- Include user comments as context during AI triage spec generation and spec review
- Invalidate stale spec approvals when new user comments are added after approval
- Add lastApprovedAt tracking to task metadata for approval freshness detection
- Add comprehensive tests for comment-aware triage, review, and stale approval logic
- Add changeset for published package and update README with feature documentation
- Add tokenCap to ProjectSettings with tokenCapEnabled flag, default 500K tokens
- Create TokenCapDetector class with usage tracking and cap enforcement
- Wire TokenCapDetector into executor agentWork() to auto-fail tasks exceeding cap
- Add token cap input fields to Settings Modal dashboard UI
- Add comprehensive unit tests for TokenCapDetector (174 lines)
When a configured primary/fallback model isn't found in the registry,
check if the provider has any known models and construct a model
on-the-fly using that provider as a template (mirroring the pi CLI's
buildFallbackModel logic). This lets any valid provider model ID work
(e.g. any OpenRouter model string) without requiring it to be in the
built-in or custom model list. Only throw if the provider itself is
completely unknown.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pass the user's models.json path to ModelRegistry so custom and
non-built-in models (e.g. openrouter/qwen/qwen3.6-plus:free) are
available when resolving the configured primary/fallback model.
The pi CLI was already passing this path; kb was not, causing
'model not found in registry' errors for any model not built-in.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Fix formatResetAt to compute day difference from calendar midnight boundaries instead of raw millisecond division
- Prevents off-by-one day counts that caused inconsistent formatting near 7-day boundary
- Add comprehensive unit tests for formatResetAt boundary conditions (7-day edge, today, beyond 7 days)
- Fix pre-existing scheduler test assertion to use expect.objectContaining for worktree/baseBranch fields
- Executor skips redundant moveTask call on manual retry (was causing race with scheduler)
- Scheduler now triggers scheduling on todo column transitions to pick up retried tasks
- Add tests for retry race condition fix covering both executor and scheduler paths
- Add optional chaining (?.) on task.description.slice() to prevent null-reference crash
- Fallback to '(untitled)' when both title and description are absent
- Add changeset for @fusion/engine patch
- Clear prInfo, issueInfo, and modelPresetId in unarchiveTask to avoid stale badge/model data
- Clear prInfo, issueInfo, and modelPresetId in moveToDone after successful merge
- Add regression tests for transient field clearing in both code paths
- Fix existing tests to use public API instead of private methods
- Add spawn settings (maxSpawnedAgentsPerParent, maxSpawnedAgentsGlobal) to ProjectSettings type
- Implement AgentStore access and spawned agent tracking in TaskExecutor
- Add spawn_agent tool with child agent execution in isolated worktrees
- Wire up child termination on parent session end and system prompt generation
- Add comprehensive test suite (554 lines) covering spawning, limits, termination, and error handling
- Document agent spawning architecture, IPC protocol, and usage in AGENTS.md
- When a slice is activated and has autoAdvance enabled, automatically triage all features in that slice
- Add auto-triage logic to mission-store activateSlice method
- Add comprehensive tests for auto-triage behavior covering edge cases
- Fix module import paths across mission-related files (mission.ts, extension.ts, mission-integration.test.ts, mission-routes.ts, scheduler.ts)
When tasks fail and retry, the old worktree lingered on disk and the new
worktree got a random name causing worktrunk to report it as unassigned.
Now: retry clears worktree/branch fields, removes old worktree from disk,
and createWorktree returns the actual branch name (including -2 suffixes)
so task.branch always matches reality.
Also fixes pre-existing test failures: Terminal WebSocket mock sessions
missing lastActivityAt, and TerminalModal keyboard overlap tests leaking
cached _initialViewportHeight between tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Strengthen TRIAGE_SYSTEM_PROMPT with explicit subtask guidance rules and examples
- Strengthen buildSpecificationPrompt to reinforce proper subtask decomposition
- Add undersplit detection to spec reviewer to flag oversized or overly broad tasks
- Add test assertions for improved prompts covering undersplit detection and subtask guidance
- Add createAgentWithTaskDone helper for deterministic test patterns with task_done tool
- Rewrite restart integration tests to use the new helper instead of verbose mock setup
- Document task_done testing pattern in README with code example
- Reduce test boilerplate by encapsulating agent creation and tool stubbing
- Add integration tests verifying export/import contract for merger restart scenarios
- Enhance regression coverage for the full merger import chain
- Add developer note in README on keeping runtime exports in sync with dist artifacts
- All tests passing, build verified
- Add ContextLimitDetector to detect agent loops via repeated tool call patterns
- Implement compact-and-resume strategy: summarize conversation and restart agent from current step
- Add loop recovery to StuckTaskDetector with configurable attempt tracking and retry limits
- Extend executor with automatic loop recovery on context limit detection
- Add loop recovery support to pi executor with same compact-and-resume pattern
- Add comprehensive tests for context-limit-detector, stuck-task-detector loop detection, executor, and pi recovery
- Add changeset for patch bump to @gsxdsm/fusion
- Update README with loop detection and recovery documentation
- Add scanOrphanedBranches utility to worktree-pool for detecting fusion/* branches with no matching task
- Add cleanupOrphanedBranches to SelfHealingManager with dry-run support and task re-registration
- Wire branch cleanup into deleteTask and archiveTask so branches are removed when tasks are deleted or archived
- Add comprehensive tests for scanning, cleanup, and integration with delete/archive flows
- Add mission store methods for pausing, stopping, and resuming missions with proper state transitions
- Implement feature triage flow that evaluates and classifies mission features
- Add scheduler blocked-task check to prevent scheduling when dependencies are unmet
- Create dashboard mission management UI with pause/stop/resume controls
- Add mission API routes for triage, pause, stop, and resume operations
- Add e2e tests for mission routes and unit tests for mission store and scheduler
- Add phase field to workflow step definitions (pre-merge vs post-merge) with persistence and API
- Execute pre-merge steps in executor before merge; post-merge steps in merger after successful merge
- Pre-merge failures block merge and keep task in in-review; post-merge failures are logged only
- Expose phase controls and phase-aware results in the dashboard UI
- Add changeset for the published @gsxdsm/fusion package
- Add memoryEnabled boolean setting to core types and default settings
- Add Memory section toggle to SettingsModal dashboard UI
- Integrate memoryEnabled check in executor and triage engine prompts
- Add tests for settings UI, executor prompt behavior, and triage prompt behavior
- Update SettingsModal section count test for the new Memory section
Add @mariozechner/pi-ai mock to restart.integration.test.ts and
project-runtime.test.ts to fix module linking errors. Fix call count
assertions in executor and restart tests to account for the
retry-without-task_done behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add getFusionDir() method to TaskStore for resolving agent database path
- Update in-process-runtime to use getFusionDir() instead of hardcoded path
- Update dashboard routes AgentStore initialization to use getFusionDir()
- Add changeset for published package patch bump
When the stuck task detector killed a task and immediately called
moveTask("todo"), the scheduler could re-dispatch the task before the
old execution's finally block cleared this.executing. The new execute()
call hit the guard and silently returned, stranding the task in
"in-progress" with no active session or worktree (seen on FN-810/FN-912).
Move the requeue responsibility from StuckTaskDetector.killAndRetry to
the executor's finally block, which runs after this.executing.delete().
The beforeRequeue budget check now runs before session.dispose() and its
result is passed via StuckTaskEvent.shouldRequeue → markStuckAborted().
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix resolveBaseBranch to use stored branch name and consistent fusion/ prefix
for both explicit deps and blockedBy paths (was using kb/ for blockedBy)
- Add main branch checkout verification in merger before squash merge to prevent
feature code from landing on wrong branch lineage
- Align all branch prefix references from stale kb/ to fusion/ across executor,
merger, store, and routes
- Fix executor test OOM by mocking merger fully, adding fake timers to retry
tests, and switching vitest pool to vmThreads
- Update all test assertions to use fusion/ branch prefix
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix branch name assertion in executor test: use kb/ prefix instead of fusion/
- Add comprehensive tests for stale worktree and branch reference recovery (267 lines)
- Document stale worktree/branch recovery procedure in README
Instead of immediately failing when an agent finishes without calling
task_done, spawn a fresh session with a recovery prompt that asks the
agent to review the worktree state and complete the task. Only fail
if the retry also doesn't produce a task_done call.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The merger was using `git diff HEAD..branch --stat` which includes
artifacts from other tasks when branches fork from older main commits.
Switch to `git diff $(merge-base)..branch --stat` so commit messages
only describe the branch's own changes.
Also surface the "files changed" button on done task cards using
mergeDetails, opening the same ChangedFilesModal with commit-backed
diffs (matching the Changes tab in the task modal).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add explicit no-progress loop detection signals to executor agent sessions
- Enhance stuck-task detector with configurable strategies (timeout, no-progress, combined)
- Add comprehensive test coverage for stuck-task detector (371 lines of tests)
- Remove modelFilter test utilities (198 lines of dead test code)
- Simplify modelFilter.ts by removing unused filtering logic
- Clean up dashboard server startup and minor README fix
- Extend WorkflowStep type with optional modelProvider, modelId, validatorModelProvider, validatorModelId fields
- Add API endpoints (PATCH /api/workflow-steps/:id) to update model overrides on workflow steps
- Update executor to use per-step model overrides when running workflow step agents
- Add AgentDetailView support for displaying workflow step model configuration
- Add store and executor tests for new model override behavior
- Auto-resume in-progress tasks that were paused and then unpaused, picking up from their last step
- Add event-driven scheduler trigger on task unpause so resumed tasks are immediately scheduled
- Update Store and TaskStore with unpause handling, type changes for resume tracking
- Refactor executor, merger, and scheduler tests for improved reliability
- Remove dead code: mission-interview, pr-comment-handler, taskStuck util, OpenRouter model sync changeset
- Update AGENTS.md documentation for pause/unpause behavior
- Fix executor bug where tasks paused during execution with a graceful session exit were silently dropped instead of moved back to todo
- When session.dispose() resolves the prompt without throwing, the paused-aborted flag now triggers moveTask(todo) so the scheduler can resume after unpause
- Add executor test suite covering pause via graceful exit, pause via abort error, and pause during session disposal
- Document corrected pause/unpause behavior for in-progress tasks in README and AGENTS.md
- Thread real project context (projectName, projectId) into ntfy notification deep links
- Dashboard deep-link URL now honors project context with /project/:name/task/:id pattern
- Update App.tsx routing to handle project-scoped deep link navigation
- Add project info to NtfyNotifier types and wire through notification payload
- Add changeset for published @gsxdsm/fusion package
- Add tests for notifier project context and dashboard routing
- Log the selected model provider and ID during triage so the agent log records which model was used
- Add agent-log entry capturing triage model choice for observability
- Add comprehensive tests for triage model logging behavior
- Update AgentsView component and dashboard docs for consistency
- Remove obsolete AgentsView test
