auto-healing

name	auto-healing
description	Automated fix loop for failed Quality/Reviewer checks. Parses failures, builds escalating prompts, dispatches through the active harness up to 3 times.
version	1.0.0
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["auto-healing","retry","fix-loop","quality","testing","escalation"],"related_skills":["implementer","quality","reviewer","systematic-debugger"]}}

Auto-Healing

Automated retry loop when Quality or Reviewer checks fail. Replaces manual re-dispatch.

When to trigger

After ANY of these:

• Test suite fails after a coding engine dispatch
• Linter/formatter check fails (ruff, eslint, flake8)
• Type checker fails (tsc, mypy, pyright)
• Build fails (go build, cargo build)

Do NOT trigger for:

• Architecture or design review feedback (those need human judgment)
• Security review findings (those need security skill review)
• First-time implementation (use the coding engine directly)

Dispatch

This runs full check-fix-recheck cycles and routinely exceeds a blocking terminal timeout — run it in the background and act on the completion notification:

terminal(
    command="python3 ~/.hermes-coder/scripts/auto_healer.py --repo '<project-dir>' --check '<test command>' --engine claude-code --tier <tier> --json",
    workdir="~/.hermes-coder",
    background=true,
    notify_on_complete=true
)

notify_on_complete=true is mandatory — without it the run is silent and you end up in a poll-and-narrate loop that wastes turns and spams chat. Launch, END YOUR TURN, and act on the single completion notification. Do not poll process in a loop; do not narrate waits.

--engine is always claude-code for fix passes. Never pass a model ID from memory — models come only from coding.* in config.yaml. If the claude dispatch fails, diagnose the harness; do not downgrade code fixes to a Gemini engine.

--tier is the tier the failing code was implemented at (the same tier triage set for the dispatch: standard, elevated, …). Early heal attempts run at that tier — healing opus-authored code on the standard (sonnet) tier wastes attempts — and the final attempt always bumps to premium automatically. Omitting it defaults to standard, which is only right for XS/S work.

The --check command is the exact command that failed:

• pytest: "pytest -x" or "pytest tests/test_specific.py -x"
• ruff: "ruff check ."
• eslint: "npx eslint src/"
• tsc: "npx tsc --noEmit"
• go: "go vet ./..."

Reading the output

JSON output (--json flag):

• status: "clean" (checks already passing), "healed" (fixed), or "escalated" (all attempts failed)
• attempts: array of attempt details (prompt summary, output, success)
• remaining_failures: structured list of still-broken items (only when escalated)
• escalation_reason: why it gave up (only when escalated)

Escalation behavior

The auto-healer uses 3-tier escalation matching the implementer skill's framework:

1. Attempt 1 (targeted): Fix only the specific issues listed
2. Attempt 2 (context-enriched): Adds file context, prior attempt analysis
3. Attempt 3 (simplified): Breaks remaining failures into single-file fixes

If all 3 attempts fail, the status is "escalated". At that point:

• Report the structured findings to the user
• Include the remaining failures and what was tried
• Do NOT attempt more fixes without user guidance

Exit codes

Code	Meaning
0	Healed or already clean
1	Escalated (all attempts failed)
2	Invalid arguments
3	Check command itself errored (infrastructure problem)