fastForwardBranch used git merge --ff-only which fails when the clone
has uncommitted files. This caused the ff to be silently skipped, the
merge to proceed on stale main, and the push to fail (non-fast-forward).
Switched to update-ref which only moves the branch pointer without
touching the working tree.
- Add `errands` table to schema with status enum and relations to agents/projects
- Add `errand` mode to agents.mode enum
- Add push rollback in orchestrator: if push fails after merge, reset to previousRef to preserve the review diff
- Extend MergeResult with previousRef for rollback support; update branch-manager and simple-git-branch-manager
- Add orchestrator tests for push rollback behaviour
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When merge_and_push failed at the push step, the local defaultBranch ref
was left pointing at the merge commit. This made the three-dot diff
(defaultBranch...initiativeBranch) return empty because main already
contained all changes — causing the review tab to show "no changes."
Now mergeBranch returns the previous ref, and approveInitiative restores
it on push failure. Also repaired the corrupted clone state.
Three bugs causing empty phase diffs when server restarts during agent
execution:
1. Startup ordering race: reconcileAfterRestart() emitted agent:stopped
before orchestrator registered listeners — events lost. Moved
reconciliation to after orchestrator.start().
2. Stuck in_progress tasks: recoverDispatchQueues() only re-queued
pending tasks. Added recovery for in_progress tasks whose agents
are dead (not running/waiting_for_input).
3. Branch force-reset destroys work: git branch -f wiped commits when
a second agent was dispatched for the same task. Now checks if the
branch has commits beyond baseBranch before resetting.
Also adds:
- agent:crashed handler with auto-retry (MAX_TASK_RETRIES=3)
- retryCount column on tasks table + migration
- retryCount reset on manual retryBlockedTask()
The sticky ReviewHeader had shadow-[0_-50px_0_0_hsl(var(--background))]
which painted a 50px background-color rectangle upward, overlapping the
tab navigation bar (only ~12px away). The header's bg-card is already
opaque, making the shadow unnecessary for scroll coverage.
Two additional fixes to prevent agents from writing .cw/output/ in the
wrong location:
1. Always create .cw/output/ at the agent workdir root during spawn,
even when no inputContext is provided. This gives the agent a visible
anchor directory so it doesn't create one inside a project subdir.
2. Add absolute output path instruction to the workspace layout prompt
for multi-project agents, explicitly telling them to write .cw/output/
relative to the workdir root, not their current cd location.
Instead of showing a manual "Re-check Mergeability" button after the
conflict agent finishes, auto-dismiss the agent and trigger mergeability
re-check immediately when the state transitions to completed.
Throws if branch and baseBranch are identical, preventing
git branch -f from force-resetting shared branches (like
the initiative branch) when accidentally passed as both.
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
approveInitiative was merging and pushing with a stale local
defaultBranch, causing "rejected (fetch first)" when origin/main
had advanced since the last project sync. Now fetches remote and
fast-forwards the target branch before merging.
Add update method to ReviewCommentRepository, updateReviewComment tRPC
procedure, and inline edit UI in CommentThread. Edit button appears on
user-authored comments (not agent comments) when unresolved. Uses the
existing CommentForm with a new initialValue prop.
overflow-hidden creates a scroll container that breaks sticky positioning
for file headers. overflow-clip provides the same visual clipping for
rounded corners without affecting the scroll/sticky context.
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.
Sets --review-header-h on the card wrapper from measured header height.
FileCard reads it for sticky top offset so file headers dock just below
the review header instead of overlapping it.
simple-git's .raw() resolves successfully even on exit code 1,
returning stdout content. git merge-tree --write-tree outputs
CONFLICT markers to stdout (not stderr), so the catch block
never fired and conflicts were reported as clean merges.
Removed wrapper divs that broke sticky positioning. ReviewHeader now
accepts a ref prop directly, with sticky top-0 on its own root element.
Card wrapper restored as the tall containing block so both header and
sidebar have room to stick within it.
merge_requires_approval (initiatives) and requires_approval (tasks)
were removed from schema.ts in the task-approval removal but left in
the DB because 0030 assumed SQLite couldn't DROP COLUMN. SQLite 3.35+
supports it. These orphaned columns caused the old stale-build approval
code path to silently set detail tasks to pending_approval, stranding
them and blocking phase completion.
Moved card border/rounding onto the sticky header wrapper itself so it
scrolls flush to top-0 with no gap. The body grid gets its own border-x
and border-b to preserve the card appearance. ResizeObserver now measures
border-box size for accurate sidebar offset.
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.
Removed overflow-hidden from grid container and changed sidebar to
sticky positioning with viewport-relative max-height. Sidebar content
scrolls independently while staying visible during diff navigation.
Adds data-comment-id attributes to comment thread rows so clicking a
discussion in the sidebar scrolls directly to the comment, not just the
file card. Includes a brief ring highlight on the target row.
handleTaskCompleted and handlePhaseAllTasksDone both bailed early when
initiative had no branch, silently skipping phase status transitions.
Also, merge failures would skip the phase completion check entirely.
- Decouple phase completion check from branch existence
- Wrap merge in try/catch so phase check runs even if merge fails
- Route updateTaskStatus through dispatchManager.completeTask when
completing, so the task:completed event fires for orchestration
Discussions section was showing only aggregate counts (total/resolved/unresolved)
with no way to see or navigate to individual threads. Now lists each root comment
with file:line location, body preview, resolved status, and reply count. Clicking
a discussion scrolls to its file in the diff viewer.
Use a temp branch + update-ref to avoid "already checked out" error
when merging into the default branch. Also show actual branch name
in the Merge & Push button instead of "Default".
When re-dispatching tasks, the branch from a previous run may still
exist. Instead of failing with "a branch named X already exists",
reset the existing branch to the base and check it out.
Snapshots were stale since migration 0008. Generated a schema-derived
snapshot at 0032 so drizzle-kit generate works again (zero diff on
current schema.ts). Also fixed migration 0032 to use statement-breakpoint
separator required by better-sqlite3.
- Added 0032_snapshot.json derived from current schema.ts
- Fixed 0032 SQL to use --> statement-breakpoint between statements
- Updated CLAUDE.md and database-migrations.md with correct workflow
The migration file existed but wasn't in _journal.json, so drizzle-kit's
migrator never applied it. Adds the journal entry for 0032_add_comment_threading.
Critical: review/merge tasks hit an early return in handleTaskCompleted()
that skipped the phase completion check, leaving phases stuck in
in_progress forever. Changed to an if-block wrapping only the merge step.
Also adds requestChangesOnInitiative() which creates/reuses a
"Finalization" phase for initiative-level review feedback, with dedup
guards for both phase and initiative request-changes flows.
Path-prefix routing (`localhost:9100/<id>/`) broke SPAs because absolute
asset paths (`/assets/index.js`) didn't match the `handle_path /<id>/*`
route. Subdomain routing (`<id>.localhost:9100/`) resolves this since
all paths are relative to the root. Chrome/Firefox resolve *.localhost
to 127.0.0.1 natively — no DNS setup needed.
The Caddyfile was using the host port (e.g., 9100) as the Caddy listen
address, but Docker maps host:9100 → container:80. Caddy inside the
container was listening on 9100 while Docker only forwarded to port 80,
causing all health checks to fail with "connection reset by peer".
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.
httpBatchLink batches polling queries behind the long-running
startPreview mutation, so refetchInterval never fires independently.
Replace polling with preview: event invalidation via the existing
useLiveUpdates SSE subscription — preview:building/ready/stopped/failed
events now trigger listPreviews and getPreviewStatus invalidation.
The startPreview mutation blocks server-side through the entire Docker
build + health check + seed cycle. isPending was checked first, so even
after polling detected containers running, the UI stayed on "Building..."
until the full mutation resolved. Now polled status (running/failed)
takes priority, falling back to isPending only when no status exists.
Two fixes:
- Call previewsQuery.refetch() in startPreview.onSuccess so the UI
transitions from "building" to the preview link without a page refresh.
- Switch from subdomain routing (*.localhost) to path-based routing
(localhost:<port>/<id>/) since macOS doesn't resolve wildcard
localhost subdomains.
Docker compose requires project names to be lowercase alphanumeric
with hyphens/underscores only. The default nanoid alphabet includes
uppercase and special characters, causing build failures.
Extract PreviewControls into shared component and wire up preview
start/stop to InitiativeReview header alongside Push Branch and
Merge & Push to Default buttons.
Same orphaned-changeset pattern as deletePhase: manually deleting all
tasks from a detail changeset now marks it reverted. Also added
deleteTask to the invalidation map (was missing entirely).
The allFiles useMemo was declared after two early-return branches. Approving
the last phase empties pendingReviewPhases, triggering the early return and
causing React to see fewer hooks than the previous render.
Manually deleting phases left their parent changeset as "applied",
causing the Plan tab to show a stale "Created N phases" banner with
no phases visible.
- deletePhase now checks if all phases from a changeset are gone and
marks it reverted
- PlanSection filters out dismissed agents so dismissed banners stay
hidden
- revertChangeSet marks reverted before entity deletion to prevent
ghost state on partial failure
- deletePhase invalidation now includes listChangeSets
Both phaseQueue and taskQueue are in-memory Maps lost on restart. Now
the orchestrator's start() method scans active initiatives and:
- Re-queues approved phases into the phase dispatch queue
- Re-queues pending tasks for in_progress phases into the task dispatch queue
- Triggers a dispatch cycle if anything was recovered
This fixes stuck phases/tasks after server restarts.
