Add a dedicated "Asking questions" section to the errand prompt so the
agent knows it can pause, ask for clarification, and wait for the user
to reply via the UI chat input or `cw errand chat`. Previously the
prompt said "work interactively" with no guidance on the mechanism,
leaving the agent to guess.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements three primitives needed before errand tRPC procedures can be wired up:
- agentManager.sendUserMessage(agentId, message): resumes an errand agent with a
raw user message, bypassing the conversations table and conversationResumeLocks.
Throws on missing agent, invalid status, or absent sessionId.
- writeErrandManifest(options): writes .cw/input/errand.md (YAML frontmatter),
.cw/input/manifest.json (errandId/agentId/agentName/mode, no files/contextFiles),
and .cw/expected-pwd.txt to an agent workdir.
- buildErrandPrompt(description): minimal prompt for errand agents; exported from
prompts/errand.ts and re-exported from prompts/index.ts.
Also fixes a pre-existing TypeScript error in lifecycle/controller.test.ts (missing
backoffMs property in RetryPolicy mock introduced by a concurrent agent commit).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Conflict-resolution agents (and any initiative-based agent) can write
.cw/output/signal.json inside a project subdirectory (e.g.
agent-workdirs/<name>/codewalk-district/.cw/output/) rather than the
parent agent workdir. This caused two failures:
1. spawnInternal wrote spawn-diagnostic.json before registering the
agent in activeAgents and starting pollForCompletion. If the .cw/
directory didn't exist (no inputContext provided), the write threw
ENOENT, orphaning the running process with no completion monitoring.
2. resolveAgentCwd in cleanup-manager and output-handler only probed
for a workspace/ subdirectory (standalone agents) but not project
subdirectories, so reconciliation and completion handling couldn't
find signal.json and marked the agent as crashed.
Fixes:
- Move activeAgents registration and pollForCompletion setup before
the diagnostic write; make the write non-fatal with mkdir -p
- Add project subdirectory probing to resolveAgentCwd in both
cleanup-manager.ts and output-handler.ts
spawnConflictResolutionAgent was passing the initiative branch as branchName,
causing SimpleGitWorktreeManager.create() to force-reset it to the target
branch. Now spawns on a unique temp branch based off the initiative branch,
with the agent using git update-ref to advance the initiative branch after
resolving conflicts. Also fixes stale diff/commits cache after resolution.
Pre-merge mergeability check via `git merge-tree --write-tree` (dry-run, no
side effects). When conflicts exist the "Merge & Push" button is disabled and
a ConflictResolutionPanel shows conflict files with options to resolve manually
or spawn a conflict-resolution agent. Agent questions appear inline via
QuestionForm; on completion the mergeability re-checks automatically.
New server-side: MergeabilityResult type, BranchManager.checkMergeability,
conflict-resolution prompt, checkInitiativeMergeability query,
spawnConflictResolutionAgent mutation, getActiveConflictAgent query.
New frontend: useConflictAgent hook, ConflictResolutionPanel component,
mergeability badge + panel integration in InitiativeReview.
Completed phases showed "No phases pending review" because:
1. Frontend filtered only pending_review phases
2. Server rejected non-pending_review phases
3. After merge, three-dot diff returned empty (merge base moved)
Fix: store pre-merge merge base hash on phase, use it to reconstruct
diffs for completed phases. Frontend now shows both pending_review and
completed phases with read-only mode (Merged badge) for completed ones.
Prevents two bugs in the resumeForCommit flow:
1. Agent navigated to main repo instead of worktree due to relative paths
in commit prompt — now uses absolute paths from getDirtyWorktreePaths
2. git add -A staged unrelated files (screenshots, other agents' work) —
now uses git add -u to only stage tracked modified files
Task-level approval (requiresApproval, mergeRequiresApproval,
pending_approval status) was redundant with executionMode
(yolo vs review_per_phase) and blocked the orchestrator's
phase completion flow. Tasks now complete directly;
phase-level review via executionMode is the right granularity.
Removed: schema columns (left in DB, removed from Drizzle),
TaskPendingApprovalEvent, approveTask/listPendingApprovals
procedures, findPendingApproval repository method, and all
frontend approval UI.
- Add getActiveRefineAgent to spawn mutation optimistic updates and
live event invalidation rules so the refine panel reflects agent
state immediately without manual refresh
- Accept optional instruction param in buildRefinePrompt() and inject
it as <user_instruction> block so the agent knows what to focus on
- Pass input.instruction through in architect router spawn call
- Track worktree removal success in autoCleanupAfterCompletion() instead of
always returning removed:true when removeAgentWorktrees() throws
- Add removeAgentBranches() call to auto-cleanup path (agent/* branches were
never cleaned after completion)
- Add filesystem cleanup (worktrees, branches, logs) to dismiss() to prevent
resource leaks until next server restart
commitRetryCount was being deleted in cleanupAgentState(), which runs
before tryAutoCleanup() checks the count. This reset the counter to 0
on every cycle, making MAX_COMMIT_RETRIES=1 dead code. Agents would
retry commits forever.
Move commitRetryCount cleanup to stop()/delete() only, letting
tryAutoCleanup() manage it during the retry lifecycle.
Replace readFileSync/writeFileSync/mkdirSync with async equivalents from
fs/promises in default-credential-manager.ts and usage.ts to stop blocking
the Node.js event loop during credential read/write operations.
readFrontmatterFile, readFrontmatterDir, readSummary, readPhaseFiles,
readTaskFiles, readDecisionFiles, and readPageFiles all used readFileSync
and readdirSync which block the Node.js event loop during agent completion
handling. Converted to async using readFile/readdir from fs/promises and
added await at all call sites in output-handler.ts.
Removes blocking readFileSync, writeFileSync, and mkdirSync calls from the
agent spawn hot path, replacing them with async fs/promises equivalents to
avoid stalling the Node.js event loop during credential operations.
writeInputFiles, spawnDetached, and diagnostic writes now use
fs/promises (mkdir, writeFile) instead of mkdirSync/writeFileSync.
File writes in writeInputFiles are batched with Promise.all.
openSync/closeSync for child process stdio FDs remain sync as
spawn() requires the FDs immediately.
Agents were bulk-reading all context task files (39 files) because
filenames are opaque IDs and there was no way to find phase-relevant
tasks without reading every file. Now writeInputFiles generates a
context/index.json with tasksByPhase mapping phaseId to task metadata
(file, id, name, status). Prompt updated to direct agents to read
the index first.
Agents were reading all contextFiles (38 tasks + 8 phases) upfront,
wasting massive context window. Updated INPUT_FILES and SESSION_STARTUP
prompts to clearly distinguish assignment files (read all) from context
files (read only when needed).
Pass targetId to buildChatPrompt and add <scope> block that clearly
distinguishes primary target files from context files. Context entities
may be modified when necessary (e.g. dependency links) but the agent
is instructed to focus changes on the primary target.
Each execute agent has significant startup cost (context loading, codebase
exploration, baseline tests). The previous sizing guidance (sweet spot
<150 lines, 1-3 files, merge at <20 lines) produced tasks too small to
justify that overhead.
New guidance targets cohesive feature units: 200-500 lines across 3-6
files as the sweet spot, merge threshold at <100 lines, and explicit
instruction to bundle related changes (validation + route + tests = one
task, not three).
Introduces a chat loop where users send instructions to an agent that
applies changes (create/update/delete phases, tasks, pages) and stays
alive for follow-up messages. Includes schema + migration, repository
layer, chat prompt, file-io action field extension, output handler chat
mode, revert support for deletes, tRPC procedures, events, frontend
slide-over UI with inline changeset display and revert, and docs.
Detail agents define task dependencies in YAML frontmatter but they were
silently dropped — never written to the task_dependencies table. This
caused all tasks to dispatch in parallel regardless of intended ordering,
and the frontend showed no dependency information.
- Add fileIdToDbId mapping and second-pass dependency creation in
output-handler.ts (mirrors existing phase dependency pattern)
- Add task_dependency to changeset entry entityType enum
- Add listPhaseTaskDependencies tRPC procedure for batch querying
- Wire blockedBy in PhaseDetailPanel and PhaseWithTasks from real data
- Clarify dependency semantics in detail prompt
Execution agents were spawning blind — no input files, no knowledge of
what predecessor tasks accomplished. This adds three capabilities:
1. summary column on tasks table — completeTask() reads the finishing
agent's result.message and stores it on the task record
2. dispatchNext() gathers full initiative context (initiative, phase,
sibling tasks, pages) and passes it as inputContext so agents get
.cw/input/task.md, initiative.md, phase.md, and context directories
3. context/tasks/*.md files now include the summary field in frontmatter
so dependent agents can see what prior agents accomplished
When an agent asks a question via `cw ask` targeting an idle agent,
the conversation router now auto-resumes the idle agent's session so
it can answer. Previously, questions to idle agents sat unanswered
forever because target resolution only matched running agents.
Changes:
- Add `resumeForConversation()` to AgentManager interface and implement
on MultiProviderAgentManager (mirrors resumeForCommit pattern)
- Relax createConversation target resolution: prefer running, fall back
to idle (was running-only)
- Trigger auto-resume after conversation creation for idle targets
- Add concurrency lock (conversationResumeLocks Set) to prevent
double-resume race conditions
Planning modes (plan, refine) get a minimal block with just cw ask
syntax. Execution modes get the full protocol: commands table, shell
recipe for listener lifecycle, targeting guidance, when/when-not
decision criteria, good/bad examples, and answering guidelines.
Architect agents (discuss, plan, detail, refine) were producing generic
analysis disconnected from the actual codebase. They had full tool access
in their worktrees but were never instructed to explore the code.
- Add CODEBASE_EXPLORATION shared constant: read project docs, explore
structure, check existing patterns, use subagents for parallel exploration
- Inject into all 4 architect prompts after INPUT_FILES
- Strengthen discuss prompt: analysis method references codebase, examples
cite specific paths, definition_of_done requires codebase references
- Fix spawnArchitectDiscuss to pass full context (pages/phases/tasks) via
gatherInitiativeContext() — was only passing bare initiative metadata
- Update docs/agent.md with new tag ordering and shared block table
Move src/ → apps/server/ and packages/web/ → apps/web/ to adopt
standard monorepo conventions (apps/ for runnable apps, packages/
for reusable libraries). Update all config files, shared package
imports, test fixtures, and documentation to reflect new paths.
Key fixes:
- Update workspace config to ["apps/*", "packages/*"]
- Update tsconfig.json rootDir/include for apps/server/
- Add apps/web/** to vitest exclude list
- Update drizzle.config.ts schema path
- Fix ensure-schema.ts migration path detection (3 levels up in dev,
2 levels up in dist)
- Fix tests/integration/cli-server.test.ts import paths
- Update packages/shared imports to apps/server/ paths
- Update all docs/ files with new paths
