Skip to content

autoplan.md

Latest commit

 

History

History
212 lines (157 loc) · 8.68 KB

autoplan.md

File metadata and controls

212 lines (157 loc) · 8.68 KB

/autoplan — Auto-Review Pipeline

GStack Intelligence

One command, fully reviewed plan out. Loads CEO, design, and engineering review skills and runs them sequentially with auto-decisions using 6 decision principles. Surfaces taste decisions at a final approval gate.

Property Value
Trigger Comment /autoplan on an issue
Browser Required No
Default State ✅ Enabled
Results Path Test plan artifact at .github-gstack-intelligence/state/results/$SLUG/, restore point saved to .github-gstack-intelligence/state/local/

How to Use

  1. Create or open an issue containing your plan (or a link to a plan file).
  2. Comment /autoplan on the issue.
  3. The skill runs the full CEO → Design → Eng review pipeline automatically.
  4. Mechanical decisions are auto-decided silently. Taste decisions and User Challenges are surfaced at a final approval gate for your input.

Best for: When you have a plan file and want the full review gauntlet without answering 15–30 intermediate questions manually.

Tip: If you want to run individual reviews interactively, use /plan-ceo-review, /plan-design-review, or /plan-eng-review separately instead.

What It Does

The 6 Decision Principles

These rules auto-answer every intermediate question:

  1. Choose completeness — Ship the whole thing. Pick the approach that covers more edge cases.
  2. Boil lakes — Fix everything in the blast radius (files modified + direct importers). Auto-approve expansions that are in blast radius AND < 1 day effort (< 5 files, no new infra).
  3. Pragmatic — If two options fix the same thing, pick the cleaner one. 5 seconds choosing, not 5 minutes.
  4. DRY — Duplicates existing functionality? Reject. Reuse what exists.
  5. Explicit over clever — 10-line obvious fix > 200-line abstraction. Pick what a new contributor reads in 30 seconds.
  6. Bias toward action — Merge > review cycles > stale deliberation. Flag concerns but don't block.

Conflict resolution (context-dependent tiebreakers):

  • CEO phase: Completeness + Boil lakes dominate.
  • Eng phase: Explicit + Pragmatic dominate.
  • Design phase: Explicit + Completeness dominate.

Decision Classification

Every auto-decision is classified:

  • Mechanical — One clearly right answer. Auto-decided silently. Examples: run evals (always yes), reduce scope on a complete plan (always no).
  • Taste — Reasonable people could disagree. Auto-decided with recommendation but surfaced at the final gate. Sources: close approaches, borderline scope, codex disagreements.
  • User Challenge — Both models agree the user's stated direction should change. NEVER auto-decided. Always presented at the final gate with: what the user said, what both models recommend, why, what context might be missing, and cost if wrong.

Phase 0: Intake + Restore Point

  • Captures a restore point of the plan file before modifying anything.
  • Reads context: CLAUDE.md, TODOS.md, git log, design docs.
  • Detects UI scope to determine whether to run the Design phase.
  • Loads skill files from disk for each review phase.

Phase 1: CEO Review (Strategy & Scope)

Follows the full plan-ceo-review methodology:

  • Premise challenge (0A–0F) — the ONE question that is NOT auto-decided.
  • All 10+ review sections at full depth.
  • Dual voices: runs both Claude subagent and Codex for independent strategic assessment.
  • Produces CEO consensus table, dream state delta, Error & Rescue Registry, Failure Modes Registry.
  • Mode: SELECTIVE EXPANSION by default.

Phase 2: Design Review (conditional)

Runs only if UI scope was detected. Follows the full plan-design-review methodology:

  • All 7 design dimensions rated 0–10.
  • Dual voices for independent design assessment.
  • Produces design litmus scorecard.

Phase 3: Engineering Review + Dual Voices

Follows the full plan-eng-review methodology:

  • Scope challenge with actual code analysis.
  • Architecture, Code Quality, Test, and Performance reviews.
  • Dual voices for independent engineering assessment.
  • Produces ASCII dependency graph, test diagram, test plan artifact.

Decision Audit Trail

Every auto-decision is logged as a row in the plan file:

| # | Phase | Decision | Classification | Principle | Rationale | Rejected |
|---|-------|----------|---------------|-----------|-----------|----------|

Final Approval Gate

After all phases complete, surfaces:

  • All taste decisions with auto-decided recommendation and both-sides reasoning.
  • All User Challenges with full context.
  • Pre-gate verification ensures all required outputs were produced.

Cross-Phase Themes

Identifies patterns that appeared across multiple review phases (e.g., a concern raised in CEO review that also appeared in engineering review).

Example Output

## /autoplan — Fully Reviewed Plan

### Phase 1: CEO Review ✅
Mode: SELECTIVE EXPANSION
Premises: 3/3 confirmed by user
Codex: 2 concerns | Claude subagent: 3 issues
Consensus: 5/6 confirmed, 1 disagreement → surfaced at gate

### Phase 2: Design Review ✅ (UI scope detected)
Initial score: 4/10 → Final: 8/10
7 dimensions evaluated, 4 auto-fixed, 1 taste decision surfaced

### Phase 3: Engineering Review ✅
Architecture: 2 issues found, auto-resolved
Test coverage: 3 gaps identified, test plan written
Performance: 1 issue found (N+1 query), auto-fixed

### Decision Audit Trail
| # | Phase | Decision           | Classification | Principle   | Rationale                    |
|---|-------|--------------------|---------------|-------------|------------------------------|
| 1 | CEO   | Accept premise #2  | Mechanical    | P6 (action) | Evidence supports direction  |
| 2 | Eng   | Add index on users | Mechanical    | P3 (pragmatic) | Obvious perf improvement  |
| 3 | Design| Sidebar vs tabs    | Taste         | P5 (explicit)  | Both viable, different UX |

### 🚦 Final Approval Gate
**Taste Decisions (1):**
1. Navigation: sidebar (auto-decided) vs. tabs — both viable...

**User Challenges (0):**
None — models agree with user's direction.

Configuration

In config.json:

{
  "skills": {
    "autoplan": {
      "enabled": true,
      "trigger": "issue_comment"
    }
  }
}
Field Description
enabled Whether the skill is active (true/false).
trigger Event type — issue_comment means it fires when /autoplan is commented on an issue.

Requirements

  • Browser: Not required.
  • Model: Uses the model configured in config.json defaults (currently gpt-5.4). Benefits from dual-model setup (Claude + Codex) for independent outside voices.
  • Allowed tools: Bash, Read, Write, Edit, Glob, Grep, WebSearch.
  • Benefits from: Prior /office-hours output for product context.

Results

  • Plan file — Updated in-place with review findings, decision audit trail, and required outputs.
  • Restore point — Saved to .github-gstack-intelligence/state/local/projects/$SLUG/ for rollback.
  • Test plan — Written to .github-gstack-intelligence/state/results/$SLUG/.
  • CEO plan — Written to .github-gstack-intelligence/state/results/$SLUG/ceo-plans/ (for EXPANSION/SELECTIVE EXPANSION modes).
  • TODOS.md — Updated with deferred items.

Related Files

See Also

GStack Intelligence

← Back to Command Reference