mirror of
https://github.com/coder/coder.git
synced 2026-09-24 15:04:27 +08:00
efd5ce366808643b49bf4a263f02be515e3491d4
17
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
4b7494be72 |
feat: harden chat generation runtime instrumentation for billing (#27451)
Closes CODAGT-835 ## Summary `chat_messages.runtime_ms` becomes the billing source of truth for Coder Agents runtime (summed hourly by #27312), but it was built for debugging: the June refactor (#26270) silently stopped recording tool-step runtime, compaction was never measured, and interrupted turns lost their partial runtime entirely. This PR defines the billable metric, closes the paths that dropped it, and documents the definition where the data lives. ## The billable definition **`runtime_ms` is the wall-clock duration of the model invocation that produced the persisted message content**, measured from just before the provider stream opens until it is fully consumed. What counts: - Assistant generation steps, in top-level and sub-agent chats (sub-agents are ordinary chats on the same generation path). - Compaction summarization calls, persisted on the compaction assistant message (**new**). - Interrupted attempts: the message-part episode's lifetime is persisted on the partial assistant message committed by `FinishInterruption`, so partial generation time survives interruption (**new**; measured via a new `Buffer.EpisodeDuration`, which works even though the generation goroutine and the interrupt task are different tasks). What deliberately does not count (each is documented in code and docs): - **Local tool execution.** Tool wall time includes idle waits, most importantly `wait_agent` polling a sub-agent chat that already bills its own model invocations; billing the batch would double count, and excluding one tool from a concurrent batch's wall time is ill-defined. Pre-refactor instrumentation did include tool time; this makes the exclusion an explicit product definition instead of a silent regression. - **Failed model calls whose output is discarded** (retried attempts, terminal errors, content-filter refusals). They persist no content, so they bill nothing; billing errs toward undercounting. Notably a stream-silence timeout can burn 10 idle minutes before a retry, which should not be billable "active generation". If product later wants failed attempts billed, that needs a place to persist runtime on error turns (`FinishError` inserts no rows today) and is a deliberate follow-up, not instrumentation drift. - **Ancillary calls that produce no chat messages** (title generation, advisor, turn summaries) and all idle/parked time (`requires_action`, queueing). The definition is documented as `COMMENT ON COLUMN chat_messages.runtime_ms` (migration 000551, surfacing as a Go doc comment on `ChatMessage.RuntimeMs`), on `chatloop.PersistedStep.Runtime`, in the chatd architecture doc, and in the Spend Management docs page. ## Index for the hourly scan None needed: `GetTotalChatMessageRuntimeMsInRange` (#27312) filters an hour-wide `created_at` range, which the existing `idx_chat_messages_created_at` b-tree already serves; the residual `runtime_ms IS NOT NULL` filter applies to one hour of rows. A partial index would add permanent write amplification for a query that runs once an hour. > [!NOTE] > Migration 000551 is also claimed by #27312; whichever merges second renumbers via `fix_migration_numbers.sh`. ## Tests - End-to-end: the existing full-server generation test now asserts `RuntimeMs.Valid` on the committed assistant row (it previously read `.Int64` without checking `.Valid`, so it passed on NULL). - Interrupted turn: full task-level test (real DB, mock clock) asserting the partial assistant message persists the attempt's runtime. - Errored stream: asserts a failed invocation yields no step and no runtime. - Tool-using turn: asserts runtime lands on the assistant row only and tool rows stay NULL. - Compaction: asserts the summarization call duration is recorded and lands on the compaction assistant message only. - `messagepartbuffer.EpisodeDuration` unit coverage. Blocks: CODAGT-843 (B3), CODAGT-838 (D8). --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Hugo Dutka <hugo@coder.com> |
||
|
Michael Suchacz
|
404bb2f663 |
refactor(coderd/x/chatd): own provider option construction in one function (#27705)
Stacked on #27704. Provider option conversion and reasoning effort injection both create OpenAI option structs, so each of the four call sites had to pair them in the right order and pick the same transport for each. `chatprovider.ProviderOptionsForCall` now owns both steps, and the two helpers it wraps are unexported. The advisor, main generation, compaction override, and quickgen paths each collapse to one call. The ARCHITECTURE section on transport selection is updated to match: `ProviderOptionsForCall` is described as the only entry point in `chatprovider` that builds provider options for a call, delegating OpenAI conversion to `chatopenai.ProviderOptionsFromChatConfig`, and it now records that the quickgen turn status label and chat summary paths deliberately send no provider options. > Mux prepared this PR on Mike's behalf. |
||
|
Michael Suchacz
|
0e16e356b0 |
refactor(coderd/x/chatd): read the OpenAI transport from the model (#27704)
Stacked on #27703. Provider option conversion, reasoning effort injection, and file part acceptance each recomputed the OpenAI wire format from `(provider, modelID, override)`. They now read it from `chatprovider.Model`, so a decision cannot drift from the client it was built for. `ProviderOptionsFromChatConfig` takes a `Transport`, `ApplyReasoningEffort` takes a `Model`, and `AcceptsFilePartMediaType` becomes a `Model` method. `UsesResponsesAPI` and `UsesResponsesOptions` are deleted. The override extraction is unexported and reachable only from `ModelFromConfig`, which now takes the model's `ChatModelOpenAIConfig` directly, removing the six scattered extractions at call sites. That also resolves the computer-use mismatch. The computer-use model is a hardcoded default with no config row of its own: its client was built without an override while request preparation applied the chat model's. Preparation now reads the computer-use model's own transport, so the two agree without one model's client settings following a different model. Passing the chat model's `openai_config` into the computer-use client would have made them agree on the wrong value. `TestModelTransportConsumersAgree` pins the invariant in one test: the HTTP path the client actually hits, the concrete provider option struct type, the type created by reasoning effort, and text/image file acceptance. > Mux prepared this PR on Mike's behalf. |
||
|
Michael Suchacz
|
b5c9e8e471 |
refactor(coderd/x/chatd): carry the resolved OpenAI transport on a model wrapper (#27703)
Stacked on #27683. The OpenAI wire format is decided when the client is built, then thrown away, so downstream sites recompute it from `(provider, modelID, override)`. Any disagreement fails silently: the SDK type-asserts the concrete provider options struct and discards every OpenAI option, and text attachments are dropped because Responses natively accepts only images and PDFs. This adds `chatprovider.Model`, which pairs a fantasy client with the transport resolved from that client's own identity. Its fields are unexported and only the constructor sets the transport, deriving it from the client, so no caller can pick a transport that disagrees with the client it wraps. `chatopenai.Transport`'s zero value is invalid and panics when read rather than defaulting to a wire format, following the existing precedent for construction invariants. `Model` is threaded through construction, the resolve paths, and the four struct fields that store a model for later request preparation. Terminal call sites keep taking `fantasy.LanguageModel` and receive `LanguageModel()`, which avoids a new package edge from `chatloop` and `chatadvisor` into `chatprovider`. No decisions move yet. The consumers still recompute the transport, and `UsesResponsesAPI` now delegates to `TransportFor` so the two agree by construction. #27704 makes the consumers read it from the model. > 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. |
||
|
Michael Suchacz
|
fc24c27dfd |
fix: reserve chat hook dispatch capacity for running turns (#27656)
## Context Follow-up fix from live UAT of the merged chat lifecycle hooks stack (#27430). Its companion UAT fix (#27655) has merged, so this targets `main` directly. ## Why? UAT measured a burst of 1,500 concurrent chat creations against a consumer with 1.2s latency. 255 were admitted and 1,245 got `502 hook_dispatch_failed (over_capacity)`, which is correct fail-closed behavior. The collateral wasn't: the same burst failed 24 `stop` dispatches, parking chats that had already been admitted and had already executed tools. One 256-slot semaphore served every event, so new-work admission could take every slot and kill turns in flight. Callers now classify each dispatch as admission or generation, and admission draws from a 192-slot gate held *before* the shared pool. At least 64 shared slots stay reachable only by dispatches for work a chat already admitted. The dispatcher is per `coderd` replica, so these limits are per replica, not deployment-wide, and the docs say so. **The caller classifies, not the event type.** Event type isn't a reliable proxy in either direction: a subagent spawn dispatches `user_prompt_submit` from inside a running turn, and the edit path dispatches `session_start` at admission time. `CapacityClassUnset` is rejected in `Dispatch`, so a new call site fails closed rather than silently inheriting a share. **Acquisition order is load-bearing.** Admission takes its own gate first. Taking a shared slot first would let admissions queued on the gate occupy the very capacity the reserve protects. `acquireCapacity` is the only path that takes either pool, so the order can't be bypassed. ## What this does not guarantee Nothing bounds how many turns generate concurrently, so the 192/64 split is a judgement call, not a derived ceiling. This stops an *admission* burst from consuming every slot; it does not make the remainder sufficient. A large enough generation load can still exhaust the reserve and error a running chat. The docs say so explicitly rather than promising a guarantee the code doesn't deliver. Generation can now take all 256 slots, so generation traffic starves admission harder than before. That's the intended priority: rejecting a new prompt is recoverable, ending a turn that already ran tools is not. ## Testing Red-green proved both new tests. Removing the release-on-failure path fails `RefusedSharedAcquireReleasesAdmission` deterministically; removing the expired-deadline check fails `ExpiredDeadlineRefusesFreeSlot` in 18/30 runs. That deadline check fixes a real race found in review. `acquire` previously shared one `time.Timer` across both acquires. Because `select` picks a ready case at random, an admission dispatch could take a slot after its capacity deadline had passed. Measured over 300 trials: 135 late acquisitions, worst overshoot 2.1ms. `acquire` now takes an absolute deadline and refuses an expired one before selecting, which measures 0/300. Go: `coderd/x/agenthooks/...` and `coderd/x/chatd/...`, plus `-race -count=3` on the dispatcher. > Mux opened this PR on Mike's behalf. |
||
|
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. |
||
|
Michael Suchacz
|
91c7232d97 |
feat: add chat suffix messages, idle failure, and content update support (#27428)
Adds generic chat state and query capabilities that the lifecycle hooks integration (#27429) builds on. Part of the lifecycle hooks stack (#27401, #27429, #27430). - `chatstate`: `EditMessage` accepts caller-provided suffix messages inserted after the replacement in the same transaction, transitions can carry a typed error kind, and `FinishError` is also allowed from waiting chats so admission-time failures can park an idle chat in error. - `chatstate`: `ValidateToolResults` holds the submitted-tool-result rules (duplicate, invalid JSON, missing, unexpected) in one place, so `CompleteRequiresAction` and API-level prechecks reject the same payloads with the same typed causes. - `database`: `InsertChat` accepts an optional caller-provided ID. No hook-specific state or behavior is introduced here; these primitives are usable by any caller. An earlier revision added a message-content rewrite primitive so a `pre_tool_use` override could update an already-committed tool call. Message content is immutable by design, and @hugodutka pushed back on changing that. The rewrite is gone: #27429 now dispatches the hook before the assistant message is stored, so the stored input is the one that runs and nothing needs updating. > This PR was written by Mux, an AI coding agent, on Mike's behalf. |
||
|
Michael Suchacz
|
02fd1cc691 | feat: allow spawn_agent model and reasoning effort override (#27385) | ||
|
Michael Suchacz
|
3227cac217 |
feat: add manual chat compaction via /compact (#27081)
Adds a user-triggered `/compact` action for Coder Agents chats: typing
`/compact` in the composer (or picking it from the `/` trigger menu)
summarizes the conversation so far to free up context window space.
## How it works
- New `POST /api/experimental/chats/{chat}/compact` endpoint
(owner-only, RBAC `ActionUpdate`, excluded from the public API reference
via `x-apidocgen skip`). It marks the chat with a durable one-shot
`chats.compaction_requested_at` signal and moves it `waiting -> running`
via a new `RequestCompaction` state transition; no message row is
inserted. AI Gateway attribution needs no per-request key: generation
preparation resolves the owner's synthetic API key (#27170) like any
other turn.
- `RequestCompaction` hands off chat ownership (clears
`worker_id`/`runner_id`) so a worker acquisition hint is published;
since the transition changes no history, the previous runner could
otherwise miss the request under reordered pubsub delivery.
- The background chat worker picks the chat up like any other turn. A
pending manual request takes precedence over turn completion in the
generation decision, and forces compaction even below the automatic
threshold (and when compaction is disabled via threshold=100). The
commit step consumes the request marker in the same transaction; any
transition that ends the turn clears stale markers.
- The summary triplet reuses the automatic-compaction path, now tagged
with a `source` (`automatic` | `manual`) that is plumbed through
streamed progress parts, persisted tool JSON, and the UI label
("Summarized (manual)").
- Validation order: busy chats reject with 409 (state-machine conflict),
empty/already-compacted chats with 409 "nothing to compact", archived
chats with 400; the owner usage-limit check runs last so no-op requests
surface the specific conflict instead of a limit error.
- Web UI: the `/` trigger menu now has a built-in "Commands" group
listing `/compact`; submit intercepts exactly `/compact` and calls the
endpoint instead of sending a message. A personal or workspace skill
named `compact` takes precedence over the built-in command; while skill
collisions are still resolving, an exact `/compact` submission is
blocked with a retryable hint instead of leaking as message text.
History and queued-message edits are never intercepted. After
compaction, the context usage indicator resets to its unknown state
until the next assistant response reports fresh usage, instead of
showing the stale pre-compaction number.
- codersdk: `ExperimentalClient.CompactChat`.
Worker-path execution (rather than compacting synchronously in the
handler) reuses the existing lock fencing, live "Summarizing..."
streaming, retry accounting, restart resilience, and debug-run
observability. Rationale documented in `coderd/x/chatd/ARCHITECTURE.md`.
## Testing
- State machine: transition-matrix coverage for `RequestCompaction`,
marker lifecycle tests (carried by lease renewals/queue appends, cleared
by terminal transitions, consumed by commit), ownership handoff +
acquisition hint assertions.
- Worker: decision-ordering and forced-compaction unit tests;
active-server end-to-end test (manual compact below threshold produces a
`source=manual` summary, returns to `waiting`, no assistant follow-up;
busy chat rejected).
- API: success, archived, non-owner, RBAC-denied, empty-chat, no-daemon
cases; usage-limit ordering (at-limit owners still get
state/nothing-to-compact conflicts for no-op requests, with marker
rollback).
- Frontend: Storybook play tests for the Commands menu group, submit
intercept, skill-name collision, queued-edit passthrough, and
manual/automatic tool rendering; unit tests for command availability
resolution and the post-compaction context usage reset.
> This PR was created by Mux, an AI coding agent, working on Mike's
behalf.
|
||
|
Michael Suchacz
|
4ed6fcced7 |
refactor(coderd): stop storing chat gateway key IDs and drop the columns (#27171)
> Mux is working on behalf of Mike. ## Summary Stop reading and writing the legacy `api_key_id` columns on chat messages and queued messages, and drop the columns in the same PR. Runtime AI Gateway attribution continues to use the per-user synthetic key introduced by #27170. With the columns gone, `sqlc` generates `database.ChatMessage` and `database.ChatQueuedMessage` without `api_key_id`, so no transitional query scaffolding is needed. Migration `000548` drops the `api_key_id` columns. #27170 already removed their foreign keys, so the down migration re-adds nullable text columns without constraints. Previous column values cannot be restored. Also moves the model config validation in `CreateChat` above the message-building work so a disabled or invalid model fails fast. On main this mattered more: the old ordering minted a synthetic API key before rejecting the request. Deploy note: replicas still running the previous release write `api_key_id` on insert, so chat message inserts on old replicas fail during the rolling window after the column drop. This was previously split across two PRs to avoid that window; per review feedback the split added more churn than it was worth for an experimental surface. Depends on #27170 (merged). |
||
|
Michael Suchacz
|
5f5efa49da |
fix: stop sending adaptive thinking to pre-4.6 Anthropic models (#27314)
## Problem
A chat model config with `reasoning_effort` set on a pre-4.6 Anthropic
model (for example `claude-haiku-4-5`) fails every generation with HTTP
400 `adaptive thinking is not supported on this model`, surfaced in chat
as "Anthropic returned an unexpected error." The fantasy Anthropic
provider always serialized effort as `thinking: {type: "adaptive"}` plus
`output_config.effort`, a shape only Claude 4.6+ accepts.
## Changes
- Bump the coder/fantasy pin to include coder/fantasy#47: the provider
now converts effort into `{type: "enabled", budget_tokens}` on models
older than Claude 4.6, with the budget derived from the call's
`max_tokens` (aibridge-mirroring ratios, 1024-token API floor; below the
floor thinking is omitted, which keeps small-budget calls like title
generation working). Adaptive-capable models keep the current shape, and
Opus 4.5 keeps `output_config.effort` alongside the derived budget since
it supports effort without adaptive thinking. Models older than Claude
3.7 predate extended thinking, so effort sends no thinking at all there.
`minimal` is normalized to `low`, `xhigh` falls back to `max` on
adaptive models that predate the xhigh tier (Claude 4.7+); effort `none`
disables thinking, including an explicit `thinking: {type: "disabled"}`
on Claude 5+ models that otherwise run adaptive thinking by default. The
Bedrock provider wraps the Anthropic one, so both are covered, and
Vertex-style `@date` model IDs parse correctly.
- `TestActiveServer_CompactionModelOverride` previously codified the
buggy shape (asserting `output_config.effort` sent to
`claude-3-5-haiku-latest`). The summary-routing subtest is now a
three-case table: pre-thinking override models (Claude 3.5) expect no
thinking, legacy budget-thinking ones (Haiku 4.5) expect enabled
thinking with the derived budget, adaptive-capable ones still expect
`output_config.effort`.
- New regression test `TestActiveServer_AnthropicModelReasoningEffort`:
a `claude-haiku-4-5` config with `reasoning_effort` produces enabled
thinking with the derived budget and no `output_config` on the wire, and
a `claude-sonnet-5` config with effort `none` sends an explicit thinking
disable.
- `chattest.AnthropicRequest` gains a `Thinking` field so tests can
assert the thinking config.
- One-sentence note in the chatd ARCHITECTURE reasoning-effort section.
No chatd production code changes: `ApplyReasoningEffort` keeps setting
`Effort`, which is now valid for every Anthropic model.
## Validation
- `go test ./coderd/x/chatd/...` passes (19 packages).
- Fork PR validated separately: full fantasy test suite plus new
provider unit tests (version gating incl. Vertex/Bedrock IDs, budget
derivation, floor behavior, normalization, effort `none` incl. Claude 5+
disable, Opus 4.5 effort preservation, sampling-param stripping),
golangci-lint clean.
Closes
[CODAGT-812](https://linear.app/codercom/issue/CODAGT-812/reasoning-effort-on-pre-46-anthropic-models-fails-generations-with).
> This PR was authored by Mux, an AI coding agent, acting on Mike's
behalf.
|
||
|
Michael Suchacz
|
997b5d0843 |
feat: add synthetic gateway keys (#27170)
> Mux is working on behalf of Mike. ## Summary Add a per-user synthetic API key for chatd AI Gateway attribution. Chatd resolves the key from the chat owner, extends it before expiry, and discards the generated bearer token so the key is never a usable credential. There is no mapping table. The key is resolved from `api_keys` by a deterministic token name (`chatd_<owner_id>_session_token`), mirroring the provisionerd session token model, with three deltas that chatd needs: - **Login type guard**: token names are unvalidated user input, so a user can create a bearer token with the colliding name. The lookup excludes `login_type = 'token'` rows, so chatd never picks up (or extends) a real user token. Synthetic keys are minted with the owner's login type, which is never `token`. - **In-place expiry extension instead of delete-and-reinsert**: chat generations have no stop boundary, and an in-flight generation may have already delegated the current key ID to aibridged. Extending `expires_at` keeps the key ID stable forever. - **Advisory-lock mint**: the unique index on token names is partial (`WHERE login_type = 'token'`), so nothing DB-enforces uniqueness for synthetic keys. A per-user advisory lock serializes concurrent mints. Keys carry a minimal scope (`api_key:read`) as defense in depth; the delegated gateway path never evaluates scopes and the secret is discarded at mint. Migration 000544 removes the foreign keys from the legacy message and queue `api_key_id` columns while chatd continues stamping them for rolling compatibility. Stale IDs are tolerated because routing uses `chats.owner_id`. Individual key deletion, delete-all, and password reset remove the key without changing chat history or queue versions, and the next lookup remints it. Suspension does not delete the key; delegated gateway authorization rejects inactive users at request time. This is the first PR in a three-PR rollout and must be fully deployed before #27171. Refs https://linear.app/codercom/issue/CODAGT-561/maintain-synthetic-api-key-per-user-per-chat |
||
|
Michael Suchacz
|
a567f6a89f | feat: allow admins to override the chat compaction model (#27151) | ||
|
Ethan
|
ea4554025e |
fix(coderd): stop manual title generation from writing to chat_messages (#27087)
Coder Agents chats could get stuck showing "Thinking" forever when a title regenerate/propose request ran while a generation was in flight. Manual title generation recorded token cost by inserting a hidden assistant message into `chat_messages` and immediately soft-deleting it. Triggers on that table sync `chats.history_version` to `snapshot_version`, so this out-of-band write broke the `history_version` fence of an in-flight generation task, killing it without a replacement and leaving the chat stuck in `running`. Remove the accounting path entirely; AI Gateway already records title-call usage in `aibridge_interceptions`/`aibridge_token_usages`. The manual title endpoints no longer write to `chat_messages` at all, and new regression tests assert `history_version` stays untouched. Note this intentionally drops title-generation cost from chatd's chat-level cost surfaces; it still counts against the user's AI budget via AI Gateway. Closes CODAGT-595 |
||
|
Danielle Maywood
|
d66e4d794f | feat: add configurable reasoning effort to Coder agents (#26974) | ||
|
Hugo Dutka
|
91543c391d |
chore: add chatd ARCHITECTURE.md and mention it in AGENTS.md (#26478)
Closes [CODAGT-610](https://linear.app/codercom/issue/CODAGT-610/add-an-architecturemd-file-to-chatd). Adds an ARCHITECTURE.md file which describes the architecture of the chatd subsystem. It's meant for reading by both humans, who would like to understand chatd better, and agents. It's an edited version of the chatd stabilization RFC. |