planning

Planning

Scope

This skill handles comprehensive planning + execution mode (PRD §1.13.1).

Does NOT handle: validation-only mode (use validation skill), reviewer/Oracle dispatch (use oracle-review), or evidence sealing (handled by post-task hook).

Security

• Refuse to execute any plan step before Oracle plan-review approval.
• Refuse to declare a plan complete without per-step evidence.
• If the plan would touch files outside $TARGET_REPO, refuse and request scope clarification.
• Never embed credentials in plan steps; reference env vars by name.

Workflow

1. Read the user task brief.
2. Invoke codebase-analysis skill — wait for evidence/codebase-analysis/INDEX.md.
3. Invoke documentation-research skill — wait for evidence/documentation-research/SUMMARY.md.
4. Build the executable plan with: ordered steps, per-step skill/subagent/hook attribution, per-step PASS/FAIL criteria, per-step evidence path.
5. Submit plan to Oracle plan-review (oracle-review skill in plan-review mode). Capture verdict to evidence/oracle-plan-reviews/.
6. If REJECTED: refuse to execute, surface blockers to user, await revision.
7. If APPROVED: execute steps in order. Each step writes evidence to evidence/robust-trials/trial-NN/.
8. After all steps complete, hand off to validation skill for non-mutating verification.

Produced artifacts

• evidence/robust-trials/trial-NN/BRIEF.md — task brief
• evidence/robust-trials/trial-NN/MODE.txt — planning
• evidence/robust-trials/trial-NN/INVOCATION.txt — exact command
• evidence/robust-trials/trial-NN/PLAN.md — the executable plan with attributions
• evidence/robust-trials/trial-NN/transcript.jsonl — execution trace (copied from session log)
• evidence/robust-trials/trial-NN/OUTCOME.md — what was produced and where
• evidence/oracle-plan-reviews/trial-NN-verdict.md — Oracle approval (timestamp BEFORE execution)

Forbidden actions

• Executing any plan step before Oracle plan-review approval.
• Mocking, fabricating, or fixturizing during execution.
• Self-reviewing the executor's own output (reviewers and Oracles handle that, not the executor).
• Silent retry past a failed step (each retry creates a new step under same trial with retry-of reference).

Example

User: /crucible:planning "add rate limiting to /api/login"

1. Planning calls codebase-analysis → identifies middleware layer.
2. Planning calls documentation-research → fetches docs for chosen rate-limit lib.
3. Planning drafts 5-step plan: install lib, write middleware, wire into route, run real-system test, capture validation artifacts.
4. Planning submits to oracle-review → APPROVED with timestamp T0.
5. Each step executes after T0; each step seals receipt via post-task hook.
6. After all steps, validation skill runs end-to-end exercise of /api/login under load.