Running Execution
Execution is where Trinity's AI agents build your project. The coordinator manages the process, assigning stories to workers that execute them through the agent pipeline.
Execution is release-scoped — you must create a release and link at least one PRD before you can start. Each release gets its own coordinator, worker pool, and job queue.
Starting Execution
- Navigate to Run
- Select which release to execute (if you have multiple)
- Click Start
The run dialog lets you configure this execution:
- Story — pick a specific story or let Trinity choose the next best
- Workers — 1-5 parallel workers (each in its own git worktree)
- Max Stories — stop after N stories, or run unlimited
- Mode — Manual (override defaults) or Autopilot (auto-skip gates, make assumptions)
Manual Mode Overrides
In manual mode you can override project defaults for this run:
- Auto-clarify, Auto-create PR, Auto-merge, Squash merge, Delete branch after merge
- Auto-approve quality checkpoints — quality checkpoints run full QA but auto-approve the gate
- Checkpoint pipeline tuning: audit perspectives (2-7) and max fix loops (2-5)
Autopilot Mode
Runs without stopping. Makes reasonable assumptions for unclear requirements and auto-skips all gates.
The Coordinator
Each release gets its own coordinator instance. The coordinator:
- Scans for stories that are ready to execute
- Assigns stories to available workers
- Monitors worker health and progress
- Auto-stops after consecutive idle scans (no more runnable stories)
Coordinator States
- Running — actively scanning and assigning work
- Stopped — manually stopped or auto-stopped after idle
- Starting — initializing workers and scanning
Workers
Workers are the processes that actually execute stories. Each worker:
- Claims a job from the queue (atomic, preventing duplicates)
- Creates a git worktree for isolated code changes
- Runs the story through the 4-phase agent pipeline
- Handles PR creation and merging on success
Worker Configuration
- Max workers — how many stories can execute in parallel (configurable in project settings)
- Worker health — the coordinator monitors workers and marks stale ones as dead
Worker Status
- Idle — waiting for a story assignment
- Busy — executing a story
- Dead — crashed or timed out (cleaned up on restart)
The Agent Pipeline
Each story flows through four phases:
1. Analyst (Read-Only)
The analyst reads the codebase and plans the implementation:
- Analyzes project structure and relevant files
- Creates an implementation plan
- Detects required services
- Checks execution gates (may pause here)
This phase doesn't modify any code.
2. Implementer (Doer)
The implementer writes the code:
- Follows the analyst's plan
- Creates or modifies files
- Writes tests
- Does NOT commit (that's the Documenter's job)
3. Auditor (Simplify Loop)
The auditor reviews and improves the code through multiple passes:
- Number of passes scales with story difficulty and surface area (1-7 passes)
- Each pass simplifies, refactors, and catches issues
- Early exit if a pass finds nothing to improve
- Final pass includes a full code review and build verification
4. Documenter (Doer)
The documenter wraps up:
- Commits all changes
- Creates pull requests
- Writes learnings to the knowledge base
- Updates audit tracking if applicable
- Triggers a daily recap update
Git Worktrees
Each running story gets its own git worktree — an isolated copy of the repository. This means:
- Multiple stories can run in parallel without conflicts
- Failed stories don't affect other work
- The main branch stays clean until code is merged
Worktrees are created in ~/.trinity/projects/{slug}/ and cleaned up after merge.
Monitoring Execution
Run Page
The Run page shows:
- Active stories — what's currently executing and which pipeline phase
- Queue — stories waiting for workers
- Gates — stories paused for your input
- Completed — recently finished stories
- Failed — stories that encountered errors
Task Indicator
The sidebar task indicator shows what's happening:
- Which release is executing
- Current pipeline phase for active stories
- Gate notifications
Stopping Execution
Click Stop on the Run page to stop the coordinator. This:
- Stops scanning for new stories
- Lets currently running stories finish
- Workers complete their current job before going idle
It does NOT cancel in-progress stories. To force-stop everything, use the kill option.
Concurrent Release Execution
You can run multiple releases simultaneously. Each gets independent:
- Coordinator instance
- Worker pool
- Job queue
This is useful for parallel releases that don't depend on each other.
Retry and Recovery
If a story fails:
- Check the agent handoffs for error details
- The story can be retried (reset to pending)
- The coordinator will pick it up again on the next scan
If Trinity crashes or restarts:
- Stale execution state is automatically cleaned up
- Running coordinators are stopped
- Orphaned tasks are failed with recovery messages
- You can restart execution normally