sequential

What This Does

Runs a series of stages in a defined order, where each stage writes its output to .agent-pipeline/ and the next stage reads that output as context. This is the standard execution mode for the implement-test-review pipeline.

Usage

/sequential implement test review         # Run three stages in order
/sequential implement test review eval docs  # Full pipeline
/sequential --from test                   # Resume pipeline from the test stage

How It Works

1. Parse stages. Extract the ordered list of stages from the input. Valid stage names are:

   | Stage | Agent | Artefact | |------------|--------------------------|---------------------------------| | implement | implementer | .agent-pipeline/implement.md | | test | tester | .agent-pipeline/test.md | | review | reviewer | .agent-pipeline/review.md | | eval | eval-agent | .agent-pipeline/eval.md | | security | security-auditor | .agent-pipeline/security.md | | docs | documentation-specialist | .agent-pipeline/docs.md |

2. Initialise pipeline directory. Run scripts/run-pipeline.sh init to create .agent-pipeline/ and write pipeline.json with the stage list and initial status.

3. Execute stages in order. For each stage: a. Mark the stage as in_progress in pipeline.json. b. Gather context from all previous stages' artefacts. c. Delegate to the appropriate agent via the Task tool, passing:

   • The task description
   • All previous stage artefacts as context
   • The feature file from .claude/context/features/ if applicable d. The agent writes its output to the corresponding artefact file. e. Mark the stage as completed in pipeline.json.

4. Handle failures. If a stage fails:

   • Mark it as failed in pipeline.json.
   • Stop the pipeline immediately. Do not proceed to subsequent stages.
   • Report which stage failed, the error details, and which stages were skipped.

5. Resume support. If --from <stage> is passed:

   • Skip all stages before the specified one.
   • Read existing artefacts from previous stages as context.
   • Continue execution from the specified stage.

6. Report summary:

   ## Pipeline Report
   
   **Stages:** 5
   **Completed:** 5
   **Failed:** 0
   
   | Stage     | Status    | Duration |
   |-----------|-----------|----------|
   | implement | COMPLETED | 30s      |
   | test      | COMPLETED | 15s      |
   | review    | COMPLETED | 20s      |
   | eval      | COMPLETED | 10s      |
   | docs      | COMPLETED | 12s      |
   
   **Total Duration:** 87s
   

Pipeline State

The pipeline state is tracked in .agent-pipeline/pipeline.json:

{
  "started_at": "2026-02-01T12:00:00Z",
  "stages": [
    { "name": "implement", "status": "completed", "duration_ms": 30000 },
    { "name": "test", "status": "completed", "duration_ms": 15000 },
    { "name": "review", "status": "in_progress", "duration_ms": null }
  ]
}

Error Handling

• If no stages are provided, default to the full pipeline: implement test review eval security docs.
• If an invalid stage name is given, report the error and list valid stages.
• If .agent-pipeline/ already exists and --from is not used, prompt whether to overwrite or resume.
• On failure, the pipeline can be resumed with /sequential --from <stage>.