mirror of
https://github.com/coder/coder.git
synced 2026-10-31 02:43:21 +08:00
c3cd692f867f235b86663f5d1f0607dc6174490e
1750
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Michael Suchacz
|
8eb1e43f0b |
feat: apply reasoning effort and surface thinking blocks for Google chat models (#28273)
## Problem
Selecting a reasoning effort on a Google (Gemini) chat model either got
dropped or broke the request outright, and even when the request
succeeded, Gemini chats never showed thinking blocks:
- The native `fantasygoogle` path had no case in `applyReasoningEffort`
at all, so effort was silently ignored. With only `include_thoughts`
configured, Gemini 3+ models default to no thinking, so chats produced
zero reasoning tokens.
- The path chat generation actually uses today routes Google providers
through aibridge as openai-compat clients and forwarded the raw effort
string as `reasoning_effort`. Google's OpenAI-compatible endpoint
validates that value against the model's supported thinking levels
instead of clamping, so out-of-range efforts fail the whole request with
HTTP 400. Verified against the live Gemini API: `gemini-3.1-pro-preview`
accepts low/medium/high and rejects none/minimal/xhigh/max. Because the
config's default effort is sent whenever the user selects nothing, a
config whose default falls outside the model's supported set failed
every no-selection message.
- Google's OpenAI-compatible endpoint never emits thought text for
`reasoning_effort` requests. Thoughts require
`extra_body.google.thinking_config` with `include_thoughts`, which is
mutually exclusive with `reasoning_effort`, and even then the thought
deltas come back inline in `content` (marked
`extra_content.google.thought`, wrapped in `<thought>` markers) rather
than in the `reasoning_content` field OpenAI-compatible clients parse.
So no reasoning parts were ever produced, persisted, or rendered for
Gemini chats (verified live in UAT).
## Fix
Reasoning effort:
- Map the global reasoning effort scale onto Google's `thinking_level`
in `applyReasoningEffort`, gated to Gemini 3+ models and clamped to each
model's supported subset (3.0 Pro launched with LOW/HIGH only, 3.1 Pro
added MEDIUM, Flash accepts all four through 3.6 while 3.7 and the
latest flash aliases drop MINIMAL, image models accept HIGH plus MINIMAL
for flash). Pre-Gemini-3 models reject `thinking_level`, so effort
degrades to a no-op for them and a config-pinned level is dropped at
call time instead of failing every generation.
- Add `thinking_level` to `ChatModelGoogleThinkingConfig` so admins can
pin a level for configs without user-selectable effort. A resolved
per-turn effort overrides the pinned level; an explicit
`thinking_budget` wins over both since the Google API rejects requests
that set `thinking_budget` and `thinking_level` together, and config
validation now rejects that combination up front. The admin panel picks
up the new field automatically through the generated model-options
schema.
Thinking blocks on the openai-compat path (the path chat generation
uses), translated in both directions at the existing Gemini transport
seam:
- chatd's OpenAI-compat request patch swaps `reasoning_effort` for an
equivalent `extra_body.google.thinking_config` with `include_thoughts`
enabled: `thinking_level` on Gemini 3+ (clamped to the model's supported
subset), `thinking_budget` on earlier models (effort `none` maps to
budget 0, which disables thinking). Non-Gemini models and non-Google
routes are untouched.
- The same transport rewrites Google's responses so thought output
surfaces as `reasoning_content`, which fantasy already converts into
reasoning parts: streaming bodies are rewritten per SSE line (preserving
streaming latency), JSON bodies in place, with the
`<thought>`/`</thought>` markers stripped at the transitions.
- aibridge's chat-completions interception preserves the `extra_body`
passthrough that openai-go's typed params silently drop, forwarding it
to Google upstreams only, so the chatd -> aibridge -> Google route
carries the thinking config end to end.
## Validation
- `go test ./coderd/x/chatd/... ./aibridge/... ./internal/googleopenai
./codersdk` pass; golangci-lint clean on touched packages.
- New unit tests cover the effort-to-level mapping for every effort
value, per-model clamping on both paths, the effort-to-thinking_config
request swap (level, budget, pinned-config precedence, non-Gemini
passthrough), SSE and JSON thought-response rewriting (fixtures mirror
live captures, including tool-call and no-thought streams), an
end-to-end fake-SSE test through the fantasy pipeline proving reasoning
deltas emerge with markers stripped, and aibridge `extra_body`
preservation (including proof the typed round trip drops it).
- Verified live against `generativelanguage.googleapis.com` on
`gemini-3-flash-preview` and `gemini-2.5-flash` through the real
chatprovider stack: reasoning deltas stream through with markers
stripped and answer text intact.
- Live UAT on a dogfood deployment (remote Coder Agents chat, 4 rounds):
effort clamping verified end to end (rounds 1-2), the missing-thoughts
limitation isolated (round 3), and the fix validated (round 4):
expandable, readable thinking blocks render for
`gemini-3.1-pro-preview`, `gemini-3-flash-preview`, and
`gemini-2.5-flash` at effort=high; no `<thought>` marker leaks;
effort=none produces no block and no 400; multi-turn history replay
works with a thinking block on each turn.
- `make gen` regenerated `typesGenerated.ts` and
`chatModelOptionsGenerated.json`; `pnpm -C site lint:types` passes.
> 🤖 Mux acted on Mike's behalf to create this pull request.
<!-- mux-attribution: model=claude-opus-4-6 thinking=high -->
|
||
|
Susana Ferreira
|
50ecd6959f |
feat: add a source filter to the AI model price list (#28329)
## Description With prices from the price book and prices set through the API both in the table, `list` showed only the price in effect and gave no way to tell where it came from. There was also no way to see the price book's entry for a model that had been overridden, since the listing resolves to the custom price. This adds a `source` column to the output and a `--source` filter for narrowing to one source, or to every row with `all`. Note: `--source` selects rows rather than models. A model carrying both a price book price and a custom one appears under either value, so the two filtered listings do not sum to the unfiltered one. `--source all` reports both rows at once. ## Changes - Add `Source` to `codersdk.AIModelPrice` and `AIModelPricesFilter`, and map it in `db2sdk`, which previously dropped it. - Filter `GetAIModelPrices` by source before resolution, so the price book's row for an overridden model stays reachable. - Accept `all` as a source, which reports every price a model holds instead of resolving to one. - Accept a `source` query parameter on the list endpoint, rejected with a 400 when outside the enum. - Add `--source` to the list command and a `source` column to its output. - Document the column and the filter, and lead the price section with `coder exp ai-model-prices list`. - Distinguish a deployment's default prices from the price book that ships with each release. Closes https://linear.app/codercom/issue/AIGOV-593 > [!NOTE] > Initially generated by Claude Opus 5, modified and reviewed by @ssncferreira |
||
|
Danielle Maywood
|
4a7a0a3348 | feat: label process_output rows with the process command (#28300) | ||
|
Michael Suchacz
|
2f996d765a |
feat: remove task management MCP tools (#28354)
Removes the six task-management tools from the MCP tool surface in `codersdk/toolsdk`: - `coder_create_task` - `coder_delete_task` - `coder_list_tasks` - `coder_get_task_status` - `coder_send_task_input` - `coder_get_task_logs` Both the coderd-hosted MCP server (`coderd/mcp`) and the CLI `coder exp mcp server` derive their tool list from `toolsdk.All`, so removing the registry entries removes the tools from both servers with no server-side changes. The MCP server documentation (`docs/ai-coder/mcp-server.md`) is updated to drop the task-management entry from the tool overview. ## What stays - `coder_report_task` is intentionally kept: it is the agent-side status-reporting mechanism (`WithTaskReporter`), already excluded from the hosted MCP server, and required by `coder exp mcp server`, Claude Code configuration, and scaletest. - The `codersdk` Task SDK methods (`CreateTask`, `DeleteTask`, `Tasks`, `TaskLogs`, etc.) and the Tasks feature (CLI, UI, API) are untouched; this PR only removes MCP tool registration. - The task-only `taskIDDescription` helper is deleted; `userDescription` stays (used by `coder_create_workspace`). The code change is deletion-only: 970 removed lines across `toolsdk.go` and its test file, no added behavior. ## Validation - `go build ./...`, `go test ./coderd/mcp/`, and targeted `golangci-lint` pass. `go test ./codersdk/toolsdk/` passes all tests; the package-level goleak failure (agentssh goroutines) reproduces identically on unmodified `main` and is unrelated to this diff. - `pnpm run lint-docs` passes on the docs change. - Remote dogfood UAT ran against the toolsdk change and passed: `tools/list` on both the hosted MCP endpoint and `coder exp mcp server` confirms the six tools are absent and the remaining surface is intact (34 hosted tools; CLI adds `search`/`fetch`), a `tools/call` of `coder_get_authenticated_user` succeeds, and calling a removed tool returns `-32602 unknown tool`. `coder_report_task` still registers in the CLI server when `CODER_MCP_APP_STATUS_SLUG` is set. - The failing `Pixel / Review` commit status is unrelated: this PR contains no `site/` changes, and the same status fails on the last 8 `main` commits including this PR's merge base (`676b9bb06bf`). > Mux acted on Mike's behalf to create this pull request. |
||
|
david-fraley
|
8203b2ebaf | feat!: hide Coder Tasks behind the enable-ai-tasks flag (#28008) | ||
|
Michael Suchacz
|
7ca7c30f40 |
feat: add group and user ACLs to MCP server configs (#27944)
Adds group and user ACLs to org-scoped MCP server configs so
organizations can restrict specific MCP servers to subsets of members,
mirroring the template ACL pattern.
## Summary
- Migration adds `group_acl`/`user_acl` JSONB columns (nested `{"<id>":
{"permissions": [...]}}` shape) and seeds every existing config with its
organization's Everyone group read entry (the Everyone group ID equals
the org ID), so member access is unchanged by default. Creation seeds
the same entry.
- The blanket org-member read grant from the base PR is replaced by ACL
evaluation: Rego requires org membership for every ACL grant, and
`GetAuthorizedMCPServerConfigs` compiles ACL-aware SQL filters.
- New `ActionShare` (org admins) gates `GET|PATCH
/api/experimental/organizations/{organization}/mcp-servers/{mcpserverconfig}/acl`
(nested under the organization like the rest of the config surface).
PATCH validates principals against the config's organization, merges
sparse updates under a row lock, stamps `updated_by`/`updated_at`, and
is audited as a Write with `Old` captured before authorization. A config
deleted concurrently between the middleware fetch and the locked
re-fetch is concealed as 404, matching the update and delete handlers.
ACL columns are tracked in the audit table.
- ACL management is available in all editions (no enterprise
entitlement), documented in the MCP servers page. Revoking an ACL does
not retro-strip already-selected configs from existing chats; new
selection is blocked at chat create.
- Force On respects the ACL: the forced set is loaded as the chat owner,
so a `force_on` server whose ACL denies the owner never attaches at
create, send, or generation time.
- No rolling-upgrade machinery: upgrades run in scheduled maintenance
downtime, so the migration only backfills existing rows; the API sets
the Everyone read grant explicitly on every insert.
Stacked on #27943. Part of the MCP org-separation stack.
Closes https://linear.app/codercom/issue/CODAGT-712
UAT: verified on a dogfood instance with two members and a custom group:
Everyone-seed default visibility, group grant with Everyone removal
(non-member loses list/fetch/selection), foreign-principal rejection,
non-admin share denial (audited 403 / concealed 404), user_acl restore,
and audited ACL diffs.
> Mux (AI agent) authored this PR on Mike's behalf.
<!-- mux-attribution: model=claude-fable-5 thinking=high -->
|
||
|
Michael Suchacz
|
299e72ad30 |
feat: audit MCP server config changes (#27943)
Adds enterprise audit logging for MCP server config create, update, and
delete, with strict secret redaction. MCP configs hold credentials
(OAuth2 client secrets, API keys, custom headers), so admin changes to
them need an audit trail.
## Summary
- `enterprise/audit/table.go` gains an `MCPServerConfig` entry
enumerating every column: `oauth2_client_secret`, `api_key_value`, and
`custom_headers` are `ActionSecret` (never appear in diffs); dbcrypt
`*_key_id` bookkeeping, IDs, and timestamps are ignored; the remaining
config fields, including the endpoint URL fields, are tracked so
auditors can see which endpoints a config points at.
- Type registration in `coderd/audit` (diff, request, resource target
with org attribution), `codersdk/audit.go`, and a `resource_type` enum
migration.
- Handlers wire `audit.InitRequest`: create records `New`; update and
delete record `Old` from the param middleware before the
write-authorization check, so a readable-but-not-writable caller
produces an audited 403 while read-denied callers stay concealed as
unaudited 404s.
- Tests: create/update/delete audit entries, write-denied and
delete-denied 403 auditing, cross-org concealment producing zero
entries, and a serializer-level regression test proving none of the
three secret classes can reach a serialized diff.
- Review round: MCP config audit entries link to
`/ai/settings/mcp-servers/{id}`, audit table comments are trimmed per
review, and a fault-injection test pins that a config row surviving a
failed post-discovery credential update still gets its creation audit
entry.
Stacked on #27942 (org-scoped MCP configs). Part of the MCP
org-separation stack.
Closes https://linear.app/codercom/issue/CODAGT-717
UAT: verified on a trial-licensed dogfood instance: audit entries for
the full CRUD lifecycle with correct actor/org/target, redacted secrets
in the update and OAuth2 create diffs, and a full plaintext scan of the
audit dump finding zero secret leaks.
> Mux (AI agent) authored this PR on Mike's behalf.
<!-- mux-attribution: model=claude-fable-5 thinking=high -->
|
||
|
Michael SuchaczandMathias Fredriksson
|
443e3b9b80 |
feat!: org-scope MCP server configs with RBAC (#27942)
Moves `mcp_server_configs` from deployment scope to organization scope
so each organization fully controls the MCP servers its members can use
with Coder Agents.
## Summary
- Migration: adds `organization_id` (NOT NULL, FK) and keeps existing
rows as the default organization's originals with credentials intact.
Other organizations start with no MCP servers and configure their own;
nothing is copied across organizations. Chats outside the default
organization keep any now-cross-organization `mcp_server_ids` entries;
the runtime already ignores IDs that do not resolve in the chat's
organization, so no data rewrite is needed. Slug uniqueness becomes
`(organization_id, slug)`.
- RBAC: new org-scoped `ResourceMCPServerConfig` with regosql converter
and `GetAuthorizedMCPServerConfigs`; org admins get in-org CRUD, org
members get read (replaced by ACL evaluation in the follow-up ACL PR in
this stack).
- API: all config routes nest under the organization, matching
templates: `POST|GET
/api/experimental/organizations/{organization}/mcp-servers` and
`GET|PATCH|DELETE .../mcp-servers/{mcpserverconfig}` (plus
`oauth2/connect`), resolved by a read-only param middleware that
conceals read-denied and cross-organization access as 404. Two routes
stay on the frozen `/api/experimental/mcp/servers/{mcpServer}` block:
the OAuth2 callback (the redirect URI baked into existing AS-side client
registrations) and `oauth2/disconnect`, which must remain reachable by
users removed from the organization so they can still revoke their token
grant.
- Chat runtime: selection validation and generation resolve configs
strictly by IDs, enabled state, and the chat's organization in SQL;
requested duplicates are normalized; invalid or cross-org IDs are
rejected with the precise ID list. IDs already persisted on a chat are
exempt from message-time rejection so disabling a selected server never
blocks sends; generation skips servers that are no longer usable.
- Frontend: API layer and admin settings pages target the new endpoints.
The admin page manages the default organization's servers; the org
picker is tracked separately (CODAGT-714).
- Security hardening from review: OAuth user grants are additionally
bound to `oauth2_revocation_url` (changing it invalidates grants, and a
racing OAuth callback gets 409 instead of recreating a grant).
Stack-wide SSRF protection for MCP config-directed traffic was split
into its own PR at the top of this stack (#28242) to keep this diff
reviewable; this PR keeps main's existing discovery IP-range guard.
OAuth2 auto-discovery now completes before the config row is inserted: a
failed discovery persists nothing, and there is no provisional row that
concurrent updates could race against.
Two follow-up PRs in this stack were split out to keep this diff
reviewable: #28064 completes the swagger annotations for the moved
routes (main already ships these experimental MCP handlers unannotated),
and #28065 carries hardening fixes and regression pins on top of the
cutover.
- Force On enforcement (landed on main mid-review) is org-scoped: the
forced set is read per chat organization
(`GetForcedMCPServerConfigsByOrganization`), so another organization's
`force_on` server never attaches to a chat.
## Breaking changes (experimental API)
The MCP server config endpoints move from the deployment-scoped
`/api/experimental/mcp/servers` block to organization-nested paths:
`POST|GET /api/experimental/organizations/{organization}/mcp-servers`
and `GET|PATCH|DELETE
/api/experimental/organizations/{organization}/mcp-servers/{mcpserverconfig}`
(plus `oauth2/connect`). The old paths are removed, so API consumers
must supply an organization. Two routes intentionally stay on the frozen
`/api/experimental/mcp/servers/{mcpServer}` block: the OAuth2 callback
(its redirect URI is baked into existing AS-side client registrations)
and `oauth2/disconnect` (must remain reachable by users removed from the
organization). These endpoints are under `/api/experimental`, so no
deprecation window is provided.
## Rolling upgrades
During a rolling deploy, an old replica creating an MCP config can fail
the new `NOT NULL organization_id` constraint until it is upgraded
(reads are unaffected: old binaries' generated queries select their own
column lists). This matches the repo's existing precedent for additive
NOT NULL migrations (000562) and affects only the admin config-create
path in the upgrade window.
Upgrades are expected to run in scheduled maintenance downtime with the
database locked during migration, so the migration ships no
rolling-upgrade compatibility machinery. The down migration deletes
organization-created configs (their chat references are cleaned by the
000510 delete trigger) and restores deployment-wide slug uniqueness.
Part of the MCP org-separation stack (CODAGT-711 -> CODAGT-717 audit ->
CODAGT-712 ACLs -> CODAGT-806 token RBAC).
Closes https://linear.app/codercom/issue/CODAGT-711
UAT: validated end to end on a two-org dogfood deployment, including a
real pre-migration to post-migration upgrade, cross-org isolation
(404s), same-slug-two-orgs, chat selection gating, and a live MCP tool
call through the org-scoped generation path. The migration was later
revised to keep existing rows in the default organization only (no
per-organization copies); that revision is covered by the migration test
suite.
> Mux (AI agent) authored this PR on Mike's behalf.
<!-- mux-attribution: model=claude-fable-5 thinking=high -->
---------
Co-authored-by: Mathias Fredriksson <mafredri@gmail.com>
|
||
|
Ben Potter
|
8a7e8d9d5b |
refactor: move chat prompt sanitization into codersdk (#28283)
Exports `SanitizePromptText` (and its helpers) from `codersdk` instead of `coderd/x/chatd`, so API consumers can sanitize prompt text exactly the way the server stores it. ## Why coder/terraform-provider-coderd#412 adds a `coderd_chat_system_prompt` resource whose `system_prompt` attribute compares values by their sanitized forms (otherwise the trailing newline from `file("system-prompt.md")` is perpetual drift, since the server stores the sanitized value). That currently requires a mirrored copy of the sanitizer in the provider, which can silently rot. Per review there ([discussion](https://github.com/coder/terraform-provider-coderd/pull/412#discussion_r3801110795)), the sanitizer should be exported from `codersdk` and imported instead. ## What - `coderd/x/chatd/sanitize.go` → `codersdk/promptsanitize.go` (pure move; package + doc-comment note about why it lives in codersdk) - `coderd/x/chatd/sanitize_test.go` → `codersdk/promptsanitize_test.go` - Callsites updated (`exp_chats.go`, `chatd.go`, `subagent.go`, `context_prompt.go`); no wrapper left behind, single source of truth - No behavior change > Generated by Coder Agents on behalf of @bpmct |
||
|
Bobby Ho
|
663f41ffa9 |
feat: derive OAuth2 client type from token_endpoint_auth_method (#28043)
Adds an OAuth2 client type (public vs confidential, RFC 7591 §2) derived from the requested auth method instead of hardcoded confidential. The type is stored and guarded here, but no endpoint enforces on it yet; public behavior at the token endpoint follows in the next PR in the stack. - Client type is derived once and reused by both registration and redirect URI validation, so they can't disagree - IsPublic() fails closed: an unrecognized or missing value reads as confidential - RFC 7592 update (PUT) now rejects moving a client between public and confidential (400) instead of silently flipping it when the auth method is omitted - Discovery still doesn't advertise "none"; follows once the token endpoint honors it ### Behavior by client shape `client_type` is derived from `token_endpoint_auth_method` at POST and pinned at PUT. RFC 7592 GET/PUT authenticate with the registration access token, not the client secret, so neither endpoint reads a secret. | Registered with | Stored `client_type` / method | GET reports | PUT that flips the method | |------------------------------------|---------------------------------------|-----------------------|-----------------------------------------| | omitted, or `client_secret_basic` | `confidential` / `client_secret_basic` | `client_secret_basic` | `none` → 400 `invalid_client_metadata` | | `none` (new) | `public` / `none` | `none` | `client_secret_*` → 400 `invalid_client_metadata` | | `none` (before this PR) | `confidential` / `none` | `none` | either → 200, type stays `confidential` | - PUT still replaces every other RFC 7591 field. `client_type` is the only pinned one; the method may move within a type (`client_secret_basic` ↔ `client_secret_post`). - Row 3 is the only shape where the two columns disagree. The guard fires only on a method change that crosses the type line, so those clients keep managing themselves instead of being locked out of their own configuration endpoint. - The token endpoint does not consult `client_type` yet, so every client still authenticates with a secret and registration still issues one. Split out of #27873, second in the stack (on top of #28041). Refs https://linear.app/codercom/issue/ENG-3029/oauth2-support-public-client |
||
|
Jon Ayers
|
821d91fabd | fix: log tailnet tunnel authorization decisions (#27819) | ||
|
Mathias FredrikssonandMichael Suchacz
|
d3f08b1983 |
feat: audit chat system instructions changes (#27668)
Adds an audit record for administrative events on the deployment-wide
chat instruction settings (system prompt, the include-default toggle,
and the plan-mode instructions), per CODAGT-719 and operator decision
D5. Each endpoint records under a stable identity: resource type
`chat_instruction_settings`, a fixed resource ID and a human-readable
target ("System prompt", "Plan mode instructions"), so two changes to
one setting share an ID and history-by-setting works. A real change
exports a Write entry with the old-to-new text visible; a
value-identical PUT still upserts and still returns 204 but records
nothing.
Attempts are recorded, not only transitions. Identity is assigned before
the authorization check, so a denied PUT exports a 403 row with an empty
diff (no request content reaches it), a validation failure exports a 400
row, and a write failure exports a 500 row, each with an empty diff; an
operator can tell "nothing changed" from "something changed and capture
degraded" by the status code.
The write path stays authoritative. The advisory lock and, on plan-mode,
the transaction exist only to serve change-detection; if any of that
machinery fails (lock, begin, commit, rollback), the handler runs main's
idempotent write path directly and derives the response from it, so a
member-visible failure of audit-only infrastructure can never replace
main's successful response. Accepted consequence: when the lock cannot
be taken, two concurrent identical writes can produce two rows instead
of one. That is audit degradation, which is allowed; changing a member's
response is not. Write failures keep the exact response the endpoint
produced before this wiring (transaction error for the system prompt,
which was always transactional; the raw write error for plan mode, which
was not), and the full transaction error is logged so rollback failures
cannot vanish.
<details>
<summary>CODAGT-66 plan entry: S1 (verbatim)</summary>
**S1 `feat: audit chat system instructions changes`** (CODAGT-719; base:
main)
- Struct: `database.ChatSystemPromptSettings{ID uuid.UUID; SystemPrompt
string; IncludeDefaultSystemPrompt bool; PlanModeInstructions string}`
in `coderd/database/types.go` (ticket-sketched shape; one struct, both
endpoints).
- Registration: union entry (diff.go), table.go entry (`id`
ActionIgnore, other three ActionTrack), `AuditActionMap` Write-only;
four request.go cases (`ResourceTarget` "", `ResourceID` from struct,
`ResourceType` new enum value `chat_system_prompt_settings`,
`ResourceRequiresOrgID` false with the "Artificial ID / deployment
singleton" comment convention).
- Migration: `ALTER TYPE resource_type ADD VALUE IF NOT EXISTS
'chat_system_prompt_settings';` comment-only no-op down (000558 shape);
number picked at push per the numbering constraint.
- codersdk: constant + prose `FriendlyString` ("chat system prompt
settings"); `TestAuditDBEnumsCovered` forces both. `coderd/audit.go`
presentation switches: rely on safe defaults (no link, generic
description); no FE changes (filter label falls back to capitalized
value; acceptable per precedent).
- Wiring `putChatSystemPrompt` and `putChatPlanModeInstructions`:
InitRequest with Action Write; artificial `ID: uuid.New()` on `New` only
when a change is detected; no-op suppression by leaving both aReq sides
unset (nil resource IDs skip the log, request.go skip rule); the write
path itself stays byte-identical (upserts still run unconditionally).
- `putChatSystemPrompt` (writes two keys conditionally in one existing
tx): inside that tx, read the pair via `GetChatSystemPromptConfig` for
`Old`, perform the conditional writes exactly as today, then RE-READ the
pair for `New`. The re-read is load-bearing:
`include_default_system_prompt` is computed from the toggle row AND the
prompt, so a prompt-only write can flip the effective value without the
request carrying the pointer. `PlanModeInstructions` stays zero on both
sides.
- `putChatPlanModeInstructions` (no tx exists today): wrap its
read-upsert in `InTx` (behavior-preserving: same single write);
`Old`/`New` populate only `PlanModeInstructions`; the two system-prompt
fields stay zero on both sides; no cross-key reads.
- Change detection compares the populated payload fields only (never the
artificial ID).
- Tests: handler-level coderdtest with `audit.NewMock()` asserting Write
entry on change and NO entry on a value-identical PUT, for both
endpoints (this also exercises `ResourceRequiresOrgID` end to end); the
fallback-flip case (no explicit include-default row, nonempty prompt set
to empty, effective boolean flips: entry emitted with the boolean diff);
diff assertions (old->new prompt text tracked, not secret) in
`enterprise/audit/diff_internal_test.go`; `TestAuditableResources`
passes by construction.
- Bookkeeping at PR open: correct CODAGT-719's no-op premise ("matches
the existing 204-on-unchanged behavior" does not exist on main;
suppression is new, write path unchanged).
- Review focus: Old capture and the New re-read inside the tx (three of
four existing singletons never set Old; do not copy them; and the
computed include-default value makes a naive New construction wrong);
the skip-on-no-op mechanism; prompt text deliberately visible in diffs.
</details>
Note: the plan excerpt above predates operator decision D5 (2026-07-30),
which this PR implements: the resource type is
`chat_instruction_settings` (not `chat_system_prompt_settings`), each
setting carries a stable ID and a display-name target (not a per-write
artificial ID and an empty target), no-op suppression runs through
`InitRequestWithCancel` (not the nil-ID skip), and attempts (denied,
failed, capture-degraded) record rows with real statuses and empty
diffs. Ticket bookkeeping for CODAGT-719 was corrected on Linear at
kickoff: the ticket's "matches the existing 204-on-unchanged behavior"
premise does not exist on main; suppression is new, and the write path
is unchanged.
> 🤖 This PR was created with the help of Coder Agents, and _will be_
reviewed by a human. 🏂🏻
---------
Co-authored-by: Michael Suchacz <203725896+ibetitsmike@users.noreply.github.com>
|
||
|
Michael Suchacz
|
affeeaf9c8 |
feat: extend agent chat MCP tools for remote UAT evidence loops (#28233)
Extends the Agent-chat MCP tools so an unattended UAT evidence loop can
fetch artifacts, monitor long runs, and find prior runs without burning
model context.
## Backend
- New `chat_files_token` crypto key feature (migration 000571) with
rotator support and a dedicated signing keycache on coderd.
- `POST /api/experimental/chats/files/{file}/download-url`
(authenticated) mints a short-lived (5 min) signed URL and returns it
with `sha256`, `size_bytes`, `name`, `mime_type`, and `expires_at`.
- `GET /api/experimental/chats/files/{file}/download?token=` (no session
token) redeems the signed URL: verifies the JWS, requires the token's
`file_id` to match the path, and re-checks the minting user's RBAC
access live at redemption. Clients can `curl -o` artifacts with zero
credentials in the URL consumer.
- `ChatFileMetadata` gains `size_bytes` (via `octet_length`, no bytes
fetched).
## MCP tools (`codersdk/toolsdk`)
- `coder_download_chat_file`: by `file_id` or `chat_id`+`file_name`;
returns the signed URL plus checksum and size instead of base64.
- `coder_await_chat`: blocks (bounded `wait_secs`, 1-120) until a chat
leaves `running`/`interrupting`, using the existing watch stream with
subscribe-before-read.
- `coder_list_chats`: label, query, and limit filtering; chat
projections now include labels.
- `coder_get_chat_messages`: `after_id` forward cursor with
`next_after_id` (exact incremental reads), plus per-message `files`
metadata so artifact-bearing messages are identifiable.
- `coder_get_chat`: file listings now include `size_bytes` and
`created_at`.
- `coder_list_templates`: exposes `agents_allowed` for pre-flight
checks.
## Testing
- coderd: mint/redeem happy path with an unauthenticated client,
expired/tampered/file-mismatched tokens, auth still required on the
plain file endpoint, non-owner mint rejection.
- toolsdk: harness + integration coverage for all new/changed tools,
including signed-URL redemption with checksum verification,
forward-cursor exactness, await transition/timeout paths, and label
filtering.
- Remote dogfood UAT (dev.coder.com Coder Agent) passed all six
acceptance scenarios end to end over both MCP transports.
Note: `go test ./codersdk/toolsdk/` has a pre-existing goleak flake on
main (leaked `agentssh` non-PTY session goroutines from SSH exec tests;
reproduced 3/3 on clean `b4971bc49f1`). It is unrelated to this diff.
> Mux acted on Mike's behalf to create this PR.
<!-- mux-attribution: model=claude-sonnet-4-6 thinking=high -->
|
||
|
Michael Suchacz
|
7724ee281a |
feat: defer MCP tool schemas behind a find_tools search (#28225)
## Summary When the `mcp-tool-search` experiment is enabled, chatd stops inlining connected MCP tool schemas into every generation. It instead exposes a built-in `find_tools` tool whose description carries a compact catalog of the deferred tools, and only ships full JSON schemas for tools the model has activated by searching or by calling them directly. Closes [CODAGT-760](https://linear.app/coder/issue/CODAGT-760). ## Problem Tool-heavy agent configurations (GitHub, Linear, Notion, and dev-tooling MCP servers) inline over 100k tokens of tool schema definitions into every generation. Initial uncached requests reached ~216k tokens with time-to-first-token close to nine minutes, while the model typically invokes only a handful of tools per turn. ## How it works - `decideMCPToolSearch` defers external and workspace `.mcp.json` MCP tools whenever the experiment is enabled. Native, dynamic, provider, skill, and transport tools are never deferred. - `find_tools` embeds a server-grouped catalog in its tool description (degrading to names-only, then counts-only, then a constant-size summary past a context-scaled size cap) and scores keyword matches across tool names, descriptions, parameter schemas, and server metadata. Queries can scope to one server with a `server:` prefix, and exact `names` arguments always activate. - Activation state is ephemeral: it is re-derived each generation from surviving chat history (`find_tools` results and direct calls to deferred tools), so activations naturally lapse when compaction summarizes them away. Aggregate activated schema weight is capped at 10% of the context window, shedding the least recently activated schemas first; `find_tools` shares that budget across parallel calls in one step. No new persistence. - Deferred tools stay registered for execution, so the model can call a cataloged tool directly without searching first; the schema is activated for subsequent steps. - Fail-open: the experiment being disabled, an empty candidate set, or an MCP tool named `find_tools` all disable deferral, leaving today's behavior byte-identical on the wire. - Prometheus counters/histograms track `find_tools` calls, matches, activations, and deferred token weight. - The conversation timeline renders `find_tools` calls with a collapsed search summary and expandable match list, falling back to the generic renderer on malformed payloads. ## Validation - Unit tests for the catalog, matcher, experiment-gated decision, and activation derivation; end-to-end chatd generation tests covering search-then-call, direct-call activation, experiment-off wire parity, compaction lapse, and subagent tool gating. - Storybook interaction tests for the timeline rendering and malformed-payload fallback. - Remote dogfood UAT on dev.coder.com passed: deferral with a real MCP server and Anthropic model, direct calls without prior search, activation persistence across turns, experiment-off parity, and clean UI/console. > Disclosure: Mux (AI agent) authored this PR on Mike's behalf. |
||
|
Michael Suchacz
|
119f2b1dd9 |
feat: limit concurrent chat agents with pooled admission (#27902)
Limits concurrent chat generation on capped deployments to 5 root chats and 10 delegated subagent chats. The pools are deployment-wide and independent, so delegated work can continue while root capacity is full. The default caps live in AGPL code. Enterprise contributes only a licensing unlock, so unlicensed deployments stay capped and cannot fail open. Licensed deployments are uncapped while Agent Hours usage stays below an explicit hard limit. Deployments without a hard limit remain uncapped, and reaching the Agent Hours allocation only triggers warnings. Admission happens before a worker takes chat ownership. Capped deployments serialize admission across replicas with a transaction-scoped advisory lock and derive active and queued state from current ownership plus fresh runner heartbeats, rather than persisted queue markers or per-replica state. The acquisition query returns a bounded, pool-interleaved candidate set instead of ranking the whole backlog; a migration replaces the acquisition index with a pool-aware one. Refused chats stay running but unowned, and interrupt requests bypass admission so users can stop queued or over-cap chats. The single-chat API derives `queued_for_capacity` from live pool state; list endpoints do not report it. The UI polls that value every 5 seconds while a chat is running and shows a callout when the chat is waiting for capacity. Updates the administrator documentation and deployment-wide Prometheus gauges for active and queued agents. Replica-level values must be aggregated with `max`, not `sum`. > Mux updated this PR on Mike's behalf. |
||
|
Jaayden Halko
|
fa8ffe4eda |
feat: report agent runtime hours usage in entitlements (#27985)
Populate `FeatureAgentRuntimeHours.Actual` on every entitlements refresh for licenses that grant the feature. A new `GetTotalUsageHBAgentRuntimeV1` query sums `runtime_ms` over the license's usage period, reading `usage_events` directly: `hb_agent_runtime_v1` is exactly one row per hourly bucket deployment-wide with `created_at` at the bucket start, enforced by the unique partial index introduced in #27983. The measurement reuses the shared `measureUsage` policy from #27984 through a new `AgentRuntimeMsFn` closure (usage publisher subject): failures publish the stable `LicenseAgentRuntimeUsageUnavailableErrorText` and log the cause. Usage is floored to whole hours, matching the unit of the `agent_runtime_hours_*` claims, and at most one warning is emitted per refresh: reaching the allocation supersedes the advisory soft limit. The dashboard renders the soft-limit advisory muted without a sales link and treats the runtime usage-unavailable text as a diagnostic. **Precise usage.** `Feature.ActualMs` (JSON `actual_ms`), set only for `agent_runtime_hours`, carries the exact stored milliseconds backing the floored `Actual` so clients can render fractional hours (e.g. `10.3`). It has the same freshness as `Actual`; the whole-hour warning thresholds are unchanged. **Unlimited licenses.** A license minted with the unlimited (`-1`) allocation decodes to an enabled feature with a nil `Limit` (#27984), so the warning write-back now guards the allocation dereference: no thresholds can exist for an unlimited license, so no runtime hours warning is ever emitted, while `Actual` is still measured and published. `Feature.Compare` is unchanged; for usage-period features the issued-at/end dates decide first, so a metered feature outranks an unlimited one only on an exact timestamp tie, an edge pinned by a `TestFeatureComparison` case and documented on `decodeAgentRuntimeHours`. **Grandfathered premium licenses.** Premium licenses without `agent_runtime_hours_*` claims are now granted the feature disabled with a zero limit over the license term, identical to an explicit `allocation: 0`: usage is measured and published for every Premium deployment, and chatd's pooled admission (#27902) caps concurrent agentic chats until a license with a positive allocation is added. The default carries a fixed early `UsagePeriod.IssuedAt` (2026-08-01, the same mechanism as the managed-agents default) so any license actually carrying the claims outranks it in the `AddFeature` merge regardless of the licenses' relative issue dates; the constant must stay earlier than the earliest legitimately issued claim-bearing license. Zero allocations (explicit or grandfathered) emit no deployment-wide warning banner: those deployments are steered by the in-page upgrade CTA and the concurrency cap. Enterprise licenses are unchanged. Part 3 of a 3-PR stack splitting up #27796 (see there for review history). Stack: #27983 → #27984 → this PR. Closes CODAGT-852. |
||
|
Jaayden Halko
|
d15800b494 |
feat: tolerate unusable runtime hours claims and decode -1 allocation as unlimited (#27984)
Two coupled changes to the license/entitlements layer, preparing for runtime-hours usage reporting. **Tolerate unusable runtime hour claims.** Unusable `agent_runtime_hours_*` claim combinations no longer reject the whole license: rejecting a signed license over a cosmetic threshold claim would drop the deployment to unlicensed. `decodeAgentRuntimeHours` drops the unusable claims, surfaces the stable `LicenseAgentRuntimeHoursClaimsIgnoredWarningText` (deduplicated across licenses), and logs the affected license and claims through the new `FeatureArguments.Logger`; `validateAgentRuntimeHours` and its license-invalidating errors are removed. The dashboard recognizes the stable diagnostic text and renders it muted, with a "License notices" heading instead of the exceedance heading and without a sales link. **Unlimited allocation.** An `agent_runtime_hours_allocation` claim of exactly `-1` (`AgentRuntimeHoursUnlimitedAllocation`, mirrored in coder/license) is reserved to mean unlimited: it decodes to an enabled feature with no `limit` in `/api/v2/entitlements`, the shape the UI already renders as "Unlimited". Threshold claims alongside it have nothing to threshold against, so they are dropped with the claims-ignored warning, and any other negative allocation remains unusable. The issuer-side counterpart (refusing to mint `-1` together with threshold claims) is coder/license#49. The managed agent measurement path is intentionally untouched: managed agents are deprecated and slated for removal, so the shared usage-measurement failure policy (`measureUsage`) now lands in #27985 next to its runtime-hours consumer instead of converting a doomed call site here. Part 2 of a 3-PR stack splitting up #27796 (see there for review history). Stack: #27983 → this PR → #27985. |
||
|
Asher
|
b5d18bb9c9 | feat: add redirect URL override for external auth (#28082) | ||
|
Michael Suchacz
|
8d4d0b35dd |
feat: add Coder Agents chat tools to the MCP toolsdk (#28025)
Exposes the experimental Coder Agents chats API through the MCP tool registry, so MCP clients (the hosted `/api/experimental/mcp/http` server and `coder exp mcp server`) can start and drive server-side coding agents. New tools in `codersdk/toolsdk`, all thin wrappers over existing `codersdk.ExperimentalClient` methods: | Tool | Wraps | |---|---| | `coder_create_chat` | `CreateChat` (prompt, optional org, model config, labels) | | `coder_get_chat` | `GetChat` (status, last error, last turn summary, workspace, files) | | `coder_get_chat_messages` | `GetChatMessages` (user-facing parts, chronological, cursor pagination, queued prompts) | | `coder_send_chat_message` | `CreateChatMessage` (queue or interrupt busy behavior) | | `coder_interrupt_chat` | `InterruptChat` | | `coder_archive_chat` | `UpdateChat` with `archived: true` | | `coder_list_chat_model_configs` | `ListChatModelConfigs` (enabled configs with default flag) | Both MCP servers register tools from `toolsdk.All`, so no additional wiring is needed. Responses are trimmed to what an MCP caller needs (IDs as strings, user-facing transcripts) rather than full SDK payloads. No new endpoints and no database changes. Also adds MCP [prompts](https://modelcontextprotocol.io/specification/2026-07-28/server/prompts) for the chat workflows, defined once in `codersdk/toolsdk` and registered by both servers: | Prompt | Purpose | |---|---| | `coder_agents_delegate` | delegate a task to a Coder Agents chat and monitor it to completion | | `coder_agents_check` | check the status and recent activity of an existing chat | Each prompt declares the tools its workflow needs; the stdio server skips prompts whose tools are excluded by `--allowed-tools`. Tests run the tools against a chat-enabled coderdtest instance (fake OpenAI-compatible provider plus in-process AI bridge), covering the full lifecycle, an interrupt against a blocked turn, pagination cursors, permission-dependent model config filtering, and argument validation. Prompt coverage spans SDK rendering, the hosted `prompts/list`/`prompts/get` round trip, and the stdio server including allowlist gating. > Mux created this PR on Mike's behalf. |
||
|
Susana Ferreira
|
2d9b6eda8f |
feat: add experimental CLI to price unpriced AI models (#27926)
## Description AI Gateway computes the cost of an interception from `ai_model_prices`, which is seeded on every server start from a price book embedded in the binary. A model the price book does not cover records a NULL cost, so its spend is invisible to cost reporting and is not enforced against budgets. The only fix was to wait for a Coder release that added the model. This adds an experimental CLI, backed by an experimental HTTP endpoint, for pricing those models. Models the price book already covers are rejected, because the seeder re-applies the book on every start and would overwrite an operator price. Support for custom pricing will be handled in https://linear.app/codercom/issue/AIGOV-589/extend-experimental-cli-command-to-set-custom-ai-model-prices. ## Commands ``` coder exp ai-model-prices list [--provider] [--model] coder exp ai-model-prices update [file|-] [--provider] [--model] [--input-price] [--output-price] [--cache-read-price] [--cache-write-price] [--yes] ``` ## Changes - Add `GET` and `POST /api/experimental/ai/model-prices`, gated behind the AI Bridge entitlement and the existing `ai_model_price` RBAC resource. - Add a `GetAIModelPrices` query with optional `provider` and `model` filters applied in SQL. - Validate the whole request before writing anything, so one bad entry cannot leave the table half updated, and report every problem at once. - Reject prices for models the embedded price book already covers, through a new `prices.IsDefaultPriced`. - Add the `coder exp ai-model-prices` command with `list` and `update`. `update` accepts a JSON document or the single-model flags and prints a plan, asking to confirm unless the document is piped in or `--yes` is passed. - Consolidate the supported provider list into `coderd/aibridge/prices/providers` so the price generator and the server share one definition. - Add `codersdk` types and client methods for both endpoints, and bound the request body at 1 MiB. - Document the command in the AI Gateway cost controls page. Closes https://linear.app/codercom/issue/AIGOV-567/experimental-cli-command-to-set-prices-for-unpriced-ai-models > [!NOTE] > Initially generated by Claude Opus 5, modified and reviewed by @ssncferreira |
||
|
Steven Masley
|
f0c17291b3 |
feat: unhide --oidc-redirect-url server option (#28072)
Unhides the `--oidc-redirect-url` / `CODER_OIDC_REDIRECT_URL` server option so it appears in `coder server --help` and the deployment configuration docs. - Removed `Hidden: true` from the option in `codersdk/deployment.go` - Regenerated CLI golden files and docs via `make gen` --- > Generated with Coder Agents on behalf of @Emyrk |
||
|
Bobby Ho
|
209d1ca498 |
fix: reject PKCE code_verifier below RFC 7636 length floor (#28003)
The token endpoint accepted any non-empty `code_verifier`, so a one-character verifier was enough to authenticate. RFC 7636 §4.1 requires 43 to 128 characters from the unreserved set. That fix plus the related gaps review surfaced in the same path: - Enforce the length and charset floor on the verifier before the S256 comparison runs. - Validate the challenge at the authorize endpoint too. It was only checked for non-emptiness, so a malformed challenge was stored and then failed late at token exchange, blaming the wrong parameter. - A malformed verifier now returns `invalid_request` (RFC 6749 §5.2); a well-formed but wrong one still returns `invalid_grant` (RFC 7636 §4.6). Both looked identical before, so a client had no way to tell a syntax error from a hash mismatch and would retry the same bad verifier forever. - Revoke the authorization code when a PKCE check fails. Without that, a leaked code could be replayed with unlimited verifier guesses for its remaining lifetime, and RFC 6749 §10.5 requires codes to be single use. - Fix verifier generation in `scripts/oauth2/*.sh` and the docs example. They deleted reserved base64 characters instead of translating them to the URL-safe alphabet, so most runs produced verifiers under the new floor. Also carries #28041, which merged into this branch: public clients may register bare custom schemes such as `vscode://` again, with `mailto`, `tel`, and `sms` rejected. Split out of #27873 (public OAuth2 client support). PKCE is already mandatory for every client, so this stands on its own. <details> <summary>Manual verification</summary> Ran against a local dev server on this branch, using a session token and a throwaway app from `scripts/oauth2/setup-test-app.sh`. 1. Happy path unchanged: HTTP 200, verifier length 43. 2. `code_verifier=short`, and a 43-character verifier ending in `!`: both HTTP 400 `invalid_request`, so charset is enforced and not just length. 3. `code_challenge=tooshort` at authorize: HTTP 400 `invalid_request`, no code issued. An empty challenge still hits the older "required and cannot be empty" message. 4. Well-formed but wrong verifier: HTTP 400 `invalid_grant`, distinct from the cases above. 5. Retrying that same code with the correct verifier: HTTP 400, code already revoked by the failed check. 6. `generate-pkce.sh` produces a 43-character verifier (20 out of 20 runs); the docs example produces 128. 7. `scripts/oauth2/test-mcp-oauth2.sh` passes end to end. The two bearer-token failures in its output are a pre-existing script bug (`09c50559f3`, July 2025) that reuses a resource-scoped token against the real API, not a regression here. </details> |
||
|
Michael Suchacz
|
1458d27d78 |
fix: allow manual chat compaction from the error state (#28022)
A chat that fails generation with a context overflow (for example `Input
length 262625 exceeds the maximum allowed input length of 262112
tokens`) is stuck in a catch-22: `POST /chats/{id}/compact` returns 409
because the `RequestCompaction` transition is only allowed from the
waiting state, and the only other way out of the error state is sending
or editing a message, which re-runs generation with the same oversized
prompt and fails again. Compaction is exactly the recovery a
context-overflowed chat needs, and it is unreachable exactly when it is
needed.
Three semantic changes:
- Allow `RequestCompaction` from the error states: `E0 -> R0` and `E1 ->
R1` (queued messages are preserved and processed after the compaction
turn).
- Clear `last_error` in `Tx.RequestCompaction`, matching the
architecture rule that transitions leaving `E0`/`E1` clear the stored
error. Without this a successful compaction would land in waiting with a
stale persisted error.
- Grant the compaction turn a fresh history epoch: a
`grant_history_epoch` flag on `UpdateChatExecutionState` sets
`history_version = snapshot_version`, resets `generation_attempt`, and
clears `retry_state` in the same atomic update that clears `last_error`
(mirroring the `chat_messages` trigger postcondition). The transition
inserts no history, so without this the turn inherits the failed turn's
spent retry budget, and resetting the counter alone could collide with
message part episode keys still retained on the erroring replica.
No frontend change is required: the chat input is already enabled in the
error state and `/compact` submission already handles both the success
and 409 paths. Also updates ARCHITECTURE.md (transition matrix,
endpoint, and manual compaction sections), the endpoint's swagger
description, and SDK comments.
> Mux created this PR on Mike's behalf.
<!-- mux-attribution: model=claude-sonnet-4-6 thinking=high -->
|
||
|
J. Scott Miller
|
866e676320 |
feat: invalidate provisioner daemon sessions on key deletion (#26532)
## Summary Closes PLAT-305. When a provisioner key is deleted, the associated daemon kept operating on its existing WebSocket connection, because authentication was only checked at connection establishment and deletion was a bare `DELETE` with no session invalidation. This adds four layers of defense so a deleted key promptly stops doing work: 1. **Publish on delete.** `deleteProvisionerKey` publishes to a new per-key pubsub channel (`coderd/pubsub.ProvisionerKeyDeletedChannel`) after a successful delete. Publish errors are logged but still return `204`, since layer 3 is the durable backstop. 2. **Subscribe and tear down.** The daemon serve handler subscribes to its key's channel and terminates the DRPC session on a deletion event. Termination is deferred while a job claimed by the session is active: the daemon may finish and report the in-flight job (`UpdateJob`/`CompleteJob` have no key check), and the last active job's completion performs the cancellation. Because Postgres `LISTEN`/`NOTIFY` does not buffer for non-listeners, the handler also performs a synchronous key-existence re-check immediately after subscribing to close the race between auth and subscription. The subscription uses `SubscribeWithErr` so that an `ErrDroppedMessages` signal (emitted when the pubsub listener reconnects) triggers the same key re-check, closing the listener-outage window in which a deletion notification could be missed. 3. **Backstop on acquire.** `AcquireJob` and `AcquireJobWithCancel` verify the key still exists before waiting for a job, and the `Acquirer` claims jobs in a transaction that first locks the worker's deletable key (`LockProvisionerKeyByIDForShare`, a `FOR KEY SHARE` row lock held until commit) before running the `AcquireProvisionerJob` claim, so a claim cannot commit after the key's deletion. This guards against a missed pubsub message. A missing key row surfaces as its own result rather than overloading the claim query's no-rows response: the acquire terminates with `ErrProvisionerKeyDeleted` (terminating the session, with the same active-job deferral) and hands the consumed wakeup to another waiting daemon in the same domain, rather than silently re-parking and starving peers of job postings. 4. **Heartbeat watchdog.** The per-session heartbeat loop (1m interval) also re-checks the key, so even a session whose deletion notification was silently lost terminates within one heartbeat interval instead of living until the connection breaks (same active-job deferral as layer 2). Reserved keys skip the check. A job that is claimed but never delivered (the session or connection dies between the database claim and the stream send) is marked failed immediately on a fresh context, instead of staying assigned to the worker until the job reaper. Reserved keys (built-in, user-auth, PSK) are exempt throughout, since they are not deletable rows. The acquire-time lookup runs as `dbauthz.AsSystemReadProvisionerDaemons`, because the provisionerd role cannot read provisioner keys and a provisioner key's RBAC object is a provisioner daemon. A single key can back many daemons (and span HA replicas), so the per-key channel fans out to invalidate all of them at once. Per-key channels keep the `LISTEN` count proportional to distinct keys rather than waking every daemon on unrelated deletions. ### Known limitations - **`UpdateJob`/`CompleteJob` intentionally have no key check.** By the time those RPCs arrive the work has already run; rejecting completion would strand a build in "running" (until the job reaper fails it) with real infrastructure left unreconciled. Session termination is deferred while a job is active so the completion can be reported; the daemon may not receive the final RPC response when the deferred termination fires, but the job's outcome is already persisted. - **After termination, the daemon process redials and receives 401s until restarted.** The dial-time exit logic only triggers on 403, and the auth middleware returns 401 for an invalid key; this dial behavior predates this PR and is tracked as a follow-up in [PLAT-452](https://linear.app/codercom/issue/PLAT-452) (return 403 for invalid provisioner keys). ## Tests - `coderd/provisionerdserver`: `TestAcquireJob_ProvisionerKeyDeleted` (both RPC variants), `TestAcquireJob_ReservedProvisionerKey`, `TestHeartbeat_ProvisionerKeyDeleted` (heartbeat watchdog cancels the session after key deletion), `TestAcquirer_ProvisionerKeyDeleted` (a dead-key acquiree exits terminally and its clearance is promoted to a peer in the same domain), and `TestTerminateSession_Deferral` (termination is immediate when idle and deferred until the last active job finishes). - `coderd/database`: `TestAcquireProvisionerJob/ProvisionerKeyLock` covers the lock query against real Postgres: it returns the key ID while the row exists and no rows once it is deleted. The lock-then-claim composition is pinned by `TestAcquirer_ProvisionerKeyDeleted`. - `enterprise/coderd`: `TestProvisionerDaemonServe/KeyDeletionClosesSession` asserts an active session closes after its key is deleted. `KeyDeletedDuringSetupClosesSession` covers the post-subscribe re-check when a key is deleted between auth and subscription, and `DroppedMessageClosesSession` covers the `ErrDroppedMessages` re-check when a deletion is missed during a listener outage. ## Validation - `make` pre-commit (gen/fmt/lint/build) passed via git hooks. - Targeted tests pass; existing acquire tests pass with no regression. - Manual: brought up a dev deployment (coder-in-coder) with a Premium license, created a deletable provisioner key, and started an external daemon with `coder provisionerd start`. Confirmed it authenticated via the key and connected, appearing as `idle` in both `coder provisioner list` (with the key name) and the organization Provisioners UI. - Manual, idle teardown: deleted the key while the daemon was idle. The server logged `provisioner key deleted, terminating session`, the daemon's session closed immediately, and it dropped from `coder provisioner list` (then entered the known 401 redial loop, PLAT-452). - Manual, deferred termination: ran a workspace build (tagged template, `sleep 45` in `local-exec`) pinned to the external daemon and deleted the key mid-build. The server logged `deferring session cancellation until active jobs finish`; the heartbeat watchdog re-checked mid-build and re-deferred rather than force-killing. The build ran to completion (`Apply complete`, workspace `Started`) and only then did `canceling session after job completion` fire. The documented caveat reproduced: the daemon lost the final `CompleteJob` ack, and the build outcome was still persisted correctly. <details> <summary>Implementation plan and design decisions</summary> ### Design - **Per-key vs global channel:** chose per-key (`provisioner_key_deleted:<keyID>`) so daemons do not wake on unrelated deletions. The cost is one `LISTEN` per distinct key per replica on the shared listener connection, which is negligible against Coder's existing channels. - **Missing-key behavior on acquire:** returns an error that tears down the acquire rather than silently returning an empty job. - **Subscribe-startup race:** ordering is `authorize -> UpsertProvisionerDaemon -> Subscribe -> GetProvisionerKeyByID`. The post-subscribe re-check handles a deletion that committed before the `LISTEN` registered (Postgres does not buffer notifications for non-listeners; the in-process buffer only smooths bursts and drops on overflow). - **`NewServer` change:** `KeyID` was added to `provisionerdserver.Options` to avoid a positional signature change across call sites. The in-memory (built-in) daemon leaves it unset and is therefore exempt. ### Files - `coderd/pubsub/provisionerkeydeleted.go` (new) — channel helper. - `enterprise/coderd/provisionerkeys.go` — publish on delete. - `enterprise/coderd/provisionerdaemons.go` — subscribe, re-check, cancel session; pass `KeyID`. - `coderd/provisionerdserver/provisionerdserver.go` — `KeyID` option and acquire-time existence check. </details> --- This pull request was created by Coder Agents on behalf of @jscottmiller. |
||
|
Jaayden Halko
|
5bdabc95c8 |
fix(codersdk/licenses.go): read trial claim instead of misspelled trail (#28014)
## What this fixes `codersdk.License.Trial()` looked up the JWT claim `"trail"` (a typo) instead of `"trial"`, the actual claim name set by license issuance (`enterprise/coderd/license/license.go`). Since no license contains a `"trail"` claim, the method always returned `false`. The only consumer is `coder licenses list` (via `cli/cliutil/license.go`), so the CLI never reported a license as a trial even when it was one. The web UI is unaffected because it reads `claims.trial` directly. ## Changes - Read the `"trial"` claim in `License.Trial()`. ## Testing - `go build` / `go vet` on `codersdk` and `cli/cliutil`. - `go test -run 'TestLicensesListFake|TestLicensesListReal' ./enterprise/cli` passes. |
||
|
Andrew Aquino
|
6e07e2610f |
feat: add paginated API endpoint for groups (#27603)
backend-only changes from #27271; see that PR for summary of changes + implementation details |
||
|
Steven Masley
|
9a57dfa642 |
feat: include agent metadata in workspace list responses (#27934)
Closes #27933. Related: #27897 (single-agent GET). Agent metadata is only readable via a per-agent watch stream, so reading it across N workspaces costs N+1 requests. This adds a batch read to the list endpoint: ```text GET /api/v2/workspaces?q=param:"pool=demo" include_agent_metadata:task_status ``` - New `include_agent_metadata` search key, repeatable and key-scoped. It expands the response, it does not filter workspaces. - `GetWorkspaces` aggregates the requested keys as JSON behind a `CASE`: without opt-in the response is unchanged and the subquery never runs. Runs only for the returned page, inside the same authorized query. - Agents in the response gain `metadata` (`[]codersdk.WorkspaceAgentMetadata`, `omitempty`), mapped by the `workspace_agent_id` each element carries. The collection script is omitted; it can be long. - `codersdk.WorkspaceFilter` gains `IncludeAgentMetadata []string`. - No wildcard, no schema change, no migration. --- Authored by Coder Agents on behalf of @Emyrk. |
||
|
dylanhuff-at-coder
|
9b27d12929 |
chore: forbid direct response body JSON decode in codersdk (#27859)
Add a ruleguard rule forbidding direct `json.NewDecoder(res.Body).Decode(...)` on `*http.Response` in codersdk packages, so new typed endpoints use `codersdk.ReadBodyAsJSON` and keep returning structured errors for non-JSON bodies. The rule matches both the chained call form and decoders assigned to a variable first. Intentional raw-body paths carry documented `//nolint:gocritic` exceptions: the 16 agent-direct HTTP decodes in `workspacesdk/agentconn.go` route through a single `decodeAgentJSON` helper (agent-direct over tailnet, so `ReadBodyAsJSON`'s reverse proxy/SSO error guidance does not apply), and the Azure IMDS attested-document decode in `agentsdk/azure.go` keeps an inline exception. The two `UseNumber` decoders in `licenses.go` are migrated to a new `codersdk.ReadBodyAsJSONUseNumber`, so `coder licenses add/list` also return structured errors for non-JSON bodies instead of `invalid character '<' looking for beginning of value`. Note for local verification: golangci-lint caches results, so run `golangci-lint cache clean` after modifying `scripts/rules.go` or the rule may silently not fire. Final PR of the stack on #27804, #27857, and #27858. Refs #27044. Stack plan Inventory (full-tree audit): 280 migratable call sites across 47 files; 17 excluded (16 agent-direct HTTP sites in `workspacesdk/agentconn.go`, 1 Azure IMDS decode in `agentsdk/azure.go`). 1. **#27857** `refactor(codersdk): use ReadBodyAsJSON in typed endpoints`: mechanical migration of all sites except `chats.go` (224 sites, 46 files). 2. **#27858** `refactor(codersdk): use shared error helpers in chat endpoints`: migrate the 56 `chats.go` sites and consolidate the duplicated `readRawBodyAsError`/`newResponseError` helpers onto the shared `client.go` error path, with regression tests for the 409 usage-limit flow. 3. **#27859** `chore: forbid direct response body JSON decode in codersdk`: ruleguard rule with documented exceptions for the intentional raw-body paths, plus `ReadBodyAsJSONUseNumber` for the `licenses.go` decoders. Reviewed and updated by Coder Agents on behalf of @dylanhuff-at-coder. |
||
|
Ethan
|
d2f9280138 |
chore: remove legacy chat template allowlist (#27515)
Relates to CODAGT-713 Depends on #27514 Removes the legacy deployment-wide allowlist now that the API and frontend use per-template `agents_allowed`: the experimental `/template-allowlist` routes, SDK methods and generated types, site config queries, frontend bindings, and the now-unused `xjson` utility. Migration `000563` deletes the obsolete `agents_template_allowlist` value. It's irreversible for deployments that configured an allowlist, which I think is fine, since `000562` already drops `agents_allowed` on the way down, and this release ships `000548` and `000555` with the same property. Two side effects of the model change worth writing down, both from #27514 rather than here. The value used to need `ActionRead` on `ResourceDeploymentConfig` to read and deployment config update to write. `AgentsAllowed` is now a plain field on the template response, readable by anyone who can read the template, and it's set with a template update, so org admins manage it themselves. That's the delegation we wanted, and it's tracked in the audit log. The rest of the stack adds `--agents-allowed` to the CLI and updates the platform controls docs. |
||
|
Ethan
|
0ac23e3ee1 |
feat: add per-template Coder Agents access control (#27285)
Relates to CODAGT-713 Depends on #27284 This makes the per-template `agents_allowed` field authoritative in the API and chatd. It adds optional create and metadata update fields with the intended default and omission semantics, supports `agents-allowed:` template search, includes the value in telemetry, and makes `list_templates`, `read_template`, and `create_workspace` read the template row directly. Existing-workspace retries remain idempotent, and blocked same-organisation templates return an actionable message. The experimental `/template-allowlist` routes remain temporarily because the shipped AI Settings page still calls them, but they no longer control chatd enforcement. #27514 moves that page to per-template metadata, #27515 removes the legacy storage, routes, SDK types, and utility, #27517 adds the CLI flags, and #27518 updates the platform controls documentation for the per-template model, directly addressing CRF-5 and CRF-6. The stack is intended to merge as a unit. |
||
|
dylanhuff-at-coder
|
ed10064748 |
refactor(codersdk): use shared error helpers in chat endpoints (#27858)
Migrate chat endpoint response decoding to `ReadBodyAsJSON` and consolidate `ReadBodyAsError` construction through `newResponseError`, so empty-body and non-JSON errors consistently include the request method and URL. Stacked on #27857, with the lint rule following in #27859. Refs #27044. Reviewed and updated by Coder Agents on behalf of @dylanhuff-at-coder. |
||
|
Michael Suchacz
|
4b9880afa6 |
feat: add --chat-hook-allow-insecure to allow plain HTTP chat hook URLs (#27896)
Adds a hidden `--chat-hook-allow-insecure` / `CODER_CHAT_HOOK_ALLOW_INSECURE` deployment option (default `false`) that allows the chat lifecycle hook URL to use plain HTTP for any host. The HTTPS requirement is enforced at two points, and the flag relaxes both: `DeploymentValues.Validate()` rejects `http` hook URLs at startup, and the hook dispatcher's `validateHookURL` allows `http` only for loopback hosts. With the flag set, any-host `http` is accepted; the host, fragment/userinfo, secret, and timeout checks are unchanged, and non-http(s) schemes still fail. This removes the need for an HTTPS reverse proxy when testing a hook consumer on a trusted network. Following security review feedback, the flag description and docs state that plain HTTP lets an on-path attacker forge hook responses (which control agent execution), and `coder server` logs a startup warning (with a redacted hook URL) when hooks run over plain HTTP. Docs, generated API types, and the server config golden are updated accordingly. > Mux acted on Mike's behalf to create this PR. |
||
|
dylanhuff-at-coder
|
76ae64391a |
refactor(codersdk): use ReadBodyAsJSON in typed endpoints (#27857)
This PR migrates 224 typed JSON response sites across 46 files to `codersdk.ReadBodyAsJSON`, so invalid 2xx bodies return structured errors while preserving URL credential redaction. It intentionally excludes agent-direct HTTP, Azure IMDS, `UseNumber`, and chat paths; stacked on coder/coder#27804, with chat and lint follow-ups in coder/coder#27858 and coder/coder#27859. Refs coder/coder#27044. Reviewed and updated by Coder Agents on behalf of @dylanhuff-at-coder. |
||
|
Nick Vigilante
|
9dcb75cd56 |
chore: add docs inline-HTML linter and backtick generated placeholders (#27399)
## What Adds CI enforcement that fails when docs Markdown contains invalid inline HTML the docs site silently drops or mangles, and fixes the remaining generated-doc placeholders at their source. This is the tooling half of the docs-HTML audit. The hand-written fixes it guards landed in #27298 (kept small and separate so it reviewed fast); this PR carries everything that touches code, CI, or generated output. ## Changes **Linter (`scripts/docshtmlcheck`), wired into `make lint` via `lint/docs-html`.** Markdown-aware: parses each file with goldmark and inspects only raw-HTML nodes, so angle brackets in fenced code blocks, inline code, HTML comments, and `<https://…>` / `<user@host>` autolinks are ignored. Flags swallowed placeholders (`<region>`), void-element end tags (`</br>`), unregistered or incorrectly capitalized component tags (`<Image>`), and unclosed container tags (a `<div class="tabs">` that leaks its wrapper). The one intentional renderer component, `<children>`, is allowed but still balance-checked. **Generator-source placeholder fixes (regenerated via `make gen`).** - `codersdk/chats.go`: backtick `<server>__` in the `ChatContextTool.Name` doc comment (it becomes the Swagger description, so it was swallowed in `reference/api/{chats,schemas}.md`). - `codersdk/deployment.go`: backtick `<region>` in the AWS Bedrock region flag help (swallowed in `reference/cli/server.md`); also updates `coder server --help` output and the golden files. **Temporary allowlist.** `docs/reference/cli/agent-firewall.md`'s `<host>` / `<glob>` come from the external `github.com/coder/boundary` CLI help (still `v0.10.0` on `main`), so they are suppressed on that one file. The suppression is self-clearing: if an allowlisted tag stops appearing on a scanned file, the linter reports `stale-allowlist-entry` and fails until the dead entry is removed, so a dead entry cannot silently mask a later regression of that tag on that page. (An entry whose file is deleted outright is never rescanned, but a missing file yields no findings, so nothing hides behind it either.) ## Review feedback addressed This tool + generator work was reviewed by Coder Agents Review while it was bundled into #27298. Addressed here: - **P1:** tokenize each raw-HTML node as a whole instead of per source line, so a tag whose attributes wrap across lines is no longer torn in half. This fixes both the missed multi-line unclosed `<div>` (a leaked wrapper that passed with exit 0) and the spurious `stray-end-tag` on valid multi-line tags. Each token maps back to its own source line. - Normalize allowlist lookup/report paths to a canonical repo-relative form, so the escape hatch no longer silently misses under absolute / `./` paths. - Route generated-page findings to the generator source. - Add `<search>` to the allowed set; reword the unknown-element message to note that a real element can be added to `allowedElements`. - Self-clearing allowlist guard (above); rename `optionalEndTag(s)` and `kindUnclosed(Tag)`; adopt `slices`/`maps` idioms; move the lint banner to the Makefile recipe; stop aliasing the input slice in `filterAllowed`. - New tests: multi-line tokenization (both classes), interleaved nesting, a pinned line number, `collectMarkdown`, and the stale-allowlist guard. ### Round 2 (Coder Agents Review on this PR) A second `/coder-agents-review` pass on this PR raised 16 findings; addressed in `fix(docshtmlcheck): catch self-closing containers and capitalized tags`: - **P2:** self-closing container tags (`<div class="tabs"/>`) were ignored by the HTML5 parser and leaked their wrapper like the open spelling; the balance check now tracks self-closing tokens too (CRF-1). - **P2:** a capitalized component tag whose lowercase name is a real element (`<Table>`, `<Section>`) slipped through on the `allowedElements` lookup. The tokenizer lowercases tag names, so the check now reads the raw token and reports any capitalized name as a component reference (CRF-2). - Narrowed the `:` / `@` autolink skip to a real URI scheme or a dotted `local@domain`, so `<region:id>` and `<user@host>` stay checked (CRF-3). - Stale-allowlist findings now report against the linter source with no line, and count separately from invalid-HTML issues in the footer (CRF-7, CRF-11). - Comment / README / Makefile wording synced to the honest capitalized-tag behavior; added the deleted-file allowlist caveat and a note that `allowedElements` is hand-maintained against the renderer (CRF-14, CRF-17, CRF-9). - Internal cleanups (`pop` -> `matchEndTag`, extracted `unclosedFinding`) and new tests: self-closing, capitalized open/close, colon/at placeholders, a non-first-token line assertion, `isGeneratedDoc`, and the stale message (CRF-12, CRF-13, CRF-1/2/3/4/5/16). Two findings resolved without a code change: - **CRF-8** (also wire `lint/docs-html` into `lint-light`): declined. `lint-light` is the Go-free fast path; `lint/docs-html` needs the Go toolchain, so it stays in the full `make lint`, which CI runs. Adding it would pull Go into the light path for no coverage gain. - **CRF-9** (`allowedElements` <-> renderer coupling): documented with a maintenance note in the `allowedElements` comment and tracked in DOCS-597 for a cross-repo sync/check decision. Deferred (note, no current trigger): raw-text element interiors (`<script>` / `<style>`) are not scanned for nested tags. No docs page relies on this today; noted for follow-up. ## Merge order #27298 (the hand-written fixes this PR guards) has merged, and this branch is rebased on `main`, so `make lint/docs-html` now reports 0 findings and the `lint` check passes. The two PRs are independent (disjoint files, no stacking). ## Verification - `go test ./scripts/docshtmlcheck/`, `go vet`, `gofmt -l`, `golangci-lint run`: clean. - `make lint/docs-html` (branch rebased on `main`): 0 findings. ## Linear - DOCS-584: https://linear.app/codercom/issue/DOCS-584/add-ci-check-that-fails-on-invalid-inline-html-in-docs - DOCS-551: https://linear.app/codercom/issue/DOCS-551/backtick-placeholder-syntax-in-generated-reference-docs-cli-help - DOCS-597 (follow-up, from CRF-9): https://linear.app/codercom/issue/DOCS-597/track-docshtmlcheck-allowedelements-drift-vs-docs-renderer-component > This PR was created with AI assistance (Coder Agents). |
||
|
dylanhuff-at-coder
|
db68c6c9fe |
fix: add codersdk JSON response decoder for typed API endpoints (#27804)
`coder whoami` and `coder list` can surface low-level JSON decode errors when a reverse proxy, SSO portal, or incorrect Coder URL returns HTML with a successful HTTP status. Add a shared SDK JSON response decoder and use it for the user and workspace list endpoints so these commands return a structured, actionable API response error instead. Refs #27044 Reviewed and updated by Coder Agents on behalf of @dylanhuff-at-coder. |
||
|
Yevhenii Shcherbina
|
11427066a1 |
fix: require bedrock model fields for the invoke-model protocol (#27846)
Implements: https://linear.app/codercom/issue/AIGOV-564/aibridge-bedrock-provider-skipped-404-on-all-routes-when-settings-omit Improves validation when creating and updating AI providers: a Bedrock provider using the `invoke-model` protocol now requires `model` and `small_fast_model`. This brings API validation in sync with the UI, which already required both fields. |
||
|
Michael Suchacz
|
6b8f820493 |
feat: remove native chat cost tracking in favor of AI Gateway cost data (#27330)
## Stack Context This stack makes AI Gateway data and budgets the source of truth for AI spend controls. 1. Re-back the per-chat cost endpoint with AI Gateway data (#27328, merged). 2. Remove native chat usage limits (#27329, merged). 3. **This PR, now based on `main`:** remove native chat cost tracking and its dedicated admin UI. ## Summary Removes native per-message price calculation, model pricing fields, cost persistence, aggregate cost queries, and admin cost API types. It also deletes the Analytics and Spend pages plus their legacy redirects. The AI Gateway-backed per-chat cost row and compact budget indicators remain. The spend documentation is renamed to `spend-management.md` and updated for the remaining surfaces, group budget APIs, CSV export, upgrade handling for native pricing and cost history, and the absence of a deployment-wide spend dashboard. The per-chat cost API documents that data follows AI Gateway retention and reports zero after all matching requests are purged. No schema is dropped in this release. `chat_messages.total_cost_micros` remains nullable and unwritten so replicas from the previous release can continue inserting messages during rolling upgrades. #27600 tracks removal after the compatibility window. > Mux prepared this PR on Mike's behalf. |
||
|
Michael Suchacz
|
f0e6ac64b3 |
feat: remove native chat usage limits in favor of AI Gateway budgets (#27329)
## Stack Context This stack makes AI Gateway data and budgets the source of truth for AI spend controls. 1. Re-back the per-chat cost endpoint with AI Gateway data (#27328, merged). 2. **This PR:** remove native chat usage limits. 3. Remove native chat cost tracking and its dedicated admin UI (#27330). ## Summary Removes the native usage-limit API, SDK types, SQL, and chat enforcement for deployment, user, and group chat limits. Compact AI Gateway budget indicators remain in the Agents sidebar, user menu, and group settings. Gateway budget rejections and provider quota failures continue to classify as usage-limit errors, including a 409 response for synchronous title generation. Budget-period labels now use the API's UTC boundaries, so users see the same dates in every browser timezone. The documentation explains the AI Gateway replacement, its licensing requirements, and the differences from native limits. No schema is dropped in this release. The usage-limit table, index, user and group columns, constraints, audit mappings, and generated scan fields remain for mixed-version rolling upgrades. #27600 tracks their removal after the compatibility window. ## Breaking change Native day, week, and month chat spend limits are removed and are not migrated. AI Gateway budgets are month-based, group-scoped with per-user overrides, and require the AI Gateway entitlement. Deployments without that entitlement no longer have chat spend enforcement. > Mux prepared this PR on Mike's behalf. |
||
|
Michael Suchacz
|
c6cee10e8b |
feat: add per-model OpenAI Responses API toggle (#27683)
chatd hardcoded `WithUseResponsesAPI()`, so the provider SDK's static known-model list decided whether an OpenAI model spoke the Responses API or Chat Completions. A model absent from that list silently fell back to Chat Completions until the fantasy fork was patched. This exposes the SDK's `WithResponsesAPIFunc` hook as a per-model setting, `openai_config.use_responses_api`, stored in the existing `chat_model_configs.options` JSONB. Unset keeps the known-model list, `true` forces Responses, `false` forces Chat Completions. There is no migration. It sits in a new construction-time `openai_config` section rather than in `provider_options.openai` because it selects the API when the client is built, while `provider_options` holds per-request parameters. That placement is also load-bearing: a config setting only this field would otherwise materialize an OpenAI request-options struct and turn on provider-side response storage, since `Store` defaults to true there. Three places independently decided the transport and would silently disagree with the client actually built: | Site | Effect when it disagrees | | --- | --- | | `ModelFromConfig` | the transport being overridden | | `AcceptsFilePartMediaType` | text attachments dropped, since Responses natively accepts only images and PDFs | | `UsesResponsesOptions` | the SDK type-asserts the concrete options struct, so every OpenAI option is discarded | They share one predicate here, `chatopenai.UsesResponsesAPI`, with the override threaded to each. The rest of the stack removes that threading by resolving the transport once and carrying it. Compaction overrides and the quickgen debug model built clients without `ConfigOptions`, so they now pass it and pick up both this setting and the existing Anthropic beta headers. The toggle also makes transport-conditional option handling admin-switchable, so two hardening changes ride along. `ServiceTierFromChat` now maps every tier the codersdk enum advertises (`auto`, `default`, `flex`, `scale`, `priority`); it previously returned nil for `default` and `scale`, so flipping a model to Responses silently dropped a configured `service_tier` that the API accepts (fantasy forwards the value unchanged). And a new `TestProviderOptionsTransportParity` pins, per `provider_options.openai` field, which transport honors it, against a table in ARCHITECTURE.md, so a field honored on one transport and silently ignored on the other fails the test unless recorded as intentional. Review rounds also caught two lifecycle gaps around the new field. `isZeroChatModelCallConfig` now inspects `OpenAIConfig`, so a stored options blob whose only setting is this toggle survives into GET/list responses instead of reading as `model_config: null`; `TestIsZeroChatModelCallConfigCoversEveryField` sets each config field in isolation and fails if any field is invisible to the zero check. And the model editor's update path sends an explicit empty `model_config` when an edit clears the last field, since an omitted property preserves the stored options server-side; covered by the `EditClearingLastOptionSendsEmptyConfig` story. Azure keeps following the known-model list, because the Azure provider exposes no equivalent hook. The model editor renders Azure with the OpenAI option schema, so instead of shipping a visible but inert control, the option schema generator gains a `providers` struct tag that it emits as `visible_for_providers`. Gating uses the raw provider type rather than the alias table, so the control appears only for openai-typed providers. No hand-written frontend field: the editor renders it from the generated schema. Closes https://linear.app/codercom/issue/CODAGT-874/add-completionsresponses-api-toggle-in-model-editor > Mux prepared this PR on Mike's behalf. |
||
|
Steven Masley
|
52423eb87b |
feat: promote MinimumImplicitMember experiment to GA (#27472)
Promotes the `minimum-implicit-member` experiment to GA and removes it.
## What changes
- The `minimum-implicit-member` experiment constant, its
`RoleOptions.MinimumImplicitMember` toggle, and the global
`rbac.MinimumImplicitMember()` accessor are deleted. The minimal-member
behavior is now the only behavior: `organization-member` and
`organization-service-account` carry only the floor (read-self records,
notifications, and similar) and grant **no workspace permissions**.
Workspace access lives exclusively on the
`organization-workspace-access` role.
- The experiment gate on customizing `default_org_member_roles` (`PATCH
/organizations/{org}`) is removed; the built-in-roles-only validation
remains.
- The dashboard's Default Roles section and the implied-roles display on
the members page are no longer experiment-gated.
- Admin docs: new "Default member roles" section in
`docs/admin/users/organizations.md`, cross-linked from
`groups-roles.md`.
## Why this is safe for existing deployments
Migration `000516` (shipped earlier) backfilled
`default_org_member_roles` with `['organization-workspace-access']` on
every organization. Members therefore keep exactly the effective
permissions they had with the experiment off; the workspace elevation
flows through the default role instead of being baked into
`organization-member`.
**Rollback caveat:** rolling back past this release restores the bundled
elevation, silently re-granting workspace access to members of
organizations that cleared their default roles.
## Review
Deep-review R1 findings are addressed in `chore: address deep-review
findings` (copy fixes, read-only Default Roles for viewers, removable
overlapping explicit grants, RBAC prose restoration, test
de-tautologizing, docs). Point-by-point disposition is in the PR
comments.
---
Generated by Coder Agents on behalf of @Emyrk.
|
||
|
Michael Suchacz
|
df1c0f9710 |
feat: show what a chat lifecycle hook changed (#27655)
## Stack Context Follow-up fixes from live UAT of the merged chat lifecycle hooks stack (#27430). Two PRs: 1. **This PR**: make hook effects visible and correctly attributed in the transcript. 2. [`mike/chat-hooks-uat/dispatch-capacity`]: reserve dispatch capacity so an admission burst can't fail running turns. ## Why? UAT found three ways the transcript misrepresented what a lifecycle hook did. All three are user-visible and share the same surface (`chathooks/effects.go`, `codersdk.ChatMessagePart`, the conversation timeline), so they're reviewed together. **A prompt `input_override` silently discarded attachments.** `ComposeUserPromptContent` replaced the entire submitted part list with one text part, dropping `file` and `file-reference` parts along with their `chat_file_links`. The user saw their attachments vanish with no explanation. The override now replaces submitted *text* parts only and preserves non-text parts in order. A consumer that wants to block attachments uses `deny`, which is the documented mechanism for refusing a submission. **Every user-visible `system` row was labelled "Lifecycle hook".** The timeline keyed the notice off `role === "system"`. That was correct only by accident, because the hook `user_message` was the sole client-visible system row. The backend now emits the notice as a typed `hook-notice` part and the timeline renders on that, so a future system row can't be mislabelled as a policy notice. **Nothing marked a tool call the hook had rewritten.** A consumer could replace tool input via `input_override` and the transcript showed the rewritten input as if the model had produced it. `ChatMessagePart` gains `hook_rewritten`, set from `preflight.Overrides` on the same path that already carries `ToolCallCreatedAt`, and the tool row renders a "Modified by policy" badge. `ToolCall.PolicyProvider` renders the badge itself, at four wrap sites: the `Tool` dispatch wrapper, the `ReadFilesTool` aggregate and its per-file rows, and `ReadFileTimelineBlock` (grouped and single `read_file` rows bypass `Tool`). Renderer props do not include the flag; descendants consume it through the provider context. The badge is emitted by the provider rather than by the shared header because several renderer branches return early without one, including the auth-required `execute` card, a completed `ask_user_question`, and an empty question payload. Those branches would drop the attribution with no type or runtime error, and the gap is not greppable: every renderer file contains a header somewhere, only individual branches do not. Emitting at the provider removes the possibility instead of enumerating the cases. A rewritten call is wrapped in a group labelled by its badge, so one rewritten file inside a merged read is attributed on its own rather than inheriting the group's badge. `HeaderButton` still appends the policy wording to an explicit `ariaLabel`, since an explicit `aria-label` replaces the name computed from descendants. Provider-executed calls are excluded from attribution. Hooks never see them, and duplicate tool-call ID rejection deliberately skips them, so a reused ID would otherwise mark a provider-executed call as policy-rewritten. ## Testing Go: `coderd/x/chatd/...`, `coderd/x/agenthooks/...`, `codersdk/...`, and `coderd -run 'Hook|Chat'`. Frontend: `tsc` plus every `AgentsPage` story; the only failures are `MCP Tool Completed` and `Scroll To Bottom Button Works With Inverse Scroll`, both of which fail on trunk. A registry-wide story asserts every registered renderer shows the badge, verified against three inverted toggles: removing the badge, hiding it with `display:none`, and skipping the provider for one renderer (which names that renderer). Storybook also covers the rewritten subagent spawn, a completed empty question payload, a non-hook system message, and a failed `read_file` guarding the accessible name. > Mux opened this PR on Mike's behalf. |
||
|
Ethan
|
4dbb3a236c |
test: fix tailnet connection teardown flake (#27768)
Closes https://github.com/coder/internal/issues/1620 Closes ENG-3043 The callback cleanup added in #20687 stops new node callbacks, but it can still race with one that's already in flight. That callback may call `UpdatePeers` after the destination `tailnet.Conn` closes, so the test fails with `connection closed` even though the redirect behaviour is correct. To fix, we'll just ignore `tailnet.ErrConnClosed` in the asynchronous `stitch` helpers, whilst continuing to assert on every other error. |
||
|
Sas SwartandClaude Opus 4.8
|
8886a5749a |
feat: add network calls list to AI session threads API (#27425)
The AI session threads API returned only a network call *summary* (total/blocked counts + top domains). This adds the per-call list so the session detail can render individual Agent Firewall network calls. `ListAIBridgeSessionNetworkCalls` reuses the same sequence-number windowing as the existing summary and includes all protocols. The list is exposed as `network_call_logs` on the threads response and is capped server-side at 100 rows. The summary (`network_calls.total`/`blocked`) remains authoritative for whole-session totals: the list length and its blocked count equal the summary only when a session has at most 100 calls, and are truncated beyond that. ### PR map (merge strictly bottom-up) This change is a 4-PR stack. Each PR depends on all the ones below it, so merge in this exact order: 1. #27417 — backend network summary 2. #27418 — frontend summary rows 3. #27425 — backend per-call list `network_call_logs` 4. #27426 — frontend network-calls panel Refs AIGOV-464 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|
|
3f3fd1c4d7 |
feat: show network request summary on AI session detail card (#27418)
Frontend for the AI session network summary. Adds Network calls, Blocked network requests, and Top domains rows to the Session summary card on the individual AI session detail page, driven by the network fields on the session threads response. Renders "Disabled" when network monitoring was not active and "No activity" when there were no calls. Covered by Storybook stories for each state. ### PR map (merge strictly bottom-up) This change is a 4-PR stack. Each PR depends on all the ones below it, so merge in this exact order: 1. #27417 — backend network summary 2. #27418 — frontend summary rows 3. #27425 — backend per-call list `network_call_logs` 4. #27426 — frontend network-calls panel Refs AIGOV-463 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Cian Johnston <cian@coder.com> |
||
|
|
841a1765f7 |
feat: add network calls summary to AI session threads API (#27417)
Backend for the AI session network summary. Exposes total/blocked
network calls and top destination domains on the session threads
endpoint (`GET /api/v2/ai-gateway/sessions/{id}`).
Total and blocked reuse the existing Agent Firewall aggregation from the
sessions list query, so the numbers match the sessions table. Top
domains are a new server-side aggregation
(`GetAIBridgeSessionTopDomains`) over boundary logs, using the same
interception-window correlation. There is no network-error state,
matching the current data model.
Frontend consuming these fields is in a separate stacked PR.
### PR map (merge strictly bottom-up)
This change is a 4-PR stack. Each PR depends on all the ones below it,
so merge in this exact order:
1. #27417 — backend network summary (base `main`)
2. #27418 — frontend summary rows (base #27417)
3. #27425 — backend per-call list `network_call_logs` (base #27418)
4. #27426 — frontend network-calls panel (base #27425)
Refs AIGOV-463
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Cian Johnston <cian@coder.com>
|
||
|
Michael Suchacz
|
95a2c2ba02 |
feat: back the per-chat cost endpoint with AI Gateway data (#27328)
## Stack Context
This stack removes native chat cost tracking and native chat usage
limits, making the AI Gateway the single source of AI spend data and
budget enforcement.
1. **This PR:** re-back the per-chat cost endpoint with AI Gateway data.
2. Remove native chat usage limits end to end, rewiring the sidebar
indicator to gateway spend.
3. Remove native chat cost tracking end to end, deleting the
Analytics/Spend cost UI.
## What?
`GET /api/experimental/chats/{chat}/cost` summed
`chat_messages.total_cost_micros`, which native chat cost tracking
maintained. It now aggregates AI Gateway interception data instead, and
has no native fallback.
- New `GetAIBridgeChatCost` query, authorized through the root chat so
members can read their own chat's cost without gaining access to raw
interception rows.
- Response fields renamed: `priced_message_count` -> `request_count`,
`unpriced_messages_having_usage_count` -> `unpriced_request_count`.
- The chat summary sidebar keys its cost cache by root chat, and hides
the cost row where the AI Gateway is off or unlicensed. The root cost is
invalidated when a chat leaves an active status and when a generated
title lands, since title generation bills its own gateway request.
`GetChatModelUsageCostByChatID` and the rest of native cost tracking are
untouched here; PR 3 removes them.
## Why?
Native cost tracking duplicates what the AI Gateway already records, and
the two disagree. Repointing the endpoint first means the cost UI keeps
working while the native implementation is deleted later in the stack.
Two behaviour changes follow from gateway semantics and are intentional:
- **Requests, not messages.** The gateway records interceptions, so
counts are requests. Title-generation traffic now counts.
- **Whole-tree totals.** The gateway records the *spawning* chat's ID as
the interception session ID, so a subagent's requests are attributed to
its immediate parent, not always the root. Only a whole chat tree can be
summed, so the query resolves the root and aggregates the tree, and
every chat in a tree reports the same total. Native returned per-subtree
totals.
## Attribution and counting semantics
The aggregate groups token usage per interception before counting, so
the reported numbers are per request even though a request records one
usage row per provider response:
- `RequestCount` counts finished `Coder Agents` interceptions in the
tree, including unpriced ones.
- `UnpricedRequestCount` counts requests with at least one usage row the
gateway could not price. It is a subset of `RequestCount`.
- `TotalCostMicros` omits only unpriced usage, so a partially priced
request still contributes its priced portion. The sidebar therefore says
`Excludes unpriced usage from N request(s)` rather than claiming whole
requests were dropped.
A recorded cost of zero is a free request, not an unpriced one. Usage
without an effective group is excluded, matching what never reached
`ai_user_daily_spend`.
## Authorization
Reads go through `ExtractChatParam` plus `ResourceChat`, with no
cost-specific RBAC widening. `TestGetChatCost/MemberCanReadOwnChat`
covers a scoped `agents-access` member reading their own chat's cost,
and `MemberCannotReadOtherUsersChat` still asserts 404 for a non-owner.
Plain members without `agents-access` cannot create or read chats at
all, so they never reach this endpoint.
## Known limitation
AI Gateway data has its own retention period, 60 days by default and
configured independently of chat retention, so spend for requests older
than that is no longer reported. A chat whose gateway records have all
been purged reports zero cost, which is indistinguishable from genuinely
free usage under this contract. The endpoint documents the caveat;
#27330 documents it on the Spend Management page.
In-flight interceptions are excluded, since cost is only known once the
response is recorded. A chat's cost therefore lags the active turn by
one request.
## Rebase note
Rebased onto `main` after #27579 removed the `ai-gateway-cost-control`
experiment. The per-chat cost row is now gated on the `aibridge` feature
alone, matching how #27579 degated the other cost-control surfaces.
> Mux prepared this PR on Mike's behalf.
|
||
|
Susana Ferreira
|
3deecb481e |
chore: remove ai-gateway-cost-control experiment flag (#27579)
## Description Closes [AIGOV-443](https://linear.app/codercom/issue/AIGOV-443/remove-ai-gateway-cost-control-experiment-flag-once-feature-is-stable). The AI Gateway cost control feature is planned for GA on the upcoming release, so this removes the `ExperimentAIGatewayCostControl` experiment and all of its gating. The cost control API endpoints remain gated by the `FeatureAIBridge` license feature (the AI Governance add-on), so this only drops the experiment layer. ## Changes - **`codersdk/deployment.go`**: remove the `ExperimentAIGatewayCostControl` const, its `DisplayName()` case, and its `ExperimentsKnown` entry. - **`enterprise/coderd/coderd.go`**: remove the `httpmw.RequireExperiment(...)` gating from the AI cost control routes. They keep `RequireFeatureMW(codersdk.FeatureAIBridge)`. Affected endpoints: - `GET /organizations/{organization}/groups/ai/spend` - `GET /organizations/{organization}/groups/{groupName}/members/ai/spend` - `GET /organizations/{organization}/ai/spend/export` - `GET /groups/{group}/members/ai/spend` - `GET /groups/{group}/ai/spend` - `GET/PUT/DELETE /users/{user}/ai/budget/override` and `GET /users/{user}/ai/spend` - **`enterprise/coderd/aibridge_test.go`**: drop the experiment from test setup and remove the now-obsolete `RequiresExperiment` negative-path tests. - **Frontend (`site/src/...`)**: remove the `ai-gateway-cost-control` experiment checks from the cost control UI (Groups pages, user dropdown) and their stories/mocks. The feature is now driven solely by the `aibridge` feature visibility. - **Generated**: regenerated `coderd/apidoc/*`, `docs/reference/api/schemas.md`, and `site/src/api/typesGenerated.ts`. ## Out of scope The dogfood `CODER_EXPERIMENTS` config lives in a separate infra repo, not `coder/coder`. Leaving `ai-gateway-cost-control` there is harmless: unknown experiment values are logged as `"ignoring unknown experiment"` at startup and otherwise ignored, so no ordering dependency or breakage. That cleanup can be a follow-up. <details> <summary>Implementation notes</summary> - Verified how unknown experiments are handled in `coderd/coderd.go` `ReadExperiments`: unknown values produce a warning log and are inert, so removing the definition before the dogfood config is updated is safe. - Noticed the group `ai/budget` routes (`/groups/{group}/ai/budget`) were already gated only by `FeatureAIBridge`, never by the experiment. After this change all cost control routes are uniformly feature-gated, resolving that inconsistency. - Removed an obsolete `RequiresExperiment` subtest in `TestUserAISpendStatus` that only asserted a 403 from the experiment gate; with the gate gone it would no longer be blocked pre-RBAC. </details> --- _This PR was created by Coder Agents on behalf of @ssncferreira._ |
||
|
Susana Ferreira
|
d6a5c8e9f8 |
refactor: make user AI budget and spend endpoints consistent (#27611)
## Description
Makes the user AI cost control endpoints consistent.
## Changes
- Replaces the flat `spend_limit_micros` and `limit_source` fields on
`GET /users/{user}/ai/spend` with a nested `effective_budget`, reusing
the type behind `group_budget`. The flat pair made it possible to encode
a limit without a source.
- Renames `AIGroupBudget` to `AIBudgetLimit`, since it also carries
`user_override` limits and is no longer group-specific. The type name is
not part of the wire format.
- Moves `/users/{user}/ai/budget` to `/users/{user}/ai/budget/override`.
The endpoint only ever managed the per-user override, which the type,
the handlers, and the operation IDs all already said; the path was the
only place that didn't.
> [!NOTE]
> Initially generated by Claude Opus 4.7, modified and reviewed by
@ssncferreira
|
||
|
Susana Ferreira
|
e71249a821 |
fix: ai cost control cap configurable AI spend limit (#27640)
## Problem A configured AI spend limit was only validated as `gte=0`, with no upper bound. The group spend query multiplies the per-member limit by the number of attributed members, so a large enough limit overflows `bigint` and fails the whole query, returning an error for every group in the request rather than just the misconfigured one. ## Changes - Add `MaxAISpendLimitMicros`, $1,000,000 per member per budget period. - Reject group budgets and per-user overrides above the maximum with a 400 naming the limit. - Bound both budget forms in the UI so they show the valid range before submitting. Follow-up https://github.com/coder/coder/pull/27589#discussion_r3668956350 Depends on https://github.com/coder/coder/pull/27589 > [!NOTE] > Initially generated by Claude Opus 5, modified and reviewed by @ssncferreira |
||
|
Michael Suchacz
|
c17bed25e0 |
feat: wire chat lifecycle hooks into chatd (#27429)
Wires chat lifecycle hooks into chatd, gated by the `agent-lifecycle-hooks` experiment. Part of the lifecycle hooks stack (#27401, #27428, #27430). See `docs/admin/setup/chat-lifecycle-hooks.md` for the consumer-facing contract. ## Summary When a hook URL is configured, chatd dispatches `session_start`, `user_prompt_submit`, `pre_tool_use`, `post_tool_use`, `pre_compact`, `post_compact`, and `stop` events to the consumer and applies its responses. ## Design - **Stateless**: Coder stores no hook dispatch or decision state. Delivery is at least once; consumers deduplicate on stable payload identifiers (chat ID, event type, tool-use ID) and answer duplicates with the same decision. - **Admission-time prompt effects**: `user_prompt_submit` dispatches exactly once per submission (create, send, queue, edit, subagent spawn) and folds its effects into the stored prompt as typed message parts: original-or-overridden user parts, then model-only `hook-context`, then a user-visible `hook-notice`. Hook context is stripped from every client-facing conversion; hook notices are excluded from model prompts. The server rejects hook parts in client-submitted content. - **Tool gating**: `pre_tool_use` allow can override tool input; deny becomes a synthetic denied tool result, with any returned model context persisted as a model-only transcript row so it never reaches clients. The denial text identifies an external policy (the deployment's lifecycle hook) as the source and marks the decision as persistent, so the model explains the denial instead of retrying it or misreporting it as an infrastructure failure. - **Fail closed**: a dispatch failure rejects the triggering request or moves the chat to the error state in the same transaction as the affected step, so a runnable state is never published with unapproved content. - **Admission before persistence**: `pre_tool_use` is dispatched for the calls the model produced, before the assistant message is stored. See "Staged tool admission" below. - **Fresh dispatch per tool call**: every non-provider-executed tool call is decided by its own `pre_tool_use` dispatch; Coder never reuses an earlier decision on the consumer's behalf. Retries re-dispatch the same logical event. ## Structure All hook dispatch flows through one seam: entry points build a `chathooks.Chat` (chat identity) and a `chathooks.Message` (event details) and call `Trigger.Trigger`, the only component that talks to the dispatcher. The integration lives in the `coderd/x/chatd/chathooks` subpackage, split by responsibility: - `trigger.go`: the trigger seam; builds the wire envelope per event, normalizes deny into a typed error, and holds the package's single enabled-check. - `effects.go`: pure conversion of hook results into transcript rows and prompt parts. - `errors.go`: failure classification (dispatch error messages, denial mapping, tool-result dispatch-failure scanning). - `tooluse.go`: the tool-call gate (`pre_tool_use` preflight, `post_tool_use` payloads, applying admitted input to the step). Server-bound glue stays in `coderd/x/chatd/hook_server.go`: the chat-parking dispatch error handlers, the step-commit row insertion wrappers, and the dynamic post-tool-use state loader, which depends on chatd validation types. This PR adopts the `codersdk/x/agenthooks` and `coderd/x/agenthooks/dispatch` import paths introduced at the tip of #27401; intermediate commits still reference the pre-move paths and are not individually buildable. ## Staged tool admission `pre_tool_use` originally ran at tool execution time, which is after the assistant message carrying the tool call was already committed. An `input_override` therefore had to rewrite stored message content in place. @hugodutka pointed out that chatd treats message content as immutable, and that the rewrite was a shortcut rather than a requirement. It was also a correctness problem in its own right: the rewrite only updated the database, so the transcript could show one input while a different one had executed. The hook now runs before the step is persisted: ```text provider stream ends (tool calls complete, in memory) -> pre_tool_use dispatch per call -> ONE transaction: assistant row with admitted inputs, synthetic denials, hook rows -> execute ``` The step is inserted once, carrying the input the tool runs with. `UpdateChatMessageContentByID` and `Tx.UpdateMessageContent` are deleted from #27428, so message content stays immutable. Two consequences, both intentional: - **Clients converge rather than wait.** Tool-call parts still stream live, so a rewritten call briefly shows the model's proposed input before the committed message replaces it. The chat store already clears stream state when an assistant message arrives, so the stored input wins with no frontend change and no added latency before tool cards appear. - **A call already in history was already admitted.** Execution consumes the stored input instead of dispatching a second decision, which keeps one dispatch and one set of hook effects per call. A consumer policy change between admission and execution applies to later calls, not to calls already admitted. The per-chat debug endpoint still records the provider's original tool input. Its purpose is to report provider behavior, and it requires an explicit per-chat debug flag; the invariant here covers the transcript. ## Configuration Adds `chat-hook-url`, `chat-hook-secret`, `chat-hook-timeout`, and `chat-hook-enabled` deployment options with startup validation. The flags are hidden from `coder server --help` while the feature is experimental; the setup guide documents them. ## Tool input validation Built-in tool arguments reach a consumer as raw JSON with key spelling preserved, but the tools decode those bytes with Go, which matches struct fields case-insensitively and keeps the last match. A policy reading `path` could therefore authorize one value while the tool executed another, and a lone case variant such as `{"PATH":"/secret"}` was invisible to a policy checking for `path`. Coder now rejects a built-in tool call whose input repeats a key or spells a schema property with different capitalization, before the `pre_tool_use` dispatch, so a consumer is never asked to authorize bytes whose meaning depends on the reader. Rejected calls produce an error result the model can retry; unambiguous calls in the same batch still run. A consumer-authored `input_override` is rechecked after the dispatch and fails the turn closed, because the model cannot correct it. Dynamic and MCP inputs are excluded because the client and the workspace agent execute those calls rather than coderd. Two paths needed more than a schema check. Execution resolves a deprecated tool name to its canonical tool, so validation resolves aliases first. The `edit_files` decoder also reads `search` and `replace`, which its schema does not advertise, so those aliases are now matched exactly and their case variants ignored. A hook denial now returns a structured 403 carrying `kind: "hook_denied"`, mirroring the dispatch-failure response that already carries its own kind. Without it a client cannot tell a policy decision apart from a generic failure, and the chat UI titled a denial "Request failed". Adding a kind needs no migration: `ChatErrorKind` is persisted only inside the JSONB `chats.last_error` column, whose decoder accepts unknown kinds. The hook docs also correct the tool-input convergence window. A batch dispatches sequentially before the assistant row commits, so the original input stays visible for a span that scales with the number of tool calls in the step rather than a single hook timeout. > This PR was written by Mux, an AI coding agent, on Mike's behalf. |