* feat(review): persist and publish native code-review findings
Adds the backend core for reviewing a task's working diff before a PR
exists: task-scoped review runs and anchored, advisory findings that the
Changes/Review panel can render as inline comments.
- task_review_runs / task_review_findings schema, models, and repository,
with supersede-on-re-review and a restart sweep that cancels in-flight
runs rather than leaving them looking alive.
- ReviewService as the single write path, publishing bus events that the
WS gateway forwards to the client.
- internal/review: reviewer-identity resolution (agent profile > code-review
utility agent > user default, failing closed on passthrough-only profiles),
change-set collection, prompt batching, response parsing, anchor
extraction, and run orchestration.
- publish_review_findings_kandev MCP tool plus the task.review.* WS actions.
- Go djb2 that matches apps/web/lib/utils/hash.ts, asserted from shared
fixtures on both sides so stale detection agrees across the wire.
Spec: docs/specs/native-code-review/spec.md
Plan: docs/plans/native-code-review/plan.md
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(review): add run_code_review workflow step and wire the review runner
Makes a review pass triggerable both ways and actually runnable end to end.
- run_code_review on_enter action across models, engine compile, and an
orchestrator callback that logs rather than propagates failures, so a task
never gets stuck in a step because no reviewer was configured.
- The step's optional agent_profile_id is what lets a step review with a
different model than it implemented with; export/import swap it for a
portable agent name/model/mode so it survives a cross-workspace round trip
instead of carrying a dangling id.
- Concrete adapters for reviewer identity, changed-file collection, prompt
rendering from the code-review utility agent's editable prompt, and
inference (session-bound agentctl, falling back to the host instance so a
review works on a task whose agent is not running).
- Boot sweep cancels review runs interrupted by a restart; the E2E reset
clears review state before deleting the tasks that own it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(web): add review findings state, helpers, and live updates
Frontend plumbing for native code review, ahead of the UI that renders it.
- Review types, WS message map, and an API client that preserves the backend's
machine-readable failure code so the surface can show an actionable message.
- A review store slice that merges findings by id (a re-publish refreshes an
issue rather than listing it twice) and keeps a bounded run history.
- Findings helpers: composite file keys matching reviewFileKey, staleness via
the same per-file diff hash that drives review marks, and best-effort
re-anchoring from the stored anchor text so a moved finding follows the code
it was written about instead of pointing at whatever now occupies its line.
- Hooks for backfill, resolve/dismiss/reopen with optimistic rollback, and
sending a finding to the agent as follow-up context without changing its
status — handing a finding to an agent is not the human accepting it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(web): render review findings inline in the Changes diff
Findings now appear where the code is, in the same annotation surface as
anchored review comments, with the run control that produces them.
- ReviewFindingCard: severity chip, category, Markdown body, and the three
advisory actions. A suggestion is labelled "not applied automatically"
because it is display-only in this iteration. Resolved and dismissed
findings collapse to one muted line with an Undo, so a reviewed diff does
not stay cluttered while the human can still change their mind.
- A review-finding annotation type threaded through the diff viewer, anchored
from the same per-file diff hash the backend stamped. A finding whose diff
moved is re-anchored from its stored text when that text survives, and
otherwise reported as unanchored rather than placed against unrelated code.
- Review changes control with idle / running / cancel / failed states, and an
inline Settings link for the one failure the user can actually fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(web): make the review control reachable from the Review toolbar
Mounts the Review changes control next to the walkthrough action and lets an
inline finding wire its own actions.
The finding card container now reads the active task and session from the store
instead of taking four callbacks threaded through FileDiffViewer and DiffViewer;
an inline finding annotation only exists when the active task has findings, so
the store is always present and every other diff consumer is spared the
plumbing. Staleness likewise hashes the diff the viewer already holds, which is
already normalized, so it matches the hash the backend stamped without passing
it down.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(web): surface stale review findings above their file's diff
A finding whose diff moved and whose anchor text is gone had nowhere to render,
so it was silently dropped. It now appears in a collapsed per-file banner above
the diff: the finding may still be valid, and placing it at its stored line
would point at whatever code now occupies that line instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs(review): mark the landed native code-review plan tasks done
Tasks 01-09 (backend persistence through inline diff rendering) are
implemented, tested, and committed. Tasks 10-13 — findings overview and step
editor, mobile parity, E2E, and public docs — remain pending.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* test(review): cover the on-demand review path end to end
Adds the mock-agent review responder and a Playwright spec, so both the happy
path and the no-reviewer path are verified against a real backend.
The mock agent answers any prompt carrying the review sentinel with fenced JSON,
deriving its anchors from the prompt's own diff rather than hardcoding them, so
findings land on files and lines the fixture really changed. It includes one
deliberately malformed entry to exercise the rejected-entry path.
Writing it caught an off-by-one shared by three diff walkers: splitting on "\n"
leaves an empty element after a trailing newline, which carries no diff marker
and was being counted as a context line, pushing new-side numbering one past the
end of a hunk. Interior blank lines still count, since a whitespace-stripped
blank context line legitimately appears as "".
Also adds a listInferenceAgents helper: default_utility_agent_id needs the
registered agent-type id, and an agent row UUID is rejected as not
inference-capable.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(review): close review-run races and stale client findings
Addresses the three blockers and three suggestions from the PR #1957 automated
reviews. Two independent reviewers found each blocker separately.
- Launch claimed the task slot only after inserting the run row, so two
concurrent launches both created a run and the loser returned a pending row
with no goroutine behind it — a review that never finished until the next
restart. The slot is now claimed before any DB write, and freed on every
early return.
- Cancelling only marked the row. The goroutine ran on the shared background
context, finished inference, and its completion overwrote the cancelled
status and published findings the user had declined. Runs now carry a cancel
handle, the WS action routes through the runner so the live context is
cancelled, CompleteRun/FailRun refuse to resurrect a terminal run, and a
cancelled pass neither publishes nor re-reports itself as failed.
- The backend deletes a superseded finding and inserts a replacement with a new
id, but the publish event carried only the new one, so every open browser
showed both at the same anchor until a reload. The event now names the
superseded ids and the slice drops them.
- Run() waited on the shared WaitGroup, blocking on every other task's review;
it now waits on its own pass only.
- createReviewSlice is typed over the `set` it actually uses, so call sites no
longer pass superfluous arguments (CodeQL js/superfluous-trailing-arguments).
- A GetAgentProfile fault no longer collapses into "profile not found", which
would have told the user to configure an agent they already have.
The cancel bug exposed a hole in the original tests: they asserted cancel could
not override a terminal run, but never the reverse. Both directions are covered
now, along with the concurrent-launch race, per-run waiting, and the supersede
path on both sides of the wire. Test fakes now mirror the service's terminal
guard rather than hiding it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(web): wire permission approval UI for Kandev MCP tools
KandevToolMessage adapter ignored the permissionMessage prop, so MCP tools
that block on a permission request rendered the result row but no Approve/
Deny buttons — users could not unblock the agent. Extract the shared
permission handlers into use-permission-handlers.ts and reuse them from
both ToolCallMessage and KandevToolMessage. KandevRow now also shows an
amber pending clock icon and hides the result-count summary while waiting
for approval so the row reads correctly before the result arrives.
Adds a mock-agent scenarioKandevMCPPermission emitting
mcp__kandev__list_workspaces_kandev plus a Playwright e2e and unit tests
covering the pending UI and the absence of "0 workspaces" while pending.
* fix(web): collapse tool + permission status into resolved Kandev row
After approve/reject the orchestrator stamps the decision onto the
permission_request message immediately, but the tool_call's own
meta.status stays "pending" until the next tool_call update lands.
KandevToolMessage previously passed meta.status straight to the renderer,
so the row kept the amber pending clock and the hidden-summary state for
the brief window between the two events. resolveKandevStatus now prefers
the permission decision once present, mapping approved -> complete and
rejected -> error.
Drops a redundant Playwright `.first()` after a toHaveCount(1) assertion
on the same locator. Adds unit coverage for the resolved-decision state.
* fix(web): keep Kandev tool status independent of permission decision
The previous fix collapsed an approved permission into tool status
"complete" and a rejected permission into "error". That was wrong on two
counts: an approved tool that's still running rendered as already done
(green check + visible result summary built from an empty payload), and
any later tool_call error was silently masked because the status had
already been forced to "complete".
Replace the status override with a separate display overlay. KandevRow
now reads a `KandevPermissionUIContext` ("pending" | "rejected" |
undefined) and uses it only to drive the waiting-on-user affordances —
amber clock, hidden summary — and a rejection mark that wins over the
tool's own success state. The renderer always receives the real
tool_call status, so a tool that's approved-but-pending still reads as
pending and a tool that errors after approval still renders the red X.
Adds explicit regression tests for the two race-window failure modes
plus the new rejected overlay.
* test(web): tolerate duplicate-render backend race in Kandev MCP e2e
CI shard 1 sometimes renders two approve buttons for the same Kandev MCP
permission because the orchestrator can create a second generic tool_call
message for the same tool_call_id alongside the kandev-routed one. Both
buttons resolve the same pending_id, so clicking the first unblocks the
agent either way.
The bug the test guards against is "no approve button at all" — the
strict toHaveCount(1) assertion was tying the test to that separate
backend issue. Relax to at-least-one and click the first.
* fix: persist Deny clicks as rejected, not approved
Backend RespondToPermission previously derived status purely from cancelled:
cancelled=true -> "rejected", otherwise "approved". Deny clicks sent
cancelled=false with a reject_* option_id, so they were recorded as
"approved". Add an explicit rejected flag from the WS request so the
backend can record "rejected" without taking the cancellation path (which
races against permission_cancelled -> "expired").
Also refactor permission-request-message.tsx to consume the shared
usePermissionResponseHandlers hook (the duplicated inline handler had the
same bug) and introduce a typed models.PermissionStatus enum on the
backend so status strings are not scattered literals.
* test(web): scope Kandev MCP permission e2e to the Kandev row
Previously the test clicked the first matching permission-approve button,
which could be a duplicate generic ToolCallMessage fallback rendered for
the same pending_id by a backend race -- a green test no longer proved the
custom Kandev approval UI rendered.
Add a `data-testid="kandev-tool-permission"` wrapper around the Kandev
action row and a `kandevPermissionApproveButtons` page-object locator that
scopes through it, then assert `toHaveCount(1)` so the test only passes
when the Kandev custom row actually mounts.
* chore(orchestrator): log rejected flag in RespondToPermission
Debug log was missing the new rejected param, making it impossible to
distinguish explicit Deny from dialog cancellation when triaging
permission-flow issues.
* feat(backend): multi-question support for ask_user_question_kandev
Reshape the MCP tool from a single prompt+options pair to a questions[]
array (1-4 entries, single-choice each, custom text fallback). The
clarification HTTP handler gates the response until every question has an
answer, fans out one chat message per question (sharing pending_id, with
question_index/question_total in metadata), and returns a JSON map keyed
by question id back to the agent. The orchestrator resume prompt now
formats multi-line Q1/A1, Q2/A2 summaries when the bundle has more than
one question. Mock-agent gets a clarification-multi scenario and
ask_user_question_test.go covers the new schema + map response. No
back-compat shim — local-first tool, single-PR breaking change.
* feat(web): refresh clarification overlay with multi-question + modern UI
Each per-question chat message gets its own card with a "Question N of M"
progress chip, a ✓ Answered badge once committed, and shadcn-aligned
option buttons with kbd hints. A new useClarificationGroup hook coordinates
the per-question commits (decision A) and posts every answer in a single
batch once the user finishes the bundle; X rejects the whole group. The
single-question case keeps the 1-9 / Esc keyboard shortcuts. Playwright
e2e adds 6 multi-question scenarios (rendering, full happy path, partial
state, mixed option/custom, skip-rejects-bundle, sibling-state) on top of
the existing single-question regression suite.
* fix(web): clarification carousel UX with stepper navigation + race fix
Replaces the vertically stacked card layout with a horizontal stepper:
numbered circle dots up top show step state (active/answered/pending),
the single active question card sits in the middle, and Back / Next /
Submit buttons sit at the bottom. Picking an option auto-advances to the
next step (single-question case auto-submits, no explicit Submit button).
The stepper steps are clickable so users can revise prior answers.
Keyboard shortcuts: 1-9 for option pick, ArrowLeft/Right for navigation,
Esc for skip.
Also fixes a race that surfaced as a 400 BadRequest on the
/api/v1/clarification/:id/respond endpoint when the user clicked an
option before all N messages of a multi-question bundle had streamed in:
findPendingClarificationGroup now gates on metadata.question_total and
returns an empty array until the full bundle has arrived.
Playwright suite expanded from 6 to 12 multi-question scenarios:
stepper rendering, auto-advance, Back, stepper jump, Submit gating,
happy path, mixed custom-text, revise-via-jump, skip-from-mid-step,
number-key shortcuts, Esc skip, first-step Back disabled.
* fix(clarification): address CodeRabbit/Greptile/Claude review comments
Backend:
- Single source of truth for question validation: rename/export
NormalizeAndValidateQuestions and call it from both the HTTP and the
WebSocket-side MCP handlers, so duplicate question ids and bad option
counts can't slip through either path.
- httpRespond now validates that every answer references a real question
id from the bundle and that no id is answered twice (regression tests
added).
- questionsFromMessages reorders by metadata.question_index so the
reconstructed bundle preserves the original question order.
- adapters.CreateClarificationRequestMessages rolls back already-created
messages when a later write in the bundle fails.
- extractQuestionAnswers always returns a JSON map (even on rejection,
with an envelope { rejected, reject_reason }) and simplifyAnswer
preserves both selected_option and custom_text when both are present.
- MCP tool description: rewrite the response example to match the actual
output shape; add minItems / maxItems to the questions and options
arrays in the JSON Schema.
- Test helper: drop the local contains() shim, use strings.Contains.
Frontend:
- useMemo([props]) was a no-op (props is always a fresh ref); split into
[message, messages] so it actually memoizes.
- Drop the dead "expired" SubmitState branch — the backend never returns
410 from this endpoint.
- Remove the asymmetric option-list dimming: the agent now receives both
selected_option and custom_text together, so a half-grey state was
misleading.
- New tests: pending-clarification.test.ts covers
findPendingClarificationGroup including the question_total gate, and
use-clarification-group.test.ts covers the hook's record/submit/skip
flows and the override-merge path used by the single-question case.
* fix(web): treat clarification 409 as success so duplicate submit doesn't stick the overlay
* fix(clarification): unblock agent on persistence failure + nil-guard rejected branch
- Add Store.CancelRequest(pendingID) and call it from both the WebSocket
MCP handler and the HTTP create-request handler when message creation
fails. Without it, the agent's WaitForResponse parked for the full
2-hour MCP timeout while no clarification cards ever reached the user.
Both handlers now return an explicit error response on failure.
- applyAnswersToMessages now guards h.repo == nil in the rejected branch
(parity with sibling expectedQuestionIDs).
- extractQuestionAnswers stubs for unanswered questions in a rejected
bundle now carry { answered: false, rejected: true } so agents reading
per-question payloads see the rejection without inspecting the
envelope.
- Hook test: added 409-as-ok regression coverage for both submit and
skip paths so the previous fix can't drift.
- Store test: covers CancelRequest unblocking a parked WaitForResponse.
* fix: drop unused e2e helper + use channel sync in CancelRequest test
* refactor(web): drop dead 'message' prop from clarification overlay
Both call sites pass 'messages' (the bundle group), so the legacy
single-message prop was unused. Tightening the prop type also lets the
useMemo depend on a single value.
* fix(web): ignore modifier-keyed events in clarification carousel shortcuts
Bare 1..9 / ArrowLeft / ArrowRight / Esc still drive the carousel, but
Cmd/Ctrl+1..9 (browser tab switch) and Alt+ArrowLeft (browser back) now
pass through as the user expects.
* fix(web): exclude permission_request from turn-complete safety net
The session.turn.completed handler iterated over messages and forced
metadata.status="complete" on anything carrying tool_call_id. Permission
prompts also carry tool_call_id (for FE pairing with their tool_call),
so this loop overwrote the user's "approved" decision on every prompt
in the turn — and the FE component re-rendered the Approve/Deny buttons.
* fix(backend): protect permission_request rows from tool_call sweeps
Three permission_request invariants the prior code violated, mirroring
the FE bug fix:
- CompletePendingToolCallsForTurn: SQL UPDATE filtered by tool_call_id
presence, also catching permission_request rows. Add type filter so
the safety net only touches actual tool messages.
- GetMessageByToolCallID: returned the oldest row carrying that
tool_call_id, which races to the permission_request when timing
shifts. Exclude permission_request type so a tool_update can never
land on the user's approve/reject row.
- applyToolCallMessageUpdate: defensive guard that refuses to overwrite
a permission_request, making the invariant explicit at the layer that
does the writing.
Adds a regression test for CompletePendingToolCallsForTurn covering
approved/rejected/pending permission rows.
* test(e2e): cover approval-reappear regression with mock-agent scenario
Adds a multi-permission mock-agent scenario (three sequential bash
tools, each requiring approval) and an E2E test that approves all
three and asserts no permission action row remains visible after
the agent's turn ends, and that the resolved state survives a reload.
Supporting test infrastructure:
- data-testid hooks on PermissionActionRow.
- permissionActionRows() / permissionApproveButtons() on SessionPage.
- backend fixture force-disables AGENTCTL_AUTO_APPROVE_PERMISSIONS=false
so a developer's local AUTO_APPROVE=true env doesn't silently bypass
the permission UI in tests that exercise the approve/reject flow.
* test(backend): cover GetMessageByToolCallID permission_request exclusion
Addresses claude-review feedback on the approvals fix:
- Adds a regression test that seeds both a permission_request and a
tool_call with the same tool_call_id (permission_request created
first, so an ORDER BY created_at ASC sweep without the type filter
would race to it) and asserts the lookup returns the tool_call row.
- Bumps the applyToolCallMessageUpdate invariant guard from Warn to
Error: reaching that branch means a caller bug, not a transient
condition, and Warn would silently swallow the violation.
* fix(e2e): scope auto-approve override to permission-approval test only
The previous fixture force-disabled AGENTCTL_AUTO_APPROVE_PERMISSIONS
globally, which broke session-isolation.spec.ts and other tests that
use scenarios with permission-required tools (e.g. /e2e:read-and-edit)
and rely on auto-approve to let the agent finish without UI clicks.
Default the env to "true" in the fixture; let permission-approval.spec
flip it to "false" via a backend.restart() override in beforeAll, then
restore in afterAll. Adds an envOverrides parameter to backend.restart.
* feat(quick-chat): eager-init agent on profile selection
Switch quick-chat backend from `IntentPrepare` to `IntentStart` so the
agent process boots immediately when the user picks a profile, instead
of waiting for the first prompt. The ACP `session/new` handshake fires
before the user types anything, so the slash menu, mode dropdown, and
model selector are populated up-front. Fixes the empty-slash-menu UX
in quick chat for ACP agents (OpenCode, Claude, Codex, etc.).
Frontend picker shows a per-card spinner + disables sibling cards
during boot. Rapid profile switches (and other picker-bypassing
actions like opening a new tab, switching tabs, or closing the modal)
supersede the in-flight start via a monotonic request id; the
resolved task is cleaned up via `deleteQuickChatTask` so we never
orphan an ephemeral task or its agent process.
Backend rollback uses a fresh background context so `DeleteTask`
runs even when the request context has already been cancelled by
the client.
`mock-agent` now emits `available_commands_update` from `NewSession`
and `LoadSession` (matching real OpenCode/Claude) so the eager-init
pipeline is exercised in e2e. Existing prompt-time emit kept as an
idempotent fallback.
* test+cleanup(quick-chat): unit tests for useAgentSelection; minor cleanups
- Add 5 unit tests for useAgentSelection covering happy path, rapid-pick
supersession, reset() during in-flight, and error toast vs silent-on-superseded.
Exercises the orphan-task cleanup that the eager-init flow depends on.
(claude bot review)
- Drop redundant isDisabled alias in AgentPickerView. (claude bot review)
- Drop unused ctx parameter from emitAvailableCommandsAfterDelay — it was
always called with context.Background() so the Done() branch was dead
code. (claude bot review)
- Re-format with prettier after the rebase.
* fix(quick-chat): use constants.TaskDeleteTimeout for rollback ctx
Match the other DeleteTask call sites in task_http_handlers.go (line 851)
instead of an ad-hoc 10s timeout. (claude bot review)
* fix: include uncommitted changes in review dialog cumulative diff
The review dialog only showed uncommitted changes instead of the full
diff from the task's base commit. Two issues combined:
1. Backend used `git diff base..HEAD` (committed only); changed to
`git diff base` which compares against the working tree.
2. Frontend gave priority to git status over cumulative diff; reversed
the merge order so cumulative diff takes precedence.
Adds E2E test with a mock-agent scenario that commits changes then
leaves uncommitted modifications, verifying the review dialog shows
the complete base→working-tree diff.
* fix(e2e): remove incorrect BASE_CONTENT assertion in review cumulative diff test
The file is created after the base commit, so the entire file shows as
added in the cumulative diff. BASE_CONTENT was replaced and doesn't
appear. The key assertion is that COMMITTED_CHANGE is visible — proving
the diff includes committed changes (not just the uncommitted portion).
* fix: address CodeRabbit review feedback
- Update 3 docstrings to reflect working-tree diff (not HEAD)
- Generate scenario list dynamically in error message
- Use data-testid selector in E2E test for resilience
* style: format E2E test file
* fix(e2e): use toContainText and click file to ensure diff is scrollable
Pierre diff viewer content inside a scroll container may not be
considered "visible" by Playwright. Switch to toContainText which
checks text presence without element-level visibility. Also click
the file in the tree to ensure its diff section is scrolled into view.