* fix(team): resolve stuck "reconciling" state and skip resume when teammates never spawned
Addresses #54.
When a team launch fails to bootstrap teammates, the team gets stuck showing
"Last launch is still reconciling" indefinitely, and retrying with --resume
reconnects the lead but does not re-spawn the dead teammates. The only
workaround was enabling "Clear context (fresh session)", which loses the
lead's prior conversation context.
Two root causes addressed:
1. createPersistedLaunchSnapshot counted members still in 'starting' state
(agentToolAccepted=false) as 'pending' regardless of launchPhase. When
launchPhase was 'finished' with never-spawned members, the aggregate
state stayed as 'partial_pending' forever, rendered as "still reconciling".
Fix: when launchPhase != 'active', promote such members to
'failed_to_start' so the aggregate becomes 'partial_failure'
("Launch failed partway"), which correctly signals a terminal state.
2. TeamProvisioningService._launchTeamInner always used --resume when a
previous leadSessionId existed, even if the previous launch had no
teammates successfully spawned. The CLI's deterministic reconnect path
restores lead context but does not re-spawn dead teammates, so the team
stays broken across relaunches. Fix: before adding --resume, read the
persisted launch state. If every expected teammate is 'starting' (never
spawned) or 'failed_to_start', skip --resume so the CLI performs a full
fresh bootstrap that spawns all teammates.
Verified manually on Linux: a team stuck in "still reconciling" correctly
transitions to "failed partway" after the first fix, and the next Launch
(without "Clear context") fully bootstraps and brings teammates online.
* fix(team): narrow skip resume to never-spawned teammates
---------
Co-authored-by: 777genius <quantjumppro@gmail.com>
P1: Poller no longer overwrites nextCursor/hasMore — those belong
to the "Load older" flow. Both poller and loadOlder now dedup
messages by messageId or timestamp+from fingerprint.
P1: Cursor is now compound (timestamp|messageId) with stable
tie-breaking sort. Messages sharing the same timestamp at page
boundaries are no longer lost.
P2: getMessagesPage now applies the same enrichment as getTeamData:
leadSessionId propagation and slash-command-result annotation.
P3: Added 3 tests for getMessagesPage covering pagination, cursor
stability with same-timestamp messages, and slash command annotation.
Two different interval sets with the same length would produce
the same cache key, returning stale results. Serialize each
interval's startedAt~completedAt into the key.
capTimerMap was cancelling oldest pending throttle/debounce timers
when the Map exceeded a size limit. Since the schedulers don't
replay dropped keys, this silently lost refresh callbacks — leaving
session/team state stale after high-volume file change events.
These Maps are self-cleaning (entries delete themselves when their
callback fires) and hold ~100 bytes per entry. Even 200 entries
is negligible memory. Removing the cap fixes the data freshness
issue without any memory concern.
Returning messages: [] broke the slash command annotation test and
any code relying on getTeamData.messages (notifications, dedup).
Keep a small batch (50 newest) in getTeamData for compatibility.
Full message history is available via getMessagesPage() API.
New cursor-based IPC endpoint for the messages timeline panel:
- team:getMessagesPage(teamName, { beforeTimestamp?, limit? })
returns { messages, nextCursor, hasMore }
- Cursor is timestamp-based — stable under live message insertion
- Default page size 50, max 200
getTeamData no longer includes messages in its response, eliminating
the ~1MB messages payload from every team refresh. Messages are now
fetched independently by MessagesPanel.
MessagesPanel changes:
- Fetches initial page on mount via getMessagesPage API
- Auto-refreshes newest page every 5s when team is alive
- "Load older messages" button for pagination
- Falls back to prop messages if API fails (graceful degradation)
The TeamDataWorkerClient logged console.warn when the worker file
was not found, which is expected during tests (no build output).
The test setup treats unexpected warnings as failures.
Downgrade to logger.debug for the "not found" message and remove
the eager warning from resolveWorkerPath().
- Move endRefreshing() outside the !cancelled guard in finally block
so it always runs even when the effect is torn down mid-refresh.
- Only call load() when isTabActive or on first load — prevents
unnecessary fetches when a hidden tab's effect re-runs.
Critical:
- findLogsForTask: call worker directly instead of wrapping in
wrapTeamHandler, so failures propagate to catch and trigger the
main-thread fallback correctly.
- TeamDataWorkerClient: scope error/exit handlers to the specific
worker instance to prevent a stale worker's exit event from
rejecting pending requests on a replacement worker (race condition).
Major:
- TeamDataWorkerClient: validate teamName and taskId before forwarding
to worker thread (input sanitization).
- team-data-worker: include status, since, and intervals length in
cache key to prevent stale results after filter changes.
- team-data-worker: move logsInFlight.delete() into .finally() so
rejected lookups don't poison the in-flight map permanently.
- MemberLogsTab: reset refreshCountRef and refreshing state in effect
cleanup to prevent the refresh indicator from latching on permanently
when the effect tears down mid-refresh (e.g. on tab switch).
The user inbox (user.json) contains real teammate-to-user messages
generated by Claude Code CLI. Filtering it as a system inbox was
incorrect — it broke message aggregation for user-directed messages.
Only the broadcast inbox (*.json) needs to be excluded since '*'
is not a valid member name and causes a phantom member in the UI.
The broadcast inbox file (inboxes/*.json) was being parsed by
listInboxNames() as a member named "*", which appeared in the UI
as a phantom team member. Since "*" fails the MEMBER_NAME_PATTERN
validation, it could not be removed through the UI.
Filter system inbox names (*, user) from listInboxNames() so they
are not treated as real team members.
Main process — worker thread for team data:
- New team-data-worker thread handles getTeamData and findLogsForTask,
isolating heavy file I/O (scanning 300+ subagent JSONL files) from
Electron's main event loop. getTeamData dropped from ~2000ms on the
main thread to ~110ms via the worker.
- Worker-side dedup and 10s result cache for findLogsForTask prevents
redundant scans when the same task is queried multiple times.
- Discovery cache TTL raised from 5s to 30s — avoids re-scanning the
entire project directory on every call.
- Message cap at 200 in TeamDataService to keep IPC payloads under 1MB
(was sending 2200+ messages / ~3MB, stalling Chromium IPC serialization).
- IPC handlers fall back to main-thread execution if the worker is
unavailable (graceful degradation).
Renderer — useShallow and memoization (55 files):
- Added useShallow to store selectors across 55 renderer files. Batched
individual useStore() calls (e.g. 17 calls in ExtensionStoreView,
10 in ConnectionSection) into single useShallow selectors, cutting
unnecessary re-render checks on every store update.
- MemberLogsTab: three 5-second polling intervals now pause when the
parent tab is hidden (display:none). Previously 5 hidden tabs × 3
intervals = 15 polling timers firing continuously.
- KanbanColumn wrapped in React.memo to skip re-renders when props
haven't changed.
- MemberList: memoized activeMembers/removedMembers/colorMap; replaced
O(n×m) per-member task scan with a pre-computed reviewer map.
- Bounded timer Maps in store initialization to prevent unbounded growth
of debounce/throttle tracking maps during long sessions.
- Eliminated the enrichMemberBranches method from TeamDataService to simplify member branch enrichment logic.
- Updated TeamDetailView to utilize live branch tracking for both lead and member worktrees, enhancing the accuracy of displayed member branches.
- Adjusted various references to ensure membersWithLiveBranches is used consistently across the component.
- Enhanced text processing to clean up raw task ID hashes and replace pipe separators with dashes.
- Ensured consistent formatting for comments and inbox messages.
- Cross-team messages now show ghost nodes (dashed hexagons) for external teams
- Ghost nodes have purple color, link icon, and connect to lead via message edge
- Particles flow between ghost node and lead with cross-team message labels
- Cross-team popover shows external team name
- Task click opens full KanbanTaskCard with glow effects and action buttons
- All kanban task actions wired through CustomEvent to TeamDetailView
- Updated task opacity logic to simplify conditions.
- Added comment count and unread count badges to task pills for better visibility.
- Improved layout for unassigned tasks, including a section header and overflow badge.
- Enhanced task interaction by restricting drag functionality to member and lead nodes only.
- Introduced new task action event listeners for better task management in the UI.
- Preserved known task change presence across refreshes to maintain state consistency.
seedTeammateOperationalPermissionRules already pre-writes MCP tool
rules to settings.local.json before spawning the CLI. But standard
file tools (Write, Edit, NotebookEdit) were missing.
FACT: Teammates requesting Write get setMode: acceptEdits suggestion
but we can't change subprocess session mode. Pre-seeding these tools
as allow rules prevents the permission prompt entirely.
Belt-and-suspenders approach:
1. Settings file: handles all FUTURE calls (teammate finds rule on retry)
2. control_response via stdin: may unblock CURRENT waiting prompt
(now includes updatedInput: {} which was the previous ZodError fix)
Without #2, approved teammates stay stuck until team restart because
the CLI doesn't hot-reload settings.local.json for pending prompts.
FACT: Write/Edit permission_requests have permission_suggestions with
type "setMode" (not "addRules"): { type: "setMode", mode: "acceptEdits" }
Our code only handled "addRules", so Write/Edit approvals were no-ops.
Translate setMode suggestions to settings rules:
- acceptEdits → add Edit, Write, NotebookEdit to allow list
- bypassPermissions → add all common tools to allow list
- Replaced inline drawing logic for task comments with a new `drawCommentBubble` function to enhance readability and maintainability.
- The new function encapsulates the drawing of a speech-bubble icon, including the rounded rectangle body, tail, and inner dots to suggest text.
Particle direction:
- Added `reverse` flag to GraphParticle — when true, particle flies
from target → source (reverse of edge direction)
- Messages FROM teammate TO lead now fly member→lead (was lead→member)
- draw-particles.ts swaps from/to nodes when reverse=true
Reverted system message filter:
- Removed #isSystemMessage — all messages shown as particles again
(user wants to see idle_notification etc.)
- Added mcpToolCatalog to the agent-teams-controller, exporting new types and constants for MCP tool groups and names.
- Updated tools registration to utilize AGENT_TEAMS_MCP_TOOL_GROUPS for streamlined tool management.
- Enhanced tests to validate the new operational permissions and ensure correct tool registration behavior.
When user approves any mcp__agent-teams__* tool, also add all other
agent-teams tools to settings.local.json preemptively. This prevents
teammates from getting stuck on subsequent tool calls (task_get,
task_start, task_complete, etc.) since each generates a separate
permission_request and the teammate blocks until resolved.
FACT: Settings file approach only prevents FUTURE blocks, not current
ones. Pre-adding all tools on first approval covers the common case.
Remove tool description from permission_request noise label - tool name
alone is clear enough. MCP tools have long technical descriptions that
flood the Messages panel.
