsPage/components/ChatConversation/messageParsing.ts` and `ConversationTimeline.tsx` — tool/result merge and rendering flow.

</details>

### Workstream map and ownership

| Workstream | Primary owner | Main files | Can run in parallel? | Done when |
|---|---|---|---|---|
| 0. Integration + gating | Integration lead | `coderd/x/chatd/chatd.go` | No; central merge lane | Tool registration, prompt sync, and model selection are wired together |
| 1. Advisor runtime + tool | Backend agent | new `coderd/x/chatd/chatadvisor/`, new `coderd/x/chatd/chattool/advisor.go` | Yes | Tool can perform a tool-less advisor call in memory and return structured results |
| 2. Planning-only execution policy | Chatloop agent | `coderd/x/chatd/chatloop/chatloop.go`, related tests | Yes | Mixed `advisor` + action-tool batches are rejected cleanly and deterministically |
| 3. Metrics/usage/config | Backend/telemetry agent | `chatd.go`, `chatloop/metrics.go`, optional config plumbing | Partially; coordinate with integration lead | Advisor usage is separately visible in metadata/metrics and limits are enforced |
| 4. Frontend rendering | Frontend agent | `site/.../tools/Tool.tsx`, new `AdvisorTool.tsx`, stories | Yes after result schema stabilizes | Advisor renders as a readable card and story tests pass |
| 5. Dogfood + QA evidence | QA agent | dev server, Storybook, dogfood output | After backend + UI are usable | Repro videos, screenshots, and a concise QA report exist |

### Parallelization rules

- **Do not split `coderd/x/chatd/chatd.go` across multiple execution agents without an integration lead.** That file owns prompt building, tool registration, model resolution, and cost persistence.
- Workstreams 1 and 2 can be developed in parallel and then stacked onto the integration branch.
- Workstream 4 should begin once the backend result schema is agreed on, even if the backend is still behind a feature flag.
- Any agent that needs to re-check Mux behavior should clone `coder/mux` into a temporary directory (for example, `$(mktemp -d)/mux`) and inspect it read-only; do not vendor or copy code from Mux directly.

## Phase 0 — Preflight and guardrails

### Goals

- Align the team on the smallest shippable architecture.
- Prevent scope creep into MCP/dynamic-tool/sub-agent variants.
- Decide upfront what is MVP vs. follow-up.

### Tasks

1. **Confirm the MVP boundary.**
   - Ship a built-in advisor tool first.
   - Do **not** make MCP, dynamic tools, or sub-agents the primary implementation.
   - Do **not** add transient streaming phases in the first backend PR unless they fall out almost for free.

2. **Confirm local workflow hygiene before coding.**
   - Ensure the repo is using the project git hooks from `scripts/githooks`.
   - Do not bypass hooks with `--no-verify`.
   - Use `./scripts/develop.sh` for the full dev server rather than manual build/run commands.

3. **Lock the model-selection policy.**
   - **Recommended MVP:** advisor uses the same resolved provider/model/cost config as the current chat, with advisor-specific max-output and usage caps.
   - **Follow-up only if required:** add a separate `AdvisorModelConfigID`-style override that resolves through the existing `configCache`/model-config path. Do not invent a new free-form `provider:model` parser if chatd already stores provider/model separately.

4. **Lock the persistence policy.**
   - **Recommended MVP:** no DB migration. Persist advisor-visible metadata in the tool result and record separate metrics in memory/Prometheus.
   - **Only if product/billing explicitly asks for queryable advisor cost:** add a later DB migration or usage table, following the normal `queries/*.sql` + `make gen` workflow.

5. **Create an execution ADR note in the work item or tracking doc.**
   - Capture: built-in tool, tool-less nested call, root-chat-only rollout, exclusive execution policy, MVP no-DB-migration default.

### Quality gate

- Everyone on the team can state the same answers to these questions:
  - Is advisor a built-in tool? **Yes.**
  - Can advisor run with action tools in the same batch? **No.**
  - Does advisor get tools of its own? **No.**
  - Is a DB migration required for MVP? **No, unless billing insists.**

## Phase 1 — Build the advisor runtime and tool wrapper

### Goals

Create the core advisor implementation in a way that is easy to test and keeps `chattool/` thin.

### Files to add

- `coderd/x/chatd/chatadvisor/types.go`
- `coderd/x/chatd/chatadvisor/guidance.go`
- `coderd/x/chatd/chatadvisor/handoff.go`
- `coderd/x/chatd/chatadvisor/runtime.go`
- `coderd/x/chatd/chatadvisor/runner.go`
- `coderd/x/chatd/chattool/advisor.go`

### Responsibilities by file

1. **`types.go`**
   - Define the input/result schema used by the tool and UI.
   - Keep the result shape close to Mux so the UI and model both have predictable cases.
   - Recommended result variants:
     - `advice`
     - `limit_reached`
     - `error`

   Recommended shape:

   ```go
   type AdvisorArgs struct {
       Question string `json:"question"`
   }

   type AdvisorResult struct {
       Type          string              `json:"type"`
       Advice        string              `json:"advice,omitempty"`
       Error         string              `json:"error,omitempty"`
       AdvisorModel  string              `json:"advisor_model,omitempty"`
       RemainingUses int                 `json:"remaining_uses,omitempty"`
       Usage         *AdvisorUsageResult `json:"usage,omitempty"`
   }
   ```

2. **`guidance.go`**
   - Hold two strings:
     - the nested advisor system prompt;
     - the parent-agent guidance block to inject into the outer system prompt.
   - The nested advisor prompt must say, in plain language:
     - you are advising the parent agent;
     - you do not address the end user directly;
     - you do not claim actions happened;
     - you return concise strategic guidance and tradeoffs.

3. **`runtime.go`**
   - Define the per-run runtime state.
   - Recommended fields:
     - resolved model + model config;
     - provider keys/options reused from the outer chat;
     - `MaxUsesPerRun`;
     - `MaxOutputTokens`;
     - atomic/current call counter;
     - callback(s) to obtain the current prompt snapshot and current-step snapshot;
     - optional metrics/usage hook.
   - Add fail-fast validation for impossible config: nil model, non-positive limits, empty prompt builders, etc.

4. **`handoff.go`**
   - Build the advisor handoff message from:
     - the explicit question;
     - the exact prompt/messages the parent model just used;
     - the current step's text/reasoning snapshot, if available;
     - the most recent relevant tool outputs, if they are already in the prompt snapshot.
   - **Important:** use the already-prepared outer prompt tail, not a fresh DB reload. That keeps the advisor aligned with compaction and the exact context the outer model saw.
   - Apply hard truncation budgets with recent-context bias.

5. **`runner.go`**
   - Execute the nested advisor call.
   - **Recommended implementation:** call `chatloop.Run()` in an in-memory, one-step mode:
     - `Tools: nil`
     - `ProviderTools: nil`
     - `MaxSteps: 1`
     - `PersistStep`: capture the assistant output in memory instead of writing DB rows
   - Reuse the existing provider/model/cost path instead of building a second provider runner.
   - Assert that no tool definitions are passed to the nested call.

6. **`chattool/advisor.go`**
   - Keep this file thin and consistent with other built-ins.
   - Responsibilities:
     - decode `AdvisorArgs`;
     - validate `Question` is non-empty and bounded;
     - call the `chatadvisor` runner;
     - return a structured tool response.

### Defensive programming requirements

- Assert `Question` is non-empty after trimming.
- Assert runtime limits are positive.
- Assert the nested advisor call runs with zero tools/provider tools.
- Assert `AdvisorResult.Type` is one of the known variants before returning.
- Assert remaining uses never goes negative.

### Acceptance criteria

- A unit test can call the advisor tool with a fake model and receive a stable `advice` result.
- The nested advisor call is impossible to run with tools accidentally attached.
- The core logic lives in `chatadvisor/`, not embedded inside `chatd.go`.

## Phase 2 — Wire advisor into chatd and keep prompt/tool availability in sync

### Goals

Register the tool in the right place, expose it only when eligible, and inject system guidance only when the tool is present.

### Files to modify

- `coderd/x/chatd/chatd.go`
- optionally a small helper file if `chatd.go` becomes too crowded

### Tasks

1. **Compute one eligibility boolean in `processChat()`.**
   Recommended inputs:
   - server-level advisor enabled flag;
   - root chat only (`chat.ParentChatID == uuid.Nil` or equivalent existing root/child check);
   - a usable resolved model/provider exists;
   - optional experiment/workspace/org gate if product wants staged rollout.

2. **Create the runtime once per outer chat run.**
   - Use the model/config/keys resolved by `resolveChatModel()`.
   - Reuse provider options from the current chat's `ChatModelCallConfig`.
   - Set `MaxUsesPerRun` and `MaxOutputTokens` from advisor config defaults.

3. **Register the tool in the built-in tool block.**
   - Insert after the skill tools and before MCP tools in `processChat()`.
   - Record `builtinToolNames["advisor"] = true` so metrics stay bounded.

4. **Inject advisor guidance into the outer system prompt using the same boolean.**
   - Use `chatprompt.InsertSystem()` in the same prompt assembly path that already injects user/system instructions.
   - Place the block near the existing instruction insertion, before plan-path/skill context blocks.
   - Wrap the guidance in an explicit tag like `<advisor-guidance>` so it is easy to spot in tests and future refactors.

5. **Keep advisor out of child chats for the first release.**
   - That avoids recursion/cost blowups with `spawn_agent` / `wait_agent` flows.
   - Document this explicitly in the rollout notes and tests.

### Acceptance criteria

- If advisor is disabled, neither the tool nor the prompt guidance appears.
- If advisor is enabled, both the tool and the prompt guidance appear.
- Root chats can use advisor; child chats cannot.
- Built-in tool names include `advisor` so metrics do not collapse it into the generic `mcp` label.

## Phase 3 — Enforce planning-only execution policy in `chatloop`

### Goals

Prevent the model from calling `advisor` and action tools in the same execution batch.

### Files to modify

- `coderd/x/chatd/chatloop/chatloop.go`
- related chatloop tests

### Recommended implementation

Keep the MVP small; do **not** build a general policy engine yet.

1. Add a minimal field to `chatloop.RunOptions`, for example:

   ```go
   ExclusiveToolName *string
   ```

2. In `Run()` / `executeTools()`, detect the case where the exclusive tool appears in the same local-tool batch as any other locally executed tool.

3. When that happens, synthesize structured tool-result errors for the affected calls instead of executing anything in the batch.
   - `advisor` should receive a clear error like: _advisor must be called by itself before action tools_.
   - The sibling action tools should receive a paired policy error like: _this tool was skipped because advisor must run alone_.

4. Let the outer model see those tool errors and retry cleanly.
   - This is simpler and safer than partial execution or hidden deferral.
   - It preserves deterministic transcript history for debugging.

5. Pass the just-finished step snapshot into the tool execution context.
   - The advisor runtime should be able to see the current step's text/reasoning content, because that is often the best hint about what the outer model is trying to decide.

### Why this is the right fit

- It matches the intended semantics: advisor is consulted **before** taking action.
- It avoids subtle race conditions caused by concurrent built-in tool execution.
- It keeps the behavior easy to test with fake models.

### Acceptance criteria

- A model-emitted batch containing only `advisor` succeeds.
- A model-emitted batch containing `advisor` plus any other locally executed tool returns deterministic policy errors and executes nothing.
- Non-advisor tool execution stays unchanged for normal chats.

## Phase 4 — Usage limits, metrics, and configuration

### Goals

Make advisor safe to operate without over-designing billing/storage in the first release.

### Files to modify

- `coderd/x/chatd/chatd.go`
- `coderd/x/chatd/chatloop/metrics.go` as needed
- `coderd/x/chatd/chatd.go` `Config` struct and constructor path
- optional follow-up config/db files only if a separate advisor model or persistent billing is required

### Tasks

1. **Add explicit server config knobs for MVP.**
   Recommended fields on `chatd.Config` or a nested advisor config struct:
   - `AdvisorEnabled bool`
   - `AdvisorMaxUsesPerRun int`
   - `AdvisorMaxOutputTokens int64`

2. **Track usage per outer run.**
   - Reset the counter for each `processChat()` invocation.
   - Return `remaining_uses` in the tool result.
   - Return `limit_reached` when the cap is exhausted.

3. **Expose advisor usage metadata in the tool result.**
   - Include model name and token/cost summary if available.
   - Use the same `callConfig.Cost` calculation path as the outer chat for MVP if advisor reuses the same model.

4. **Record server-side metrics.**
   - Count advisor invocations, failures, and latency.
   - Ensure they show up under the built-in tool label `advisor`.

5. **Optional decision gate: separate advisor model.**
   - If product insists on a stronger/different advisor model, add a follow-up config hook that resolves another existing chat model config through the same `configCache` path.
   - Keep that out of the first landing PR unless it is required for acceptance.

6. **Optional decision gate: queryable advisor cost.**
   - If this becomes required, spin a follow-up DB task:
     - update `coderd/database/queries/*.sql`;
     - add migration files;
     - run `make gen`;
     - update audit mappings if a new auditable type/field is introduced.

### Acceptance criteria

- Advisor calls are capped per outer run.
- Limit exhaustion is user-visible in the tool result.
- Metrics distinguish advisor calls from other built-in tools.
- MVP does not require a schema migration unless explicitly approved.

## Phase 5 — Frontend rendering and Storybook coverage

### Goals

Make advisor feel intentional in the Agents UI without blocking the backend on fancy streaming UI.

### Files to modify

- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx`
- new `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`
- Storybook story file(s) in the same tools directory

### Delivery strategy

1. **Intermediate milestone during backend bring-up:** rely on the existing generic tool renderer if needed.
   - This is acceptable only as a short-lived integration checkpoint.

2. **Release milestone:** add a dedicated lightweight `AdvisorTool` renderer.
   - Reuse existing primitives:
     - `ToolCollapsible`
     - `ToolIcon`
     - `Response` for markdown/prose rendering
     - `ScrollArea` if the advice can be long
   - Keep styling light and consistent with the Agents page.
   - Do not add unnecessary React memoization in `site/src/pages/AgentsPage/`; that area is already React-Compiler aware.

3. **Render the structured result states cleanly.**
   - `advice` — readable prose/markdown with optional metadata footer.
   - `limit_reached` — warning-style message.
   - `error` — error state with visible fallback text.
   - `running` — existing tool loading state/spinner is enough for MVP.

4. **Add Storybook coverage instead of ad-hoc component tests.**
   Recommended stories:
   - successful advice;
   - running/loading;
   - limit reached;
   - error.

5. **Keep the UI contract narrow.**
   - Prefer one text field like `advice` plus small metadata rather than a deeply nested schema.
   - That keeps the UI resilient to prompt iteration.

### Acceptance criteria

- The advisor tool card renders readable content rather than raw quoted JSON in the final release branch.
- Running, limit, and error states are visibly distinct.
- Storybook stories and play assertions cover the new states.
- Existing tool rendering flows remain unchanged.

## Phase 6 — Automated tests and validation gates

### Backend tests to add

1. **Advisor runtime/tool tests**
   - question validation;
   - tool-less nested execution assertion;
   - success result shaping;
   - limit-reached result shaping;
   - error result shaping.

2. **Prompt/gating tests in chatd**
   - advisor disabled ⇒ no tool, no guidance;
   - advisor enabled/root chat ⇒ tool + guidance;
   - child chat ⇒ advisor absent.

3. **Chatloop policy tests**
   - advisor alone runs;
   - advisor + action tool mixed batch returns deterministic policy errors;
   - non-advisor tools still execute normally.

4. **Usage/metrics tests**
   - per-run cap resets correctly;
   - builtin tool labeling includes `advisor`;
   - returned metadata includes model/usage summary when available.

### Frontend tests to add

- Storybook `play()` assertions for the advisor renderer states.
- Verify expand/collapse behavior and visible fallback text.
- Verify the message timeline still renders adjacent tools correctly.

### Recommended command sequence

Run these as the implementation matures, not only at the end:

1. Backend-focused gate after phases 1–4:
   - `make test RUN=TestAdvisor`
   - `make test RUN=TestChatloopAdvisor`
   - `make lint`

2. Frontend-focused gate after phase 5:
   - `pnpm test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`
   - `pnpm lint`
   - `pnpm format`

3. Final repo gate before handoff:
   - `make pre-commit`
   - run any additional targeted `make test RUN=...` selections covering touched chatd paths

> Use the exact new test names the implementing agents create; the names above are recommended anchors, not existing tests.

## Dogfooding plan

### Principle

Dogfood the change as a real agent feature, not just a unit-tested backend. Per the dogfood and `agent-browser` skills, the reviewer should get **watchable repro videos** plus screenshots that make the behavior obvious without reading logs.

### Required setup

1. Start the full dev environment with:
   - `./scripts/develop.sh`
2. If the frontend renderer changes, also start Storybook from `site/` with:
   - `pnpm storybook --no-open`
3. Use `agent-browser` directly — **never `npx agent-browser`**.
4. Use named browser sessions and an output folder such as:
   - `./dogfood-output/advisor/`
   - with subfolders `screenshots/` and `videos/`

### Evidence protocol

For every interactive scenario below:

1. Start video recording **before** the action.
2. Capture step-by-step screenshots at human pace.
3. Capture one annotated screenshot of the final state.
4. Stop the recording.
5. Note the exact pass/fail observation in the QA report.

For static UI states (for example Storybook error/limit cards), an annotated screenshot is sufficient; video is optional but still encouraged by this project’s review preference.

### Dogfood scenarios

#### Scenario A — Happy path in the real Agents UI

**Goal:** prove that a root agent chat can invoke advisor and produce a readable recommendation before taking further action.

Steps:

1. Open the Agents page with an advisor-enabled root chat.
2. Start a repro video.
3. Send a prompt that should reasonably trigger strategic planning, such as an architecture or multi-tradeoff question.
4. Capture screenshots of:
   - the prompt before send;
   - the running advisor state;
   - the completed advisor card and the assistant’s follow-up response.
5. Stop recording.

Pass criteria:

- advisor appears in the timeline;
- the rendered result is readable;
- the assistant can continue after consuming the advisor output.

#### Scenario B — Advisor unavailable path

**Goal:** prove the feature is truly gated.

Suggested variants (at least one is required, both are better):

- feature flag/config off;
- child/sub-agent chat.

Evidence:

- annotated screenshot of the chat/tool state showing advisor is absent;
- short video if toggling the gate live is part of the repro.

Pass criteria:

- no advisor tool is available;
- no advisor-specific prompt behavior leaks through.

#### Scenario C — UI states in Storybook

**Goal:** prove the renderer handles non-happy states cleanly.

Required story states:

- success/advice;
- running;
- limit reached;
- error.

Evidence:

- one screenshot per state;
- at least one short video showing collapse/expand behavior.

Pass criteria:

- success renders readable advice;
- limit/error have visible fallback text;
- the component behaves like the other tool cards.

#### Scenario D — Regression sweep of nearby tools

**Goal:** ensure advisor does not break the surrounding chat timeline.

Check at minimum:

- another existing built-in tool still renders correctly near advisor;
- sub-agent/tool cards still expand/collapse normally;
- no obvious console errors appear in the Agents page during the advisor flow.

Evidence:

- screenshots of adjacent tool cards;
- console/error capture if anything suspicious appears.

### `agent-browser` usage notes for the QA agent

- Prefer `agent-browser batch` for 2+ sequential commands when no intermediate parsing is needed.
- Use `snapshot -i` to discover interactive refs.
- Re-snapshot after navigation or major DOM changes.
- Avoid `wait --load networkidle` unless the page is known to go idle; prefer explicit element/text waits or short fixed waits.
- Record videos at human pace and include pauses that a reviewer can follow.

## Rollout plan

### Initial rollout

- Gate behind a server-side advisor-enabled flag.
- Enable only for selected internal/root agent chats first.
- Watch metrics for:
  - invocation count;
  - failure rate;
  - latency;
  - obvious retry loops.

### Expansion conditions

Expand beyond the initial rollout only after the following are true:

- mixed-batch policy behavior is stable;
- cost impact is understood;
- frontend UX is readable in production-like dogfood;
- no recursion surprises have appeared with sub-agent flows.

### Explicit non-goals for the first release

- advisor inside child/sub-agent chats;
- provider-agnostic streaming phase UI;
- MCP-based external advisor implementation;
- mandatory DB-backed advisor cost reporting.

## Final acceptance checklist

- [ ] `advisor` is a built-in chatd tool, not an MCP/dynamic-tool substitute.
- [ ] The nested advisor call is tool-less and bounded to one in-memory step.
- [ ] One eligibility boolean controls both tool registration and prompt guidance injection.
- [ ] Root chats can use advisor; child chats cannot in the initial rollout.
- [ ] Mixed advisor/action batches produce deterministic policy errors instead of partial execution.
- [ ] Per-run usage caps and limit-reached behavior work.
- [ ] Advisor usage is visible in metadata/metrics without forcing a DB migration for MVP.
- [ ] The Agents UI has a readable advisor card and Storybook coverage.
- [ ] Dogfooding produced screenshots and repro videos for the required scenarios.
- [ ] Validation commands (`make lint`, targeted `make test`, Storybook tests, `make pre-commit`) passed before handoff.

## Suggested PR split

1. **PR 1 — Backend foundation**
   - `chatadvisor/` package
   - `chattool/advisor.go`
   - `chatloop` exclusive policy
   - chatd gating/prompt sync
   - backend tests

2. **PR 2 — Frontend + QA**
   - advisor renderer
   - stories/play assertions
   - dogfood artifacts and QA notes

3. **PR 3 — Optional follow-ups only if demanded by stakeholders**
   - separate advisor model override
   - persistent advisor billing/queryability
   - transient phase-stream UX


</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-7` • Thinking: `max`_
 2026-04-30 15:41:09 +02:00
Thomas Kosiewski 07e86ae565 feat(site/src/pages/AgentsPage): add advisor chat tool renderer (#24623) 
## Summary

Add a dedicated lightweight advisor chat card to the Agents page so advisor results render as readable prose instead of raw JSON. Includes Storybook coverage for the renderer's states.

## Motivation

Once PR 4 lands, the backend can emit advisor tool results, but the generic tool renderer shows them as quoted JSON. A small dedicated renderer turns the recommendation into something an operator or reviewer can actually read, without blocking the backend on full streaming UI.

## Changes

- `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`: renders the `AdvisorResult` states:
  - `advice` as markdown prose via the existing `Response` primitive,
  - `limit_reached` as a warning card,
  - `error` as an error card with fallback text,
  - `running` via the existing tool loading spinner.
  Reuses `ToolCollapsible`, `ToolIcon`, and `Response`. Question is shown only in the card subtitle (no duplicate label in the body).
- Metadata footer surfaces `advisor_model` and `remaining_uses` when present.
- `Tool.tsx` dispatcher routes the `advisor` tool to the new renderer.
- Storybook stories for each state (success/running/limit/error) with `play()` assertions for expand/collapse behavior.
- No new React memoization; the Agents page is already React-Compiler-aware.

## Stack context

This is **PR 5 of 6** in the advisor feature stack, stacked on top of PR 4.

## Scope / non-goals

- No changes to the backend schema.
- No changes to adjacent tool renderers.

## Validation

- `pnpm --filter site lint:types`
- `pnpm --filter site test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`

---

<details>
<summary>📋 Implementation Plan (shared across the advisor stack)</summary>

# Plan: Add a Mux-style advisor tool to coder agents/chatd

## Outcome

Add a first-class `advisor` tool to agent chats in `coderd/x/chatd` that feels native to Coder:

- it is a built-in server-side tool, not an MCP/dynamic-tool workaround;
- it performs a nested **tool-less** model call for strategic advice;
- it is exposed only when eligible, and the prompt mentions it only when it is actually available;
- it is treated as a **planning-only** tool so it does not run alongside action tools in the same batch;
- it tracks usage/cost separately enough for operators to reason about it;
- it has a minimally polished UI in the Agents page;
- and it ships with explicit dogfooding evidence, including screenshots and repro videos.

## Design decisions to lock before coding

1. **Primary architecture:** native built-in tool in `chattool/`, backed by a small `chatadvisor` package.
2. **Nested model execution:** reuse chatd's existing model/provider stack for a one-step, tool-less advisor call rather than inventing a new provider pathway.
3. **Execution policy:** treat `advisor` as an exclusive/planning-only tool; mixed batches must return structured policy errors and force the model to retry cleanly.
4. **Availability:** initial rollout is for root agent chats only; disable for child/sub-agent chats until recursion/cost policy is proven.
5. **Prompt sync:** use one eligibility boolean to drive both tool registration and advisor guidance injection.
6. **Persistence/cost split:** MVP should keep advisor usage visible in result metadata and server metrics; only add DB schema if product/billing explicitly needs queryable advisor-specific cost.
7. **UI scope:** generic tool rendering is an acceptable temporary milestone during backend bring-up, but the release candidate should include a dedicated lightweight advisor renderer.

## Delivery model

The work should be executed as coordinated workstreams with one integration owner and parallel contributors for low-conflict areas. The integration owner should own `coderd/x/chatd/chatd.go` because prompt assembly, tool registration, and model resolution all converge there.

## Detailed workstreams

### Repo evidence used for this plan

<details>
<summary>Mux reference and current chatd seams</summary>

**Mux reference implementation**

- `src/node/services/tools/advisor.ts` — native advisor tool implementation.
- `src/common/constants/advisor.ts` — advisor prompt/constants and truncation policy.
- `src/common/utils/tools/tools.ts` — conditional tool registration.
- `src/node/services/streamContextBuilder.ts` — injects advisor guidance only when the tool is available.

**Current chatd seams**

- `coderd/x/chatd/chatd.go`
  - `processChat()` — tool assembly, prompt assembly, and chatloop invocation.
  - `resolveChatModel()` — current model/provider/key resolution seam.
  - `type Config struct` — server-level chatd configuration surface.
- `coderd/x/chatd/chatloop/chatloop.go`
  - `Run()` — main streaming/model loop.
  - `executeTools()` — built-in tool execution/batching seam.
- `coderd/x/chatd/chattool/` — built-in tool implementations.
- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx` — tool renderer dispatch.
- `site/src/pages/AgentsPage/components/ChatConversation/messageParsing.ts` and `ConversationTimeline.tsx` — tool/result merge and rendering flow.

</details>

### Workstream map and ownership

| Workstream | Primary owner | Main files | Can run in parallel? | Done when |
|---|---|---|---|---|
| 0. Integration + gating | Integration lead | `coderd/x/chatd/chatd.go` | No; central merge lane | Tool registration, prompt sync, and model selection are wired together |
| 1. Advisor runtime + tool | Backend agent | new `coderd/x/chatd/chatadvisor/`, new `coderd/x/chatd/chattool/advisor.go` | Yes | Tool can perform a tool-less advisor call in memory and return structured results |
| 2. Planning-only execution policy | Chatloop agent | `coderd/x/chatd/chatloop/chatloop.go`, related tests | Yes | Mixed `advisor` + action-tool batches are rejected cleanly and deterministically |
| 3. Metrics/usage/config | Backend/telemetry agent | `chatd.go`, `chatloop/metrics.go`, optional config plumbing | Partially; coordinate with integration lead | Advisor usage is separately visible in metadata/metrics and limits are enforced |
| 4. Frontend rendering | Frontend agent | `site/.../tools/Tool.tsx`, new `AdvisorTool.tsx`, stories | Yes after result schema stabilizes | Advisor renders as a readable card and story tests pass |
| 5. Dogfood + QA evidence | QA agent | dev server, Storybook, dogfood output | After backend + UI are usable | Repro videos, screenshots, and a concise QA report exist |

### Parallelization rules

- **Do not split `coderd/x/chatd/chatd.go` across multiple execution agents without an integration lead.** That file owns prompt building, tool registration, model resolution, and cost persistence.
- Workstreams 1 and 2 can be developed in parallel and then stacked onto the integration branch.
- Workstream 4 should begin once the backend result schema is agreed on, even if the backend is still behind a feature flag.
- Any agent that needs to re-check Mux behavior should clone `coder/mux` into a temporary directory (for example, `$(mktemp -d)/mux`) and inspect it read-only; do not vendor or copy code from Mux directly.

## Phase 0 — Preflight and guardrails

### Goals

- Align the team on the smallest shippable architecture.
- Prevent scope creep into MCP/dynamic-tool/sub-agent variants.
- Decide upfront what is MVP vs. follow-up.

### Tasks

1. **Confirm the MVP boundary.**
   - Ship a built-in advisor tool first.
   - Do **not** make MCP, dynamic tools, or sub-agents the primary implementation.
   - Do **not** add transient streaming phases in the first backend PR unless they fall out almost for free.

2. **Confirm local workflow hygiene before coding.**
   - Ensure the repo is using the project git hooks from `scripts/githooks`.
   - Do not bypass hooks with `--no-verify`.
   - Use `./scripts/develop.sh` for the full dev server rather than manual build/run commands.

3. **Lock the model-selection policy.**
   - **Recommended MVP:** advisor uses the same resolved provider/model/cost config as the current chat, with advisor-specific max-output and usage caps.
   - **Follow-up only if required:** add a separate `AdvisorModelConfigID`-style override that resolves through the existing `configCache`/model-config path. Do not invent a new free-form `provider:model` parser if chatd already stores provider/model separately.

4. **Lock the persistence policy.**
   - **Recommended MVP:** no DB migration. Persist advisor-visible metadata in the tool result and record separate metrics in memory/Prometheus.
   - **Only if product/billing explicitly asks for queryable advisor cost:** add a later DB migration or usage table, following the normal `queries/*.sql` + `make gen` workflow.

5. **Create an execution ADR note in the work item or tracking doc.**
   - Capture: built-in tool, tool-less nested call, root-chat-only rollout, exclusive execution policy, MVP no-DB-migration default.

### Quality gate

- Everyone on the team can state the same answers to these questions:
  - Is advisor a built-in tool? **Yes.**
  - Can advisor run with action tools in the same batch? **No.**
  - Does advisor get tools of its own? **No.**
  - Is a DB migration required for MVP? **No, unless billing insists.**

## Phase 1 — Build the advisor runtime and tool wrapper

### Goals

Create the core advisor implementation in a way that is easy to test and keeps `chattool/` thin.

### Files to add

- `coderd/x/chatd/chatadvisor/types.go`
- `coderd/x/chatd/chatadvisor/guidance.go`
- `coderd/x/chatd/chatadvisor/handoff.go`
- `coderd/x/chatd/chatadvisor/runtime.go`
- `coderd/x/chatd/chatadvisor/runner.go`
- `coderd/x/chatd/chattool/advisor.go`

### Responsibilities by file

1. **`types.go`**
   - Define the input/result schema used by the tool and UI.
   - Keep the result shape close to Mux so the UI and model both have predictable cases.
   - Recommended result variants:
     - `advice`
     - `limit_reached`
     - `error`

   Recommended shape:

   ```go
   type AdvisorArgs struct {
       Question string `json:"question"`
   }

   type AdvisorResult struct {
       Type          string              `json:"type"`
       Advice        string              `json:"advice,omitempty"`
       Error         string              `json:"error,omitempty"`
       AdvisorModel  string              `json:"advisor_model,omitempty"`
       RemainingUses int                 `json:"remaining_uses,omitempty"`
       Usage         *AdvisorUsageResult `json:"usage,omitempty"`
   }
   ```

2. **`guidance.go`**
   - Hold two strings:
     - the nested advisor system prompt;
     - the parent-agent guidance block to inject into the outer system prompt.
   - The nested advisor prompt must say, in plain language:
     - you are advising the parent agent;
     - you do not address the end user directly;
     - you do not claim actions happened;
     - you return concise strategic guidance and tradeoffs.

3. **`runtime.go`**
   - Define the per-run runtime state.
   - Recommended fields:
     - resolved model + model config;
     - provider keys/options reused from the outer chat;
     - `MaxUsesPerRun`;
     - `MaxOutputTokens`;
     - atomic/current call counter;
     - callback(s) to obtain the current prompt snapshot and current-step snapshot;
     - optional metrics/usage hook.
   - Add fail-fast validation for impossible config: nil model, non-positive limits, empty prompt builders, etc.

4. **`handoff.go`**
   - Build the advisor handoff message from:
     - the explicit question;
     - the exact prompt/messages the parent model just used;
     - the current step's text/reasoning snapshot, if available;
     - the most recent relevant tool outputs, if they are already in the prompt snapshot.
   - **Important:** use the already-prepared outer prompt tail, not a fresh DB reload. That keeps the advisor aligned with compaction and the exact context the outer model saw.
   - Apply hard truncation budgets with recent-context bias.

5. **`runner.go`**
   - Execute the nested advisor call.
   - **Recommended implementation:** call `chatloop.Run()` in an in-memory, one-step mode:
     - `Tools: nil`
     - `ProviderTools: nil`
     - `MaxSteps: 1`
     - `PersistStep`: capture the assistant output in memory instead of writing DB rows
   - Reuse the existing provider/model/cost path instead of building a second provider runner.
   - Assert that no tool definitions are passed to the nested call.

6. **`chattool/advisor.go`**
   - Keep this file thin and consistent with other built-ins.
   - Responsibilities:
     - decode `AdvisorArgs`;
     - validate `Question` is non-empty and bounded;
     - call the `chatadvisor` runner;
     - return a structured tool response.

### Defensive programming requirements

- Assert `Question` is non-empty after trimming.
- Assert runtime limits are positive.
- Assert the nested advisor call runs with zero tools/provider tools.
- Assert `AdvisorResult.Type` is one of the known variants before returning.
- Assert remaining uses never goes negative.

### Acceptance criteria

- A unit test can call the advisor tool with a fake model and receive a stable `advice` result.
- The nested advisor call is impossible to run with tools accidentally attached.
- The core logic lives in `chatadvisor/`, not embedded inside `chatd.go`.

## Phase 2 — Wire advisor into chatd and keep prompt/tool availability in sync

### Goals

Register the tool in the right place, expose it only when eligible, and inject system guidance only when the tool is present.

### Files to modify

- `coderd/x/chatd/chatd.go`
- optionally a small helper file if `chatd.go` becomes too crowded

### Tasks

1. **Compute one eligibility boolean in `processChat()`.**
   Recommended inputs:
   - server-level advisor enabled flag;
   - root chat only (`chat.ParentChatID == uuid.Nil` or equivalent existing root/child check);
   - a usable resolved model/provider exists;
   - optional experiment/workspace/org gate if product wants staged rollout.

2. **Create the runtime once per outer chat run.**
   - Use the model/config/keys resolved by `resolveChatModel()`.
   - Reuse provider options from the current chat's `ChatModelCallConfig`.
   - Set `MaxUsesPerRun` and `MaxOutputTokens` from advisor config defaults.

3. **Register the tool in the built-in tool block.**
   - Insert after the skill tools and before MCP tools in `processChat()`.
   - Record `builtinToolNames["advisor"] = true` so metrics stay bounded.

4. **Inject advisor guidance into the outer system prompt using the same boolean.**
   - Use `chatprompt.InsertSystem()` in the same prompt assembly path that already injects user/system instructions.
   - Place the block near the existing instruction insertion, before plan-path/skill context blocks.
   - Wrap the guidance in an explicit tag like `<advisor-guidance>` so it is easy to spot in tests and future refactors.

5. **Keep advisor out of child chats for the first release.**
   - That avoids recursion/cost blowups with `spawn_agent` / `wait_agent` flows.
   - Document this explicitly in the rollout notes and tests.

### Acceptance criteria

- If advisor is disabled, neither the tool nor the prompt guidance appears.
- If advisor is enabled, both the tool and the prompt guidance appear.
- Root chats can use advisor; child chats cannot.
- Built-in tool names include `advisor` so metrics do not collapse it into the generic `mcp` label.

## Phase 3 — Enforce planning-only execution policy in `chatloop`

### Goals

Prevent the model from calling `advisor` and action tools in the same execution batch.

### Files to modify

- `coderd/x/chatd/chatloop/chatloop.go`
- related chatloop tests

### Recommended implementation

Keep the MVP small; do **not** build a general policy engine yet.

1. Add a minimal field to `chatloop.RunOptions`, for example:

   ```go
   ExclusiveToolName *string
   ```

2. In `Run()` / `executeTools()`, detect the case where the exclusive tool appears in the same local-tool batch as any other locally executed tool.

3. When that happens, synthesize structured tool-result errors for the affected calls instead of executing anything in the batch.
   - `advisor` should receive a clear error like: _advisor must be called by itself before action tools_.
   - The sibling action tools should receive a paired policy error like: _this tool was skipped because advisor must run alone_.

4. Let the outer model see those tool errors and retry cleanly.
   - This is simpler and safer than partial execution or hidden deferral.
   - It preserves deterministic transcript history for debugging.

5. Pass the just-finished step snapshot into the tool execution context.
   - The advisor runtime should be able to see the current step's text/reasoning content, because that is often the best hint about what the outer model is trying to decide.

### Why this is the right fit

- It matches the intended semantics: advisor is consulted **before** taking action.
- It avoids subtle race conditions caused by concurrent built-in tool execution.
- It keeps the behavior easy to test with fake models.

### Acceptance criteria

- A model-emitted batch containing only `advisor` succeeds.
- A model-emitted batch containing `advisor` plus any other locally executed tool returns deterministic policy errors and executes nothing.
- Non-advisor tool execution stays unchanged for normal chats.

## Phase 4 — Usage limits, metrics, and configuration

### Goals

Make advisor safe to operate without over-designing billing/storage in the first release.

### Files to modify

- `coderd/x/chatd/chatd.go`
- `coderd/x/chatd/chatloop/metrics.go` as needed
- `coderd/x/chatd/chatd.go` `Config` struct and constructor path
- optional follow-up config/db files only if a separate advisor model or persistent billing is required

### Tasks

1. **Add explicit server config knobs for MVP.**
   Recommended fields on `chatd.Config` or a nested advisor config struct:
   - `AdvisorEnabled bool`
   - `AdvisorMaxUsesPerRun int`
   - `AdvisorMaxOutputTokens int64`

2. **Track usage per outer run.**
   - Reset the counter for each `processChat()` invocation.
   - Return `remaining_uses` in the tool result.
   - Return `limit_reached` when the cap is exhausted.

3. **Expose advisor usage metadata in the tool result.**
   - Include model name and token/cost summary if available.
   - Use the same `callConfig.Cost` calculation path as the outer chat for MVP if advisor reuses the same model.

4. **Record server-side metrics.**
   - Count advisor invocations, failures, and latency.
   - Ensure they show up under the built-in tool label `advisor`.

5. **Optional decision gate: separate advisor model.**
   - If product insists on a stronger/different advisor model, add a follow-up config hook that resolves another existing chat model config through the same `configCache` path.
   - Keep that out of the first landing PR unless it is required for acceptance.

6. **Optional decision gate: queryable advisor cost.**
   - If this becomes required, spin a follow-up DB task:
     - update `coderd/database/queries/*.sql`;
     - add migration files;
     - run `make gen`;
     - update audit mappings if a new auditable type/field is introduced.

### Acceptance criteria

- Advisor calls are capped per outer run.
- Limit exhaustion is user-visible in the tool result.
- Metrics distinguish advisor calls from other built-in tools.
- MVP does not require a schema migration unless explicitly approved.

## Phase 5 — Frontend rendering and Storybook coverage

### Goals

Make advisor feel intentional in the Agents UI without blocking the backend on fancy streaming UI.

### Files to modify

- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx`
- new `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`
- Storybook story file(s) in the same tools directory

### Delivery strategy

1. **Intermediate milestone during backend bring-up:** rely on the existing generic tool renderer if needed.
   - This is acceptable only as a short-lived integration checkpoint.

2. **Release milestone:** add a dedicated lightweight `AdvisorTool` renderer.
   - Reuse existing primitives:
     - `ToolCollapsible`
     - `ToolIcon`
     - `Response` for markdown/prose rendering
     - `ScrollArea` if the advice can be long
   - Keep styling light and consistent with the Agents page.
   - Do not add unnecessary React memoization in `site/src/pages/AgentsPage/`; that area is already React-Compiler aware.

3. **Render the structured result states cleanly.**
   - `advice` — readable prose/markdown with optional metadata footer.
   - `limit_reached` — warning-style message.
   - `error` — error state with visible fallback text.
   - `running` — existing tool loading state/spinner is enough for MVP.

4. **Add Storybook coverage instead of ad-hoc component tests.**
   Recommended stories:
   - successful advice;
   - running/loading;
   - limit reached;
   - error.

5. **Keep the UI contract narrow.**
   - Prefer one text field like `advice` plus small metadata rather than a deeply nested schema.
   - That keeps the UI resilient to prompt iteration.

### Acceptance criteria

- The advisor tool card renders readable content rather than raw quoted JSON in the final release branch.
- Running, limit, and error states are visibly distinct.
- Storybook stories and play assertions cover the new states.
- Existing tool rendering flows remain unchanged.

## Phase 6 — Automated tests and validation gates

### Backend tests to add

1. **Advisor runtime/tool tests**
   - question validation;
   - tool-less nested execution assertion;
   - success result shaping;
   - limit-reached result shaping;
   - error result shaping.

2. **Prompt/gating tests in chatd**
   - advisor disabled ⇒ no tool, no guidance;
   - advisor enabled/root chat ⇒ tool + guidance;
   - child chat ⇒ advisor absent.

3. **Chatloop policy tests**
   - advisor alone runs;
   - advisor + action tool mixed batch returns deterministic policy errors;
   - non-advisor tools still execute normally.

4. **Usage/metrics tests**
   - per-run cap resets correctly;
   - builtin tool labeling includes `advisor`;
   - returned metadata includes model/usage summary when available.

### Frontend tests to add

- Storybook `play()` assertions for the advisor renderer states.
- Verify expand/collapse behavior and visible fallback text.
- Verify the message timeline still renders adjacent tools correctly.

### Recommended command sequence

Run these as the implementation matures, not only at the end:

1. Backend-focused gate after phases 1–4:
   - `make test RUN=TestAdvisor`
   - `make test RUN=TestChatloopAdvisor`
   - `make lint`

2. Frontend-focused gate after phase 5:
   - `pnpm test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`
   - `pnpm lint`
   - `pnpm format`

3. Final repo gate before handoff:
   - `make pre-commit`
   - run any additional targeted `make test RUN=...` selections covering touched chatd paths

> Use the exact new test names the implementing agents create; the names above are recommended anchors, not existing tests.

## Dogfooding plan

### Principle

Dogfood the change as a real agent feature, not just a unit-tested backend. Per the dogfood and `agent-browser` skills, the reviewer should get **watchable repro videos** plus screenshots that make the behavior obvious without reading logs.

### Required setup

1. Start the full dev environment with:
   - `./scripts/develop.sh`
2. If the frontend renderer changes, also start Storybook from `site/` with:
   - `pnpm storybook --no-open`
3. Use `agent-browser` directly — **never `npx agent-browser`**.
4. Use named browser sessions and an output folder such as:
   - `./dogfood-output/advisor/`
   - with subfolders `screenshots/` and `videos/`

### Evidence protocol

For every interactive scenario below:

1. Start video recording **before** the action.
2. Capture step-by-step screenshots at human pace.
3. Capture one annotated screenshot of the final state.
4. Stop the recording.
5. Note the exact pass/fail observation in the QA report.

For static UI states (for example Storybook error/limit cards), an annotated screenshot is sufficient; video is optional but still encouraged by this project’s review preference.

### Dogfood scenarios

#### Scenario A — Happy path in the real Agents UI

**Goal:** prove that a root agent chat can invoke advisor and produce a readable recommendation before taking further action.

Steps:

1. Open the Agents page with an advisor-enabled root chat.
2. Start a repro video.
3. Send a prompt that should reasonably trigger strategic planning, such as an architecture or multi-tradeoff question.
4. Capture screenshots of:
   - the prompt before send;
   - the running advisor state;
   - the completed advisor card and the assistant’s follow-up response.
5. Stop recording.

Pass criteria:

- advisor appears in the timeline;
- the rendered result is readable;
- the assistant can continue after consuming the advisor output.

#### Scenario B — Advisor unavailable path

**Goal:** prove the feature is truly gated.

Suggested variants (at least one is required, both are better):

- feature flag/config off;
- child/sub-agent chat.

Evidence:

- annotated screenshot of the chat/tool state showing advisor is absent;
- short video if toggling the gate live is part of the repro.

Pass criteria:

- no advisor tool is available;
- no advisor-specific prompt behavior leaks through.

#### Scenario C — UI states in Storybook

**Goal:** prove the renderer handles non-happy states cleanly.

Required story states:

- success/advice;
- running;
- limit reached;
- error.

Evidence:

- one screenshot per state;
- at least one short video showing collapse/expand behavior.

Pass criteria:

- success renders readable advice;
- limit/error have visible fallback text;
- the component behaves like the other tool cards.

#### Scenario D — Regression sweep of nearby tools

**Goal:** ensure advisor does not break the surrounding chat timeline.

Check at minimum:

- another existing built-in tool still renders correctly near advisor;
- sub-agent/tool cards still expand/collapse normally;
- no obvious console errors appear in the Agents page during the advisor flow.

Evidence:

- screenshots of adjacent tool cards;
- console/error capture if anything suspicious appears.

### `agent-browser` usage notes for the QA agent

- Prefer `agent-browser batch` for 2+ sequential commands when no intermediate parsing is needed.
- Use `snapshot -i` to discover interactive refs.
- Re-snapshot after navigation or major DOM changes.
- Avoid `wait --load networkidle` unless the page is known to go idle; prefer explicit element/text waits or short fixed waits.
- Record videos at human pace and include pauses that a reviewer can follow.

## Rollout plan

### Initial rollout

- Gate behind a server-side advisor-enabled flag.
- Enable only for selected internal/root agent chats first.
- Watch metrics for:
  - invocation count;
  - failure rate;
  - latency;
  - obvious retry loops.

### Expansion conditions

Expand beyond the initial rollout only after the following are true:

- mixed-batch policy behavior is stable;
- cost impact is understood;
- frontend UX is readable in production-like dogfood;
- no recursion surprises have appeared with sub-agent flows.

### Explicit non-goals for the first release

- advisor inside child/sub-agent chats;
- provider-agnostic streaming phase UI;
- MCP-based external advisor implementation;
- mandatory DB-backed advisor cost reporting.

## Final acceptance checklist

- [ ] `advisor` is a built-in chatd tool, not an MCP/dynamic-tool substitute.
- [ ] The nested advisor call is tool-less and bounded to one in-memory step.
- [ ] One eligibility boolean controls both tool registration and prompt guidance injection.
- [ ] Root chats can use advisor; child chats cannot in the initial rollout.
- [ ] Mixed advisor/action batches produce deterministic policy errors instead of partial execution.
- [ ] Per-run usage caps and limit-reached behavior work.
- [ ] Advisor usage is visible in metadata/metrics without forcing a DB migration for MVP.
- [ ] The Agents UI has a readable advisor card and Storybook coverage.
- [ ] Dogfooding produced screenshots and repro videos for the required scenarios.
- [ ] Validation commands (`make lint`, targeted `make test`, Storybook tests, `make pre-commit`) passed before handoff.

## Suggested PR split

1. **PR 1 — Backend foundation**
   - `chatadvisor/` package
   - `chattool/advisor.go`
   - `chatloop` exclusive policy
   - chatd gating/prompt sync
   - backend tests

2. **PR 2 — Frontend + QA**
   - advisor renderer
   - stories/play assertions
   - dogfood artifacts and QA notes

3. **PR 3 — Optional follow-ups only if demanded by stakeholders**
   - separate advisor model override
   - persistent advisor billing/queryability
   - transient phase-stream UX


</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-7` • Thinking: `max`_
 2026-04-30 15:17:32 +02:00
Thomas Kosiewski 17409a515c feat(coderd): wire advisor runtime to admin config (#24622) 
## Summary

Wire the advisor runtime into `chatd`: read the admin config on every `runChat`, gate tool registration and system-prompt guidance on a **single eligibility boolean**, register the `advisor` built-in tool, and apply the exclusive-tool policy from PR 1.

## Motivation

This is the integration seam where PRs 1–3 come together into an actual user-visible feature. Gating is deliberately root-chat-only for the initial rollout; child/sub-agent chats still do not see the tool or the guidance block.

## Changes

### `coderd/x/chatd/chatd.go`
- `loadAdvisorConfig(ctx, logger)` reads the admin config (from PR 3) on each run. If `ModelConfigID` is set, it resolves the override model via `configCache.ModelConfigByID`; otherwise it falls back to the outer chat's model and provider options. Reasoning effort is plumbed into provider options via `applyAdvisorReasoningEffort`.
- One computed `advisorEligible` boolean drives **both** tool registration (after skill tools, before MCP tools) and guidance injection via `chatprompt.InsertSystem(prompt, chatadvisor.ParentGuidanceBlock)`.
- `setAdvisorPromptSnapshot` closures capture the outer prompt state at the right points in the lifecycle (`renderPlanPathPrompt`, `ReloadMessages`, `PrepareMessages`) so the advisor handoff uses the same context the outer model saw.
- `ExclusiveToolNames["advisor"] = true` is passed to `chatloop.Run()` so mixed batches are rejected cleanly (PR 1 machinery).
- `builtinToolNames["advisor"] = true` so metrics keep advisor distinct from the generic `mcp` label.

### Child-chat guard
- Child/sub-agent chats deliberately do not see the advisor tool or guidance block, to avoid recursion/cost blowups until the pattern is proven. This is covered by `TestAdvisorGating_ChildChat` (currently skipped pending a rewrite against the new `plan`/`explore` subagent infrastructure; core gating logic is still exercised by `TestAdvisorGating_Disabled` and `TestAdvisorGating_RootChat`).

## Stack context

This is **PR 4 of 6** in the advisor feature stack. It depends on PRs 1–3.

## Scope / non-goals

- No frontend changes. The feature is invocable via the backend but renders generically until PR 5.
- No separate provider runner; the nested advisor call reuses the existing model/provider path.
- No DB migration.

## Validation

- `go test ./coderd/x/chatd/... -run TestAdvisor`
- `go build ./...`
- `make lint`

---

<details>
<summary>📋 Implementation Plan (shared across the advisor stack)</summary>

# Plan: Add a Mux-style advisor tool to coder agents/chatd

## Outcome

Add a first-class `advisor` tool to agent chats in `coderd/x/chatd` that feels native to Coder:

- it is a built-in server-side tool, not an MCP/dynamic-tool workaround;
- it performs a nested **tool-less** model call for strategic advice;
- it is exposed only when eligible, and the prompt mentions it only when it is actually available;
- it is treated as a **planning-only** tool so it does not run alongside action tools in the same batch;
- it tracks usage/cost separately enough for operators to reason about it;
- it has a minimally polished UI in the Agents page;
- and it ships with explicit dogfooding evidence, including screenshots and repro videos.

## Design decisions to lock before coding

1. **Primary architecture:** native built-in tool in `chattool/`, backed by a small `chatadvisor` package.
2. **Nested model execution:** reuse chatd's existing model/provider stack for a one-step, tool-less advisor call rather than inventing a new provider pathway.
3. **Execution policy:** treat `advisor` as an exclusive/planning-only tool; mixed batches must return structured policy errors and force the model to retry cleanly.
4. **Availability:** initial rollout is for root agent chats only; disable for child/sub-agent chats until recursion/cost policy is proven.
5. **Prompt sync:** use one eligibility boolean to drive both tool registration and advisor guidance injection.
6. **Persistence/cost split:** MVP should keep advisor usage visible in result metadata and server metrics; only add DB schema if product/billing explicitly needs queryable advisor-specific cost.
7. **UI scope:** generic tool rendering is an acceptable temporary milestone during backend bring-up, but the release candidate should include a dedicated lightweight advisor renderer.

## Delivery model

The work should be executed as coordinated workstreams with one integration owner and parallel contributors for low-conflict areas. The integration owner should own `coderd/x/chatd/chatd.go` because prompt assembly, tool registration, and model resolution all converge there.

## Detailed workstreams

### Repo evidence used for this plan

<details>
<summary>Mux reference and current chatd seams</summary>

**Mux reference implementation**

- `src/node/services/tools/advisor.ts` — native advisor tool implementation.
- `src/common/constants/advisor.ts` — advisor prompt/constants and truncation policy.
- `src/common/utils/tools/tools.ts` — conditional tool registration.
- `src/node/services/streamContextBuilder.ts` — injects advisor guidance only when the tool is available.

**Current chatd seams**

- `coderd/x/chatd/chatd.go`
  - `processChat()` — tool assembly, prompt assembly, and chatloop invocation.
  - `resolveChatModel()` — current model/provider/key resolution seam.
  - `type Config struct` — server-level chatd configuration surface.
- `coderd/x/chatd/chatloop/chatloop.go`
  - `Run()` — main streaming/model loop.
  - `executeTools()` — built-in tool execution/batching seam.
- `coderd/x/chatd/chattool/` — built-in tool implementations.
- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx` — tool renderer dispatch.
- `site/src/pages/AgentsPage/components/ChatConversation/messageParsing.ts` and `ConversationTimeline.tsx` — tool/result merge and rendering flow.

</details>

### Workstream map and ownership

| Workstream | Primary owner | Main files | Can run in parallel? | Done when |
|---|---|---|---|---|
| 0. Integration + gating | Integration lead | `coderd/x/chatd/chatd.go` | No; central merge lane | Tool registration, prompt sync, and model selection are wired together |
| 1. Advisor runtime + tool | Backend agent | new `coderd/x/chatd/chatadvisor/`, new `coderd/x/chatd/chattool/advisor.go` | Yes | Tool can perform a tool-less advisor call in memory and return structured results |
| 2. Planning-only execution policy | Chatloop agent | `coderd/x/chatd/chatloop/chatloop.go`, related tests | Yes | Mixed `advisor` + action-tool batches are rejected cleanly and deterministically |
| 3. Metrics/usage/config | Backend/telemetry agent | `chatd.go`, `chatloop/metrics.go`, optional config plumbing | Partially; coordinate with integration lead | Advisor usage is separately visible in metadata/metrics and limits are enforced |
| 4. Frontend rendering | Frontend agent | `site/.../tools/Tool.tsx`, new `AdvisorTool.tsx`, stories | Yes after result schema stabilizes | Advisor renders as a readable card and story tests pass |
| 5. Dogfood + QA evidence | QA agent | dev server, Storybook, dogfood output | After backend + UI are usable | Repro videos, screenshots, and a concise QA report exist |

### Parallelization rules

- **Do not split `coderd/x/chatd/chatd.go` across multiple execution agents without an integration lead.** That file owns prompt building, tool registration, model resolution, and cost persistence.
- Workstreams 1 and 2 can be developed in parallel and then stacked onto the integration branch.
- Workstream 4 should begin once the backend result schema is agreed on, even if the backend is still behind a feature flag.
- Any agent that needs to re-check Mux behavior should clone `coder/mux` into a temporary directory (for example, `$(mktemp -d)/mux`) and inspect it read-only; do not vendor or copy code from Mux directly.

## Phase 0 — Preflight and guardrails

### Goals

- Align the team on the smallest shippable architecture.
- Prevent scope creep into MCP/dynamic-tool/sub-agent variants.
- Decide upfront what is MVP vs. follow-up.

### Tasks

1. **Confirm the MVP boundary.**
   - Ship a built-in advisor tool first.
   - Do **not** make MCP, dynamic tools, or sub-agents the primary implementation.
   - Do **not** add transient streaming phases in the first backend PR unless they fall out almost for free.

2. **Confirm local workflow hygiene before coding.**
   - Ensure the repo is using the project git hooks from `scripts/githooks`.
   - Do not bypass hooks with `--no-verify`.
   - Use `./scripts/develop.sh` for the full dev server rather than manual build/run commands.

3. **Lock the model-selection policy.**
   - **Recommended MVP:** advisor uses the same resolved provider/model/cost config as the current chat, with advisor-specific max-output and usage caps.
   - **Follow-up only if required:** add a separate `AdvisorModelConfigID`-style override that resolves through the existing `configCache`/model-config path. Do not invent a new free-form `provider:model` parser if chatd already stores provider/model separately.

4. **Lock the persistence policy.**
   - **Recommended MVP:** no DB migration. Persist advisor-visible metadata in the tool result and record separate metrics in memory/Prometheus.
   - **Only if product/billing explicitly asks for queryable advisor cost:** add a later DB migration or usage table, following the normal `queries/*.sql` + `make gen` workflow.

5. **Create an execution ADR note in the work item or tracking doc.**
   - Capture: built-in tool, tool-less nested call, root-chat-only rollout, exclusive execution policy, MVP no-DB-migration default.

### Quality gate

- Everyone on the team can state the same answers to these questions:
  - Is advisor a built-in tool? **Yes.**
  - Can advisor run with action tools in the same batch? **No.**
  - Does advisor get tools of its own? **No.**
  - Is a DB migration required for MVP? **No, unless billing insists.**

## Phase 1 — Build the advisor runtime and tool wrapper

### Goals

Create the core advisor implementation in a way that is easy to test and keeps `chattool/` thin.

### Files to add

- `coderd/x/chatd/chatadvisor/types.go`
- `coderd/x/chatd/chatadvisor/guidance.go`
- `coderd/x/chatd/chatadvisor/handoff.go`
- `coderd/x/chatd/chatadvisor/runtime.go`
- `coderd/x/chatd/chatadvisor/runner.go`
- `coderd/x/chatd/chattool/advisor.go`

### Responsibilities by file

1. **`types.go`**
   - Define the input/result schema used by the tool and UI.
   - Keep the result shape close to Mux so the UI and model both have predictable cases.
   - Recommended result variants:
     - `advice`
     - `limit_reached`
     - `error`

   Recommended shape:

   ```go
   type AdvisorArgs struct {
       Question string `json:"question"`
   }

   type AdvisorResult struct {
       Type          string              `json:"type"`
       Advice        string              `json:"advice,omitempty"`
       Error         string              `json:"error,omitempty"`
       AdvisorModel  string              `json:"advisor_model,omitempty"`
       RemainingUses int                 `json:"remaining_uses,omitempty"`
       Usage         *AdvisorUsageResult `json:"usage,omitempty"`
   }
   ```

2. **`guidance.go`**
   - Hold two strings:
     - the nested advisor system prompt;
     - the parent-agent guidance block to inject into the outer system prompt.
   - The nested advisor prompt must say, in plain language:
     - you are advising the parent agent;
     - you do not address the end user directly;
     - you do not claim actions happened;
     - you return concise strategic guidance and tradeoffs.

3. **`runtime.go`**
   - Define the per-run runtime state.
   - Recommended fields:
     - resolved model + model config;
     - provider keys/options reused from the outer chat;
     - `MaxUsesPerRun`;
     - `MaxOutputTokens`;
     - atomic/current call counter;
     - callback(s) to obtain the current prompt snapshot and current-step snapshot;
     - optional metrics/usage hook.
   - Add fail-fast validation for impossible config: nil model, non-positive limits, empty prompt builders, etc.

4. **`handoff.go`**
   - Build the advisor handoff message from:
     - the explicit question;
     - the exact prompt/messages the parent model just used;
     - the current step's text/reasoning snapshot, if available;
     - the most recent relevant tool outputs, if they are already in the prompt snapshot.
   - **Important:** use the already-prepared outer prompt tail, not a fresh DB reload. That keeps the advisor aligned with compaction and the exact context the outer model saw.
   - Apply hard truncation budgets with recent-context bias.

5. **`runner.go`**
   - Execute the nested advisor call.
   - **Recommended implementation:** call `chatloop.Run()` in an in-memory, one-step mode:
     - `Tools: nil`
     - `ProviderTools: nil`
     - `MaxSteps: 1`
     - `PersistStep`: capture the assistant output in memory instead of writing DB rows
   - Reuse the existing provider/model/cost path instead of building a second provider runner.
   - Assert that no tool definitions are passed to the nested call.

6. **`chattool/advisor.go`**
   - Keep this file thin and consistent with other built-ins.
   - Responsibilities:
     - decode `AdvisorArgs`;
     - validate `Question` is non-empty and bounded;
     - call the `chatadvisor` runner;
     - return a structured tool response.

### Defensive programming requirements

- Assert `Question` is non-empty after trimming.
- Assert runtime limits are positive.
- Assert the nested advisor call runs with zero tools/provider tools.
- Assert `AdvisorResult.Type` is one of the known variants before returning.
- Assert remaining uses never goes negative.

### Acceptance criteria

- A unit test can call the advisor tool with a fake model and receive a stable `advice` result.
- The nested advisor call is impossible to run with tools accidentally attached.
- The core logic lives in `chatadvisor/`, not embedded inside `chatd.go`.

## Phase 2 — Wire advisor into chatd and keep prompt/tool availability in sync

### Goals

Register the tool in the right place, expose it only when eligible, and inject system guidance only when the tool is present.

### Files to modify

- `coderd/x/chatd/chatd.go`
- optionally a small helper file if `chatd.go` becomes too crowded

### Tasks

1. **Compute one eligibility boolean in `processChat()`.**
   Recommended inputs:
   - server-level advisor enabled flag;
   - root chat only (`chat.ParentChatID == uuid.Nil` or equivalent existing root/child check);
   - a usable resolved model/provider exists;
   - optional experiment/workspace/org gate if product wants staged rollout.

2. **Create the runtime once per outer chat run.**
   - Use the model/config/keys resolved by `resolveChatModel()`.
   - Reuse provider options from the current chat's `ChatModelCallConfig`.
   - Set `MaxUsesPerRun` and `MaxOutputTokens` from advisor config defaults.

3. **Register the tool in the built-in tool block.**
   - Insert after the skill tools and before MCP tools in `processChat()`.
   - Record `builtinToolNames["advisor"] = true` so metrics stay bounded.

4. **Inject advisor guidance into the outer system prompt using the same boolean.**
   - Use `chatprompt.InsertSystem()` in the same prompt assembly path that already injects user/system instructions.
   - Place the block near the existing instruction insertion, before plan-path/skill context blocks.
   - Wrap the guidance in an explicit tag like `<advisor-guidance>` so it is easy to spot in tests and future refactors.

5. **Keep advisor out of child chats for the first release.**
   - That avoids recursion/cost blowups with `spawn_agent` / `wait_agent` flows.
   - Document this explicitly in the rollout notes and tests.

### Acceptance criteria

- If advisor is disabled, neither the tool nor the prompt guidance appears.
- If advisor is enabled, both the tool and the prompt guidance appear.
- Root chats can use advisor; child chats cannot.
- Built-in tool names include `advisor` so metrics do not collapse it into the generic `mcp` label.

## Phase 3 — Enforce planning-only execution policy in `chatloop`

### Goals

Prevent the model from calling `advisor` and action tools in the same execution batch.

### Files to modify

- `coderd/x/chatd/chatloop/chatloop.go`
- related chatloop tests

### Recommended implementation

Keep the MVP small; do **not** build a general policy engine yet.

1. Add a minimal field to `chatloop.RunOptions`, for example:

   ```go
   ExclusiveToolName *string
   ```

2. In `Run()` / `executeTools()`, detect the case where the exclusive tool appears in the same local-tool batch as any other locally executed tool.

3. When that happens, synthesize structured tool-result errors for the affected calls instead of executing anything in the batch.
   - `advisor` should receive a clear error like: _advisor must be called by itself before action tools_.
   - The sibling action tools should receive a paired policy error like: _this tool was skipped because advisor must run alone_.

4. Let the outer model see those tool errors and retry cleanly.
   - This is simpler and safer than partial execution or hidden deferral.
   - It preserves deterministic transcript history for debugging.

5. Pass the just-finished step snapshot into the tool execution context.
   - The advisor runtime should be able to see the current step's text/reasoning content, because that is often the best hint about what the outer model is trying to decide.

### Why this is the right fit

- It matches the intended semantics: advisor is consulted **before** taking action.
- It avoids subtle race conditions caused by concurrent built-in tool execution.
- It keeps the behavior easy to test with fake models.

### Acceptance criteria

- A model-emitted batch containing only `advisor` succeeds.
- A model-emitted batch containing `advisor` plus any other locally executed tool returns deterministic policy errors and executes nothing.
- Non-advisor tool execution stays unchanged for normal chats.

## Phase 4 — Usage limits, metrics, and configuration

### Goals

Make advisor safe to operate without over-designing billing/storage in the first release.

### Files to modify

- `coderd/x/chatd/chatd.go`
- `coderd/x/chatd/chatloop/metrics.go` as needed
- `coderd/x/chatd/chatd.go` `Config` struct and constructor path
- optional follow-up config/db files only if a separate advisor model or persistent billing is required

### Tasks

1. **Add explicit server config knobs for MVP.**
   Recommended fields on `chatd.Config` or a nested advisor config struct:
   - `AdvisorEnabled bool`
   - `AdvisorMaxUsesPerRun int`
   - `AdvisorMaxOutputTokens int64`

2. **Track usage per outer run.**
   - Reset the counter for each `processChat()` invocation.
   - Return `remaining_uses` in the tool result.
   - Return `limit_reached` when the cap is exhausted.

3. **Expose advisor usage metadata in the tool result.**
   - Include model name and token/cost summary if available.
   - Use the same `callConfig.Cost` calculation path as the outer chat for MVP if advisor reuses the same model.

4. **Record server-side metrics.**
   - Count advisor invocations, failures, and latency.
   - Ensure they show up under the built-in tool label `advisor`.

5. **Optional decision gate: separate advisor model.**
   - If product insists on a stronger/different advisor model, add a follow-up config hook that resolves another existing chat model config through the same `configCache` path.
   - Keep that out of the first landing PR unless it is required for acceptance.

6. **Optional decision gate: queryable advisor cost.**
   - If this becomes required, spin a follow-up DB task:
     - update `coderd/database/queries/*.sql`;
     - add migration files;
     - run `make gen`;
     - update audit mappings if a new auditable type/field is introduced.

### Acceptance criteria

- Advisor calls are capped per outer run.
- Limit exhaustion is user-visible in the tool result.
- Metrics distinguish advisor calls from other built-in tools.
- MVP does not require a schema migration unless explicitly approved.

## Phase 5 — Frontend rendering and Storybook coverage

### Goals

Make advisor feel intentional in the Agents UI without blocking the backend on fancy streaming UI.

### Files to modify

- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx`
- new `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`
- Storybook story file(s) in the same tools directory

### Delivery strategy

1. **Intermediate milestone during backend bring-up:** rely on the existing generic tool renderer if needed.
   - This is acceptable only as a short-lived integration checkpoint.

2. **Release milestone:** add a dedicated lightweight `AdvisorTool` renderer.
   - Reuse existing primitives:
     - `ToolCollapsible`
     - `ToolIcon`
     - `Response` for markdown/prose rendering
     - `ScrollArea` if the advice can be long
   - Keep styling light and consistent with the Agents page.
   - Do not add unnecessary React memoization in `site/src/pages/AgentsPage/`; that area is already React-Compiler aware.

3. **Render the structured result states cleanly.**
   - `advice` — readable prose/markdown with optional metadata footer.
   - `limit_reached` — warning-style message.
   - `error` — error state with visible fallback text.
   - `running` — existing tool loading state/spinner is enough for MVP.

4. **Add Storybook coverage instead of ad-hoc component tests.**
   Recommended stories:
   - successful advice;
   - running/loading;
   - limit reached;
   - error.

5. **Keep the UI contract narrow.**
   - Prefer one text field like `advice` plus small metadata rather than a deeply nested schema.
   - That keeps the UI resilient to prompt iteration.

### Acceptance criteria

- The advisor tool card renders readable content rather than raw quoted JSON in the final release branch.
- Running, limit, and error states are visibly distinct.
- Storybook stories and play assertions cover the new states.
- Existing tool rendering flows remain unchanged.

## Phase 6 — Automated tests and validation gates

### Backend tests to add

1. **Advisor runtime/tool tests**
   - question validation;
   - tool-less nested execution assertion;
   - success result shaping;
   - limit-reached result shaping;
   - error result shaping.

2. **Prompt/gating tests in chatd**
   - advisor disabled ⇒ no tool, no guidance;
   - advisor enabled/root chat ⇒ tool + guidance;
   - child chat ⇒ advisor absent.

3. **Chatloop policy tests**
   - advisor alone runs;
   - advisor + action tool mixed batch returns deterministic policy errors;
   - non-advisor tools still execute normally.

4. **Usage/metrics tests**
   - per-run cap resets correctly;
   - builtin tool labeling includes `advisor`;
   - returned metadata includes model/usage summary when available.

### Frontend tests to add

- Storybook `play()` assertions for the advisor renderer states.
- Verify expand/collapse behavior and visible fallback text.
- Verify the message timeline still renders adjacent tools correctly.

### Recommended command sequence

Run these as the implementation matures, not only at the end:

1. Backend-focused gate after phases 1–4:
   - `make test RUN=TestAdvisor`
   - `make test RUN=TestChatloopAdvisor`
   - `make lint`

2. Frontend-focused gate after phase 5:
   - `pnpm test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`
   - `pnpm lint`
   - `pnpm format`

3. Final repo gate before handoff:
   - `make pre-commit`
   - run any additional targeted `make test RUN=...` selections covering touched chatd paths

> Use the exact new test names the implementing agents create; the names above are recommended anchors, not existing tests.

## Dogfooding plan

### Principle

Dogfood the change as a real agent feature, not just a unit-tested backend. Per the dogfood and `agent-browser` skills, the reviewer should get **watchable repro videos** plus screenshots that make the behavior obvious without reading logs.

### Required setup

1. Start the full dev environment with:
   - `./scripts/develop.sh`
2. If the frontend renderer changes, also start Storybook from `site/` with:
   - `pnpm storybook --no-open`
3. Use `agent-browser` directly — **never `npx agent-browser`**.
4. Use named browser sessions and an output folder such as:
   - `./dogfood-output/advisor/`
   - with subfolders `screenshots/` and `videos/`

### Evidence protocol

For every interactive scenario below:

1. Start video recording **before** the action.
2. Capture step-by-step screenshots at human pace.
3. Capture one annotated screenshot of the final state.
4. Stop the recording.
5. Note the exact pass/fail observation in the QA report.

For static UI states (for example Storybook error/limit cards), an annotated screenshot is sufficient; video is optional but still encouraged by this project’s review preference.

### Dogfood scenarios

#### Scenario A — Happy path in the real Agents UI

**Goal:** prove that a root agent chat can invoke advisor and produce a readable recommendation before taking further action.

Steps:

1. Open the Agents page with an advisor-enabled root chat.
2. Start a repro video.
3. Send a prompt that should reasonably trigger strategic planning, such as an architecture or multi-tradeoff question.
4. Capture screenshots of:
   - the prompt before send;
   - the running advisor state;
   - the completed advisor card and the assistant’s follow-up response.
5. Stop recording.

Pass criteria:

- advisor appears in the timeline;
- the rendered result is readable;
- the assistant can continue after consuming the advisor output.

#### Scenario B — Advisor unavailable path

**Goal:** prove the feature is truly gated.

Suggested variants (at least one is required, both are better):

- feature flag/config off;
- child/sub-agent chat.

Evidence:

- annotated screenshot of the chat/tool state showing advisor is absent;
- short video if toggling the gate live is part of the repro.

Pass criteria:

- no advisor tool is available;
- no advisor-specific prompt behavior leaks through.

#### Scenario C — UI states in Storybook

**Goal:** prove the renderer handles non-happy states cleanly.

Required story states:

- success/advice;
- running;
- limit reached;
- error.

Evidence:

- one screenshot per state;
- at least one short video showing collapse/expand behavior.

Pass criteria:

- success renders readable advice;
- limit/error have visible fallback text;
- the component behaves like the other tool cards.

#### Scenario D — Regression sweep of nearby tools

**Goal:** ensure advisor does not break the surrounding chat timeline.

Check at minimum:

- another existing built-in tool still renders correctly near advisor;
- sub-agent/tool cards still expand/collapse normally;
- no obvious console errors appear in the Agents page during the advisor flow.

Evidence:

- screenshots of adjacent tool cards;
- console/error capture if anything suspicious appears.

### `agent-browser` usage notes for the QA agent

- Prefer `agent-browser batch` for 2+ sequential commands when no intermediate parsing is needed.
- Use `snapshot -i` to discover interactive refs.
- Re-snapshot after navigation or major DOM changes.
- Avoid `wait --load networkidle` unless the page is known to go idle; prefer explicit element/text waits or short fixed waits.
- Record videos at human pace and include pauses that a reviewer can follow.

## Rollout plan

### Initial rollout

- Gate behind a server-side advisor-enabled flag.
- Enable only for selected internal/root agent chats first.
- Watch metrics for:
  - invocation count;
  - failure rate;
  - latency;
  - obvious retry loops.

### Expansion conditions

Expand beyond the initial rollout only after the following are true:

- mixed-batch policy behavior is stable;
- cost impact is understood;
- frontend UX is readable in production-like dogfood;
- no recursion surprises have appeared with sub-agent flows.

### Explicit non-goals for the first release

- advisor inside child/sub-agent chats;
- provider-agnostic streaming phase UI;
- MCP-based external advisor implementation;
- mandatory DB-backed advisor cost reporting.

## Final acceptance checklist

- [ ] `advisor` is a built-in chatd tool, not an MCP/dynamic-tool substitute.
- [ ] The nested advisor call is tool-less and bounded to one in-memory step.
- [ ] One eligibility boolean controls both tool registration and prompt guidance injection.
- [ ] Root chats can use advisor; child chats cannot in the initial rollout.
- [ ] Mixed advisor/action batches produce deterministic policy errors instead of partial execution.
- [ ] Per-run usage caps and limit-reached behavior work.
- [ ] Advisor usage is visible in metadata/metrics without forcing a DB migration for MVP.
- [ ] The Agents UI has a readable advisor card and Storybook coverage.
- [ ] Dogfooding produced screenshots and repro videos for the required scenarios.
- [ ] Validation commands (`make lint`, targeted `make test`, Storybook tests, `make pre-commit`) passed before handoff.

## Suggested PR split

1. **PR 1 — Backend foundation**
   - `chatadvisor/` package
   - `chattool/advisor.go`
   - `chatloop` exclusive policy
   - chatd gating/prompt sync
   - backend tests

2. **PR 2 — Frontend + QA**
   - advisor renderer
   - stories/play assertions
   - dogfood artifacts and QA notes

3. **PR 3 — Optional follow-ups only if demanded by stakeholders**
   - separate advisor model override
   - persistent advisor billing/queryability
   - transient phase-stream UX


</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-7` • Thinking: `max`_
 2026-04-30 15:07:33 +02:00
Thomas Kosiewski 06bad73df4 feat: add admin-configurable advisor API, SDK, and queries (#24621) 
## Summary

Add the **admin-configurable advisor configuration**: database-backed storage, SDK types, and the experimental HTTP handlers that back the admin settings UI (later PRs). Follows the same "site-configs" pattern as Virtual Desktop.

## Motivation

The advisor needs runtime-tunable knobs (enable/disable, per-run cap, max output tokens, reasoning effort, optional model override) without a service restart or redeploy. Using the existing `site_configs` K/V table keeps this pattern consistent with other admin features and avoids a bespoke schema.

## Changes

### Database (`coderd/database/queries/siteconfig.sql`)
- `GetChatAdvisorConfig` returns the stored JSON blob (default `'{}'`) under key `agents_advisor_config`.
- `UpsertChatAdvisorConfig` uses the standard `INSERT ... ON CONFLICT` pattern.
- Regenerated via `make gen` (queries.sql.go + mocks).

### SDK (`codersdk/chats.go`)
- `AdvisorConfig` type with `Enabled`, `MaxUsesPerRun`, `MaxOutputTokens`, `ReasoningEffort` (`""` / `low` / `medium` / `high`), `ModelConfigID uuid.UUID`.
- Client methods: `ChatAdvisorConfig(ctx)` / `UpdateChatAdvisorConfig(ctx, cfg)`.

### API (`coderd/exp_chats.go`)
- `GET /api/experimental/chats/config/advisor`: reads current config; relies on `ActorFromContext` validation.
- `PUT /api/experimental/chats/config/advisor`: requires `policy.ActionUpdate` on `rbac.ResourceDeploymentConfig`.
- Handlers unmarshal `{}` to a typed zero value and re-marshal on upsert for schema stability.
- Tests in `exp_chats_test.go` cover empty defaults, round-trip update, unauthorized update, and invalid body.

## Stack context

This is **PR 3 of 6** in the advisor feature stack. Consumed by:
- PR 4 (`feat/advisor-04-chatd-runtime`), which reads this config on every `runChat`.
- PR 6 (`feat/advisor-06-admin-settings-ui`), which renders the admin form.

## Scope / non-goals

- No `chatd` read path (lands in PR 4).
- No UI (lands in PR 6).
- `agents_advisor_config` remains a single-row JSON blob; we intentionally do not shard per-org/per-template yet.

## Validation

- `make gen`
- `go test ./coderd/database/... -run TestChatAdvisor`
- `go test ./coderd/... -run TestChatAdvisorConfig`
- `make lint`

---

<details>
<summary>📋 Implementation Plan (shared across the advisor stack)</summary>

# Plan: Add a Mux-style advisor tool to coder agents/chatd

## Outcome

Add a first-class `advisor` tool to agent chats in `coderd/x/chatd` that feels native to Coder:

- it is a built-in server-side tool, not an MCP/dynamic-tool workaround;
- it performs a nested **tool-less** model call for strategic advice;
- it is exposed only when eligible, and the prompt mentions it only when it is actually available;
- it is treated as a **planning-only** tool so it does not run alongside action tools in the same batch;
- it tracks usage/cost separately enough for operators to reason about it;
- it has a minimally polished UI in the Agents page;
- and it ships with explicit dogfooding evidence, including screenshots and repro videos.

## Design decisions to lock before coding

1. **Primary architecture:** native built-in tool in `chattool/`, backed by a small `chatadvisor` package.
2. **Nested model execution:** reuse chatd's existing model/provider stack for a one-step, tool-less advisor call rather than inventing a new provider pathway.
3. **Execution policy:** treat `advisor` as an exclusive/planning-only tool; mixed batches must return structured policy errors and force the model to retry cleanly.
4. **Availability:** initial rollout is for root agent chats only; disable for child/sub-agent chats until recursion/cost policy is proven.
5. **Prompt sync:** use one eligibility boolean to drive both tool registration and advisor guidance injection.
6. **Persistence/cost split:** MVP should keep advisor usage visible in result metadata and server metrics; only add DB schema if product/billing explicitly needs queryable advisor-specific cost.
7. **UI scope:** generic tool rendering is an acceptable temporary milestone during backend bring-up, but the release candidate should include a dedicated lightweight advisor renderer.

## Delivery model

The work should be executed as coordinated workstreams with one integration owner and parallel contributors for low-conflict areas. The integration owner should own `coderd/x/chatd/chatd.go` because prompt assembly, tool registration, and model resolution all converge there.

## Detailed workstreams

### Repo evidence used for this plan

<details>
<summary>Mux reference and current chatd seams</summary>

**Mux reference implementation**

- `src/node/services/tools/advisor.ts` — native advisor tool implementation.
- `src/common/constants/advisor.ts` — advisor prompt/constants and truncation policy.
- `src/common/utils/tools/tools.ts` — conditional tool registration.
- `src/node/services/streamContextBuilder.ts` — injects advisor guidance only when the tool is available.

**Current chatd seams**

- `coderd/x/chatd/chatd.go`
  - `processChat()` — tool assembly, prompt assembly, and chatloop invocation.
  - `resolveChatModel()` — current model/provider/key resolution seam.
  - `type Config struct` — server-level chatd configuration surface.
- `coderd/x/chatd/chatloop/chatloop.go`
  - `Run()` — main streaming/model loop.
  - `executeTools()` — built-in tool execution/batching seam.
- `coderd/x/chatd/chattool/` — built-in tool implementations.
- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx` — tool renderer dispatch.
- `site/src/pages/AgentsPage/components/ChatConversation/messageParsing.ts` and `ConversationTimeline.tsx` — tool/result merge and rendering flow.

</details>

### Workstream map and ownership

| Workstream | Primary owner | Main files | Can run in parallel? | Done when |
|---|---|---|---|---|
| 0. Integration + gating | Integration lead | `coderd/x/chatd/chatd.go` | No; central merge lane | Tool registration, prompt sync, and model selection are wired together |
| 1. Advisor runtime + tool | Backend agent | new `coderd/x/chatd/chatadvisor/`, new `coderd/x/chatd/chattool/advisor.go` | Yes | Tool can perform a tool-less advisor call in memory and return structured results |
| 2. Planning-only execution policy | Chatloop agent | `coderd/x/chatd/chatloop/chatloop.go`, related tests | Yes | Mixed `advisor` + action-tool batches are rejected cleanly and deterministically |
| 3. Metrics/usage/config | Backend/telemetry agent | `chatd.go`, `chatloop/metrics.go`, optional config plumbing | Partially; coordinate with integration lead | Advisor usage is separately visible in metadata/metrics and limits are enforced |
| 4. Frontend rendering | Frontend agent | `site/.../tools/Tool.tsx`, new `AdvisorTool.tsx`, stories | Yes after result schema stabilizes | Advisor renders as a readable card and story tests pass |
| 5. Dogfood + QA evidence | QA agent | dev server, Storybook, dogfood output | After backend + UI are usable | Repro videos, screenshots, and a concise QA report exist |

### Parallelization rules

- **Do not split `coderd/x/chatd/chatd.go` across multiple execution agents without an integration lead.** That file owns prompt building, tool registration, model resolution, and cost persistence.
- Workstreams 1 and 2 can be developed in parallel and then stacked onto the integration branch.
- Workstream 4 should begin once the backend result schema is agreed on, even if the backend is still behind a feature flag.
- Any agent that needs to re-check Mux behavior should clone `coder/mux` into a temporary directory (for example, `$(mktemp -d)/mux`) and inspect it read-only; do not vendor or copy code from Mux directly.

## Phase 0 — Preflight and guardrails

### Goals

- Align the team on the smallest shippable architecture.
- Prevent scope creep into MCP/dynamic-tool/sub-agent variants.
- Decide upfront what is MVP vs. follow-up.

### Tasks

1. **Confirm the MVP boundary.**
   - Ship a built-in advisor tool first.
   - Do **not** make MCP, dynamic tools, or sub-agents the primary implementation.
   - Do **not** add transient streaming phases in the first backend PR unless they fall out almost for free.

2. **Confirm local workflow hygiene before coding.**
   - Ensure the repo is using the project git hooks from `scripts/githooks`.
   - Do not bypass hooks with `--no-verify`.
   - Use `./scripts/develop.sh` for the full dev server rather than manual build/run commands.

3. **Lock the model-selection policy.**
   - **Recommended MVP:** advisor uses the same resolved provider/model/cost config as the current chat, with advisor-specific max-output and usage caps.
   - **Follow-up only if required:** add a separate `AdvisorModelConfigID`-style override that resolves through the existing `configCache`/model-config path. Do not invent a new free-form `provider:model` parser if chatd already stores provider/model separately.

4. **Lock the persistence policy.**
   - **Recommended MVP:** no DB migration. Persist advisor-visible metadata in the tool result and record separate metrics in memory/Prometheus.
   - **Only if product/billing explicitly asks for queryable advisor cost:** add a later DB migration or usage table, following the normal `queries/*.sql` + `make gen` workflow.

5. **Create an execution ADR note in the work item or tracking doc.**
   - Capture: built-in tool, tool-less nested call, root-chat-only rollout, exclusive execution policy, MVP no-DB-migration default.

### Quality gate

- Everyone on the team can state the same answers to these questions:
  - Is advisor a built-in tool? **Yes.**
  - Can advisor run with action tools in the same batch? **No.**
  - Does advisor get tools of its own? **No.**
  - Is a DB migration required for MVP? **No, unless billing insists.**

## Phase 1 — Build the advisor runtime and tool wrapper

### Goals

Create the core advisor implementation in a way that is easy to test and keeps `chattool/` thin.

### Files to add

- `coderd/x/chatd/chatadvisor/types.go`
- `coderd/x/chatd/chatadvisor/guidance.go`
- `coderd/x/chatd/chatadvisor/handoff.go`
- `coderd/x/chatd/chatadvisor/runtime.go`
- `coderd/x/chatd/chatadvisor/runner.go`
- `coderd/x/chatd/chattool/advisor.go`

### Responsibilities by file

1. **`types.go`**
   - Define the input/result schema used by the tool and UI.
   - Keep the result shape close to Mux so the UI and model both have predictable cases.
   - Recommended result variants:
     - `advice`
     - `limit_reached`
     - `error`

   Recommended shape:

   ```go
   type AdvisorArgs struct {
       Question string `json:"question"`
   }

   type AdvisorResult struct {
       Type          string              `json:"type"`
       Advice        string              `json:"advice,omitempty"`
       Error         string              `json:"error,omitempty"`
       AdvisorModel  string              `json:"advisor_model,omitempty"`
       RemainingUses int                 `json:"remaining_uses,omitempty"`
       Usage         *AdvisorUsageResult `json:"usage,omitempty"`
   }
   ```

2. **`guidance.go`**
   - Hold two strings:
     - the nested advisor system prompt;
     - the parent-agent guidance block to inject into the outer system prompt.
   - The nested advisor prompt must say, in plain language:
     - you are advising the parent agent;
     - you do not address the end user directly;
     - you do not claim actions happened;
     - you return concise strategic guidance and tradeoffs.

3. **`runtime.go`**
   - Define the per-run runtime state.
   - Recommended fields:
     - resolved model + model config;
     - provider keys/options reused from the outer chat;
     - `MaxUsesPerRun`;
     - `MaxOutputTokens`;
     - atomic/current call counter;
     - callback(s) to obtain the current prompt snapshot and current-step snapshot;
     - optional metrics/usage hook.
   - Add fail-fast validation for impossible config: nil model, non-positive limits, empty prompt builders, etc.

4. **`handoff.go`**
   - Build the advisor handoff message from:
     - the explicit question;
     - the exact prompt/messages the parent model just used;
     - the current step's text/reasoning snapshot, if available;
     - the most recent relevant tool outputs, if they are already in the prompt snapshot.
   - **Important:** use the already-prepared outer prompt tail, not a fresh DB reload. That keeps the advisor aligned with compaction and the exact context the outer model saw.
   - Apply hard truncation budgets with recent-context bias.

5. **`runner.go`**
   - Execute the nested advisor call.
   - **Recommended implementation:** call `chatloop.Run()` in an in-memory, one-step mode:
     - `Tools: nil`
     - `ProviderTools: nil`
     - `MaxSteps: 1`
     - `PersistStep`: capture the assistant output in memory instead of writing DB rows
   - Reuse the existing provider/model/cost path instead of building a second provider runner.
   - Assert that no tool definitions are passed to the nested call.

6. **`chattool/advisor.go`**
   - Keep this file thin and consistent with other built-ins.
   - Responsibilities:
     - decode `AdvisorArgs`;
     - validate `Question` is non-empty and bounded;
     - call the `chatadvisor` runner;
     - return a structured tool response.

### Defensive programming requirements

- Assert `Question` is non-empty after trimming.
- Assert runtime limits are positive.
- Assert the nested advisor call runs with zero tools/provider tools.
- Assert `AdvisorResult.Type` is one of the known variants before returning.
- Assert remaining uses never goes negative.

### Acceptance criteria

- A unit test can call the advisor tool with a fake model and receive a stable `advice` result.
- The nested advisor call is impossible to run with tools accidentally attached.
- The core logic lives in `chatadvisor/`, not embedded inside `chatd.go`.

## Phase 2 — Wire advisor into chatd and keep prompt/tool availability in sync

### Goals

Register the tool in the right place, expose it only when eligible, and inject system guidance only when the tool is present.

### Files to modify

- `coderd/x/chatd/chatd.go`
- optionally a small helper file if `chatd.go` becomes too crowded

### Tasks

1. **Compute one eligibility boolean in `processChat()`.**
   Recommended inputs:
   - server-level advisor enabled flag;
   - root chat only (`chat.ParentChatID == uuid.Nil` or equivalent existing root/child check);
   - a usable resolved model/provider exists;
   - optional experiment/workspace/org gate if product wants staged rollout.

2. **Create the runtime once per outer chat run.**
   - Use the model/config/keys resolved by `resolveChatModel()`.
   - Reuse provider options from the current chat's `ChatModelCallConfig`.
   - Set `MaxUsesPerRun` and `MaxOutputTokens` from advisor config defaults.

3. **Register the tool in the built-in tool block.**
   - Insert after the skill tools and before MCP tools in `processChat()`.
   - Record `builtinToolNames["advisor"] = true` so metrics stay bounded.

4. **Inject advisor guidance into the outer system prompt using the same boolean.**
   - Use `chatprompt.InsertSystem()` in the same prompt assembly path that already injects user/system instructions.
   - Place the block near the existing instruction insertion, before plan-path/skill context blocks.
   - Wrap the guidance in an explicit tag like `<advisor-guidance>` so it is easy to spot in tests and future refactors.

5. **Keep advisor out of child chats for the first release.**
   - That avoids recursion/cost blowups with `spawn_agent` / `wait_agent` flows.
   - Document this explicitly in the rollout notes and tests.

### Acceptance criteria

- If advisor is disabled, neither the tool nor the prompt guidance appears.
- If advisor is enabled, both the tool and the prompt guidance appear.
- Root chats can use advisor; child chats cannot.
- Built-in tool names include `advisor` so metrics do not collapse it into the generic `mcp` label.

## Phase 3 — Enforce planning-only execution policy in `chatloop`

### Goals

Prevent the model from calling `advisor` and action tools in the same execution batch.

### Files to modify

- `coderd/x/chatd/chatloop/chatloop.go`
- related chatloop tests

### Recommended implementation

Keep the MVP small; do **not** build a general policy engine yet.

1. Add a minimal field to `chatloop.RunOptions`, for example:

   ```go
   ExclusiveToolName *string
   ```

2. In `Run()` / `executeTools()`, detect the case where the exclusive tool appears in the same local-tool batch as any other locally executed tool.

3. When that happens, synthesize structured tool-result errors for the affected calls instead of executing anything in the batch.
   - `advisor` should receive a clear error like: _advisor must be called by itself before action tools_.
   - The sibling action tools should receive a paired policy error like: _this tool was skipped because advisor must run alone_.

4. Let the outer model see those tool errors and retry cleanly.
   - This is simpler and safer than partial execution or hidden deferral.
   - It preserves deterministic transcript history for debugging.

5. Pass the just-finished step snapshot into the tool execution context.
   - The advisor runtime should be able to see the current step's text/reasoning content, because that is often the best hint about what the outer model is trying to decide.

### Why this is the right fit

- It matches the intended semantics: advisor is consulted **before** taking action.
- It avoids subtle race conditions caused by concurrent built-in tool execution.
- It keeps the behavior easy to test with fake models.

### Acceptance criteria

- A model-emitted batch containing only `advisor` succeeds.
- A model-emitted batch containing `advisor` plus any other locally executed tool returns deterministic policy errors and executes nothing.
- Non-advisor tool execution stays unchanged for normal chats.

## Phase 4 — Usage limits, metrics, and configuration

### Goals

Make advisor safe to operate without over-designing billing/storage in the first release.

### Files to modify

- `coderd/x/chatd/chatd.go`
- `coderd/x/chatd/chatloop/metrics.go` as needed
- `coderd/x/chatd/chatd.go` `Config` struct and constructor path
- optional follow-up config/db files only if a separate advisor model or persistent billing is required

### Tasks

1. **Add explicit server config knobs for MVP.**
   Recommended fields on `chatd.Config` or a nested advisor config struct:
   - `AdvisorEnabled bool`
   - `AdvisorMaxUsesPerRun int`
   - `AdvisorMaxOutputTokens int64`

2. **Track usage per outer run.**
   - Reset the counter for each `processChat()` invocation.
   - Return `remaining_uses` in the tool result.
   - Return `limit_reached` when the cap is exhausted.

3. **Expose advisor usage metadata in the tool result.**
   - Include model name and token/cost summary if available.
   - Use the same `callConfig.Cost` calculation path as the outer chat for MVP if advisor reuses the same model.

4. **Record server-side metrics.**
   - Count advisor invocations, failures, and latency.
   - Ensure they show up under the built-in tool label `advisor`.

5. **Optional decision gate: separate advisor model.**
   - If product insists on a stronger/different advisor model, add a follow-up config hook that resolves another existing chat model config through the same `configCache` path.
   - Keep that out of the first landing PR unless it is required for acceptance.

6. **Optional decision gate: queryable advisor cost.**
   - If this becomes required, spin a follow-up DB task:
     - update `coderd/database/queries/*.sql`;
     - add migration files;
     - run `make gen`;
     - update audit mappings if a new auditable type/field is introduced.

### Acceptance criteria

- Advisor calls are capped per outer run.
- Limit exhaustion is user-visible in the tool result.
- Metrics distinguish advisor calls from other built-in tools.
- MVP does not require a schema migration unless explicitly approved.

## Phase 5 — Frontend rendering and Storybook coverage

### Goals

Make advisor feel intentional in the Agents UI without blocking the backend on fancy streaming UI.

### Files to modify

- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx`
- new `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`
- Storybook story file(s) in the same tools directory

### Delivery strategy

1. **Intermediate milestone during backend bring-up:** rely on the existing generic tool renderer if needed.
   - This is acceptable only as a short-lived integration checkpoint.

2. **Release milestone:** add a dedicated lightweight `AdvisorTool` renderer.
   - Reuse existing primitives:
     - `ToolCollapsible`
     - `ToolIcon`
     - `Response` for markdown/prose rendering
     - `ScrollArea` if the advice can be long
   - Keep styling light and consistent with the Agents page.
   - Do not add unnecessary React memoization in `site/src/pages/AgentsPage/`; that area is already React-Compiler aware.

3. **Render the structured result states cleanly.**
   - `advice` — readable prose/markdown with optional metadata footer.
   - `limit_reached` — warning-style message.
   - `error` — error state with visible fallback text.
   - `running` — existing tool loading state/spinner is enough for MVP.

4. **Add Storybook coverage instead of ad-hoc component tests.**
   Recommended stories:
   - successful advice;
   - running/loading;
   - limit reached;
   - error.

5. **Keep the UI contract narrow.**
   - Prefer one text field like `advice` plus small metadata rather than a deeply nested schema.
   - That keeps the UI resilient to prompt iteration.

### Acceptance criteria

- The advisor tool card renders readable content rather than raw quoted JSON in the final release branch.
- Running, limit, and error states are visibly distinct.
- Storybook stories and play assertions cover the new states.
- Existing tool rendering flows remain unchanged.

## Phase 6 — Automated tests and validation gates

### Backend tests to add

1. **Advisor runtime/tool tests**
   - question validation;
   - tool-less nested execution assertion;
   - success result shaping;
   - limit-reached result shaping;
   - error result shaping.

2. **Prompt/gating tests in chatd**
   - advisor disabled ⇒ no tool, no guidance;
   - advisor enabled/root chat ⇒ tool + guidance;
   - child chat ⇒ advisor absent.

3. **Chatloop policy tests**
   - advisor alone runs;
   - advisor + action tool mixed batch returns deterministic policy errors;
   - non-advisor tools still execute normally.

4. **Usage/metrics tests**
   - per-run cap resets correctly;
   - builtin tool labeling includes `advisor`;
   - returned metadata includes model/usage summary when available.

### Frontend tests to add

- Storybook `play()` assertions for the advisor renderer states.
- Verify expand/collapse behavior and visible fallback text.
- Verify the message timeline still renders adjacent tools correctly.

### Recommended command sequence

Run these as the implementation matures, not only at the end:

1. Backend-focused gate after phases 1–4:
   - `make test RUN=TestAdvisor`
   - `make test RUN=TestChatloopAdvisor`
   - `make lint`

2. Frontend-focused gate after phase 5:
   - `pnpm test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`
   - `pnpm lint`
   - `pnpm format`

3. Final repo gate before handoff:
   - `make pre-commit`
   - run any additional targeted `make test RUN=...` selections covering touched chatd paths

> Use the exact new test names the implementing agents create; the names above are recommended anchors, not existing tests.

## Dogfooding plan

### Principle

Dogfood the change as a real agent feature, not just a unit-tested backend. Per the dogfood and `agent-browser` skills, the reviewer should get **watchable repro videos** plus screenshots that make the behavior obvious without reading logs.

### Required setup

1. Start the full dev environment with:
   - `./scripts/develop.sh`
2. If the frontend renderer changes, also start Storybook from `site/` with:
   - `pnpm storybook --no-open`
3. Use `agent-browser` directly — **never `npx agent-browser`**.
4. Use named browser sessions and an output folder such as:
   - `./dogfood-output/advisor/`
   - with subfolders `screenshots/` and `videos/`

### Evidence protocol

For every interactive scenario below:

1. Start video recording **before** the action.
2. Capture step-by-step screenshots at human pace.
3. Capture one annotated screenshot of the final state.
4. Stop the recording.
5. Note the exact pass/fail observation in the QA report.

For static UI states (for example Storybook error/limit cards), an annotated screenshot is sufficient; video is optional but still encouraged by this project’s review preference.

### Dogfood scenarios

#### Scenario A — Happy path in the real Agents UI

**Goal:** prove that a root agent chat can invoke advisor and produce a readable recommendation before taking further action.

Steps:

1. Open the Agents page with an advisor-enabled root chat.
2. Start a repro video.
3. Send a prompt that should reasonably trigger strategic planning, such as an architecture or multi-tradeoff question.
4. Capture screenshots of:
   - the prompt before send;
   - the running advisor state;
   - the completed advisor card and the assistant’s follow-up response.
5. Stop recording.

Pass criteria:

- advisor appears in the timeline;
- the rendered result is readable;
- the assistant can continue after consuming the advisor output.

#### Scenario B — Advisor unavailable path

**Goal:** prove the feature is truly gated.

Suggested variants (at least one is required, both are better):

- feature flag/config off;
- child/sub-agent chat.

Evidence:

- annotated screenshot of the chat/tool state showing advisor is absent;
- short video if toggling the gate live is part of the repro.

Pass criteria:

- no advisor tool is available;
- no advisor-specific prompt behavior leaks through.

#### Scenario C — UI states in Storybook

**Goal:** prove the renderer handles non-happy states cleanly.

Required story states:

- success/advice;
- running;
- limit reached;
- error.

Evidence:

- one screenshot per state;
- at least one short video showing collapse/expand behavior.

Pass criteria:

- success renders readable advice;
- limit/error have visible fallback text;
- the component behaves like the other tool cards.

#### Scenario D — Regression sweep of nearby tools

**Goal:** ensure advisor does not break the surrounding chat timeline.

Check at minimum:

- another existing built-in tool still renders correctly near advisor;
- sub-agent/tool cards still expand/collapse normally;
- no obvious console errors appear in the Agents page during the advisor flow.

Evidence:

- screenshots of adjacent tool cards;
- console/error capture if anything suspicious appears.

### `agent-browser` usage notes for the QA agent

- Prefer `agent-browser batch` for 2+ sequential commands when no intermediate parsing is needed.
- Use `snapshot -i` to discover interactive refs.
- Re-snapshot after navigation or major DOM changes.
- Avoid `wait --load networkidle` unless the page is known to go idle; prefer explicit element/text waits or short fixed waits.
- Record videos at human pace and include pauses that a reviewer can follow.

## Rollout plan

### Initial rollout

- Gate behind a server-side advisor-enabled flag.
- Enable only for selected internal/root agent chats first.
- Watch metrics for:
  - invocation count;
  - failure rate;
  - latency;
  - obvious retry loops.

### Expansion conditions

Expand beyond the initial rollout only after the following are true:

- mixed-batch policy behavior is stable;
- cost impact is understood;
- frontend UX is readable in production-like dogfood;
- no recursion surprises have appeared with sub-agent flows.

### Explicit non-goals for the first release

- advisor inside child/sub-agent chats;
- provider-agnostic streaming phase UI;
- MCP-based external advisor implementation;
- mandatory DB-backed advisor cost reporting.

## Final acceptance checklist

- [ ] `advisor` is a built-in chatd tool, not an MCP/dynamic-tool substitute.
- [ ] The nested advisor call is tool-less and bounded to one in-memory step.
- [ ] One eligibility boolean controls both tool registration and prompt guidance injection.
- [ ] Root chats can use advisor; child chats cannot in the initial rollout.
- [ ] Mixed advisor/action batches produce deterministic policy errors instead of partial execution.
- [ ] Per-run usage caps and limit-reached behavior work.
- [ ] Advisor usage is visible in metadata/metrics without forcing a DB migration for MVP.
- [ ] The Agents UI has a readable advisor card and Storybook coverage.
- [ ] Dogfooding produced screenshots and repro videos for the required scenarios.
- [ ] Validation commands (`make lint`, targeted `make test`, Storybook tests, `make pre-commit`) passed before handoff.

## Suggested PR split

1. **PR 1 — Backend foundation**
   - `chatadvisor/` package
   - `chattool/advisor.go`
   - `chatloop` exclusive policy
   - chatd gating/prompt sync
   - backend tests

2. **PR 2 — Frontend + QA**
   - advisor renderer
   - stories/play assertions
   - dogfood artifacts and QA notes

3. **PR 3 — Optional follow-ups only if demanded by stakeholders**
   - separate advisor model override
   - persistent advisor billing/queryability
   - transient phase-stream UX


</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-7` • Thinking: `max`_
 2026-04-30 14:53:08 +02:00
Thomas Kosiewski eaf2609bb8 feat(coderd/x/chatd/chatadvisor): add advisor runtime and tool wrapper (#24620) 
## Summary

Introduce the `coderd/x/chatd/chatadvisor` package: the self-contained runtime that performs a nested, **tool-less, single-step** model call to return strategic guidance to the parent agent. Also ships the thin `AgentTool` wrapper that the agent exposes as the `advisor` built-in.

## Motivation

The advisor is a "consult before you act" planning step. Keeping its runtime in its own package (rather than inside `chatd.go` or `chattool/`) makes the behavior easy to unit-test, keeps `chattool/` thin, and avoids import cycles with `chatprompt`.

## Changes

- `chatadvisor/types.go`: `AdvisorArgs` and `AdvisorResult` (`advice` / `limit_reached` / `error` variants) plus usage metadata. Stable JSON shape for the UI.
- `chatadvisor/guidance.go`: the nested advisor system prompt and `ParentGuidanceBlock` injected into the outer agent (tagged `<advisor-guidance>`). The nested prompt explicitly tells the advisor it is advising the parent agent, must not address the user, and must not claim actions happened.
- `chatadvisor/handoff.go`: builds the advisor input from the already-prepared outer prompt tail (not a fresh DB reload) with hard truncation budgets and a recent-context bias so the advisor sees the exact context the outer model saw.
- `chatadvisor/runner.go`: wraps `chatloop.Run()` in strict one-step mode (`Tools: nil`, `ProviderTools: nil`, `MaxSteps: 1`) so the nested call is structurally incapable of calling tools.
- `chatadvisor/tool.go`: the `AgentTool` wrapper. Validates the question (non-empty, bounded to 2000 runes), invokes the runtime, and maps the result to a JSON tool response.
- Defensive assertions throughout (question non-empty after trim, positive limits, zero tools on nested call, known result variants, non-negative remaining uses).
- Unit tests cover question validation, tool-less execution, and each result variant.

## Stack context

This is **PR 2 of 6** in the advisor feature stack. It depends on PR 1's `ExclusiveToolNames` hook; no consumer of this package exists yet (that lands in PR 4).

## Scope / non-goals

- No wiring into `chatd`.
- No HTTP/API surface.
- No UI.
- No separate provider; the runtime reuses the outer chat's resolved model/provider/keys.

## Validation

- `go test ./coderd/x/chatd/chatadvisor/...`
- `make lint`

---

<details>
<summary>📋 Implementation Plan (shared across the advisor stack)</summary>

# Plan: Add a Mux-style advisor tool to coder agents/chatd

## Outcome

Add a first-class `advisor` tool to agent chats in `coderd/x/chatd` that feels native to Coder:

- it is a built-in server-side tool, not an MCP/dynamic-tool workaround;
- it performs a nested **tool-less** model call for strategic advice;
- it is exposed only when eligible, and the prompt mentions it only when it is actually available;
- it is treated as a **planning-only** tool so it does not run alongside action tools in the same batch;
- it tracks usage/cost separately enough for operators to reason about it;
- it has a minimally polished UI in the Agents page;
- and it ships with explicit dogfooding evidence, including screenshots and repro videos.

## Design decisions to lock before coding

1. **Primary architecture:** native built-in tool in `chattool/`, backed by a small `chatadvisor` package.
2. **Nested model execution:** reuse chatd's existing model/provider stack for a one-step, tool-less advisor call rather than inventing a new provider pathway.
3. **Execution policy:** treat `advisor` as an exclusive/planning-only tool; mixed batches must return structured policy errors and force the model to retry cleanly.
4. **Availability:** initial rollout is for root agent chats only; disable for child/sub-agent chats until recursion/cost policy is proven.
5. **Prompt sync:** use one eligibility boolean to drive both tool registration and advisor guidance injection.
6. **Persistence/cost split:** MVP should keep advisor usage visible in result metadata and server metrics; only add DB schema if product/billing explicitly needs queryable advisor-specific cost.
7. **UI scope:** generic tool rendering is an acceptable temporary milestone during backend bring-up, but the release candidate should include a dedicated lightweight advisor renderer.

## Delivery model

The work should be executed as coordinated workstreams with one integration owner and parallel contributors for low-conflict areas. The integration owner should own `coderd/x/chatd/chatd.go` because prompt assembly, tool registration, and model resolution all converge there.

## Detailed workstreams

### Repo evidence used for this plan

<details>
<summary>Mux reference and current chatd seams</summary>

**Mux reference implementation**

- `src/node/services/tools/advisor.ts` — native advisor tool implementation.
- `src/common/constants/advisor.ts` — advisor prompt/constants and truncation policy.
- `src/common/utils/tools/tools.ts` — conditional tool registration.
- `src/node/services/streamContextBuilder.ts` — injects advisor guidance only when the tool is available.

**Current chatd seams**

- `coderd/x/chatd/chatd.go`
  - `processChat()` — tool assembly, prompt assembly, and chatloop invocation.
  - `resolveChatModel()` — current model/provider/key resolution seam.
  - `type Config struct` — server-level chatd configuration surface.
- `coderd/x/chatd/chatloop/chatloop.go`
  - `Run()` — main streaming/model loop.
  - `executeTools()` — built-in tool execution/batching seam.
- `coderd/x/chatd/chattool/` — built-in tool implementations.
- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx` — tool renderer dispatch.
- `site/src/pages/AgentsPage/components/ChatConversation/messageParsing.ts` and `ConversationTimeline.tsx` — tool/result merge and rendering flow.

</details>

### Workstream map and ownership

| Workstream | Primary owner | Main files | Can run in parallel? | Done when |
|---|---|---|---|---|
| 0. Integration + gating | Integration lead | `coderd/x/chatd/chatd.go` | No; central merge lane | Tool registration, prompt sync, and model selection are wired together |
| 1. Advisor runtime + tool | Backend agent | new `coderd/x/chatd/chatadvisor/`, new `coderd/x/chatd/chattool/advisor.go` | Yes | Tool can perform a tool-less advisor call in memory and return structured results |
| 2. Planning-only execution policy | Chatloop agent | `coderd/x/chatd/chatloop/chatloop.go`, related tests | Yes | Mixed `advisor` + action-tool batches are rejected cleanly and deterministically |
| 3. Metrics/usage/config | Backend/telemetry agent | `chatd.go`, `chatloop/metrics.go`, optional config plumbing | Partially; coordinate with integration lead | Advisor usage is separately visible in metadata/metrics and limits are enforced |
| 4. Frontend rendering | Frontend agent | `site/.../tools/Tool.tsx`, new `AdvisorTool.tsx`, stories | Yes after result schema stabilizes | Advisor renders as a readable card and story tests pass |
| 5. Dogfood + QA evidence | QA agent | dev server, Storybook, dogfood output | After backend + UI are usable | Repro videos, screenshots, and a concise QA report exist |

### Parallelization rules

- **Do not split `coderd/x/chatd/chatd.go` across multiple execution agents without an integration lead.** That file owns prompt building, tool registration, model resolution, and cost persistence.
- Workstreams 1 and 2 can be developed in parallel and then stacked onto the integration branch.
- Workstream 4 should begin once the backend result schema is agreed on, even if the backend is still behind a feature flag.
- Any agent that needs to re-check Mux behavior should clone `coder/mux` into a temporary directory (for example, `$(mktemp -d)/mux`) and inspect it read-only; do not vendor or copy code from Mux directly.

## Phase 0 — Preflight and guardrails

### Goals

- Align the team on the smallest shippable architecture.
- Prevent scope creep into MCP/dynamic-tool/sub-agent variants.
- Decide upfront what is MVP vs. follow-up.

### Tasks

1. **Confirm the MVP boundary.**
   - Ship a built-in advisor tool first.
   - Do **not** make MCP, dynamic tools, or sub-agents the primary implementation.
   - Do **not** add transient streaming phases in the first backend PR unless they fall out almost for free.

2. **Confirm local workflow hygiene before coding.**
   - Ensure the repo is using the project git hooks from `scripts/githooks`.
   - Do not bypass hooks with `--no-verify`.
   - Use `./scripts/develop.sh` for the full dev server rather than manual build/run commands.

3. **Lock the model-selection policy.**
   - **Recommended MVP:** advisor uses the same resolved provider/model/cost config as the current chat, with advisor-specific max-output and usage caps.
   - **Follow-up only if required:** add a separate `AdvisorModelConfigID`-style override that resolves through the existing `configCache`/model-config path. Do not invent a new free-form `provider:model` parser if chatd already stores provider/model separately.

4. **Lock the persistence policy.**
   - **Recommended MVP:** no DB migration. Persist advisor-visible metadata in the tool result and record separate metrics in memory/Prometheus.
   - **Only if product/billing explicitly asks for queryable advisor cost:** add a later DB migration or usage table, following the normal `queries/*.sql` + `make gen` workflow.

5. **Create an execution ADR note in the work item or tracking doc.**
   - Capture: built-in tool, tool-less nested call, root-chat-only rollout, exclusive execution policy, MVP no-DB-migration default.

### Quality gate

- Everyone on the team can state the same answers to these questions:
  - Is advisor a built-in tool? **Yes.**
  - Can advisor run with action tools in the same batch? **No.**
  - Does advisor get tools of its own? **No.**
  - Is a DB migration required for MVP? **No, unless billing insists.**

## Phase 1 — Build the advisor runtime and tool wrapper

### Goals

Create the core advisor implementation in a way that is easy to test and keeps `chattool/` thin.

### Files to add

- `coderd/x/chatd/chatadvisor/types.go`
- `coderd/x/chatd/chatadvisor/guidance.go`
- `coderd/x/chatd/chatadvisor/handoff.go`
- `coderd/x/chatd/chatadvisor/runtime.go`
- `coderd/x/chatd/chatadvisor/runner.go`
- `coderd/x/chatd/chattool/advisor.go`

### Responsibilities by file

1. **`types.go`**
   - Define the input/result schema used by the tool and UI.
   - Keep the result shape close to Mux so the UI and model both have predictable cases.
   - Recommended result variants:
     - `advice`
     - `limit_reached`
     - `error`

   Recommended shape:

   ```go
   type AdvisorArgs struct {
       Question string `json:"question"`
   }

   type AdvisorResult struct {
       Type          string              `json:"type"`
       Advice        string              `json:"advice,omitempty"`
       Error         string              `json:"error,omitempty"`
       AdvisorModel  string              `json:"advisor_model,omitempty"`
       RemainingUses int                 `json:"remaining_uses,omitempty"`
       Usage         *AdvisorUsageResult `json:"usage,omitempty"`
   }
   ```

2. **`guidance.go`**
   - Hold two strings:
     - the nested advisor system prompt;
     - the parent-agent guidance block to inject into the outer system prompt.
   - The nested advisor prompt must say, in plain language:
     - you are advising the parent agent;
     - you do not address the end user directly;
     - you do not claim actions happened;
     - you return concise strategic guidance and tradeoffs.

3. **`runtime.go`**
   - Define the per-run runtime state.
   - Recommended fields:
     - resolved model + model config;
     - provider keys/options reused from the outer chat;
     - `MaxUsesPerRun`;
     - `MaxOutputTokens`;
     - atomic/current call counter;
     - callback(s) to obtain the current prompt snapshot and current-step snapshot;
     - optional metrics/usage hook.
   - Add fail-fast validation for impossible config: nil model, non-positive limits, empty prompt builders, etc.

4. **`handoff.go`**
   - Build the advisor handoff message from:
     - the explicit question;
     - the exact prompt/messages the parent model just used;
     - the current step's text/reasoning snapshot, if available;
     - the most recent relevant tool outputs, if they are already in the prompt snapshot.
   - **Important:** use the already-prepared outer prompt tail, not a fresh DB reload. That keeps the advisor aligned with compaction and the exact context the outer model saw.
   - Apply hard truncation budgets with recent-context bias.

5. **`runner.go`**
   - Execute the nested advisor call.
   - **Recommended implementation:** call `chatloop.Run()` in an in-memory, one-step mode:
     - `Tools: nil`
     - `ProviderTools: nil`
     - `MaxSteps: 1`
     - `PersistStep`: capture the assistant output in memory instead of writing DB rows
   - Reuse the existing provider/model/cost path instead of building a second provider runner.
   - Assert that no tool definitions are passed to the nested call.

6. **`chattool/advisor.go`**
   - Keep this file thin and consistent with other built-ins.
   - Responsibilities:
     - decode `AdvisorArgs`;
     - validate `Question` is non-empty and bounded;
     - call the `chatadvisor` runner;
     - return a structured tool response.

### Defensive programming requirements

- Assert `Question` is non-empty after trimming.
- Assert runtime limits are positive.
- Assert the nested advisor call runs with zero tools/provider tools.
- Assert `AdvisorResult.Type` is one of the known variants before returning.
- Assert remaining uses never goes negative.

### Acceptance criteria

- A unit test can call the advisor tool with a fake model and receive a stable `advice` result.
- The nested advisor call is impossible to run with tools accidentally attached.
- The core logic lives in `chatadvisor/`, not embedded inside `chatd.go`.

## Phase 2 — Wire advisor into chatd and keep prompt/tool availability in sync

### Goals

Register the tool in the right place, expose it only when eligible, and inject system guidance only when the tool is present.

### Files to modify

- `coderd/x/chatd/chatd.go`
- optionally a small helper file if `chatd.go` becomes too crowded

### Tasks

1. **Compute one eligibility boolean in `processChat()`.**
   Recommended inputs:
   - server-level advisor enabled flag;
   - root chat only (`chat.ParentChatID == uuid.Nil` or equivalent existing root/child check);
   - a usable resolved model/provider exists;
   - optional experiment/workspace/org gate if product wants staged rollout.

2. **Create the runtime once per outer chat run.**
   - Use the model/config/keys resolved by `resolveChatModel()`.
   - Reuse provider options from the current chat's `ChatModelCallConfig`.
   - Set `MaxUsesPerRun` and `MaxOutputTokens` from advisor config defaults.

3. **Register the tool in the built-in tool block.**
   - Insert after the skill tools and before MCP tools in `processChat()`.
   - Record `builtinToolNames["advisor"] = true` so metrics stay bounded.

4. **Inject advisor guidance into the outer system prompt using the same boolean.**
   - Use `chatprompt.InsertSystem()` in the same prompt assembly path that already injects user/system instructions.
   - Place the block near the existing instruction insertion, before plan-path/skill context blocks.
   - Wrap the guidance in an explicit tag like `<advisor-guidance>` so it is easy to spot in tests and future refactors.

5. **Keep advisor out of child chats for the first release.**
   - That avoids recursion/cost blowups with `spawn_agent` / `wait_agent` flows.
   - Document this explicitly in the rollout notes and tests.

### Acceptance criteria

- If advisor is disabled, neither the tool nor the prompt guidance appears.
- If advisor is enabled, both the tool and the prompt guidance appear.
- Root chats can use advisor; child chats cannot.
- Built-in tool names include `advisor` so metrics do not collapse it into the generic `mcp` label.

## Phase 3 — Enforce planning-only execution policy in `chatloop`

### Goals

Prevent the model from calling `advisor` and action tools in the same execution batch.

### Files to modify

- `coderd/x/chatd/chatloop/chatloop.go`
- related chatloop tests

### Recommended implementation

Keep the MVP small; do **not** build a general policy engine yet.

1. Add a minimal field to `chatloop.RunOptions`, for example:

   ```go
   ExclusiveToolName *string
   ```

2. In `Run()` / `executeTools()`, detect the case where the exclusive tool appears in the same local-tool batch as any other locally executed tool.

3. When that happens, synthesize structured tool-result errors for the affected calls instead of executing anything in the batch.
   - `advisor` should receive a clear error like: _advisor must be called by itself before action tools_.
   - The sibling action tools should receive a paired policy error like: _this tool was skipped because advisor must run alone_.

4. Let the outer model see those tool errors and retry cleanly.
   - This is simpler and safer than partial execution or hidden deferral.
   - It preserves deterministic transcript history for debugging.

5. Pass the just-finished step snapshot into the tool execution context.
   - The advisor runtime should be able to see the current step's text/reasoning content, because that is often the best hint about what the outer model is trying to decide.

### Why this is the right fit

- It matches the intended semantics: advisor is consulted **before** taking action.
- It avoids subtle race conditions caused by concurrent built-in tool execution.
- It keeps the behavior easy to test with fake models.

### Acceptance criteria

- A model-emitted batch containing only `advisor` succeeds.
- A model-emitted batch containing `advisor` plus any other locally executed tool returns deterministic policy errors and executes nothing.
- Non-advisor tool execution stays unchanged for normal chats.

## Phase 4 — Usage limits, metrics, and configuration

### Goals

Make advisor safe to operate without over-designing billing/storage in the first release.

### Files to modify

- `coderd/x/chatd/chatd.go`
- `coderd/x/chatd/chatloop/metrics.go` as needed
- `coderd/x/chatd/chatd.go` `Config` struct and constructor path
- optional follow-up config/db files only if a separate advisor model or persistent billing is required

### Tasks

1. **Add explicit server config knobs for MVP.**
   Recommended fields on `chatd.Config` or a nested advisor config struct:
   - `AdvisorEnabled bool`
   - `AdvisorMaxUsesPerRun int`
   - `AdvisorMaxOutputTokens int64`

2. **Track usage per outer run.**
   - Reset the counter for each `processChat()` invocation.
   - Return `remaining_uses` in the tool result.
   - Return `limit_reached` when the cap is exhausted.

3. **Expose advisor usage metadata in the tool result.**
   - Include model name and token/cost summary if available.
   - Use the same `callConfig.Cost` calculation path as the outer chat for MVP if advisor reuses the same model.

4. **Record server-side metrics.**
   - Count advisor invocations, failures, and latency.
   - Ensure they show up under the built-in tool label `advisor`.

5. **Optional decision gate: separate advisor model.**
   - If product insists on a stronger/different advisor model, add a follow-up config hook that resolves another existing chat model config through the same `configCache` path.
   - Keep that out of the first landing PR unless it is required for acceptance.

6. **Optional decision gate: queryable advisor cost.**
   - If this becomes required, spin a follow-up DB task:
     - update `coderd/database/queries/*.sql`;
     - add migration files;
     - run `make gen`;
     - update audit mappings if a new auditable type/field is introduced.

### Acceptance criteria

- Advisor calls are capped per outer run.
- Limit exhaustion is user-visible in the tool result.
- Metrics distinguish advisor calls from other built-in tools.
- MVP does not require a schema migration unless explicitly approved.

## Phase 5 — Frontend rendering and Storybook coverage

### Goals

Make advisor feel intentional in the Agents UI without blocking the backend on fancy streaming UI.

### Files to modify

- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx`
- new `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`
- Storybook story file(s) in the same tools directory

### Delivery strategy

1. **Intermediate milestone during backend bring-up:** rely on the existing generic tool renderer if needed.
   - This is acceptable only as a short-lived integration checkpoint.

2. **Release milestone:** add a dedicated lightweight `AdvisorTool` renderer.
   - Reuse existing primitives:
     - `ToolCollapsible`
     - `ToolIcon`
     - `Response` for markdown/prose rendering
     - `ScrollArea` if the advice can be long
   - Keep styling light and consistent with the Agents page.
   - Do not add unnecessary React memoization in `site/src/pages/AgentsPage/`; that area is already React-Compiler aware.

3. **Render the structured result states cleanly.**
   - `advice` — readable prose/markdown with optional metadata footer.
   - `limit_reached` — warning-style message.
   - `error` — error state with visible fallback text.
   - `running` — existing tool loading state/spinner is enough for MVP.

4. **Add Storybook coverage instead of ad-hoc component tests.**
   Recommended stories:
   - successful advice;
   - running/loading;
   - limit reached;
   - error.

5. **Keep the UI contract narrow.**
   - Prefer one text field like `advice` plus small metadata rather than a deeply nested schema.
   - That keeps the UI resilient to prompt iteration.

### Acceptance criteria

- The advisor tool card renders readable content rather than raw quoted JSON in the final release branch.
- Running, limit, and error states are visibly distinct.
- Storybook stories and play assertions cover the new states.
- Existing tool rendering flows remain unchanged.

## Phase 6 — Automated tests and validation gates

### Backend tests to add

1. **Advisor runtime/tool tests**
   - question validation;
   - tool-less nested execution assertion;
   - success result shaping;
   - limit-reached result shaping;
   - error result shaping.

2. **Prompt/gating tests in chatd**
   - advisor disabled ⇒ no tool, no guidance;
   - advisor enabled/root chat ⇒ tool + guidance;
   - child chat ⇒ advisor absent.

3. **Chatloop policy tests**
   - advisor alone runs;
   - advisor + action tool mixed batch returns deterministic policy errors;
   - non-advisor tools still execute normally.

4. **Usage/metrics tests**
   - per-run cap resets correctly;
   - builtin tool labeling includes `advisor`;
   - returned metadata includes model/usage summary when available.

### Frontend tests to add

- Storybook `play()` assertions for the advisor renderer states.
- Verify expand/collapse behavior and visible fallback text.
- Verify the message timeline still renders adjacent tools correctly.

### Recommended command sequence

Run these as the implementation matures, not only at the end:

1. Backend-focused gate after phases 1–4:
   - `make test RUN=TestAdvisor`
   - `make test RUN=TestChatloopAdvisor`
   - `make lint`

2. Frontend-focused gate after phase 5:
   - `pnpm test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`
   - `pnpm lint`
   - `pnpm format`

3. Final repo gate before handoff:
   - `make pre-commit`
   - run any additional targeted `make test RUN=...` selections covering touched chatd paths

> Use the exact new test names the implementing agents create; the names above are recommended anchors, not existing tests.

## Dogfooding plan

### Principle

Dogfood the change as a real agent feature, not just a unit-tested backend. Per the dogfood and `agent-browser` skills, the reviewer should get **watchable repro videos** plus screenshots that make the behavior obvious without reading logs.

### Required setup

1. Start the full dev environment with:
   - `./scripts/develop.sh`
2. If the frontend renderer changes, also start Storybook from `site/` with:
   - `pnpm storybook --no-open`
3. Use `agent-browser` directly — **never `npx agent-browser`**.
4. Use named browser sessions and an output folder such as:
   - `./dogfood-output/advisor/`
   - with subfolders `screenshots/` and `videos/`

### Evidence protocol

For every interactive scenario below:

1. Start video recording **before** the action.
2. Capture step-by-step screenshots at human pace.
3. Capture one annotated screenshot of the final state.
4. Stop the recording.
5. Note the exact pass/fail observation in the QA report.

For static UI states (for example Storybook error/limit cards), an annotated screenshot is sufficient; video is optional but still encouraged by this project’s review preference.

### Dogfood scenarios

#### Scenario A — Happy path in the real Agents UI

**Goal:** prove that a root agent chat can invoke advisor and produce a readable recommendation before taking further action.

Steps:

1. Open the Agents page with an advisor-enabled root chat.
2. Start a repro video.
3. Send a prompt that should reasonably trigger strategic planning, such as an architecture or multi-tradeoff question.
4. Capture screenshots of:
   - the prompt before send;
   - the running advisor state;
   - the completed advisor card and the assistant’s follow-up response.
5. Stop recording.

Pass criteria:

- advisor appears in the timeline;
- the rendered result is readable;
- the assistant can continue after consuming the advisor output.

#### Scenario B — Advisor unavailable path

**Goal:** prove the feature is truly gated.

Suggested variants (at least one is required, both are better):

- feature flag/config off;
- child/sub-agent chat.

Evidence:

- annotated screenshot of the chat/tool state showing advisor is absent;
- short video if toggling the gate live is part of the repro.

Pass criteria:

- no advisor tool is available;
- no advisor-specific prompt behavior leaks through.

#### Scenario C — UI states in Storybook

**Goal:** prove the renderer handles non-happy states cleanly.

Required story states:

- success/advice;
- running;
- limit reached;
- error.

Evidence:

- one screenshot per state;
- at least one short video showing collapse/expand behavior.

Pass criteria:

- success renders readable advice;
- limit/error have visible fallback text;
- the component behaves like the other tool cards.

#### Scenario D — Regression sweep of nearby tools

**Goal:** ensure advisor does not break the surrounding chat timeline.

Check at minimum:

- another existing built-in tool still renders correctly near advisor;
- sub-agent/tool cards still expand/collapse normally;
- no obvious console errors appear in the Agents page during the advisor flow.

Evidence:

- screenshots of adjacent tool cards;
- console/error capture if anything suspicious appears.

### `agent-browser` usage notes for the QA agent

- Prefer `agent-browser batch` for 2+ sequential commands when no intermediate parsing is needed.
- Use `snapshot -i` to discover interactive refs.
- Re-snapshot after navigation or major DOM changes.
- Avoid `wait --load networkidle` unless the page is known to go idle; prefer explicit element/text waits or short fixed waits.
- Record videos at human pace and include pauses that a reviewer can follow.

## Rollout plan

### Initial rollout

- Gate behind a server-side advisor-enabled flag.
- Enable only for selected internal/root agent chats first.
- Watch metrics for:
  - invocation count;
  - failure rate;
  - latency;
  - obvious retry loops.

### Expansion conditions

Expand beyond the initial rollout only after the following are true:

- mixed-batch policy behavior is stable;
- cost impact is understood;
- frontend UX is readable in production-like dogfood;
- no recursion surprises have appeared with sub-agent flows.

### Explicit non-goals for the first release

- advisor inside child/sub-agent chats;
- provider-agnostic streaming phase UI;
- MCP-based external advisor implementation;
- mandatory DB-backed advisor cost reporting.

## Final acceptance checklist

- [ ] `advisor` is a built-in chatd tool, not an MCP/dynamic-tool substitute.
- [ ] The nested advisor call is tool-less and bounded to one in-memory step.
- [ ] One eligibility boolean controls both tool registration and prompt guidance injection.
- [ ] Root chats can use advisor; child chats cannot in the initial rollout.
- [ ] Mixed advisor/action batches produce deterministic policy errors instead of partial execution.
- [ ] Per-run usage caps and limit-reached behavior work.
- [ ] Advisor usage is visible in metadata/metrics without forcing a DB migration for MVP.
- [ ] The Agents UI has a readable advisor card and Storybook coverage.
- [ ] Dogfooding produced screenshots and repro videos for the required scenarios.
- [ ] Validation commands (`make lint`, targeted `make test`, Storybook tests, `make pre-commit`) passed before handoff.

## Suggested PR split

1. **PR 1 — Backend foundation**
   - `chatadvisor/` package
   - `chattool/advisor.go`
   - `chatloop` exclusive policy
   - chatd gating/prompt sync
   - backend tests

2. **PR 2 — Frontend + QA**
   - advisor renderer
   - stories/play assertions
   - dogfood artifacts and QA notes

3. **PR 3 — Optional follow-ups only if demanded by stakeholders**
   - separate advisor model override
   - persistent advisor billing/queryability
   - transient phase-stream UX


</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-7` • Thinking: `max`_
 2026-04-30 14:22:33 +02:00
Thomas Kosiewski a84534315c feat(coderd/x/chatd/chatloop): add exclusive tool execution policy (#24619) 
## Summary

Introduce an **exclusive-tool execution policy** in `chatloop.Run()` so that a designated "planning-only" tool can never execute in the same batch as any other locally executed tool. This is the generic foundation for the advisor tool landed in the stacked PRs above.

## Motivation

The advisor tool (stacked on top) is intentionally a pre-action planning step. It must be consulted *alone*, not interleaved with side-effecting tools, so the model has to reason strategically before committing to actions. Rather than building advisor-specific plumbing into `chatloop`, this PR adds a small, general-purpose policy hook that any future tool can opt into.

## Changes

- `RunOptions.ExclusiveToolNames map[string]bool` declares which tool names must run exclusively within a single tool-execution batch.
- `executeTools()` detects mixed batches: if any exclusive tool name appears next to any other locally executed tool, no tool runs. Instead, structured `ToolResultOutputContentError` entries are synthesized for every tool in the batch so the model can cleanly retry.
- Deterministic, model-facing error copy:
  - Exclusive tool gets: "must be called by itself before action tools".
  - Sibling tools get: "skipped because `<exclusive>` must run alone".
- New tests in `chatloop_test.go` cover: single exclusive batch runs normally, mixed batch returns policy errors and executes nothing, non-exclusive batches are untouched.

## Stack context

This is **PR 1 of 6** in the advisor feature stack.

```
main
 └─ feat/advisor-01-chatloop-exclusive-policy   ← this PR
     └─ feat/advisor-02-chatadvisor-pkg
         └─ feat/advisor-03-config-api
             ├─ feat/advisor-04-chatd-runtime
             │   └─ feat/advisor-05-chat-tool-renderer
             └─ feat/advisor-06-admin-settings-ui
```

## Scope / non-goals

- No advisor-specific code lives in this PR.
- Zero behavior change when `ExclusiveToolNames` is empty (current default for every existing caller).

## Validation

- `go test ./coderd/x/chatd/chatloop/... -run TestExclusive`
- `make lint`

---

<details>
<summary>📋 Implementation Plan (shared across the advisor stack)</summary>

# Plan: Add a Mux-style advisor tool to coder agents/chatd

## Outcome

Add a first-class `advisor` tool to agent chats in `coderd/x/chatd` that feels native to Coder:

- it is a built-in server-side tool, not an MCP/dynamic-tool workaround;
- it performs a nested **tool-less** model call for strategic advice;
- it is exposed only when eligible, and the prompt mentions it only when it is actually available;
- it is treated as a **planning-only** tool so it does not run alongside action tools in the same batch;
- it tracks usage/cost separately enough for operators to reason about it;
- it has a minimally polished UI in the Agents page;
- and it ships with explicit dogfooding evidence, including screenshots and repro videos.

## Design decisions to lock before coding

1. **Primary architecture:** native built-in tool in `chattool/`, backed by a small `chatadvisor` package.
2. **Nested model execution:** reuse chatd's existing model/provider stack for a one-step, tool-less advisor call rather than inventing a new provider pathway.
3. **Execution policy:** treat `advisor` as an exclusive/planning-only tool; mixed batches must return structured policy errors and force the model to retry cleanly.
4. **Availability:** initial rollout is for root agent chats only; disable for child/sub-agent chats until recursion/cost policy is proven.
5. **Prompt sync:** use one eligibility boolean to drive both tool registration and advisor guidance injection.
6. **Persistence/cost split:** MVP should keep advisor usage visible in result metadata and server metrics; only add DB schema if product/billing explicitly needs queryable advisor-specific cost.
7. **UI scope:** generic tool rendering is an acceptable temporary milestone during backend bring-up, but the release candidate should include a dedicated lightweight advisor renderer.

## Delivery model

The work should be executed as coordinated workstreams with one integration owner and parallel contributors for low-conflict areas. The integration owner should own `coderd/x/chatd/chatd.go` because prompt assembly, tool registration, and model resolution all converge there.

## Detailed workstreams

### Repo evidence used for this plan

<details>
<summary>Mux reference and current chatd seams</summary>

**Mux reference implementation**

- `src/node/services/tools/advisor.ts` — native advisor tool implementation.
- `src/common/constants/advisor.ts` — advisor prompt/constants and truncation policy.
- `src/common/utils/tools/tools.ts` — conditional tool registration.
- `src/node/services/streamContextBuilder.ts` — injects advisor guidance only when the tool is available.

**Current chatd seams**

- `coderd/x/chatd/chatd.go`
  - `processChat()` — tool assembly, prompt assembly, and chatloop invocation.
  - `resolveChatModel()` — current model/provider/key resolution seam.
  - `type Config struct` — server-level chatd configuration surface.
- `coderd/x/chatd/chatloop/chatloop.go`
  - `Run()` — main streaming/model loop.
  - `executeTools()` — built-in tool execution/batching seam.
- `coderd/x/chatd/chattool/` — built-in tool implementations.
- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx` — tool renderer dispatch.
- `site/src/pages/AgentsPage/components/ChatConversation/messageParsing.ts` and `ConversationTimeline.tsx` — tool/result merge and rendering flow.

</details>

### Workstream map and ownership

| Workstream | Primary owner | Main files | Can run in parallel? | Done when |
|---|---|---|---|---|
| 0. Integration + gating | Integration lead | `coderd/x/chatd/chatd.go` | No; central merge lane | Tool registration, prompt sync, and model selection are wired together |
| 1. Advisor runtime + tool | Backend agent | new `coderd/x/chatd/chatadvisor/`, new `coderd/x/chatd/chattool/advisor.go` | Yes | Tool can perform a tool-less advisor call in memory and return structured results |
| 2. Planning-only execution policy | Chatloop agent | `coderd/x/chatd/chatloop/chatloop.go`, related tests | Yes | Mixed `advisor` + action-tool batches are rejected cleanly and deterministically |
| 3. Metrics/usage/config | Backend/telemetry agent | `chatd.go`, `chatloop/metrics.go`, optional config plumbing | Partially; coordinate with integration lead | Advisor usage is separately visible in metadata/metrics and limits are enforced |
| 4. Frontend rendering | Frontend agent | `site/.../tools/Tool.tsx`, new `AdvisorTool.tsx`, stories | Yes after result schema stabilizes | Advisor renders as a readable card and story tests pass |
| 5. Dogfood + QA evidence | QA agent | dev server, Storybook, dogfood output | After backend + UI are usable | Repro videos, screenshots, and a concise QA report exist |

### Parallelization rules

- **Do not split `coderd/x/chatd/chatd.go` across multiple execution agents without an integration lead.** That file owns prompt building, tool registration, model resolution, and cost persistence.
- Workstreams 1 and 2 can be developed in parallel and then stacked onto the integration branch.
- Workstream 4 should begin once the backend result schema is agreed on, even if the backend is still behind a feature flag.
- Any agent that needs to re-check Mux behavior should clone `coder/mux` into a temporary directory (for example, `$(mktemp -d)/mux`) and inspect it read-only; do not vendor or copy code from Mux directly.

## Phase 0 — Preflight and guardrails

### Goals

- Align the team on the smallest shippable architecture.
- Prevent scope creep into MCP/dynamic-tool/sub-agent variants.
- Decide upfront what is MVP vs. follow-up.

### Tasks

1. **Confirm the MVP boundary.**
   - Ship a built-in advisor tool first.
   - Do **not** make MCP, dynamic tools, or sub-agents the primary implementation.
   - Do **not** add transient streaming phases in the first backend PR unless they fall out almost for free.

2. **Confirm local workflow hygiene before coding.**
   - Ensure the repo is using the project git hooks from `scripts/githooks`.
   - Do not bypass hooks with `--no-verify`.
   - Use `./scripts/develop.sh` for the full dev server rather than manual build/run commands.

3. **Lock the model-selection policy.**
   - **Recommended MVP:** advisor uses the same resolved provider/model/cost config as the current chat, with advisor-specific max-output and usage caps.
   - **Follow-up only if required:** add a separate `AdvisorModelConfigID`-style override that resolves through the existing `configCache`/model-config path. Do not invent a new free-form `provider:model` parser if chatd already stores provider/model separately.

4. **Lock the persistence policy.**
   - **Recommended MVP:** no DB migration. Persist advisor-visible metadata in the tool result and record separate metrics in memory/Prometheus.
   - **Only if product/billing explicitly asks for queryable advisor cost:** add a later DB migration or usage table, following the normal `queries/*.sql` + `make gen` workflow.

5. **Create an execution ADR note in the work item or tracking doc.**
   - Capture: built-in tool, tool-less nested call, root-chat-only rollout, exclusive execution policy, MVP no-DB-migration default.

### Quality gate

- Everyone on the team can state the same answers to these questions:
  - Is advisor a built-in tool? **Yes.**
  - Can advisor run with action tools in the same batch? **No.**
  - Does advisor get tools of its own? **No.**
  - Is a DB migration required for MVP? **No, unless billing insists.**

## Phase 1 — Build the advisor runtime and tool wrapper

### Goals

Create the core advisor implementation in a way that is easy to test and keeps `chattool/` thin.

### Files to add

- `coderd/x/chatd/chatadvisor/types.go`
- `coderd/x/chatd/chatadvisor/guidance.go`
- `coderd/x/chatd/chatadvisor/handoff.go`
- `coderd/x/chatd/chatadvisor/runtime.go`
- `coderd/x/chatd/chatadvisor/runner.go`
- `coderd/x/chatd/chattool/advisor.go`

### Responsibilities by file

1. **`types.go`**
   - Define the input/result schema used by the tool and UI.
   - Keep the result shape close to Mux so the UI and model both have predictable cases.
   - Recommended result variants:
     - `advice`
     - `limit_reached`
     - `error`

   Recommended shape:

   ```go
   type AdvisorArgs struct {
       Question string `json:"question"`
   }

   type AdvisorResult struct {
       Type          string              `json:"type"`
       Advice        string              `json:"advice,omitempty"`
       Error         string              `json:"error,omitempty"`
       AdvisorModel  string              `json:"advisor_model,omitempty"`
       RemainingUses int                 `json:"remaining_uses,omitempty"`
       Usage         *AdvisorUsageResult `json:"usage,omitempty"`
   }
   ```

2. **`guidance.go`**
   - Hold two strings:
     - the nested advisor system prompt;
     - the parent-agent guidance block to inject into the outer system prompt.
   - The nested advisor prompt must say, in plain language:
     - you are advising the parent agent;
     - you do not address the end user directly;
     - you do not claim actions happened;
     - you return concise strategic guidance and tradeoffs.

3. **`runtime.go`**
   - Define the per-run runtime state.
   - Recommended fields:
     - resolved model + model config;
     - provider keys/options reused from the outer chat;
     - `MaxUsesPerRun`;
     - `MaxOutputTokens`;
     - atomic/current call counter;
     - callback(s) to obtain the current prompt snapshot and current-step snapshot;
     - optional metrics/usage hook.
   - Add fail-fast validation for impossible config: nil model, non-positive limits, empty prompt builders, etc.

4. **`handoff.go`**
   - Build the advisor handoff message from:
     - the explicit question;
     - the exact prompt/messages the parent model just used;
     - the current step's text/reasoning snapshot, if available;
     - the most recent relevant tool outputs, if they are already in the prompt snapshot.
   - **Important:** use the already-prepared outer prompt tail, not a fresh DB reload. That keeps the advisor aligned with compaction and the exact context the outer model saw.
   - Apply hard truncation budgets with recent-context bias.

5. **`runner.go`**
   - Execute the nested advisor call.
   - **Recommended implementation:** call `chatloop.Run()` in an in-memory, one-step mode:
     - `Tools: nil`
     - `ProviderTools: nil`
     - `MaxSteps: 1`
     - `PersistStep`: capture the assistant output in memory instead of writing DB rows
   - Reuse the existing provider/model/cost path instead of building a second provider runner.
   - Assert that no tool definitions are passed to the nested call.

6. **`chattool/advisor.go`**
   - Keep this file thin and consistent with other built-ins.
   - Responsibilities:
     - decode `AdvisorArgs`;
     - validate `Question` is non-empty and bounded;
     - call the `chatadvisor` runner;
     - return a structured tool response.

### Defensive programming requirements

- Assert `Question` is non-empty after trimming.
- Assert runtime limits are positive.
- Assert the nested advisor call runs with zero tools/provider tools.
- Assert `AdvisorResult.Type` is one of the known variants before returning.
- Assert remaining uses never goes negative.

### Acceptance criteria

- A unit test can call the advisor tool with a fake model and receive a stable `advice` result.
- The nested advisor call is impossible to run with tools accidentally attached.
- The core logic lives in `chatadvisor/`, not embedded inside `chatd.go`.

## Phase 2 — Wire advisor into chatd and keep prompt/tool availability in sync

### Goals

Register the tool in the right place, expose it only when eligible, and inject system guidance only when the tool is present.

### Files to modify

- `coderd/x/chatd/chatd.go`
- optionally a small helper file if `chatd.go` becomes too crowded

### Tasks

1. **Compute one eligibility boolean in `processChat()`.**
   Recommended inputs:
   - server-level advisor enabled flag;
   - root chat only (`chat.ParentChatID == uuid.Nil` or equivalent existing root/child check);
   - a usable resolved model/provider exists;
   - optional experiment/workspace/org gate if product wants staged rollout.

2. **Create the runtime once per outer chat run.**
   - Use the model/config/keys resolved by `resolveChatModel()`.
   - Reuse provider options from the current chat's `ChatModelCallConfig`.
   - Set `MaxUsesPerRun` and `MaxOutputTokens` from advisor config defaults.

3. **Register the tool in the built-in tool block.**
   - Insert after the skill tools and before MCP tools in `processChat()`.
   - Record `builtinToolNames["advisor"] = true` so metrics stay bounded.

4. **Inject advisor guidance into the outer system prompt using the same boolean.**
   - Use `chatprompt.InsertSystem()` in the same prompt assembly path that already injects user/system instructions.
   - Place the block near the existing instruction insertion, before plan-path/skill context blocks.
   - Wrap the guidance in an explicit tag like `<advisor-guidance>` so it is easy to spot in tests and future refactors.

5. **Keep advisor out of child chats for the first release.**
   - That avoids recursion/cost blowups with `spawn_agent` / `wait_agent` flows.
   - Document this explicitly in the rollout notes and tests.

### Acceptance criteria

- If advisor is disabled, neither the tool nor the prompt guidance appears.
- If advisor is enabled, both the tool and the prompt guidance appear.
- Root chats can use advisor; child chats cannot.
- Built-in tool names include `advisor` so metrics do not collapse it into the generic `mcp` label.

## Phase 3 — Enforce planning-only execution policy in `chatloop`

### Goals

Prevent the model from calling `advisor` and action tools in the same execution batch.

### Files to modify

- `coderd/x/chatd/chatloop/chatloop.go`
- related chatloop tests

### Recommended implementation

Keep the MVP small; do **not** build a general policy engine yet.

1. Add a minimal field to `chatloop.RunOptions`, for example:

   ```go
   ExclusiveToolName *string
   ```

2. In `Run()` / `executeTools()`, detect the case where the exclusive tool appears in the same local-tool batch as any other locally executed tool.

3. When that happens, synthesize structured tool-result errors for the affected calls instead of executing anything in the batch.
   - `advisor` should receive a clear error like: _advisor must be called by itself before action tools_.
   - The sibling action tools should receive a paired policy error like: _this tool was skipped because advisor must run alone_.

4. Let the outer model see those tool errors and retry cleanly.
   - This is simpler and safer than partial execution or hidden deferral.
   - It preserves deterministic transcript history for debugging.

5. Pass the just-finished step snapshot into the tool execution context.
   - The advisor runtime should be able to see the current step's text/reasoning content, because that is often the best hint about what the outer model is trying to decide.

### Why this is the right fit

- It matches the intended semantics: advisor is consulted **before** taking action.
- It avoids subtle race conditions caused by concurrent built-in tool execution.
- It keeps the behavior easy to test with fake models.

### Acceptance criteria

- A model-emitted batch containing only `advisor` succeeds.
- A model-emitted batch containing `advisor` plus any other locally executed tool returns deterministic policy errors and executes nothing.
- Non-advisor tool execution stays unchanged for normal chats.

## Phase 4 — Usage limits, metrics, and configuration

### Goals

Make advisor safe to operate without over-designing billing/storage in the first release.

### Files to modify

- `coderd/x/chatd/chatd.go`
- `coderd/x/chatd/chatloop/metrics.go` as needed
- `coderd/x/chatd/chatd.go` `Config` struct and constructor path
- optional follow-up config/db files only if a separate advisor model or persistent billing is required

### Tasks

1. **Add explicit server config knobs for MVP.**
   Recommended fields on `chatd.Config` or a nested advisor config struct:
   - `AdvisorEnabled bool`
   - `AdvisorMaxUsesPerRun int`
   - `AdvisorMaxOutputTokens int64`

2. **Track usage per outer run.**
   - Reset the counter for each `processChat()` invocation.
   - Return `remaining_uses` in the tool result.
   - Return `limit_reached` when the cap is exhausted.

3. **Expose advisor usage metadata in the tool result.**
   - Include model name and token/cost summary if available.
   - Use the same `callConfig.Cost` calculation path as the outer chat for MVP if advisor reuses the same model.

4. **Record server-side metrics.**
   - Count advisor invocations, failures, and latency.
   - Ensure they show up under the built-in tool label `advisor`.

5. **Optional decision gate: separate advisor model.**
   - If product insists on a stronger/different advisor model, add a follow-up config hook that resolves another existing chat model config through the same `configCache` path.
   - Keep that out of the first landing PR unless it is required for acceptance.

6. **Optional decision gate: queryable advisor cost.**
   - If this becomes required, spin a follow-up DB task:
     - update `coderd/database/queries/*.sql`;
     - add migration files;
     - run `make gen`;
     - update audit mappings if a new auditable type/field is introduced.

### Acceptance criteria

- Advisor calls are capped per outer run.
- Limit exhaustion is user-visible in the tool result.
- Metrics distinguish advisor calls from other built-in tools.
- MVP does not require a schema migration unless explicitly approved.

## Phase 5 — Frontend rendering and Storybook coverage

### Goals

Make advisor feel intentional in the Agents UI without blocking the backend on fancy streaming UI.

### Files to modify

- `site/src/pages/AgentsPage/components/ChatElements/tools/Tool.tsx`
- new `site/src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.tsx`
- Storybook story file(s) in the same tools directory

### Delivery strategy

1. **Intermediate milestone during backend bring-up:** rely on the existing generic tool renderer if needed.
   - This is acceptable only as a short-lived integration checkpoint.

2. **Release milestone:** add a dedicated lightweight `AdvisorTool` renderer.
   - Reuse existing primitives:
     - `ToolCollapsible`
     - `ToolIcon`
     - `Response` for markdown/prose rendering
     - `ScrollArea` if the advice can be long
   - Keep styling light and consistent with the Agents page.
   - Do not add unnecessary React memoization in `site/src/pages/AgentsPage/`; that area is already React-Compiler aware.

3. **Render the structured result states cleanly.**
   - `advice` — readable prose/markdown with optional metadata footer.
   - `limit_reached` — warning-style message.
   - `error` — error state with visible fallback text.
   - `running` — existing tool loading state/spinner is enough for MVP.

4. **Add Storybook coverage instead of ad-hoc component tests.**
   Recommended stories:
   - successful advice;
   - running/loading;
   - limit reached;
   - error.

5. **Keep the UI contract narrow.**
   - Prefer one text field like `advice` plus small metadata rather than a deeply nested schema.
   - That keeps the UI resilient to prompt iteration.

### Acceptance criteria

- The advisor tool card renders readable content rather than raw quoted JSON in the final release branch.
- Running, limit, and error states are visibly distinct.
- Storybook stories and play assertions cover the new states.
- Existing tool rendering flows remain unchanged.

## Phase 6 — Automated tests and validation gates

### Backend tests to add

1. **Advisor runtime/tool tests**
   - question validation;
   - tool-less nested execution assertion;
   - success result shaping;
   - limit-reached result shaping;
   - error result shaping.

2. **Prompt/gating tests in chatd**
   - advisor disabled ⇒ no tool, no guidance;
   - advisor enabled/root chat ⇒ tool + guidance;
   - child chat ⇒ advisor absent.

3. **Chatloop policy tests**
   - advisor alone runs;
   - advisor + action tool mixed batch returns deterministic policy errors;
   - non-advisor tools still execute normally.

4. **Usage/metrics tests**
   - per-run cap resets correctly;
   - builtin tool labeling includes `advisor`;
   - returned metadata includes model/usage summary when available.

### Frontend tests to add

- Storybook `play()` assertions for the advisor renderer states.
- Verify expand/collapse behavior and visible fallback text.
- Verify the message timeline still renders adjacent tools correctly.

### Recommended command sequence

Run these as the implementation matures, not only at the end:

1. Backend-focused gate after phases 1–4:
   - `make test RUN=TestAdvisor`
   - `make test RUN=TestChatloopAdvisor`
   - `make lint`

2. Frontend-focused gate after phase 5:
   - `pnpm test:storybook src/pages/AgentsPage/components/ChatElements/tools/AdvisorTool.stories.tsx`
   - `pnpm lint`
   - `pnpm format`

3. Final repo gate before handoff:
   - `make pre-commit`
   - run any additional targeted `make test RUN=...` selections covering touched chatd paths

> Use the exact new test names the implementing agents create; the names above are recommended anchors, not existing tests.

## Dogfooding plan

### Principle

Dogfood the change as a real agent feature, not just a unit-tested backend. Per the dogfood and `agent-browser` skills, the reviewer should get **watchable repro videos** plus screenshots that make the behavior obvious without reading logs.

### Required setup

1. Start the full dev environment with:
   - `./scripts/develop.sh`
2. If the frontend renderer changes, also start Storybook from `site/` with:
   - `pnpm storybook --no-open`
3. Use `agent-browser` directly — **never `npx agent-browser`**.
4. Use named browser sessions and an output folder such as:
   - `./dogfood-output/advisor/`
   - with subfolders `screenshots/` and `videos/`

### Evidence protocol

For every interactive scenario below:

1. Start video recording **before** the action.
2. Capture step-by-step screenshots at human pace.
3. Capture one annotated screenshot of the final state.
4. Stop the recording.
5. Note the exact pass/fail observation in the QA report.

For static UI states (for example Storybook error/limit cards), an annotated screenshot is sufficient; video is optional but still encouraged by this project’s review preference.

### Dogfood scenarios

#### Scenario A — Happy path in the real Agents UI

**Goal:** prove that a root agent chat can invoke advisor and produce a readable recommendation before taking further action.

Steps:

1. Open the Agents page with an advisor-enabled root chat.
2. Start a repro video.
3. Send a prompt that should reasonably trigger strategic planning, such as an architecture or multi-tradeoff question.
4. Capture screenshots of:
   - the prompt before send;
   - the running advisor state;
   - the completed advisor card and the assistant’s follow-up response.
5. Stop recording.

Pass criteria:

- advisor appears in the timeline;
- the rendered result is readable;
- the assistant can continue after consuming the advisor output.

#### Scenario B — Advisor unavailable path

**Goal:** prove the feature is truly gated.

Suggested variants (at least one is required, both are better):

- feature flag/config off;
- child/sub-agent chat.

Evidence:

- annotated screenshot of the chat/tool state showing advisor is absent;
- short video if toggling the gate live is part of the repro.

Pass criteria:

- no advisor tool is available;
- no advisor-specific prompt behavior leaks through.

#### Scenario C — UI states in Storybook

**Goal:** prove the renderer handles non-happy states cleanly.

Required story states:

- success/advice;
- running;
- limit reached;
- error.

Evidence:

- one screenshot per state;
- at least one short video showing collapse/expand behavior.

Pass criteria:

- success renders readable advice;
- limit/error have visible fallback text;
- the component behaves like the other tool cards.

#### Scenario D — Regression sweep of nearby tools

**Goal:** ensure advisor does not break the surrounding chat timeline.

Check at minimum:

- another existing built-in tool still renders correctly near advisor;
- sub-agent/tool cards still expand/collapse normally;
- no obvious console errors appear in the Agents page during the advisor flow.

Evidence:

- screenshots of adjacent tool cards;
- console/error capture if anything suspicious appears.

### `agent-browser` usage notes for the QA agent

- Prefer `agent-browser batch` for 2+ sequential commands when no intermediate parsing is needed.
- Use `snapshot -i` to discover interactive refs.
- Re-snapshot after navigation or major DOM changes.
- Avoid `wait --load networkidle` unless the page is known to go idle; prefer explicit element/text waits or short fixed waits.
- Record videos at human pace and include pauses that a reviewer can follow.

## Rollout plan

### Initial rollout

- Gate behind a server-side advisor-enabled flag.
- Enable only for selected internal/root agent chats first.
- Watch metrics for:
  - invocation count;
  - failure rate;
  - latency;
  - obvious retry loops.

### Expansion conditions

Expand beyond the initial rollout only after the following are true:

- mixed-batch policy behavior is stable;
- cost impact is understood;
- frontend UX is readable in production-like dogfood;
- no recursion surprises have appeared with sub-agent flows.

### Explicit non-goals for the first release

- advisor inside child/sub-agent chats;
- provider-agnostic streaming phase UI;
- MCP-based external advisor implementation;
- mandatory DB-backed advisor cost reporting.

## Final acceptance checklist

- [ ] `advisor` is a built-in chatd tool, not an MCP/dynamic-tool substitute.
- [ ] The nested advisor call is tool-less and bounded to one in-memory step.
- [ ] One eligibility boolean controls both tool registration and prompt guidance injection.
- [ ] Root chats can use advisor; child chats cannot in the initial rollout.
- [ ] Mixed advisor/action batches produce deterministic policy errors instead of partial execution.
- [ ] Per-run usage caps and limit-reached behavior work.
- [ ] Advisor usage is visible in metadata/metrics without forcing a DB migration for MVP.
- [ ] The Agents UI has a readable advisor card and Storybook coverage.
- [ ] Dogfooding produced screenshots and repro videos for the required scenarios.
- [ ] Validation commands (`make lint`, targeted `make test`, Storybook tests, `make pre-commit`) passed before handoff.

## Suggested PR split

1. **PR 1 — Backend foundation**
   - `chatadvisor/` package
   - `chattool/advisor.go`
   - `chatloop` exclusive policy
   - chatd gating/prompt sync
   - backend tests

2. **PR 2 — Frontend + QA**
   - advisor renderer
   - stories/play assertions
   - dogfood artifacts and QA notes

3. **PR 3 — Optional follow-ups only if demanded by stakeholders**
   - separate advisor model override
   - persistent advisor billing/queryability
   - transient phase-stream UX


</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-7` • Thinking: `max`_
 2026-04-30 14:08:59 +02:00
Thomas Kosiewski ab75e46f1d fix: record debug runs for proposed chat titles (#24820) 2026-04-29 16:45:48 +02:00
Thomas Kosiewski 88b62a3359 fix(site/src/pages/AgentsPage): add copy buttons to raw attempts (#24705) 2026-04-24 20:13:10 +02:00
Thomas Kosiewski 5c316d4252 test(site/src/pages/AgentsPage): add Debug panel Storybook stories (#23921) 
## Summary

Add comprehensive Storybook stories covering the Debug panel in various states: loading, empty, single/multiple runs, expanded steps, tool calls, error states, and streaming indicators.

This is PR 9/9 in the chat debug logging stack.

### Changes

- **DebugPanel stories** (`site/src/pages/AgentsPage/components/RightPanel/DebugPanel/DebugPanel.stories.tsx`): Storybook stories with `play` functions for interaction testing, covering:
  - Loading state
  - Empty / disabled state
  - Single-step successful run
  - Multi-step run with tool calls
  - Error state with redacted headers
  - Compaction and title generation run kind badges
  - Long raw request/response payload rendering
  - Streaming / in-progress indicators

### Stack overview

1. Database schema & SDK types
2. Types, context, and model normalization
3. Recorder, transport, and redaction
4. Service and summary aggregation
5. Chat lifecycle wiring
6. HTTP handlers and API docs
7. Frontend API layer and panel utilities
8. Debug panel components and settings
9. **→ Storybook stories** (this PR)

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-6` • Thinking: `xhigh`_
 2026-04-22 15:44:32 +02:00
Thomas Kosiewski 249b71b96a feat(site): add Debug panel components and settings (#23920) 
## Summary

Add the Debug panel UI components: run list, run cards, step cards with transcript and tool-call rendering, attempt accordions, and shared primitives. Wire the panel into the AgentChatPage sidebar and add the per-chat debug logging toggle in the behavior settings page.

This is PR 8/9 in the chat debug logging stack.

### Screenshots

Settings Page

<img width="4608" height="2348" alt="CleanShot 2026-04-21 at 21 19 36@2x" src="https://github.com/user-attachments/assets/69391465-4c56-468a-9923-59576d326963" />

Conditional Debug tab

<img width="4608" height="2348" alt="CleanShot 2026-04-21 at 21 19 58@2x" src="https://github.com/user-attachments/assets/bc1e07cb-21d9-40e7-8928-6fd9a7ec7f57" />

Last request's tools and schema 

<img width="4608" height="2348" alt="CleanShot 2026-04-21 at 21 20 03@2x" src="https://github.com/user-attachments/assets/401f26af-98ce-443f-a586-424d3636d98b" />

"Raw" JSON request bodies

<img width="4608" height="2348" alt="CleanShot 2026-04-21 at 21 20 35@2x" src="https://github.com/user-attachments/assets/3605a373-9e29-4183-89e7-8b2704ff9333" />


### Changes

- **DebugPanel** (`site/src/pages/AgentsPage/components/RightPanel/DebugPanel/DebugPanel.tsx`): top-level panel component owning data fetching and subscription lifecycle.
- **DebugRunCard**: compact single-row header with capitalized provider name, status badge, compact duration (`1.3s`), and token summary (`3→5 tok`). Expandable to show child step cards.
- **DebugStepCard**: step inspector with normalized transcript rendering — system prompts, assistant text with 160-char clamping + independent "see more/less" toggle, tool calls with fully-expanded JSON payloads in `CopyableCodeBlock`.
- **DebugAttemptAccordion**: nested accordion for HTTP-level attempt details showing request/response headers and bodies.
- **Shared primitives**: `CopyableCodeBlock`, `MessageRow`, `ToolPayloadDisclosure`, `StatusBadge`.
- **Sidebar wiring** (`AgentChatPageView.tsx`): adds the Debug tab to the right panel when debug logging is enabled.
- **Behavior settings**: deployment-wide and per-user debug logging toggles on the settings page.

### Stack overview

1. Database schema & SDK types
2. Types, context, and model normalization
3. Recorder, transport, and redaction
4. Service and summary aggregation
5. Chat lifecycle wiring
6. HTTP handlers and API docs
7. Frontend API layer and panel utilities
8. **→ Debug panel components and settings** (this PR)
9. Storybook stories

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-6` • Thinking: `xhigh`_
 2026-04-22 15:34:32 +02:00
Thomas Kosiewski 8c0fe6d5f2 feat(site): add chat debug API layer and panel utilities (#23919) 
## Summary

Add API client methods, React Query builders, and unit tests for the chat debug endpoints. Add `debugPanelUtils` with coercion helpers that transform raw debug step data into structured display models for the Debug panel, and wire debug run streaming into the chat store.

This is PR 7/9 in the chat debug logging stack.

### Changes

- **API client** (`site/src/api/api.ts`): typed methods for all debug endpoints — list runs, list steps, get/set deployment logging, get/set user logging, set per-chat override.
- **React Query builders** (`site/src/api/queries/chats.ts`): `chatDebugRuns`, `chatDebugSteps`, `chatDebugLoggingConfig`, `userDebugLoggingConfig` query/mutation factories with `refetchInterval: 5000` for live polling.
- **Debug panel utilities** (`site/src/pages/AgentsPage/components/RightPanel/DebugPanel/debugPanelUtils.ts`): `coerceStepRequest` / `coerceStepResponse` that recursively parse nested JSONB into `StepRequestViewModel` / `StepResponseViewModel` with `MessagePart`, `ToolDef`, and `ToolCallPart` types. Includes `formatTokenSummary` (compact `3→5 tok` notation) and `compactDuration` helpers.
- **Unit tests**: coverage for coercion edge cases and formatting utilities.

### Stack overview

1. Database schema & SDK types
2. Types, context, and model normalization
3. Recorder, transport, and redaction
4. Service and summary aggregation
5. Chat lifecycle wiring
6. HTTP handlers and API docs
7. **→ Frontend API layer and panel utilities** (this PR)
8. Debug panel components and settings
9. Storybook stories

---
_Generated with [`mux`](https://github.com/coder/mux) • Model: `anthropic:claude-opus-4-6` • Thinking: `xhigh`_
 2026-04-22 15:24:32 +02:00
Thomas Kosiewski b7c2c59931 fix(coderd/x/chatd/chatdebug): allow Anthropic per-modality ratelimit headers (#24592) 
Previously, Anthropic's per-modality, Priority Tier, and fast-mode rate-limit headers (`Anthropic-Ratelimit-Input-Tokens-*`, `Anthropic-Ratelimit-Output-Tokens-*`, `Anthropic-Priority-Input-Tokens-*`, `Anthropic-Priority-Output-Tokens-*`, `Anthropic-Fast-Input-Tokens-*`, and `Anthropic-Fast-Output-Tokens-*`) were shown as `[REDACTED]` in the Debug panel because they contain `"token"` in the name and fell through the generic credential filter.

Add them to the allowlist in `coderd/x/chatd/chatdebug/redaction.go` alongside the existing `Anthropic-Ratelimit-Tokens-*` entries so the limits/remaining/reset values surface in the raw response view.
 2026-04-22 15:14:31 +02:00
Thomas Kosiewski 26b64fa523 fix(coderd/x/chatd/chatdebug): record SSE attempts on EOF (#24565) 
`chat_turn` debug steps persist with `attempts: []` even when the
streaming call to Anthropic completes successfully. Fantasy's
Anthropic SSE adapter iterates the response to EOF via
`for stream.Next()` and abandons the body without calling `Close()`,
so `RecordingTransport`'s Close-only recording path never fires and
the attempt is lost. Non-streaming runs (`quickgen`,
`title_generation`) go through `model.Generate(...)` and are
unaffected.

Record on `io.EOF` for `text/event-stream` bodies specifically.
Non-SSE responses stay on the Close-only path so JSON integrity,
content-length validation, and inner-`Close()` error semantics are
preserved. `record()` is already `sync.Once`-guarded, so a later
`Close()` is a no-op for recording.
 2026-04-22 15:02:02 +02:00
Thomas Kosiewski 18a30a7a10 feat: add chat debug HTTP handlers and API docs (#23918) 2026-04-20 13:34:41 +02:00
Thomas Kosiewski df7e838c21 feat(coderd): wire debug logging into chat lifecycle (#23917) 2026-04-20 12:27:16 +02:00
Thomas Kosiewski 91f9de27a1 feat(coderd): add chat debug service and summary aggregation (#23916) 2026-04-17 16:27:53 +02:00
Thomas Kosiewski 4651ca5a9a feat(coderd/x/chatd/chatdebug): add recorder, transport, and redaction (#23915) 2026-04-15 15:14:51 +02:00
Thomas KosiewskiandCian Johnston 5812f84e1c fix(coderd): validate webpush subscription endpoints (#24347) 
Co-authored-by: Cian Johnston <cian@coder.com>
 2026-04-15 11:31:43 +02:00
Thomas Kosiewski 8382e96a81 feat: add types, context, and model normalization (#23914) 2026-04-13 19:59:47 +02:00
Thomas Kosiewski 6ab30123bf feat: add chat debug log tables, queries, and SDK types (#23913) 2026-04-13 15:06:06 +02:00
Thomas Kosiewski 12f87acad6 feat(site): add terminal panel to chat sidebar (#23231) 
## Add terminal panel to chat sidebar

Extract the reusable terminal runtime from `TerminalPage` into
`modules/terminal/` and wire it into the agents chat right sidebar as a
new **Terminal** tab.

### Changes

- **`modules/terminal/WorkspaceTerminal.tsx`** — Shared xterm +
websocket terminal component (container-sized, no route dependency)
- **`modules/terminal/WorkspaceTerminalAlerts.tsx`** — Moved from
`TerminalPage/` to shared module
- **`pages/AgentsPage/TerminalPanel.tsx`** — Sidebar wrapper around
`WorkspaceTerminal`
- **`pages/AgentsPage/AgentDetailView.tsx`** — Terminal tab added (gated
on `hasWorkspace`)
- **`pages/TerminalPage/TerminalPage.tsx`** — Slimmed to page-shell
using shared component

### Demo


[dogfood-terminal-demo.webm](https://github.com/user-attachments/assets/359200dc-f8e4-4a9a-b00b-923f142dc228)

### Behavior

- Terminal tab appears only when the chat has a workspace with a
connected agent
- Connects via the existing workspace agent PTY websocket
- Resizes correctly on panel width changes, expand/collapse, and
viewport resize
- `fitAddon.fit()` guarded against pre-renderer crashes (fixes proxy
access)
- Tab switching unmounts/remounts cleanly (reconnects via session token)
- No changes to Git or Desktop panel behavior

---

_Generated with [`mux`](https://github.com/coder/mux) • Model:
`anthropic:claude-opus-4-6` • Thinking: `xhigh`_
 2026-04-01 13:38:23 +00:00
Thomas Kosiewski ea37f1ff86 feat: pass session token as query param on agent chat WebSockets (#23405) 
## Problem

When the Coder chat UI is embedded in a VS Code webview, the session
token is set via the Coder-Session-Token header for HTTP requests.
However, browsers cannot attach custom headers to WebSocket connections,
and VS Code Electron webview environment does not support cookies set
via Set-Cookie from iframe origins. This causes all chat WebSocket
connections to fail with authorization errors.

## Solution

Pass the session token as a coder_session_token query parameter on all
chat-related WebSocket connections. The backend already accepts this
parameter (see APITokenFromRequest in coderd/httpmw/apikey.go).

The token is only included when API.getSessionToken() returns a value,
which only happens in the embed bootstrap flow. Normal browser sessions
use cookies and are unaffected.

> Built with [Coder Agents](https://coder.com/agents)
 2026-03-23 15:27:55 +01:00
Thomas Kosiewski 83809bb380 feat: add token-to-cookie endpoint for embedded chat WebSocket auth (#23280) 
## Problem

The VS Code extension embeds the Coder agent chat UI in an iframe,
passing the session token via `postMessage`. HTTP requests use the
`Coder-Session-Token` header, but browser WebSocket connections **cannot
carry custom headers** — they rely on cookies. This causes all WebSocket
requests (e.g. streaming chat messages) to fail with authorization
errors in the embedded iframe.

## Solution

Add `POST /api/v2/users/me/session/token-to-cookie` — a lightweight
endpoint that converts the current (already-validated) session token
into a `Set-Cookie` response. The frontend embed bootstrap flow calls
this immediately after `API.setSessionToken(token)`, before any
WebSocket connections are opened.

### Backend (`coderd/userauth.go`, `coderd/coderd.go`)
- New handler `postSessionTokenCookie` behind `apiKeyMiddleware`.
- Reads the validated token via `httpmw.APITokenFromRequest(r)`.
- Sets an `HttpOnly` cookie with the API key's expiry, applying
site-wide cookie config (Secure, SameSite, host prefix) via
`HTTPCookies.Apply`.
- Returns `204 No Content`.

### Frontend (`site/src/pages/AgentsPage/EmbedContext.tsx`)
- `bootstrapChatEmbedSessionFn` now calls the new endpoint after setting
the header token and before fetching user/permissions.
- The cookie is in place before any WebSocket connections are opened.

## Security

- **No privilege escalation**: The token is already valid — this just
moves it from a header credential to a cookie credential.
- **POST only**: Avoids CSRF-via-navigation.
- **Same origin**: The iframe loads from the Coder server, so the cookie
applies to the correct domain.
- **HttpOnly**: The cookie is not accessible to JavaScript.

> Built with [Coder Agents](https://coder.com/agents) 🤖
 2026-03-19 16:12:31 +01:00
Thomas Kosiewski 20ac96e68d feat(site): include chatId in editor deep links (#23214) 
## Summary

- include the current agent chat ID in VS Code and Cursor deep links
opened from the agent detail page
- extend `getVSCodeHref` so `chatId` is added only when provided
- add focused tests for deep-link generation with and without `chatId`

## Testing

- `pnpm -C site run format -- src/modules/apps/apps.ts
src/modules/apps/apps.test.ts src/pages/AgentsPage/AgentDetail.tsx`
- `pnpm -C site run check -- src/modules/apps/apps.ts
src/modules/apps/apps.test.ts src/pages/AgentsPage/AgentDetail.tsx`
- `pnpm -C site exec vitest run src/modules/apps/apps.test.ts`
- `pnpm -C site run lint:types`

---
_Generated with [`mux`](https://github.com/coder/mux) • Model:
`openai:gpt-5.4` • Thinking: `high`_
 2026-03-18 14:25:38 +01:00
Thomas Kosiewski c1884148f0 feat: add VS Code iframe embed auth bootstrap (#23060) 
## VS Code iframe embed auth via postMessage + setSessionToken

Adds embed auth for VS Code iframe integration, allowing the Coder agent
chat UI to be embedded in VS Code webviews without manual login — using
direct header auth instead of cookies.

### How it works

1. **Parent frame** (VS Code webview) loads an iframe pointing to
`/agents/:agentId/embed`
2. **Embed page** detects the user is signed out and posts
`coder:vscode-ready` to the parent
3. **Parent** responds with `coder:vscode-auth-bootstrap` containing the
user's Coder API token
4. **Embed page** calls `API.setSessionToken(token)` to set the
`Coder-Session-Token` header on all subsequent axios requests
5. **Embed page** fetches user + permissions, sets them in the React
Query cache atomically, and renders the authenticated agent chat UI

No cookies, no CSRF, no backend endpoint needed. The token is passed via
postMessage and used as a header on every API request.

### What changed

**Frontend** (`site/src/pages/AgentsPage/`):
- `AgentEmbedPage.tsx` — added postMessage bootstrap directly in the
embed page: listens for `coder:vscode-auth-bootstrap`, calls
`API.setSessionToken(token)`, fetches user/permissions atomically to
avoid race conditions
- `EmbedContext.tsx` — React context signaling embed mode (from previous
commit, unchanged)
- `AgentDetail/TopBar.tsx` — conditionally hides navigation elements in
embed mode (from previous commit, unchanged)
- Both `/agents/:agentId/embed` and `/agents/:agentId/embed/session`
routes live outside `RequireAuth`

**Auth bootstrap** (`site/src/api/queries/users.ts`):
- `bootstrapChatEmbedSessionFn` now calls `API.setSessionToken(token)`
instead of posting to a backend endpoint
- Fetches user and permissions directly via `API.getAuthenticatedUser()`
and `API.checkAuthorization()`, then sets both in the query cache
atomically — this avoids a race where `isSignedIn` flips before
permissions are loaded

**Removed** (no longer needed):
- `coderd/embedauth.go` — the `POST
/api/experimental/chats/embed-session` handler
- `coderd/embedauth_test.go` — backend tests for the endpoint
- `codersdk/embedauth.go` — `EmbedSessionTokenRequest` SDK type
- `site/src/api/api.ts` — `postChatEmbedSession` method
- `docs/user-guides/workspace-access/vscode-embed-auth.md` — doc page
for the old cookie flow
- Swagger/API doc entries for the endpoint

### Why not cookies?

The initial implementation used a backend endpoint to set an HttpOnly
session cookie. This required `SameSite=None; Secure` for cross-origin
iframes, which doesn't work over HTTP in development (Chrome requires
HTTPS for `Secure` cookies). The `setSessionToken` approach bypasses
cookies entirely — the token is set as an axios default header, and
header-based auth also naturally bypasses CSRF protection.

### Dogfooding

Tested end-to-end with a VS Code extension that:
1. Registers a `/openChat` deep link handler
(`vscode://coder.coder-remote/openChat?url=...&token=...&agentId=...`)
2. Starts a local HTTP reverse proxy (to work around VS Code webview
iframe sandboxing)
3. Loads `/agents/:agentId/embed` in an iframe through the proxy
4. Relays the postMessage handshake between the iframe and the extension
host
5. The embed page receives the token, calls `setSessionToken`, and
renders the chat

Verified: chat title, messages, and input field all display correctly in
VS Code's secondary sidebar panel.
 2026-03-16 17:45:01 +01:00
Thomas Kosiewski d6306461bb feat(site): render computer tool screenshots as images in chat UI (#23074) 
Instead of showing raw base64 JSON for Anthropic's computer use tool,
render the screenshot as an inline image. The image is clickable to open
at full resolution in a new tab.

## Changes

- **ComputerTool.tsx** — New component that renders base64 image data as
an `<img>` tag
- **Tool.tsx** — Added `ComputerRenderer` handling both single-object
and array-of-blocks result shapes
- **ToolIcon.tsx** — Added `MonitorIcon` for the `computer` tool
- **ToolLabel.tsx** — Added \Screenshot\ label for the `computer` tool
 2026-03-16 12:36:18 +01:00
Thomas Kosiewski 069d3e2beb fix(coderd): require ssh access for workspace chats (#23094) 
### Motivation
- The chat creation flow associated a workspace agent for a chat if the requester could read the workspace, enabling privilege escalation where users without SSH/app-connect permissions could cause the daemon to open privileged agent connections and execute commands.
- The intent is to ensure that attaching a workspace agent to a chat only happens when the requester has the workspace SSH permission so the chat daemon cannot be abused to bypass RBAC.

### Description
- Require request-scoped authorization for workspace agent usage by changing `validateCreateChatWorkspaceSelection` to accept the `*http.Request` and calling `api.Authorize(r, policy.ActionSSH, workspace)` before selecting the workspace for a chat.
- Pass the HTTP request into the validator from `postChats` so authorization is evaluated in the request context (`postChats` now calls `validateCreateChatWorkspaceSelection(ctx, r, req)`).
- Add a regression test `WorkspaceAccessibleButNoSSH` in `coderd/chats_test.go` which creates an org-admin-scoped user (read access but no `ActionSSH`) and asserts that creating a chat with `WorkspaceID` is denied.

### Testing
- Ran `gofmt -w coderd/chats.go coderd/chats_test.go` which succeeded.
- Attempted to run repository pre-commit checks (`make pre-commit`) and targeted `go test` invocations; these checks could not be completed in this environment due to missing local tooling and environment constraints (protobuf include resolution, containerized DB access via Docker socket, and long-running golden generation tasks), so full CI/pre-commit verification and end-to-end test runs did not complete here.
- Added a focused regression unit test (`WorkspaceAccessibleButNoSSH`) to prevent reintroduction of the authorization bypass; this test is included in the change and should be executed in CI where the full toolchain and test environment are available.

------
[Codex Task](https://chatgpt.com/codex/tasks/task_b_69b432502670832e91d14e937745de46)
 2026-03-16 11:42:01 +01:00
Thomas Kosiewski 3f7f25b3ee fix(chats): enforce desktop connect authorization (#23073) 
### Motivation

- The desktop watch handler opened a VNC stream using the chat's
workspace ID while only relying on workspace read permissions, allowing
read-only users to escalate to interactive desktop access.
- Enforce connect-level authorization so only actors with
`ActionApplicationConnect` or `ActionSSH` can open the desktop stream.

### Description

- Added an explicit workspace lookup in `watchChatDesktop` using
`GetWorkspaceByID` to obtain a workspace object for authorization.
- Require the requester to be authorized for either
`policy.ActionApplicationConnect` or `policy.ActionSSH` on the workspace
before proceeding to locate agents or connect to the VNC stream, and
return `403 Forbidden` when neither permission is present.
- The change is minimal and localized to `coderd/chats.go` and does not
alter other code paths or behavior when the requester has the necessary
connect permissions.

### Testing

- Ran `gofmt -w coderd/chats.go` to format the modified file, which
succeeded.
- Attempted to run the unit test `TestWatchChatDesktop/NoWorkspace` via
`go test` in this environment but the test run did not complete within
the environment constraints and did not produce a full pass result.
- Attempted to run the repository pre-commit/gen steps but they could
not complete due to missing developer tooling and services in this
environment (e.g. `sqlc`, `mockgen`, `protoc` plugins and test services
like Docker/Postgres), so full pre-commit validation did not finish
here.
- Code review and static validation confirm the added authorization
check properly prevents read-only access from opening the desktop VNC
stream.

------
[Codex
Task](https://chatgpt.com/codex/tasks/task_b_69b46a4ac5c4832ea9d330aeba43c32d)
 2026-03-14 17:53:05 +01:00
Thomas Kosiewski e96cd5cbb2 chore(githooks): remove pre-push hook (#22956) 
## Summary
- remove the `pre-push` git hook script from the repository
- remove the `make pre-push` target and related Makefile documentation
- update contributor and agent docs so they only describe the remaining
`pre-commit` hook

## Validation
- `make pre-commit`
- `git diff --check`

---
_Generated with [`mux`](https://github.com/coder/mux) • Model:
`openai:gpt-5.4` • Thinking: `high`_
 2026-03-11 17:44:19 +01:00
Thomas Kosiewski b776a14b46 fix(coderd): harden OAuth2 provider security (#22194) 
## Summary

Harden the OAuth2 provider with multiple security fixes addressing
`coder/security#121` (CSRF session takeover) and converge on OAuth 2.1
compliance.

### Security Fixes

| Fix | Description | Commits |
|-----|-------------|---------|
| **CSRF on `/oauth2/authorize`** | Enforce CSRF protection on the
authorize endpoint POST (consent form submission) | `ba7d646`, `b94a64e`
|
| **Clickjacking: `frame-ancestors` CSP** | Prevent consent page from
being iframed (`Content-Security-Policy: frame-ancestors 'none'` +
`X-Frame-Options: DENY`) | `597aeb2` |
| **Exact redirect URI matching** | Changed from prefix matching to full
string exact matching per OAuth 2.1 §4.1.2.1 | `73d64b1`, `93897f1` |
| **Store & verify `redirect_uri`** | Store redirect_uri with auth code
in DB, verify at token exchange matches exactly (RFC 6749 §4.1.3) |
`50569b9`, `d7ca315` |
| **Mandatory PKCE** | Require `code_challenge` at authorization (for
`response_type=code`) + unconditional `code_verifier` verification at
token exchange | `d7ca315`, `1cda1a9` |
| **Reject implicit grant** | `response_type=token` now returns
`unsupported_response_type` error page (OAuth 2.1 removes implicit flow)
| `d7ca315`, `91b8863` |

### Changes by File

**`coderd/httpmw/csrf.go`** — Extended the CSRF `ExemptFunc` to enforce
CSRF on `/oauth2/authorize` in addition to `/api` routes. The consent
form POST is now CSRF-protected to prevent cross-site authorization code
theft.

**`site/site.go`** — Added `Content-Security-Policy: frame-ancestors
'none'` and `X-Frame-Options: DENY` headers to `RenderOAuthAllowPage`
(consent page only — does not affect the SPA/global CSP used by AI
tasks).

**`coderd/httpapi/queryparams.go`** — Changed `RedirectURL` from prefix
matching (`strings.HasPrefix(v.Path, base.Path)`) to full URI exact
matching (`v.String() != base.String()`), comparing scheme, host, path,
and query.

**`coderd/oauth2provider/authorize.go`** — Added PKCE enforcement:
`code_challenge` is required when `response_type=code` (via a
conditional check, not `RequiredNotEmpty`, so `response_type=token` can
reach the explicit rejection path). `ShowAuthorizePage` (GET) validates
`response_type` before rendering and returns a 400 error page for
unsupported types. `ProcessAuthorize` (POST) stores the `redirect_uri`
with the auth code when explicitly provided.

**`coderd/oauth2provider/tokens.go`** — PKCE verification is now
unconditional (not gated on `code_challenge` being present in DB). If
the stored code has a `redirect_uri`, the token endpoint verifies it
matches exactly — mismatch returns `errBadCode` → `invalid_grant`.
Missing `code_verifier` returns `invalid_grant`.

**`codersdk/oauth2.go`** — `OAuth2ProviderResponseTypeToken` constant
and `Valid()` acceptance are **kept** so the authorize handler can parse
`response_type=token` and return the proper `unsupported_response_type`
error rather than failing at parameter validation.

**`coderd/database/migrations/000421_*`** — Added `redirect_uri text`
column to `oauth2_provider_app_codes`.

### Design Decisions

**`state` parameter remains optional** — The plan initially required
`state` via `RequiredNotEmpty`, but this was reverted in `376a753` to
avoid breaking existing clients. The `state` is still hashed and stored
when provided (via `state_hash` column), securing clients that opt in.

**`response_type=token` kept in `Valid()`** — Removing it from `Valid()`
would cause the parameter parser to reject the request before the
authorize handler can return the proper `unsupported_response_type`
error. The constant is kept for correct error handling flow.

**CSP scoped to consent page only** — `frame-ancestors 'none'` is set
only on the OAuth consent page renderer, not globally. The SPA/global
CSP was previously changed to allow framing for AI tasks
([#18102](https://github.com/coder/coder/pull/18102)); this change does
not regress that.

### Out of Scope (follow-up PRs)

- Bearer tokens in query strings (needs internal caller audit)
- Scope enforcement on OAuth2 tokens
- Rate limiting on dynamic client registration


---

<details>
<summary>📋 Implementation Plan</summary>

# Plan: Harden OAuth2 Provider — Security Fixes + OAuth 2.1 Compliance

## Context & Why

Security issue `coder/security#121` reports a critical session takeover
via CSRF on the OAuth2 provider. This plan covers all remaining security
fixes from that issue **plus** convergence on OAuth 2.1 requirements.
The goal is a single PR that closes all actionable gaps.

## Current State (already committed on branch `csrf-sjx1`)

| Fix | Status | Commits |
|-----|--------|---------|
| Fix 1: CSRF on `/oauth2/authorize` | ✅ Done | `ba7d646`, `b94a64e` |
| CSRF token in consent form HTML | ✅ Done | `b94a64e` |
| `state_hash` column + storage | ✅ Done (hash stored, but state still
optional) | `9167d83`, `b94a64e` |
| Tests for CSRF + state hash | ✅ Done | `e4119b5` |

## Remaining Work

### ~~Fix 2 — Require `state` parameter~~ (DROPPED)

> **Decision:** Do not enforce `state` as required. The `state`
parameter is still hashed and stored when provided (via
`hashOAuth2State` / `state_hash` column from prior commits), but clients
are not forced to supply it. This avoids breaking existing integrations
that omit state.

**Rollback:** Remove `"state"` from the `RequiredNotEmpty` call in
`coderd/oauth2provider/authorize.go:42`:

```go
// BEFORE (current on branch)
p.RequiredNotEmpty("response_type", "client_id", "state", "code_challenge")

// AFTER
p.RequiredNotEmpty("response_type", "client_id", "code_challenge")
```

No test changes needed — tests already pass `state` voluntarily.

### Fix 4 — Exact redirect URI matching

Currently `coderd/httpapi/queryparams.go:233` uses prefix matching:

```go
// CURRENT — prefix match
if v.Host != base.Host || !strings.HasPrefix(v.Path, base.Path) {
```

OAuth 2.1 requires **exact string matching**. Change to:

```go
// AFTER — exact match (OAuth 2.1 §4.1.2.1)
if v.Host != base.Host || v.Path != base.Path {
```

**File: `coderd/httpapi/queryparams.go` — `RedirectURL` method**

Also update the error message from "must be a subset of" to "must
exactly match".

**Additionally**, store `redirect_uri` with the auth code and verify at
the token endpoint (RFC 6749 §4.1.3):

1. **New migration** (same migration file or a new `000421`): Add
`redirect_uri text` column to `oauth2_provider_app_codes`
2. **Update INSERT query** in `coderd/database/queries/oauth2.sql` to
include `redirect_uri`
3. **`coderd/oauth2provider/authorize.go`**: Store
`params.redirectURL.String()` when inserting the code
4. **`coderd/oauth2provider/tokens.go`**: After retrieving the code from
DB, verify that `redirect_uri` from the token request matches the stored
value exactly. Currently `tokens.go:103` calls `p.RedirectURL(vals,
callbackURL, "redirect_uri")` for prefix validation only — it must
compare against the stored redirect_uri from the code, not just the
app's callback URL.

<details>
<summary>Why both exact match AND store+verify?</summary>

Exact matching at the authorize endpoint prevents open redirectors
(attacker can't use a sub-path).
Storing and verifying at the token endpoint prevents code injection — an
attacker who steals a code can't exchange it with a different
redirect_uri than was originally authorized. This is required by RFC
6749 §4.1.3 and OAuth 2.1.
</details>

### Fix 7 — `frame-ancestors` CSP on consent page

The consent page can be iframed by a workspace app (same-site), which is
the attack vector. Add a `Content-Security-Policy` header to prevent
framing.

**File: `site/site.go` — `RenderOAuthAllowPage` function (~line 731)**

Before writing the response, add:

```go
func RenderOAuthAllowPage(rw http.ResponseWriter, r *http.Request, data RenderOAuthAllowData) {
    rw.Header().Set("Content-Type", "text/html; charset=utf-8")
    // Prevent the consent page from being framed to mitigate
    // clickjacking attacks (coder/security#121).
    rw.Header().Set("Content-Security-Policy", "frame-ancestors 'none'")
    rw.Header().Set("X-Frame-Options", "DENY")
    ...
```

Both headers for defense-in-depth (CSP for modern browsers,
X-Frame-Options for legacy).

### OAuth 2.1 — Mandatory PKCE

Currently PKCE is checked only when `code_challenge` was provided during
authorization (`tokens.go:258`):

```go
// CURRENT — conditional check
if dbCode.CodeChallenge.Valid && dbCode.CodeChallenge.String != "" {
    // verify PKCE
}
```

OAuth 2.1 requires PKCE for ALL authorization code flows. Change to:

**File: `coderd/oauth2provider/authorize.go`** — Add `"code_challenge"`
to required params:

```go
p.RequiredNotEmpty("response_type", "client_id", "code_challenge")
```

**File: `coderd/oauth2provider/tokens.go:257-265`** — Make PKCE
verification unconditional:

```go
// AFTER — PKCE always required (OAuth 2.1)
if req.CodeVerifier == "" {
    return codersdk.OAuth2TokenResponse{}, errInvalidPKCE
}
if !dbCode.CodeChallenge.Valid || dbCode.CodeChallenge.String == "" {
    // Code was issued without a challenge — should not happen
    // with the authorize endpoint enforcement, but defend in
    // depth.
    return codersdk.OAuth2TokenResponse{}, errInvalidPKCE
}
if !VerifyPKCE(dbCode.CodeChallenge.String, req.CodeVerifier) {
    return codersdk.OAuth2TokenResponse{}, errInvalidPKCE
}
```

**File: `codersdk/oauth2.go`** — Remove
`OAuth2ProviderResponseTypeToken` from the enum or reject it explicitly
in the authorize handler. Currently it's defined at line 216 but the
handler ignores `response_type` and always issues a code. We should
either:
- (a) Remove the `"token"` variant from the enum and reject it with
`unsupported_response_type`, OR
- (b) Add an explicit check in `ProcessAuthorize` that rejects
`response_type=token`

Option (b) is simpler and more backwards-compatible:

```go
// In ProcessAuthorize, after extracting params:
if params.responseType != codersdk.OAuth2ProviderResponseTypeCode {
    httpapi.WriteOAuth2Error(ctx, rw, http.StatusBadRequest,
        codersdk.OAuth2ErrorCodeUnsupportedResponseType,
        "Only response_type=code is supported")
    return
}
```

### OAuth 2.1 — Bearer tokens in query strings

`coderd/httpmw/apikey.go:743` accepts `access_token` from URL query
parameters. OAuth 2.1 prohibits this. However, this may be used
internally (e.g., workspace apps, DERP). Need to audit callers before
removing.

**Approach:** This is a larger change with potential breakage. Mark as a
**separate follow-up issue** rather than including in this PR. Document
the finding.

### OAuth 2.1 — Removed flows

✅ **Already compliant.** `tokens.go` only supports `authorization_code`
and `refresh_token` grant types. The implicit grant
(`response_type=token`) will be explicitly rejected per the PKCE section
above.

### OAuth 2.1 — Refresh token rotation

✅ **Already compliant.** `tokens.go:442` deletes the old API key when a
refresh token is used.

## Migration Plan

All DB changes can go in a single new migration (or extend 000420 if the
branch is rebased before merge). Columns to add:
- `redirect_uri text` on `oauth2_provider_app_codes`

The `state_hash` column is already added by migration 000420.

## Implementation Order

1. **Fix 7** — CSP headers on consent page (isolated, no deps)
2. ~~**Fix 2** — Require `state` parameter~~ (DROPPED — state stays
optional)
3. **Fix 4** — Exact redirect URI matching + store/verify redirect_uri
4. **PKCE mandatory** — Require `code_challenge` + reject
`response_type=token`
5. **Rollback** — Remove `"state"` from `RequiredNotEmpty` in
`authorize.go`
6. **Tests** — Update/add tests for all changes
7. **`make gen`** after DB changes

## Out of Scope (separate PRs)

- Bearer tokens in query strings (needs internal caller audit)
- Scope enforcement on OAuth2 tokens
- Rate limiting / quota on dynamic client registration

</details>

---
_Generated with [`mux`](https://github.com/coder/mux) • Model:
`anthropic:claude-opus-4-6` • Thinking: `xhigh`_
 2026-02-23 12:18:44 +01:00
Thomas Kosiewski dd6aec04d7 fix(coderd/oauth2provider): support client_secret_basic client auth (#21793) 2026-02-02 16:01:33 +01:00
Thomas Kosiewski cadf1352b4 feat: add scoped token support to CLI (#19985) 
<!--

If you have used AI to produce some or all of this PR, please ensure you have read our [AI Contribution guidelines](https://coder.com/docs/about/contributing/AI_CONTRIBUTING) before submitting.

-->

Add support for scoped API tokens in CLI

This PR adds CLI support for creating and viewing API tokens with scopes and allow lists. It includes:

- New `--scope` and `--allow` flags for the `tokens create` command
- A new `tokens view` command to display detailed information about a token
- Updated table columns in `tokens list` to show scopes and allow list entries
- Updated help text and examples

These changes enable users to create tokens with limited permissions through the CLI, similar to the existing functionality in the web UI.
 2025-10-27 17:07:25 +01:00
Thomas Kosiewski c6e551f538 fix: renumber api key allow list migration (#20457) 2025-10-24 11:54:51 +00:00
Thomas Kosiewski f684831f56 feat: add allow list to API keys (#19972) 
Add API key allow list to the SDK

This PR adds an allow list to API keys in the SDK. The allow list is a list of targets that the API key is allowed to access. If the allow list is empty, a default allow list with a single entry that allows access to all resources is created.

The changes include:

- Adding a default allow list when generating an API key if none is provided
- Adding allow list to the API key response in the SDK
- Converting database allow list entries to SDK format in the API response
- Adding tests to verify the default allow list behavior



Fixes #19854
 2025-10-24 12:33:56 +01:00
Thomas Kosiewski 251f787743 fix: normalize oauth2 scope parsing (#20359) 2025-10-19 13:20:48 +02:00
Thomas Kosiewski ed90ecf00e feat: add allow_list to resource-scoped API tokens (#19964) 
# Add API key allow_list for resource-scoped tokens

This PR adds support for API key allow lists, enabling tokens to be scoped to specific resources. The implementation:

1. Adds a new `allow_list` field to the `CreateTokenRequest` struct, allowing clients to specify resource-specific scopes when creating API tokens
2. Implements `APIAllowListTarget` type to represent resource targets in the format `<type>:<id>` with support for wildcards
3. Adds validation and normalization logic for allow lists to handle wildcards and deduplication
4. Integrates with RBAC by creating an `APIKeyEffectiveScope` that merges API key scopes with allow list restrictions
5. Updates API documentation and TypeScript types to reflect the new functionality

This feature enables creating tokens that are limited to specific resources (like workspaces or templates) by ID, making it possible to create more granular API tokens with limited access.
 2025-10-09 14:53:08 +02:00
Thomas Kosiewski b60ae0a0c4 refactor: add wildcard scope entries for API key scopes (#20032) 
# Add API Key Scope Wildcards

This PR adds wildcard API key scopes (`resource:*`) for all RBAC resources to ensure every resource has a matching wildcard value. It also adds all individual `resource:action`​ scopes to the API documentation and TypeScript definitions.

The changes include:

- Adding a new database migration (000377) that adds wildcard API key scopes
- Updating the API documentation to include all available scopes
- Enhancing the scope generation scripts to include all resource wildcards
- Updating the TypeScript definitions to match the expanded scope list

These changes make creating API keys with comprehensive permissions for specific resource types easier.
 2025-10-06 12:08:17 +02:00
Thomas Kosiewski 79126ab6c7 feat: implement composite API key scopes for workspaces and templates (#19945) 
# Add Composite API Key Scopes

This PR adds high-level composite API key scopes to simplify token creation with common permission sets:

- `coder:workspaces.create` - Create and update workspaces
- `coder:workspaces.operate` - Read and update workspaces
- `coder:workspaces.delete` - Read and delete workspaces
- `coder:workspaces.access` - Read, SSH, and connect to workspace applications
- `coder:templates.build` - Read templates and create/read files
- `coder:templates.author` - Full template management with insights
- `coder:apikeys.manage_self` - Manage your own API keys

These composite scopes are persisted in the database and expanded during authorization, providing a more intuitive way to grant permissions compared to the granular resource:action scopes.
 2025-09-29 13:17:08 +02:00
Thomas Kosiewski 05537c1894 feat: publish RBAC scopes in OAuth2 metadata endpoints (#19942) 
<!--

If you have used AI to produce some or all of this PR, please ensure you have read our [AI Contribution guidelines](https://coder.com/docs/about/contributing/AI_CONTRIBUTING) before submitting.

-->

Publish supported OAuth2 scopes from RBAC external scope names

This PR updates the OAuth2 metadata endpoints to publish the supported scopes from the RBAC external scope names. Previously, the `ScopesSupported` field was empty with a TODO to implement a scope system. Now, both the authorization server metadata and protected resource metadata endpoints return the list of scopes from `rbac.ExternalScopeNames()`.

The tests have been updated to verify that the correct scopes are being returned in the metadata responses.
 2025-09-26 12:15:36 +02:00
Thomas Kosiewski d0db9ec88f feat: add multi-scope support to API keys (#19917) 
# Canonicalize API Key Scopes

This PR introduces canonical API key scopes with a `coder:` namespace prefix to avoid collisions with low-level resource:action names. It:

1. Renames special API key scopes in the database:
   - `all` → `coder:all`
   - `application_connect` → `coder:application_connect`

2. Adds support for a new `scopes` field in the API key creation request, allowing multiple scopes to be specified while maintaining backward compatibility with the singular `scope` field.

3. Updates the API documentation to reflect these changes, including the new endpoint for listing public API key scopes.

4. Ensures backward compatibility by mapping between legacy and canonical scope names in relevant code paths.
 2025-09-26 11:56:34 +02:00
Thomas Kosiewski 4bda39585d feat: add external API key scopes (#19916) 
# Add support for low-level API key scopes

This PR adds support for fine-grained API key scopes based on RBAC resource:action pairs. It includes:

1. A new endpoint `/api/v2/auth/scopes` to list all public low-level API key scopes
2. Generated constants in the SDK for all public scopes
3. Tests to verify scope validation during token creation
4. Updated API documentation to reflect the expanded scope options

The implementation allows users to create API keys with specific permissions like `workspace:read` or `template:use` instead of only the legacy `all` or `application_connect` scopes.



Fixes #19847
 2025-09-26 11:43:32 +02:00
Thomas Kosiewski 47c92ad1d2 feat: add public RBAC scope catalog for user-requestable permissions (#19913) 
# Add a curated catalog of public RBAC scopes

This PR introduces a curated catalog of public RBAC scopes that are exposed to users. It adds:

- A `publicLowLevel` map in `scopes_catalog.go` that defines which resource:action pairs are user-requestable
- `IsPublicLowLevel()` function to check if a scope is in the public catalog
- `PublicLowLevelScopeNames()` function that returns a sorted list of public scopes
- Tests to verify the catalog entries are valid and properly sorted
- Updated documentation in the check-scopes README to clarify that public scopes should be added to this catalog

This change helps distinguish between internal-only scopes and those that should be exposed to users in the API.
 2025-09-26 11:30:28 +02:00
Thomas Kosiewski adb7521066 feat: generate RBAC scope name constants (#19896) 
# Generate RBAC scope name constants

This PR adds a new generated file `coderd/rbac/scopes_constants_gen.go` that contains typed constants for all RBAC scope names in the format `Scope<Resource><Action>`. For example, `ScopeWorkspaceRead` for the scope "workspace:read".

These constants make it easier to reference specific scopes in code without using string literals, improving type safety and making refactoring easier.

The PR:
- Adds a new template file `scripts/typegen/scopenames.gotmpl`
- Updates the typegen script to support generating scope name constants
- Updates the Makefile to include the new generated file in build targets
 2025-09-24 18:40:36 +02:00
Thomas Kosiewski acc0890dce feat: add lint check for API key scope enum completeness (#19862) 
Added a script/linter to ensure all `policy.RBACPermissions` entries are part of the `api_key_scope` enumerated in the `coderd/database/dump.sql` file.

Fixes #19846
 2025-09-24 18:06:16 +02:00
Thomas Kosiewski 77d19c99cd ci: fix shellcheck lint error (#19933) 2025-09-24 10:47:50 +02:00
Thomas Kosiewski fb0ce389a6 feat: implement API key scopes database migration (#19861) 
Added database migration for API key scopes.

Fixes #19845
 2025-09-22 19:26:51 +02:00
Thomas Kosiewski 6fb4cc6b82 fix: add retry logic to OAuth2 metadata tests to avoid race conditions (#19813) 
This PR adds a readiness wait to OAuth2 metadata endpoint tests to avoid rare races with server startup. Instead of immediately making HTTP requests, the tests now use `testutil.Eventually` to retry the requests until they succeed, with a short interval between attempts. This helps prevent flaky tests that might fail due to timing issues during server initialization.  
  
Fixes: https://github.com/coder/internal/issues/996
 2025-09-22 18:30:36 +02:00
Thomas Kosiewski 4d8dc221bf chore: upgrade Node.js from 20.19.4 to 22.19.0 and update dependencies (#19870) 
# Update Node.js from 20.19.4 to 22.19.0

This PR updates Node.js from v20.19.4 to v22.19.0 across the codebase. The change includes:

- Updated Node.js version in GitHub Actions setup-node workflow
- Updated Node.js version in the dogfood Dockerfile
- Changed from `pkgs.nodejs_20` to `unstablePkgs.nodejs_22` in the Nix flake
- Updated the Node.js engine version constraints in package.json files to allow Node.js 22
- Updated Playwright from v1.47.0 to v1.50.1
- Updated tzdata dependency from v1.0.44 to v1.0.46
- Updated the flake.lock file with latest nixpkgs references

The PR also improves the error message for Playwright version mismatches by showing the actual versions in the error.
 2025-09-18 23:08:55 +02:00
Thomas Kosiewski d238480c7a fix: trim whitespace from API tokens (#19814) 2025-09-15 10:02:10 +02:00
Thomas Kosiewski 088d14933c feat: ensure OAuth2 refresh tokens outlive access tokens (#19769) 2025-09-13 08:57:26 +02:00
Thomas Kosiewski 2701d5588e fix: support path parameters in OAuth2 metadata endpoints (#19729) 
Update OAuth2 metadata endpoint routes to support path suffixes

This PR updates the OAuth2 metadata endpoint routes to include a wildcard character (*) at the end of the paths. This change allows the endpoints to match requests with path suffixes, making our OAuth2 discovery implementation more flexible and compliant with the relevant RFCs.

The updated routes are:
- `/.well-known/oauth-authorization-server*` for RFC 8414 discovery
- `/.well-known/oauth-protected-resource*` for RFC 9728 discovery
 2025-09-08 14:21:57 +02:00
Thomas Kosiewski b7e026682a fix: skip bash tests on Windows (#19277) 
Adds Windows compatibility to toolsdk tests

This PR adds Windows compatibility to the toolsdk tests by:

1. Adding build constraints to exclude bash_test.go from running on
Windows
2. Skipping the WorkspaceSSHExec test on Windows platforms with a clear
message

These changes ensure tests run properly across all supported platforms.

Related to https://github.com/coder/internal/issues/798

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-08-11 11:43:25 +02:00
Thomas Kosiewski 247efc0dcc docs: add OAuth2 provider experimental feature documentation (#19165) 
# Add OAuth2 Provider Documentation

This PR adds comprehensive documentation for the experimental OAuth2
Provider feature, which allows Coder to function as an OAuth2
authorization server. The documentation covers:

- Feature overview and experimental status warning
- Setup requirements and enabling the feature
- Methods for creating OAuth2 applications (UI and API)
- Integration patterns including standard OAuth2 and PKCE flows
- Discovery endpoints and token management
- Testing and development guidance
- Troubleshooting common issues
- Security considerations and current limitations

The documentation is marked as experimental and includes appropriate
warnings about production usage.

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-08-04 20:17:47 +02:00
Thomas Kosiewski 398e80f003 feat: add timeout support to workspace bash tool (#19035) 
# Add timeout support to workspace bash tool

This PR adds a timeout feature to the workspace bash tool, allowing
users to specify a maximum execution time for commands. Key changes
include:

- Added a `timeout_ms` parameter to control command execution time
(defaults to 60 seconds, with a maximum of 5 minutes)
- Implemented a new `executeCommandWithTimeout` function that properly
handles command timeouts
- Added proper output capturing during timeout scenarios, returning all
output collected before the timeout
- Updated documentation to explain the timeout feature and provide usage
examples
- Added comprehensive tests for the timeout functionality, including
integration tests

When a command times out, the tool now returns all captured output up to
that point along with a cancellation message, making it clear to users
what happened.

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-28 11:25:43 +02:00
Thomas Kosiewski d1595781e1 fix: fix nil pointer dereference in ReportTask (#19045) 
This pull request addresses a bug related to a nil pointer dereference
in the task reporting functionality.

### Bug Fixes and Error Handling:

* Updated `RegisterTools` in `mcp.go` to skip registering the
`ReportTask` tool in the remote MCP context when a task reporter is not
configured, preventing potential nil pointer dereference panics.
* Added a check in `toolsdk.go` to ensure task reporting dependencies
are available before invoking the reporter, returning an appropriate
error if not.

### Test Coverage:

* Added `TestReportTaskNilPointerDeref` in `toolsdk_test.go` to verify
that the system does not panic when task reporting dependencies are
missing and instead returns a clear error message.
* Added `TestReportTaskWithReporter` in `toolsdk_test.go` to validate
correct behavior when a task reporter is configured, ensuring the
handler processes the request as expected.

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-28 11:23:29 +02:00
Thomas Kosiewski c6efe64a65 fix: handle nil writer in bash MCP tool (#18978) 
- Refactors the bash tool to use `io.Discard` instead of nil to avoid panics.

- Enhances panic recovery in `codersdk/toolsdk/toolsdk.go` by adding stack trace information in development builds. When a panic occurs in a tool handler:
   - In development builds: The error includes the full stack trace for easier debugging
   - In production builds: A simpler error message is shown without the stack trace
 2025-07-22 18:03:26 +02:00
Thomas Kosiewski 326c02459f feat: add workspace SSH execution tool for AI SDK (#18924) 
# Add SSH Command Execution Tool for Coder Workspaces

This PR adds a new AI tool `coder_workspace_ssh_exec` that allows executing commands in Coder workspaces via SSH. The tool provides functionality similar to the `coder ssh <workspace> <command>` CLI command.

Key features:
- Executes commands in workspaces via SSH and returns the output and exit code
- Automatically starts workspaces if they're stopped
- Waits for the agent to be ready before executing commands
- Trims leading and trailing whitespace from command output
- Supports various workspace identifier formats:
  - `workspace` (uses current user)
  - `owner/workspace`
  - `owner--workspace`
  - `workspace.agent` (specific agent)
  - `owner/workspace.agent`

The implementation includes:
- A new tool definition with schema and handler
- Helper functions for workspace and agent discovery
- Workspace name normalization to handle different input formats
- Comprehensive test coverage including integration tests

This tool enables AI assistants to execute commands in user workspaces, making it possible to automate tasks and provide more interactive assistance.

<!-- This is an auto-generated comment: release notes by coderabbit.ai -->
## Summary by CodeRabbit

* **New Features**
  * Introduced the ability to execute bash commands inside a Coder workspace via SSH, supporting multiple workspace identification formats.
* **Tests**
  * Added comprehensive unit and integration tests for executing bash commands in workspaces, including input validation, output handling, and error scenarios.
* **Chores**
  * Registered the new bash execution tool in the global tools list.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
 2025-07-21 21:24:00 +02:00
Thomas Kosiewski 1db096d8f9 chore: fix CodeRabbit config to disable review status (#18973) 
Disable review status in CodeRabbit configuration

Change-Id: I0ee266e0b284832b65762a4f7a3f26d56af53e86
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-21 20:26:01 +02:00
Thomas Kosiewski 4ac6be6d83 chore: add CodeRabbit config with disabled auto-reviews (#18949) 2025-07-21 15:51:48 +01:00
Thomas Kosiewski 0d3b7703f7 docs: remove dbmem references from documentation files (#18861) 
Change-Id: Ic33bc383d00d0e354c25a0dd6080a4307d9862b6
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-21 11:21:58 +02:00
Thomas Kosiewski fcd361d374 feat: add logo SVG and replace inline SVG with image reference (#18930) 
# Replace SVG with external logo file in OAuth2 authorization page

This PR replaces the inline SVG logo in the OAuth2 authorization page with a reference to an external SVG file. The change:

1. Adds a new `logo.svg` file in the static directory with the Coder logo
2. Updates the OAuth2 authorization page to use this external file instead of embedding the SVG directly

This approach improves maintainability by centralizing the logo in a single file and reduces duplication in the codebase.
 2025-07-21 11:04:21 +02:00
Thomas Kosiewski 7b06fc77ae refactor: simplify OAuth2 authorization flow and use 302 redirects (#18923) 
# Refactor OAuth2 Provider Authorization Flow

This PR refactors the OAuth2 provider authorization flow by:

1. Removing the `authorizeMW` middleware and directly implementing its functionality in the `ShowAuthorizePage` handler
2. Simplifying function signatures by removing unnecessary parameters:
   - Removed `db` parameter from `ShowAuthorizePage`
   - Removed `accessURL` parameter from `ProcessAuthorize`
3. Changing the redirect status code in `ProcessAuthorize` from 307 (Temporary Redirect) to 302 (Found) to improve compatibility with external OAuth2 apps and browsers. (Technical explanation: we replied with a 307 to a POST request, thus the browser performs a redirect to that URL as a POST request, but we need it to be a GET request to be compatible. Thus, we use the 302 redirect so that browsers turn it into a GET request when redirecting back to the redirect_uri.)

The changes maintain the same functionality while simplifying the code and improving compatibility with external systems.
 2025-07-20 16:22:52 +02:00
Thomas Kosiewski 071383bbe8 feat: add RFC 9728 OAuth2 resource metadata support (#18920) 
# Enhanced OAuth2 and MCP Compliance for API Authentication

This PR improves OAuth2 and MCP (Microsoft Cloud for Sovereignty)
compliance by:

1. Adding RFC 9728 compliant `WWW-Authenticate` headers with resource
metadata URLs
2. Passing the configured `AccessURL` to API key middleware for proper
audience validation
3. Creating specialized CORS handling for OAuth2 and MCP endpoints with
appropriate headers
4. Making the `state` parameter optional in OAuth2 authorization
requests

These changes ensure proper OAuth2 token audience validation against the
configured access URL and improve interoperability with OAuth2 clients
by providing better error responses and metadata discovery.

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-19 22:05:15 +02:00
Thomas Kosiewski 3dcd2acf1d fix: return 404 instead of 401 for missing OAuth2 apps (#18755) 
## Problem

Users were being automatically logged out when deleting OAuth2
applications.

## Root Cause

1. User deletes OAuth2 app successfully
2. React Query automatically refetches the app data  
3. Management API incorrectly returned **401 Unauthorized** for the
missing app
4. Frontend axios interceptor sees 401 and calls `signOut()`
5. User gets logged out unexpectedly

## Solution

- Change management API to return **404 Not Found** for missing OAuth2
apps
- OAuth2 protocol endpoints continue returning 401 per RFC 6749
- Rename `writeInvalidClient` to `writeClientNotFound` for clarity

## Additional Changes

- Add conditional OAuth2 navigation when experiment is enabled or in dev
builds
- Add `isDevBuild()` utility and `buildInfo` to dashboard context
- Minor improvements to format script and warning dialogs

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-07 19:57:32 +02:00
Thomas Kosiewski 4607e5113b refactor: organize OAuth2 provider tests into dedicated packages (#18747) 
# OAuth2 Provider Code Reorganization

This PR reorganizes the OAuth2 provider code to improve separation of concerns and maintainability. The changes include:

1. Migrating OAuth2 provider app validation tests from `coderd/oauth2_test.go` to `oauth2provider/provider_test.go`
2. Moving OAuth2 client registration validation tests to `oauth2provider/validation_test.go`
3. Adding new comprehensive test files for metadata and validation edge cases
4. Renaming `OAuth2ProviderAppSecret` to `AppSecret` for better naming consistency
5. Simplifying the main integration test in `oauth2_test.go` to focus on core functionality

The PR maintains all existing test coverage while organizing the code more logically, making it easier to understand and maintain the OAuth2 provider implementation. This reorganization will help with future enhancements to the OAuth2 provider functionality.
 2025-07-03 20:41:47 +02:00
Thomas Kosiewski c65013384a refactor: move OAuth2 provider code to dedicated package (#18746) 
# Refactor OAuth2 Provider Code into Dedicated Package

This PR refactors the OAuth2 provider functionality by moving it from the main `coderd` package into a dedicated `oauth2provider` package. The change improves code organization and maintainability without changing functionality.

Key changes:

- Created a new `oauth2provider` package to house all OAuth2 provider-related code
- Moved existing OAuth2 provider functionality from `coderd/identityprovider` to the new package
- Refactored handler functions to follow a consistent pattern of returning `http.HandlerFunc` instead of being handlers directly
- Split large files into smaller, more focused files organized by functionality:
  - `app_secrets.go` - Manages OAuth2 application secrets
  - `apps.go` - Handles OAuth2 application CRUD operations
  - `authorize.go` - Implements the authorization flow
  - `metadata.go` - Provides OAuth2 metadata endpoints
  - `registration.go` - Handles dynamic client registration
  - `revoke.go` - Implements token revocation
  - `secrets.go` - Manages secret generation and validation
  - `tokens.go` - Handles token issuance and validation

This refactoring improves code organization and makes the OAuth2 provider functionality more maintainable while preserving all existing behavior.
 2025-07-03 20:24:45 +02:00
Thomas Kosiewski 7fbb3ced5b feat: add MCP HTTP server experiment and improve experiment middleware (#18712) 
# Add MCP HTTP Server Experiment

This PR adds a new experiment flag `mcp-server-http` to enable the MCP HTTP server functionality. The changes include:

1. Added a new experiment constant `ExperimentMCPServerHTTP` with the value "mcp-server-http"
2. Added display name and documentation for the new experiment
3. Improved the experiment middleware to:
   - Support requiring multiple experiments
   - Provide better error messages with experiment display names
   - Add a development mode bypass option
4. Applied the new experiment requirement to the MCP HTTP endpoint
5. Replaced the custom OAuth2 middleware with the standard experiment middleware

The PR also improves the `Enabled()` method on the `Experiments` type by using `slices.Contains()` for better readability.
 2025-07-03 20:09:18 +02:00
Thomas Kosiewski 15551541e8 feat: add OAuth2 provider functionality as an experiment (#18692) 
# Add OAuth2 Provider Functionality as an Experiment

This PR adds a new experiment flag `oauth2` that enables OAuth2 provider functionality in Coder. When enabled, this experiment allows Coder to act as an OAuth2 provider.

The changes include:
- Added the new `ExperimentOAuth2` constant with appropriate documentation
- Updated the OAuth2 provider middleware to check for the experiment flag
- Modified the error message to indicate that the OAuth2 provider requires enabling the experiment
- Added the new experiment to the known experiments list in the SDK

Previously, OAuth2 provider functionality was only available in development mode. With this change, it can be enabled in production environments by activating the experiment.
 2025-07-03 19:44:29 +02:00
Thomas Kosiewski 494dccc510 feat: implement MCP HTTP server endpoint with authentication (#18670) 
# Add MCP HTTP server with streamable transport support

- Add MCP HTTP server with streamable transport support
- Integrate with existing toolsdk for Coder workspace operations
- Add comprehensive E2E tests with OAuth2 bearer token support
- Register MCP endpoint at /api/experimental/mcp/http with authentication
- Support RFC 6750 Bearer token authentication for MCP clients

Change-Id: Ib9024569ae452729908797c42155006aa04330af
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-03 19:27:41 +02:00
Thomas Kosiewski 60b08f0960 fix: remove unique constraint on OAuth2 provider app names (#18669) 
# Remove unique constraint on OAuth2 provider app names

This PR removes the unique constraint on the `name` field in the `oauth2_provider_apps` table to comply with RFC 7591, which only requires unique client IDs, not unique client names.

Changes include:
- Removing the unique constraint from the database schema
- Adding migration files for both up and down migrations
- Removing the name uniqueness check in the in-memory database implementation
- Updating the unique constraint constants

Change-Id: Iae7a1a06546fbc8de541a52e291f8a4510d57e8a
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-03 19:13:13 +02:00
Thomas Kosiewski 4dcf0c3e7e docs: add comprehensive development documentation (#18646) 
# Organize Development Documentation into Separate Files

This PR reorganizes the development documentation by splitting the monolithic CLAUDE.md file into multiple focused documents. The main file now provides a concise overview with essential commands and critical patterns, while importing detailed content from specialized guides.

Key improvements:
- Created separate documentation files for specific domains:
  - Database development patterns
  - OAuth2 implementation guidelines
  - Testing best practices
  - Troubleshooting common issues
  - Development workflows and guidelines
- Restructured the main CLAUDE.md to be more scannable with improved formatting
- Added quick-reference tables for common commands
- Maintained all existing content while making it more accessible
- Highlighted critical patterns that must be followed

This organization makes the documentation more maintainable and easier to navigate, allowing developers to quickly find relevant information for their specific tasks.
 2025-07-03 18:51:23 +02:00
Thomas Kosiewski 74e1d5c4b6 feat: implement OAuth2 dynamic client registration (RFC 7591/7592) (#18645) 
# Implement OAuth2 Dynamic Client Registration (RFC 7591/7592)

This PR implements OAuth2 Dynamic Client Registration according to RFC 7591 and Client Configuration Management according to RFC 7592. These standards allow OAuth2 clients to register themselves programmatically with Coder as an authorization server.

Key changes include:

1. Added database schema extensions to support RFC 7591/7592 fields in the `oauth2_provider_apps` table
2. Implemented `/oauth2/register` endpoint for dynamic client registration (RFC 7591)
3. Added client configuration management endpoints (RFC 7592):
   - GET/PUT/DELETE `/oauth2/clients/{client_id}`
   - Registration access token validation middleware

4. Added comprehensive validation for OAuth2 client metadata:
   - URI validation with support for custom schemes for native apps
   - Grant type and response type validation
   - Token endpoint authentication method validation

5. Enhanced developer documentation with:
   - RFC compliance guidelines
   - Testing best practices to avoid race conditions
   - Systematic debugging approaches for OAuth2 implementations

The implementation follows security best practices from the RFCs, including proper token handling, secure defaults, and appropriate error responses. This enables third-party applications to integrate with Coder's OAuth2 provider capabilities programmatically.
 2025-07-03 18:33:47 +02:00
Thomas Kosiewski 09c50559f3 feat: implement RFC 6750 Bearer token authentication (#18644) 
# Add RFC 6750 Bearer Token Authentication Support

This PR implements RFC 6750 Bearer Token authentication as an additional authentication method for Coder's API. This allows clients to authenticate using standard OAuth 2.0 Bearer tokens in two ways:

1. Using the `Authorization: Bearer <token>` header
2. Using the `access_token` query parameter

Key changes:

- Added support for extracting tokens from both Bearer headers and access_token query parameters
- Implemented proper WWW-Authenticate headers for 401/403 responses with appropriate error descriptions
- Added comprehensive test coverage for the new authentication methods
- Updated the OAuth2 protected resource metadata endpoint to advertise Bearer token support
- Enhanced the OAuth2 testing script to verify Bearer token functionality

These authentication methods are added as fallback options, maintaining backward compatibility with Coder's existing authentication mechanisms. The existing authentication methods (cookies, session token header, etc.) still take precedence.

This implementation follows the OAuth 2.0 Bearer Token specification (RFC 6750) and improves interoperability with standard OAuth 2.0 clients.
 2025-07-02 19:14:54 +02:00
Thomas Kosiewski 33bbf18a4b feat: add OAuth2 protected resource metadata endpoint for RFC 9728 (#18643) 
# Add OAuth2 Protected Resource Metadata Endpoint

This PR implements the OAuth2 Protected Resource Metadata endpoint according to RFC 9728. The endpoint is available at `/.well-known/oauth-protected-resource` and provides information about Coder as an OAuth2 protected resource.

Key changes:
- Added a new endpoint at `/.well-known/oauth-protected-resource` that returns metadata about Coder as an OAuth2 protected resource
- Created a new `OAuth2ProtectedResourceMetadata` struct in the SDK
- Added tests to verify the endpoint functionality
- Updated API documentation to include the new endpoint

The implementation currently returns basic metadata including the resource identifier and authorization server URL. The `scopes_supported` field is empty until a scope system based on RBAC permissions is implemented. The `bearer_methods_supported` field is omitted as Coder uses custom authentication methods rather than standard RFC 6750 bearer tokens.

A TODO has been added to implement RFC 6750 bearer token support in the future.
 2025-07-02 18:58:41 +02:00
Thomas Kosiewski 1b73b1a12f docs: add Go LSP MCP configs and tools guide for code navigation (#18613) 
# Add Code Navigation and Investigation Guide for Go LSP Tools

Added a new section to the CLAUDE.md documentation that explains how to use Go Language Server Protocol (LSP) tools when working with the Coder codebase. The guide includes:

- Commands for finding function definitions, symbol references, and getting symbol information
- Examples of LSP usage with specific commands
- Guidance on when to use LSP versus other tools like grep or bash
- A structured investigation strategy for navigating the codebase, starting with route registration and tracing through to implementations

This documentation helps developers more efficiently explore and understand the codebase structure.
 2025-07-02 18:43:35 +02:00
Thomas Kosiewski f0c9c4dbcd feat: oauth2 - add RFC 8707 resource indicators and audience validation (#18575) 
This pull request implements RFC 8707, Resource Indicators for OAuth 2.0 (https://datatracker.ietf.org/doc/html/rfc8707), to enhance the security of our OAuth 2.0 provider. 

This change enables proper audience validation and binds access tokens to their intended resource, which is crucial
  for preventing token misuse in multi-tenant environments or deployments with multiple resource servers.

##  Key Changes:


   * Resource Parameter Support: Adds support for the resource parameter in both the authorization (`/oauth2/authorize`) and token (`/oauth2/token`) endpoints, allowing clients to specify the intended resource server.
   * Audience Validation: Implements server-side validation to ensure that the resource parameter provided during the token exchange matches the one from the authorization request.
   * API Middleware Enforcement: Introduces a new validation step in the API authentication middleware (`coderd/httpmw/apikey.go`) to verify that the audience of the access token matches the resource server being accessed.
   * Database Schema Updates:
       * Adds a `resource_uri` column to the `oauth2_provider_app_codes` table to store the resource requested during authorization.
       * Adds an `audience` column to the `oauth2_provider_app_tokens` table to bind the issued token to a specific audience.
   * Enhanced PKCE: Includes a minor enhancement to the PKCE implementation to protect against timing attacks.
   * Comprehensive Testing: Adds extensive new tests to `coderd/oauth2_test.go` to cover various RFC 8707 scenarios, including valid flows, mismatched resources, and refresh token validation.

##  How it Works:


   1. An OAuth2 client specifies the target resource (e.g., https://coder.example.com) using the resource parameter in the authorization request.
   2. The authorization server stores this resource URI with the authorization code.
   3. During the token exchange, the server validates that the client provides the same resource parameter.
   4. The server issues an access token with an audience claim set to the validated resource URI.
   5. When the client uses the access token to call an API endpoint, the middleware verifies that the token's audience matches the URL of the Coder deployment, rejecting any tokens intended for a different resource.


  This ensures that a token issued for one Coder deployment cannot be used to access another, significantly strengthening our authentication security.

---

Change-Id: I3924cb2139e837e3ac0b0bd40a5aeb59637ebc1b
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-02 17:49:00 +02:00
Thomas Kosiewski 6f2834f62a feat: oauth2 - add authorization server metadata endpoint and PKCE support (#18548) 
## Summary

  This PR implements critical MCP OAuth2 compliance features for Coder's authorization server, adding PKCE support, resource parameter handling, and OAuth2 server metadata discovery. This brings Coder's OAuth2 implementation significantly closer to production readiness for MCP (Model Context Protocol)
  integrations.

  ## What's Added

  ### OAuth2 Authorization Server Metadata (RFC 8414)
  - Add `/.well-known/oauth-authorization-server` endpoint for automatic client discovery
  - Returns standardized metadata including supported grant types, response types, and PKCE methods
  - Essential for MCP client compatibility and OAuth2 standards compliance

  ### PKCE Support (RFC 7636)
  - Implement Proof Key for Code Exchange with S256 challenge method
  - Add `code_challenge` and `code_challenge_method` parameters to authorization flow
  - Add `code_verifier` validation in token exchange
  - Provides enhanced security for public clients (mobile apps, CLIs)

  ### Resource Parameter Support (RFC 8707)
  - Add `resource` parameter to authorization and token endpoints
  - Store resource URI and bind tokens to specific audiences
  - Critical for MCP's resource-bound token model

  ### Enhanced OAuth2 Error Handling
  - Add OAuth2-compliant error responses with proper error codes
  - Use standard error format: `{"error": "code", "error_description": "details"}`
  - Improve error consistency across OAuth2 endpoints

  ### Authorization UI Improvements
  - Fix authorization flow to use POST-based consent instead of GET redirects
  - Remove dependency on referer headers for security decisions
  - Improve CSRF protection with proper state parameter validation

  ## Why This Matters

  **For MCP Integration:** MCP requires OAuth2 authorization servers to support PKCE, resource parameters, and metadata discovery. Without these features, MCP clients cannot securely authenticate with Coder.

  **For Security:** PKCE prevents authorization code interception attacks, especially critical for public clients. Resource binding ensures tokens are only valid for intended services.

  **For Standards Compliance:** These are widely adopted OAuth2 extensions that improve interoperability with modern OAuth2 clients.

  ## Database Changes

  - **Migration 000343:** Adds `code_challenge`, `code_challenge_method`, `resource_uri` to `oauth2_provider_app_codes`
  - **Migration 000343:** Adds `audience` field to `oauth2_provider_app_tokens` for resource binding
  - **Audit Updates:** New OAuth2 fields properly tracked in audit system
  - **Backward Compatibility:** All changes maintain compatibility with existing OAuth2 flows

  ## Test Coverage

  - Comprehensive PKCE test suite in `coderd/identityprovider/pkce_test.go`
  - OAuth2 metadata endpoint tests in `coderd/oauth2_metadata_test.go`
  - Integration tests covering PKCE + resource parameter combinations
  - Negative tests for invalid PKCE verifiers and malformed requests

  ## Testing Instructions

  ```bash
  # Run the comprehensive OAuth2 test suite
  ./scripts/oauth2/test-mcp-oauth2.sh

  Manual Testing with Interactive Server

  # Start Coder in development mode
  ./scripts/develop.sh

  # In another terminal, set up test app and run interactive flow
  eval $(./scripts/oauth2/setup-test-app.sh)
  ./scripts/oauth2/test-manual-flow.sh
  # Opens browser with OAuth2 flow, handles callback automatically

  # Clean up when done
  ./scripts/oauth2/cleanup-test-app.sh

  Individual Component Testing

  # Test metadata endpoint
  curl -s http://localhost:3000/.well-known/oauth-authorization-server | jq .

  # Test PKCE generation
  ./scripts/oauth2/generate-pkce.sh

  # Run specific test suites
  go test -v ./coderd/identityprovider -run TestVerifyPKCE
  go test -v ./coderd -run TestOAuth2AuthorizationServerMetadata
```

  ### Breaking Changes

  None. All changes maintain backward compatibility with existing OAuth2 flows.

---

Change-Id: Ifbd0d9a543d545f9f56ecaa77ff2238542ff954a
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-07-01 15:39:29 +02:00
Thomas Kosiewski 1b1d09158d fix: pin Nix version to 2.28.4 to avoid JSON type error (#18612) 
Pin Nix version to 2.28.4 in dogfood workflow

Pins the Nix version in the dogfood workflow to 2.28.4 to avoid a JSON type error that occurs with Nix 2.29 and above.

Change-Id: Ie024d5070dbe5901952fc52463c6602363ef8886
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-06-26 18:33:44 +02:00
Thomas Kosiewski f569d9c33d feat: add separate max token lifetime for administrators (#18267) 
# Add separate token lifetime limits for administrators

This PR introduces a new configuration option `--max-admin-token-lifetime` that allows administrators to create API tokens with longer lifetimes than regular users. By default, administrators can create tokens with a lifetime of up to 7 days (168 hours), while the existing `--max-token-lifetime` setting continues to apply to regular users.

The implementation:
- Adds a new `MaximumAdminTokenDuration` field to the session configuration
- Modifies the token validation logic to check the user's role and apply the appropriate lifetime limit
- Updates the token configuration endpoint to return the correct maximum lifetime based on the user's role
- Adds tests to verify that administrators can create tokens with longer and shorter lifetimes
- Updates documentation and help text to reflect the new option

This change allows organizations to grant administrators extended token lifetimes while maintaining tighter security controls for regular users.

Fixes #17395
 2025-06-06 17:36:30 +02:00
Thomas Kosiewski b551a062d7 fix: correct environment variable name for MCP app status slug (#17948) 
Fixed environment variable name for app status slug in Claude MCP configuration from `CODER_MCP_CLAUDE_APP_STATUS_SLUG` to `CODER_MCP_APP_STATUS_SLUG` to maintain consistency with other MCP environment variables.

This also caused the User level Claude.md to not contain instructions to report its progress, so it did not receive status reports.
 2025-05-20 19:35:19 +02:00
Thomas Kosiewski 93f17bc73e fix: remove unnecessary user lookup in agent API calls (#17934) 
# Use workspace.OwnerUsername instead of fetching the owner

This PR optimizes the agent API by using the `workspace.OwnerUsername` field directly instead of making an additional database query to fetch the owner's username. The change removes the need to call `GetUserByID` in the manifest API and workspace agent RPC endpoints.

An issue arose when the agent token was scoped without access to user data (`api_key_scope = "no_user_data"`), causing the agent to fail to fetch the manifest due to an RBAC issue.

Change-Id: I3b6e7581134e2374b364ee059e3b18ece3d98b41
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-05-20 17:07:50 +02:00
Thomas Kosiewski 1bacd82e80 feat: add API key scope to restrict access to user data (#17692) 2025-05-15 15:32:52 +01:00
Thomas Kosiewski 29bce8d9e6 feat(cli): make MCP server work without user authentication (#17688) 
Part of #17649

---

# Allow MCP server to run without authentication

This PR enhances the MCP server to operate without requiring authentication, making it more flexible for environments where authentication isn't available or necessary. Key changes:

- Replaced `InitClient` with `TryInitClient` to allow the MCP server to start without credentials
- Added graceful handling when URL or authentication is missing
- Made authentication status visible in server logs
- Added logic to skip user-dependent tools when no authenticated user is present
- Made the `coder_report_task` tool available with just an agent token (no user token required)
- Added comprehensive tests to verify operation without authentication

These changes allow the MCP server to function in more environments while still using authentication when available, improving flexibility for CI/CD and other automated environments.
 2025-05-07 21:53:06 +02:00
Thomas Kosiewski abe3ad68f5 fix: add continue-on-error to SBOM generation and force flag to cosign clean (#17288) 
This PR makes the SBOM generation and attestation process more resilient
by:

1. Adding `continue-on-error: true` to the SBOM generation steps in both
CI and release workflows
2. Adding `--force=true` flag to all `cosign clean` commands to ensure
they don't fail if in a non-interactive shell (which is the case for CI)

Change-Id: Ide303c059b1a3d0e3fd77863310e99668325bc69
Signed-off-by: Thomas Kosiewski <tk@coder.com>

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-04-08 10:29:00 +02:00
Thomas Kosiewski f48a24c18e feat: add SBOM generation and attestation to GitHub workflow (#17277) 
Move SBOM generation and attestation to GitHub workflow

This PR moves the SBOM generation and attestation process from the `build_docker.sh` script to the GitHub workflow. The change:

1. Removes SBOM generation and attestation from the `build_docker.sh` script
2. Adds a new "SBOM Generation and Attestation" step in the GitHub workflow
3. Generates and attests SBOMs for both multi-arch images and latest tags when applicable

This approach ensures SBOM generation happens once for the final multi-architecture image rather than for each architecture separately.

Change-Id: I2e15d7322ddec933bbc9bd7880abba9b0842719f
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-04-07 17:54:05 +02:00
Thomas Kosiewski 7171d52279 fix: replace both colons and slashes in SBOM filename for Docker image (#16915) 
This PR fixes the SBOM filename generation in the Docker build script to
properly handle image tags that contain slashes. The current
implementation only replaces colons with underscores, but fails when
image tags include slashes (common in registry paths).

The fix updates the string replacement to handle both colons and slashes
in the image tag when generating the SBOM filename.

Change-Id: Ifd7bad6d165393e11202e5bf070a4cb26eaa6a6a
Signed-off-by: Thomas Kosiewski <tk@coder.com>

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-03-13 23:01:03 +05:00
Thomas Kosiewski 389af22dac chore: replace colons in SBOM filename for Docker image attestation (#16914) 
This PR fixes an issue in the Docker build script where the SBOM file path used the image tag directly, which could contain colons. Since colons are not valid characters in filenames on many filesystems, this replaces colons with underscores in the output filename.

Change-Id: I887f4fc255d9bfa19b6c5d23ad0a5db7352aa2af
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-03-13 18:20:43 +01:00
Thomas Kosiewski 5285c12b9e chore: update terraform to 1.11.1 in nix image (#16880) 
Followup PR to #16781, update the terraform version in our Nix devshell.

Additionally: 

1. Switches from DeterminateSystems/nix-installer-action to nixbuild/nix-quick-install-action -- quicker installer, reduces actions time from ~60 seconds to ~1 seconds.
2. Adds nix-community/cache-nix-action for better caching with garbage collection -- avoids unnecessary rebuilding on subsequent runs, reduces nix image build time from ~6 minutes to <4 minutes.
3. Adds nixpkgs-unstable input to use Terraform 1.11.1

Change-Id: I05d6dfd3f3cf1af48cf8a2d9e61b396bcd2b7191
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-03-11 16:23:33 +01:00
Thomas Kosiewski 9ded2cc7ec fix(flake.nix): synchronize playwright version in nix and package.json (#16715) 
Ensure that the version of Playwright installed with the Nix flake is
equal to the one specified in `site/package.json.` -- This assertion
ensures that `pnpm playwright:install` will not attempt to download
newer browser versions not present in the Nix image, fixing the startup
script and reducing the startup time, as `pnpm playwright:install` will
not download or install anything.

We also pre-install the required Playwright web browsers in the dogfood
Dockerfile. This change prevents us from redownloading system
dependencies and Google Chrome each time a workspace starts.

Change-Id: I8cc78e842f7d0b1d2a90a4517a186a03636c5559
Signed-off-by: Thomas Kosiewski <tk@coder.com>

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-03-11 13:49:03 +01:00
Thomas Kosiewski d0e2060692 feat(agent): add second SSH listener on port 22 (#16627) 
Fixes: https://github.com/coder/internal/issues/377

Added an additional SSH listener on port 22, so the agent now listens on both, port one and port 22.

---
Change-Id: Ifd986b260f8ac317e37d65111cd4e0bd1dc38af8
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-03-03 04:47:42 +01:00
Thomas Kosiewski 38c0e8a086 fix(agent/agentssh): ensure RSA key generation always produces valid keys (#16694) 
Modify the RSA key generation algorithm to check that GCD(e, p-1) = 1 and
GCD(e, q-1) = 1 when selecting prime numbers, ensuring that e and φ(n) 
are coprime. This prevents ModInverse from returning nil, which would 
cause private key generation to fail and result in a panic when `Precompute` is called.

Change-Id: I0a453e1e1f8c638e40e7a4b87a6d0d7299e1cb5d
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-26 11:45:35 +01:00
Thomas Kosiewski 660746462e fix(agent/agentssh): use deterministic host key for SSH server (#16626) 
Fixes: https://github.com/coder/coder/issues/16490

The Agent's SSH server now initially generates fixed host keys and, once it receives its manifest, generates and replaces that host key with the one derived from the workspace ID, ensuring consistency across agent restarts. This prevents SSH warnings and host key verification errors when connecting to workspaces through Coder Desktop.

While deterministic keys might seem insecure, the underlying Wireguard tunnel already provides encryption and anti-spoofing protection at the network layer, making this approach acceptable for our use case.

---
Change-Id: I8c7e3070324e5d558374fd6891eea9d48660e1e9
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-21 14:58:41 +01:00
Thomas Kosiewski 00e2703aca fix(flake.nix): add procps to nix dogfood image (#16607) 
Add procps to flake.nix and release name to Docker image

Adds the `procps` package to flake.nix to enable the `free` command, and includes a release name file in the Docker image at `/etc/coderniximage-release`.

Change-Id: I85432acc06a204229fa3675e0020bd3acacf775a
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-18 16:24:47 +01:00
Thomas Kosiewski 420855dc55 fix(helm): ensure coder can be deployed in a non-default namespace (#16579) 
Added namespace to all resources in the helm chart and added tests to ensure that coder can be deployed in non-default namespaces, as specified via the namespace flag in the helm command.

Ways to verify this:

- current state: 
  ```bash
  $ helm template my-coder coder -n coder --version 2.19.0 --repo https://helm.coder.com/v2 | yq '.metadata.namespace'
  null
  ---
  null
  ---
  null
  ---
  null
  ---
  null
  ```

- fixed state when checking out this PR: 
  ```bash
  $ helm template my-coder ./helm/coder -n coder --set coder.image.tag=latest | yq '.metadata.namespace'
  coder
  ---
  coder
  ---
  coder
  ---
  coder
  ---
  coder
  ```

Change-Id: Ib66d4be9bcc4984dfe15709362e1fe0dcd3e847f
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-18 12:50:35 +01:00
Thomas Kosiewski 6a67e2ede6 feat(cli/server.go): allow the use of public OIDC clients (#16489) 
Support public OIDC clients

- Enables support for public OIDC clients by only checking for a client
ID being set. This allows for confidential and public clients to be used
with Coder's OIDC authentication.
- Also adds a public client configuration to the development OIDC setup
script.

Fixes #16135

Change-Id: Iadd85d40c2faa595a0498e25d3407a1f94b5c8a8
Signed-off-by: Thomas Kosiewski <tk@coder.com>

Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-07 14:06:38 +01:00
Thomas Kosiewski acc1e67a77 ci(.github/dependabot.yaml): delete dependabot ignore rules for actions (#16443) 
Removed dependabot ignore rules for `actions/*` and `marocchino/sticky-pull-request-comment` GitHub Actions.

Change-Id: I3288c813c777e7045616f29d265c0860268ece4c
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-05 14:57:54 +01:00
Thomas Kosiewski 182bb9bdda chore: update actions/cache to v4.2.0 (#16441) 
This updates actions/cache to v4.2.0 and adds missing development
dependencies (gawk, gnutar, which, zip, gzip) to the Nix flake.

Change-Id: I1156810c9e02f0cef8e1345a1cbf2b6ba484974a
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-05 11:25:20 +01:00
Thomas Kosiewski 8c265018c4 fix(flake.nix): readd grep for agent startup script (#16429) 
Added `gnugrep` to the development shell dependencies, as its a dependency of the bootstrap script for an agent.

Change-Id: Ia56e16a831bb94af2324e33ae5274833d0123d47
Signed-off-by: Thomas Kosiewski <tk@coder.com>
 2025-02-04 17:14:03 +01:00
Thomas Kosiewski 6e2dc6ffd2 fix(flake.nix): include dev buildInputs in dogfood nix image (#16325) 2025-01-30 17:47:19 +01:00