workflow-revision-355535— agent skill

Execute revision based on review findings: $ARGUMENTS

Purpose

You are executing Stage 6 (Revision) of the development workflow. Review found issues with the implementation. Your job is to dispatch sub-agents to fix them, then re-run review until it passes — or escalate after 3 iterations.

Pre-Conditions

Before starting, verify:

1. .workflow/state.json exists and currentStage is "revision"
2. Review findings exist (either passed as argument or in stages.review.artifacts.findings)
3. The working tree has uncommitted changes from implementation (since commits happen post-review)

Process

Step 1: Classify Review Findings

Read the review findings and classify each into:

Step 2: Dispatch Fixes

For each finding that requires a sub-agent:

1. Dispatch a sub-agent with:

• The specific finding to address
• The relevant spec section
• The file(s) to modify
• Instructions to return the structured summary format (from the conductor skill)

1. Persist the returned summary to disk immediately

For findings that can be fixed inline (style/quality):

• Make the fix directly
• Note it in the revision log

For ADR conflicts:

• Present to the user with the four ADR reconciliation dispositions:

1. STILL VALID — false positive, ADR reasoning holds
2. NEEDS AMENDMENT — core decision correct, context needs updating
3. SUPERSEDED — another ADR has replaced this one
4. INVALIDATED — reasoning no longer holds, no replacement

• Wait for user decision before proceeding

Step 3: Handle Spec Updates

If any finding reveals that the spec was wrong (not the implementation), update the spec:

• If the spec assumption was based on incomplete knowledge of the codebase: update the spec to match reality. This is expected and not a failure.
• If the spec's intent was correct but the approach was wrong: update the spec with the revised approach and note what changed.
• If the ADR's reasoning is affected: escalate to user. Do not silently update the ADR.

Spec updates go in the same commit as the code fix.

Step 4: Re-Run Review

After all fixes are applied:

1. Increment stages.revision.iterations in the state file
2. Dispatch /workflows-review on the fixed code
3. Evaluate the review results:

If review passes (all claims verified, no blocking findings):

• Update state: set stages.revision.status to "completed", currentStage to "compound"

• Hand off to the conductor

If review still has findings:

• Check iteration count against maxIterations (default: 3)
• If under max: loop back to Step 1 with the new findings
• If at max: proceed to Step 5 (Escalation)

Step 5: Escalation (Max Iterations Reached)

If 3 revision iterations have not resolved all findings:

REVISION ESCALATION
====================

Iterations completed: 3/3
Remaining findings: N

Unresolved:
1. [finding]: Attempted [approach], result: [what happened]
2. [finding]: Attempted [approach], result: [what happened]

Options:
A. Accept current state (acknowledge known gaps, proceed to Compound)
B. Re-enter at Planning stage (re-plan the approach)
C. Re-enter at Dev-Docs stage (revise spec/ADR)
D. Abandon workflow (delete branch)

Recommendation: [A/B/C/D] because [reason]

Wait for user decision. Update state accordingly.

State Updates

Throughout this stage, keep the state file current:

{
  "stages": {
    "revision": {
      "status": "in_progress",
      "iterations": 1,
      "maxIterations": 3,
      "findings": [
        { "id": 1, "category": "claim_failure", "description": "...", "status": "fixed|open|escalated" }
      ]
    }
  }
}

Operational Rules

• Each fix gets its own sub-agent (1 fix = 1 commit after review passes)
• Do NOT commit during revision — commits happen only after review validates the final state
• If a fix introduces new issues, those count against the iteration budget
• Inline style fixes do not consume an iteration
• Always persist sub-agent summaries to disk before proceeding

Anti-Patterns

• Do NOT skip review after making fixes — the whole point is verification
• Do NOT increase maxIterations beyond 3 without user approval
• Do NOT fix findings by deleting the test or weakening the claim
• Do NOT modify the ADR without user approval (specs are fair game, ADRs are not)