care

---

description: Apply the FORGE C-A-R-E workflow (Context → Analyze → Respond → Evaluate) to a repair or optimization task argument-hint: <bug or perf issue>

You are operating under the FORGE framework's C-A-R-E workflow for REPAIR or OPTIMIZATION problems.

Setup (read these before starting)

1. Workflow definition: workflows/care.md
2. Master spec sections for protocol context: FORGE_MASTER.md — sections 5 (Confidence), 6 (Assumptions), 7 (Meta-Prompting), 8 (Verification), 12 (Operating Principles)

Phase 0 — Prior learnings retrieval

If learnings/_index.jsonl exists, search for matching prior REPAIR/OPTIMIZATION entries — failure patterns are especially valuable here. Surface top 3 before Phase 1 per core/omega_retrieval.md. Otherwise, note: "No learnings store yet — proceeding without prior context."

Phases 1-4 (iterative loop)

Apply C-A-R-E. Honor:

• Confidence Protocol on root-cause hypotheses and the proposed fix
• Track all assumptions; do not proceed past Analyze with unvalidated high-risk assumptions
• Iteration cap: default 3, extend to 5 only if quality is improving each pass; if plateaued after 3, escalate to R-I-S-E and re-classify
• Self-critique before Evaluate; verify no regressions before declaring complete

Closing emissions

After resolution:

1. Telemetry — Append one JSON line to telemetry/sessions.jsonl per telemetry/schema.json. Self-report honestly:

• task_id (UUID4), timestamp (ISO 8601 UTC), workflow: "care"
• classification: category (REPAIR or OPTIMIZATION), complexity, domain, confidence
• gates_passed / gates_failed against the gate names in schema.json
• assumptions_count, iterations (count actual loops — this matters for C-A-R-E), outcome
• prompt_hash: SHA-256 of $ARGUMENTS, first 16 hex chars
• raw_prompt: $ARGUMENTS only if FORGE_TELEMETRY_RAW=1, else null
• notes: brief commentary — for C-A-R-E especially, what failed during iteration is most valuable

Skip emission entirely if FORGE_TELEMETRY_DISABLED=1.

2. Learnings — Append a learning entry capturing what failed during iteration; failure patterns generalize better than success patterns. Skip if nothing non-obvious surfaced.

---

User task

$ARGUMENTS