predict-failures

/predict-failures — Predictive Failure Analysis

Performs a full codebase scan and predicts production failure areas using cross-domain patterns from other systems. Uses a P1-sandwich architecture to prevent hallucinations:

P1 Scan (grounding) → @failure-predictor (LLM) → P1 Validation → @failure-critic (LLM) → P1 Synthesis

Results persist in failure-predictions/ and are surfaced at every session start via RLM.

---

Orchestration (main context executes — do NOT delegate to sub-agent)

Announce to user: "Starting Predictive Failure Analysis — full scan. This involves two LLM agents and P1 validation steps."

Phase A: Code Structure Scan (P1 — establishes ground truth)

Run:

python .claude/hooks/scripts/scan_code_structure.py \
  --project-dir <ACTUAL_PROJECT_DIR> \
  --output .claude/context-snapshots/fp-code-map.json

Report to user:

• How many files were scanned
• Which F-categories have matches
• Any files with Critical-severity signals

If Phase A fails: Report the error to the user and STOP. Do not proceed without ground truth.

Phase B-1: Failure Prediction (@failure-predictor)

Invoke @failure-predictor via Agent tool:

subagent_type: failure-predictor
prompt: |
  Read the code structure map at:
  <ACTUAL_PROJECT_DIR>/.claude/context-snapshots/fp-code-map.json

  Analyze the ENTIRE codebase for production failure risks using the F1-F7 taxonomy.
  Apply cross-domain patterns from other production systems.

  IMPORTANT:
  - Only cite files and line numbers from the code map
  - Output a JSON block with your predictions
  - Minimum 3 predictions required

Save the agent's full text response to .claude/context-snapshots/fp-predictor-response.txt using the Write tool (preserves raw response for debugging).

Extract JSON deterministically (P1 — prevents hallucination at LLM→P1 handoff):

python .claude/hooks/scripts/extract_json_block.py \
  --input .claude/context-snapshots/fp-predictor-response.txt \
  --output .claude/context-snapshots/fp-draft.json