pipeline-phase

name: pipeline-phase description: Five-phase pipeline definitions for cooperations, mirroring lab_tools/runner_pipeline.py PHASES and APPROVAL_TRANSITIONS. Defines per-phase role, sandbox, artifact filename, checkpoint gate, and allowed transitions.

Source of truth: lab_tools/runner_pipeline.py § PHASES, § APPROVAL_TRANSITIONS.

Phases

Phase	Role	Default sandbox	Checkpoint gate	Artifact filename
`plan`	architect (product-planner)	read-only	`plan_check`	`plan.md`
`claude_review`	reviewer (deployment-reviewer)	read-only	`review_check`	`claude-review.md`
`plan_refine`	architect (product-planner)	read-only	`refine_check`	`refined-plan.md`
`implement`	implementer (frontend-ux \| email-integration \| ai-features)	workspace-write	(no gate; hands off to test phase)	`implementation-result.md`
`test`	test-runner	read-only on src, write on tests	(no gate; hands off to claude_final on green)	`test-report.md`
`claude_final`	reviewer (deployment-reviewer)	read-only	`final_check`	`claude-final-critique.md`

Approval transitions

Mirrors APPROVAL_TRANSITIONS in lab_tools verbatim:

plan_check:
  approve → claude_review
  revise  → plan
  stop    → null (halt)

review_check:
  approve → implement
  revise  → plan_refine
  stop    → null

refine_check:
  approve → claude_review
  revise  → plan_refine
  stop    → null

final_check:
  approve → null (done)
  revise  → implement
  stop    → null

Sandbox enforcement

read-only: agent may write only to .cooperations/evidence/<task-id>/. The handoff-validator.mjs hook + the agent frontmatter tools whitelist enforce this.
workspace-write: agent may write to src/, tests/, and .cooperations/evidence/<task-id>/. Still must not touch specs/, .env.local, node_modules/, .next/.
danger-full-access: not used in this project. Out of scope by demo-mode-parity.md rule.

Max review cycles

Default 2 (mirrors cooperations MAX_REVIEW_CYCLES). If plan_check revise fires twice without approve, the orchestrator escalates to navigator + human.

Capabilities per phase

Each phase may declare required skills (skill:<name>) or specialist agents. Defaults:

plan: skill:spec-drift-guard, skill:handoff-discipline.
claude_review: skill:spec-drift-guard, skill:security-token-review (if any src/server/security/* touched).
plan_refine: same as plan.
implement: depends on implementer; frontend-ux pulls skill:mobile-pwa-review, email-integration pulls skill:provider-adapter-design, ai-features pulls skill:ai-email-triage.
test: skill:smoke-path-mail.
claude_final: all of the above + skill:evidence-bundle.