Bring buildErrandPrompt() and buildErrandRevisionPrompt() in line with
the codebase pattern used by execute, plan, detail, etc. Import shared
DEVIATION_RULES and GIT_WORKFLOW constants. Add session_startup (pwd,
git status, CLAUDE.md, expected-pwd.txt), execution_rules, and
anti_patterns sections. Keep signal format inline since errands use
result.message instead of the standard questions format.
Agents sometimes write signal.json before finishing output files,
causing work to be silently discarded. Replace weak "As your final
action" with explicit CRITICAL warning explaining the completion
trigger mechanism.
In YOLO mode, if mergePhaseIntoInitiative() threw (e.g., branch
doesn't exist, merge conflict), the error propagated and completePhase()
was never called — leaving the phase permanently stuck at in_progress.
Also wrap per-phase recovery in try-catch so one failing phase doesn't
abort the entire recoverDispatchQueues() loop.
When git worktree add fails with "branch already used by worktree at
<path>", parse the stale path, force-remove it, and retry once. Fixes
blocked task retries where the old agent-workdirs directory still exists
on disk and git worktree prune alone can't clear the reference.
Errands now create worktrees via ProcessManager.createWorktreesForProjects()
into agent-workdirs/<alias>/<project.name>/ instead of repos/<project>/.cw-worktrees/<errandId>.
This makes getAgentWorkdir + resolveAgentCwd work correctly for all agent types.
Key changes:
- Extract createWorktreesForProjects() from createProjectWorktrees() in ProcessManager
- Add resolveAgentCwd() to ProcessManager (probes for .cw/output in subdirs)
- Add projectId to SpawnAgentOptions for single-project agents (errands)
- Skip auto-cleanup for errand agents (worktrees persist for merge/abandon)
- Errand router uses agentManager.delete() for cleanup instead of SimpleGitWorktreeManager
- Remove cwd parameter from sendUserMessage (resolves via worktreeId)
- Add pruneProjectRepos() to CleanupManager for errand worktree refs
Replaces the in-memory filter (agentManager.list() + filter) with a direct
repository query that LEFT JOINs tasks, phases, and initiatives to return
taskName, phaseName, initiativeName, and taskDescription alongside agent fields.
- Adds AgentWithContext interface and findWaitingWithContext() to AgentRepository port
- Implements findWaitingWithContext() in DrizzleAgentRepository using getTableColumns
- Wires agentRepository into TRPCContext, CreateContextOptions, and TrpcAdapterOptions
- Adds requireAgentRepository() helper following existing pattern
- Updates listWaitingAgents to use repository query instead of agentManager
- Adds 5 unit tests for findWaitingWithContext() covering all FK join edge cases
- Updates existing AgentRepository mocks to satisfy updated interface
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Errand agents are spawned in the errand's git worktree (managed by
SimpleGitWorktreeManager), not in agent-workdirs/<alias>/.
sendUserMessage was deriving the cwd from worktreeId which pointed to
the non-existent agent-workdirs path. Now the errand.sendMessage
procedure resolves the actual worktree path and passes it through.
When spawn() receives an explicit cwd (errands), the manager was still
creating a standalone worktree at agent-workdirs/<alias>/ and injecting
its path into the workspace layout prompt. The agent then edited files
in the wrong directory — on a different branch than the errand's.
Now when cwd is provided, we skip worktree creation entirely and use
the caller's cwd for workspace layout, .cw/output/, and all paths.
Merge two useEffects in AgentOutputViewer into one to fix race where
agentId reset clears messages after data effect sets them on remount.
Add "commit before signaling" instruction to errand prompts so
Changes tab shows diff after completion.
- Register errandProcedures in appRouter (was defined but never spread)
- Fix nullable projectId guard in errand delete/abandon procedures
- Add sendUserMessage stub to MockAgentManager in headquarters and
radar-procedures tests (AgentManager interface gained this method)
- Add missing qualityReview field to Initiative fixture in file-io test
(schema gained this column from the quality-review phase)
- Cast conflictFiles access in CLI errand resolve command
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds the fourth test case from the spec: when shouldRunQualityReview
throws, the orchestrator must not crash, must log a warning (verified
implicitly by the catch block), and must still call scheduleDispatch()
so dispatch continuity is maintained.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add errand.requestChanges procedure that re-spawns an agent in the
existing worktree with user feedback. Replace raw <pre> diff blocks
with syntax-highlighted ErrandDiffView using FileCard components.
Add Output/Changes tabs to the active errand view.
When an agent stops, check whether a quality review should run before
auto-completing the task. If shouldRunQualityReview returns run:true,
delegate to runQualityReview (which transitions task to quality_review
and spawns a review agent) instead of calling completeTask directly.
Falls back to completeTask when agentRepository or agentManager are
not injected, or when the task lacks phaseId/initiativeId context.
- Add agentManager optional param to ExecutionOrchestrator constructor
- Extract tryQualityReview() private method to compute branch names and
repo path before delegating to the quality-review service
- Pass agentManager to ExecutionOrchestrator in container.ts
- Add orchestrator integration tests for the agent:stopped quality hook
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When an execute-mode agent stops with task_complete and the initiative has
qualityReview=true, the orchestrator now spawns a fresh execute-mode agent
to run /simplify on changed .ts/.tsx/.js files before marking the task
completed. The task transitions through quality_review status as a recursion
guard so the review agent's stop event is handled normally.
- Add apps/server/execution/quality-review.ts with three exported functions:
computeQualifyingFiles, shouldRunQualityReview, runQualityReview
- Add apps/server/execution/quality-review.test.ts (28 tests)
- Update ExecutionOrchestrator to accept agentManager, replace
handleAgentStopped with quality-review-aware logic, add getRepoPathForTask
- Update orchestrator.test.ts with 3 quality-review integration tests
- Update container.ts to pass agentManager to ExecutionOrchestrator
- Update docs/dispatch-events.md to reflect new agent:stopped behavior
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds apps/server/execution/quality-review.ts with three exported functions:
- computeQualifyingFiles: diffs task branch vs base, filters out *.gen.ts and dist/ paths
- shouldRunQualityReview: evaluates all six guard conditions (task_complete, execute mode,
in_progress status, initiative membership, qualityReview flag, non-empty changeset)
and returns { run, qualifyingFiles } to avoid recomputing the diff in the orchestrator
- runQualityReview: transitions task to quality_review, spawns execute-mode review agent
on the task branch, logs the review agent ID, and falls back to completed on spawn failure
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds qualityReview: z.boolean().optional() to the updateInitiativeConfig
input schema so the field passes through to the repository layer.
Includes integration tests verifying set-true, set-false, and
omit-preserves-existing round-trip behavior.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds two new fields to the database and propagates them through the
repository layer:
- Task status enum gains 'quality_review' (between in_progress and
completed), enabling a QA gate before tasks are marked complete.
- initiatives.quality_review (INTEGER DEFAULT 0) lets an initiative be
flagged for quality-review workflow without a data migration (existing
rows default to false).
Includes:
- Schema changes in schema.ts
- Migration 0037 (ALTER TABLE initiatives ADD quality_review)
- Snapshot chain repaired: deleted stale 0036 snapshot, fixed 0035
prevId to create a linear chain (0032 → 0034 → 0035), then generated
clean 0037 snapshot
- Repository adapter already uses SELECT * / spread-update pattern so
no adapter code changes were needed
- Initiative and task repository tests extended with qualityReview /
quality_review_status describe blocks (7 new tests)
- docs/database.md updated
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three fixes for phases getting stuck when a detail task crashes and is retried:
1. detailPhase mutation (architect.ts): clean up orphaned pending/in_progress
detail tasks before creating new ones, preventing duplicates at the source
2. orchestrator recovery: detect and complete stale duplicate planning tasks
(same category+phase, one completed, one pending)
3. ensureBranch: catch "already exists" TOCTOU race instead of blocking phase
Populates the agent_metrics table from existing agent_log_chunks data after
the schema migration. Reads chunks in batches of 500, accumulates per-agent
counts in memory, then upserts with additive ON CONFLICT DO UPDATE to match
the ongoing insertChunk write-path behavior.
- apps/server/scripts/backfill-metrics.ts: core backfillMetrics(db) + CLI wrapper backfillMetricsFromPath(dbPath)
- apps/server/scripts/backfill-metrics.test.ts: 8 tests covering all chunk types, malformed JSON, isolation, empty DB, and re-run double-count behavior
- apps/server/cli/index.ts: new top-level `cw backfill-metrics [--db <path>]` command
- docs/database-migrations.md: Post-migration backfill scripts section documenting when and how to run the script
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
listForRadar previously called findByAgentIds() and JSON-parsed every chunk to
compute questionsCount, subagentsCount, and compactionsCount. Switch to
findMetricsByAgentIds() which reads the pre-computed agent_metrics table,
eliminating the chunk scan and per-row JSON.parse entirely.
Add two new test cases: agent with no metrics row returns zero counts, and
listForRadar response rows never carry chunk content.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wrap insertChunk in a synchronous better-sqlite3 transaction that upserts
agent_metrics counters atomically on every chunk insert. Malformed JSON
skips the upsert but always preserves the chunk row.
Add findMetricsByAgentIds to the interface and Drizzle adapter for
efficient bulk metric reads.
Add 8-test suite covering all write/read paths and edge cases.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When an initiative has multiple end phases (leaf nodes with no
dependents), queueAllPhases now auto-creates an Integration phase
that depends on all of them. This catches cross-phase incompatibilities
(type mismatches, conflicting exports, broken tests) before review.
Adds the agentMetrics table to SQLite schema for storing pre-computed
per-agent event counts (questions, subagents, compactions), enabling
listForRadar to fetch one row per agent instead of scanning log chunks.
Also fixes pre-existing Drizzle snapshot chain collision in meta/
(0035/0036 snapshots had wrong prevId due to parallel agent branches)
to unblock drizzle-kit generate.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ErrandRepository was instantiated in the container but never passed
from TrpcAdapterOptions into createContext(), causing all errand
procedures to throw "Errand repository not available" at runtime.
Claude CLI occasionally hangs after writing signal.json but never exits.
Add an optional signal check to pollForCompletion: after a 60s grace
period, check signal.json every 30s. If a valid completion signal is
found while the process is still alive, SIGTERM it and proceed to
normal completion handling.
errandProcedures was defined but never imported/spread into the app
router, causing "No procedure found on path errand.create". Also fixed
nullable projectId TS errors in delete/abandon and added missing
sendUserMessage to test mocks.
- Remove original task blocking in handleConflict (task is already completed by handleAgentStopped)
- Return created conflict task from handleConflict so orchestrator can queue it for dispatch
- Add dedup check to prevent duplicate resolution tasks on crash retries
- Queue conflict resolution task via dispatchManager in mergeTaskIntoPhase
- Add recovery for erroneously blocked tasks in recoverDispatchQueues
- Update tests and docs
Resolves add/add conflict in diff-cache.ts (kept typed PhaseMetaResponse/
FileDiffResponse interfaces from HEAD over unknown-typed singletons from test
branch) and content conflict in phase.ts (kept both phaseMetaCache and
fileDiffCache imports; removed auto-merged duplicate firstClone/headHash/
cacheKey/cached declarations and unreachable empty-projects guard).
Also cleans auto-merged duplicate getHeadCommitHash in orchestrator.test.ts
and simple-git-branch-manager.ts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Creates diff-cache.ts module with generic DiffCache<T> class (TTL, prefix
invalidation, env-var configuration) and exports phaseMetaCache / fileDiffCache
singletons. Wires cache into getPhaseReviewDiff via getHeadCommitHash on
BranchManager. Adds 6 unit tests for DiffCache and 5 integration tests
verifying cache hit/miss behaviour, prefix invalidation, and NOT_FOUND guard.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds DiffCache<T> module, extends BranchManager with getHeadCommitHash,
and wires phase-level caching into getPhaseReviewDiff and getFileDiff.
Cache is invalidated in ExecutionOrchestrator after each task merges into
the phase branch, ensuring stale diffs are never served after new commits.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Rewrites getPhaseReviewDiff to return file-level metadata (path, status,
additions, deletions) instead of a raw diff string, eliminating 10MB+
payloads for large repos. Adds getFileDiff for on-demand per-file hunk
content with binary detection via numstat. Multi-project initiatives
prefix file paths with the project name to avoid collisions.
Adds integration tests that use real local git repos + in-memory SQLite
to verify both procedures end-to-end (binary files, deleted files,
spaces in paths, error cases).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds FileStatEntry type and two new primitives to the BranchManager
port and SimpleGitBranchManager adapter, enabling split diff
procedures in the tRPC layer without returning raw multi-megabyte diffs.
- FileStatEntry captures path, status, additions/deletions, oldPath
(renames), and optional projectId for multi-project routing
- diffBranchesStat uses --name-status + --numstat, detects binary
files (shown as - / - in numstat), handles spaces in filenames
- diffFileSingle returns raw unified diff for a single file path
Re-recorded all 4 cassette files to reflect current prompt templates.
Added `workdir/**` to vitest exclude list to prevent test discovery
in agent worktree directories.
Integrates main branch changes (headquarters dashboard, task retry count,
agent prompt persistence, remote sync improvements) with the initiative's
errand agent feature. Both features coexist in the merged result.
Key resolutions:
- Schema: take main's errands table (nullable projectId, no conflictFiles,
with errandsRelations); migrate to 0035_faulty_human_fly
- Router: keep both errandProcedures and headquartersProcedures
- Errand prompt: take main's simpler version (no question-asking flow)
- Manager: take main's status check (running|idle only, no waiting_for_input)
- Tests: update to match removed conflictFiles field and undefined vs null
Add five new tRPC query procedures powering the Radar page's per-agent
behavioral metrics (questions asked, subagent spawns, compaction events,
inter-agent messages) plus the batch repository methods they require.
Repository changes:
- LogChunkRepository: add findByAgentIds() for batch fetching without N+1
- ConversationRepository: add countByFromAgentIds() and findByFromAgentId()
- Drizzle adapters: implement all three new methods using inArray()
- InMemoryConversationRepository (integration test): implement new methods
tRPC procedures added:
- agent.listForRadar: filtered agent list with per-agent metrics computed
from log chunks (questionsCount, subagentsCount, compactionsCount) and
conversation counts (messagesCount); supports timeRange/status/mode/initiative filters
- agent.getCompactionEvents: compact system init chunks for one agent (cap 200)
- agent.getSubagentSpawns: Agent tool_use entries with prompt preview (cap 200)
- agent.getQuestionsAsked: AskUserQuestion tool calls with questions array (cap 200)
- conversation.getByFromAgent: conversations by fromAgentId with toAgentName resolved
All 13 new unit tests pass; existing test suite unaffected.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
