mirror of
https://github.com/coder/coder.git
synced 2026-10-31 11:02:50 +08:00
5480299cefc3699bc8ed6cea4e4fa865dc49d6a6
11
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
|
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
|
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.
|
||
|
Mathias Fredriksson
|
f997afa220 |
feat(coderd/x/chatd/chatloop): retain user constraints in compaction summaries (#27230)
Compaction summaries drop or soften user-stated constraints, corrections, and prohibitions, so post-compaction assistants repeat behavior the user already corrected. Add a summary prompt bullet that instructs the summarizer to quote them, treat them as standing until revoked, and attribute rules to their true source instead of defaulting to the user. Validated offline on unseen human chats: holdout P1 delta +0.175 (arbitrated), 13/18 cases improve. This improves per-compaction retention only; it does not address deep-chain correction loss. |
||
|
Michael Suchacz
|
a567f6a89f | feat: allow admins to override the chat compaction model (#27151) | ||
|
Michael Suchacz
|
0ae4554de9 | fix(coderd/x/chatd/chatloop): remove compaction timeout (#27007) | ||
|
Hugo Dutka
|
4debd23cbb |
fix: chatd refactor (#26270)
Implements the chatd stabilization RFC. Combines: - https://github.com/coder/coder/pull/25908 - https://github.com/coder/coder/pull/25923 - https://github.com/coder/coder/pull/26109 - https://github.com/coder/coder/pull/26110 - https://github.com/coder/coder/pull/26111 - https://github.com/coder/coder/pull/26112 |
||
|
Mathias Fredriksson
|
8b7e040105 |
fix(coderd/x/chatd/chatloop): discourage doctrine in compaction summaries (#25850)
Two additions to the compaction summary prompt: 1. Error specificity: the "errors encountered" bullet now instructs the model to keep error notes specific (name the file, the error, the fix) and not generalize from a specific failure to a blanket tool-avoidance rule. This addresses the doctrine crystallization pattern where a single tool failure gets promoted to a standing "avoid tool X" rule that persists across compactions and model swaps. 2. Reproducibility: a new closing sentence instructs the model to reference reproducible content by path, command, or URL rather than inlining it. Content without a stable reproducer is still preserved inline with a brief summary. This targets summary bloat from inlined code blocks (worst case: 34k chars, 76 code blocks reproducing repo content verbatim). Refs CODAGT-331 |
||
|
Thomas Kosiewski
|
df7e838c21 | feat(coderd): wire debug logging into chat lifecycle (#23917) | ||
|
Hugo Dutka
|
84740f4619 |
fix: save media message type to db (#23427)
We had a bug where computer use base64-encoded screenshots would not be interpreted as screenshots anymore once saved to the db, loaded back into memory, and sent to Anthropic. Instead, they would be interpreted as regular text. Once a computer use agent made enough screenshots and stopped, and you tried sending it another message, you'd get an out of context error: <img width="808" height="367" alt="Screenshot 2026-03-23 at 12 02 54" src="https://github.com/user-attachments/assets/f0bf6be2-4863-47ca-a7a9-9e6d9dfceeed" /> This PR fixes that. |
||
|
Cian Johnston
|
80a172f932 |
chore: move chatd and related packages to /x/ subpackage (#23445)
- Moves `coderd/chatd/`, `coderd/gitsync/`, `enterprise/coderd/chatd/` under `x/` parent directories to signal instability - Adds `Experimental:` glue code comments in `coderd/coderd.go` > 🤖 This PR was created with the help of Coder Agents, and was reviewed by my human. 🧑💻 |