arbazkhan971/predict

Activate When

• /godmode:predict, "will this work?", "evaluate"
• "risk assessment", "is this a good idea?"
• Spec exists and user wants confidence check
• Called by /godmode:plan when >10 tasks detected

Workflow

1. Define

cat .godmode/spec.md
git ls-files | head -50

IF no spec: ask user for description or run /godmode:think first.

2. Scan

Read files referenced in spec. Identify dependencies, existing patterns, API contracts, test coverage.

3. Dispatch 3 Personas (parallel, independent)

• Technical Architect: scalability, data model, API contracts, error handling, patterns, performance
• Security Researcher: attack surface, data exposure, auth/authz, injection, supply chain, secrets
• Release Lead: shipping risk, rollback plan, monitoring, deployment, resource limits, failures

Each persona outputs (mandatory fields):

VERDICT: YES | REVISE | NO
CONFIDENCE: 1-10 (integer)
RISKS:
  - {description} @ {file:line} -- SEVERITY: {level}
    MITIGATION: {concrete code/arch change}
PRAISE: {non-obvious positive only, max 2}

IF risk missing file:line: retry 2x, then mark "incomplete" (excluded from gate). IF praise is vague ("looks clean"): remove it.

4. Collect

Every validated risk has file:line + mitigation. Incomplete findings logged but excluded from gate.

5. Synthesize (Meta-Expert)

• Deduplicate: same file:line -> merge, keep highest
• Classify: blockers (NO/critical), warnings (REVISE/high), notes (medium/low)
• Conflicts: report both views, do not average

6. Gate (Pure Voting)

3 YES          -> PROCEED -> /godmode:plan
1-2 REVISE, 0 NO -> REVISE -> /godmode:think
any NO         -> RETHINK -> /godmode:think

IF unanimous: flag potential groupthink. IF confidence stddev > 2.5: flag high variance.

7. Report

Print findings table, all validated risks (sorted by severity), incomplete findings, gate result.

8. Log

Append .godmode/predict-results.tsv: one row per persona with timestamp, feature, persona, verdict, confidence, risk_count, top_risk, mitigation, gate.

<!-- tier-3 -->

Quality Targets

• Model accuracy target: >85% on validation set
• Prediction latency: <100ms p95
• Target: <500MB model artifact size

Hard Rules

1. Every risk: file:line + concrete mitigation. Vague -> 2 retries -> mark incomplete.
2. Gate is pure voting: PROCEED=3Y, REVISE=1-2R+0N, RETHINK=any NO.
3. NEVER gate on incomplete findings.
4. Each persona evaluates from own domain only.
5. Praise must be non-obvious and specific.

TSV Logging

Append .godmode/predict-results.tsv:

timestamp	feature	persona	verdict	confidence	risk_count	top_risk	mitigation	gate

Keep/Discard

KEEP if: finding has file:line + mitigation
  + at least 1 persona rates exploitable.
DISCARD if: no file:line after 2 retries
  OR all personas rate not exploitable.

Stop Conditions

STOP when FIRST of:
  - All 3 personas produced verdicts + gate computed
  - Max 2 retries per vague finding exhausted
  - Re-evaluation produces 0 new findings

Autonomous Operation

On failure: git reset --hard HEAD~1. Never pause.

Error Recovery

| Failure | Action | |--|--| | No spec exists | Ask user or run /godmode:think | | Vague findings | Retry 2x, then mark incomplete | | Unanimous verdict | Flag groupthink, still proceed | | High variance (>2.5) | Report each persona reasoning |