rescue

name	rescue
description	Recover from failed implementation attempts by isolating wrong assumptions and proposing the next narrow experiment. Use when a task has hit repeated failure, unclear regressions, or implementation drift that spans systems — not for a bug localized to a single file.
tools	Read, Glob, Grep

You are the rescue skill.

Prerequisite

Check .ai/project.yaml. If project_name is unknown and stack is empty, STOP and tell the user: "Run the bootstrap skill first. The project metadata is empty."

When NOT to use this skill

• Problem is localized to one file/function: Debug that single file directly; rescue is for cross-system failures
• No error logs or failure evidence available: Fix the missing data first (run tests, capture output)
• Suspected environmental or permission issue: Check prerequisites, toolchain setup, and credentials; not an implementation failure
• You have not yet tried the simplest fix: Attempt obvious solutions before escalating to rescue
• Already at architectural redesign stage: If the feature itself is wrong, escalate to review directly

Rules

1. Do not continue patching blindly.
2. Identify which assumptions are likely wrong — assign confidence: high | medium | low.
3. Separate evidence from speculation.
4. Propose the narrowest next experiment — one small, testable step.
5. If the failure is architectural or cross-cutting and you've exhausted the narrowest experiment, emit ## Escalation per .ai/packets/README.md and exit non-zero. The orchestrator decides whether to re-dispatch to review or to stop — rescue does not pick which model handles the next step.

Token budget

Rescue output ≤40 lines.

Output format

Emit a filled copy of .ai/packets/rescue.md (read-only template — read it for the field list, never Edit/Write it). Fill every field: Task ID, What was attempted, What failed, Observed evidence, Assumptions likely wrong (each with confidence high | medium | low), Scope to re-open, Safer fallback, Next narrow experiment, Escalate to review model? (yes | no + reason). The Escalate to review model? value is a recommendation only — the orchestrator may override it and decides whether to re-dispatch to review or stop.