- Add executor logic to resolve and reuse dependency worktrees for warm build caches
- Add reuseWorktree method that creates a new branch in an existing worktree
- Add findWorktreeUser helper to check if a worktree is shared across tasks
- Update executor and merger cleanup to skip worktree removal when still in use
- Add comprehensive tests for worktree reuse and conditional cleanup paths
- task_done() marks all remaining steps as done and signals completion
- No more filesystem side channel for completion detection
- Removes all .DONE references from executor and prompts
- task_list: returns all non-done tasks (ID, column, description, deps)
- task_get: returns full task detail including PROMPT.md content
- Agent calls task_list to scan for duplicates, task_get to verify
- Add duplicate check instructions to specification prompt
- Read existing tasks before writing spec to detect duplicates
- Agent outputs DUPLICATE: {id} instead of PROMPT.md when duplicate found
- Parse agent output for duplicate flag and auto-close duplicate tasks
- Log duplicate relationship before deleting the duplicate task
- Extract tool parameter schemas to module-level constants for reuse
- Add Static<typeof schema> type annotations to tool execute callbacks
- Import Static type from @mariozechner/pi-ai
- Remove inline parameter schema definitions from tool objects
- Add markdown-body base styles in dashboard styles.css
- Remove backdrop and enable scroll on task detail prompt rendering
- Add TaskDetailModal unit tests with vitest jsdom environment
- Update vitest config with globals and jsdom support
- Clean up unused test files and legacy concurrency module
Extract TypeBox schemas to module-level constants and use Static<typeof schema>
to explicitly type execute callback params, resolving all TS7006 implicit-any
errors in the four create*Tool methods.
- Create AgentSemaphore class in engine/src/concurrency.ts with configurable maxConcurrent
- Integrate semaphore into TriageProcessor and TaskExecutor to gate agent spawning
- Wire semaphore into merge path and dashboard engine components
- Add comprehensive tests for concurrency, executor, and triage integration
- Update JSDoc and types for maxConcurrent configuration option
Replaces CLI-based progress reporting (hai task update/log/create)
with custom tools registered on the worker's pi session. No PATH
dependency, no shell overhead, works in any worktree.
- task_update(step, status) — update step lifecycle
- task_log(message) — log actions and decisions
- task_create(description) — create tasks for out-of-scope work
- review_step unchanged (already a tool)
- System prompt and execution prompt updated to reference tools
- Add autoMerge boolean to Settings model with default false
- Add auto-merge toggle UI to the In Review column header
- Implement server-side auto-merge behavior in dashboard command
- Fix pre-existing build errors in types, executor, and dashboard server
- Update README with auto-merge documentation
- Add groupOverlappingFiles toggle to Settings types and store with default enabled
- Parse File Scope sections from PROMPT.md to extract per-task file lists
- Implement overlap detection in scheduler to group tasks sharing files
- Add groupOverlappingFiles toggle to the Settings UI modal
- Update README and remove obsolete triage overlap logic
Triage prompt now explicitly:
- Checks for test framework before writing spec
- Creates a prerequisite task if no test framework exists
- Adds it as a dependency on the current task
- Forbids typecheck-only or manual-only testing steps
- Add maxWorktrees field to Settings interface with default of 4
- Enforce worktree limit in Scheduler.schedule() alongside maxConcurrent
- Wire setting through dashboard startup from persisted settings
- Include maxWorktrees in /api/config response
- Add Max Worktrees input to Settings Modal UI
- Backward compatible: existing configs without maxWorktrees use default
- System prompt explains review_step tool with review level table
- Execution prompt injects real task ID in all CLI examples
- Execution prompt parses review level from PROMPT.md and tells agent exactly when to call review_step
- Removed {TASK_ID} placeholders that agent saw as literal text
- reviewer.ts: spawns separate pi agent with read-only tools
and reviewer system prompt (taskplane's review format/verdicts)
- Executor registers review_step as a custom tool on the worker session
- Worker calls review_step(step, type, step_name) at step boundaries
based on review level (0=none, 1=plan, 2=plan+code, 3=full)
- Reviewer returns APPROVE/REVISE/RETHINK with structured feedback
- REVISE feedback returned inline to worker for immediate action
- Review calls logged to task via hai task log
