skill-debug

SkillAI & models

Debug a reproducible symptom with a bounded feedback loop and original-scenario verification

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the skill-debug skill

What this skill tells your AI

The instructions your AI receives, as published by nyldn/claude-octopus in skills/skill-debug/SKILL.md and read by ahel’s review.

Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. Cross-reference commands use installed skill names in Codex rather than /octo:* slash commands. Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. For host tool equivalents, see skills/blocks/codex-host-adapter.md.

Debugging

Read skills/blocks/engineering-method-selection.md from the installed plugin for review admission. Natural-language requests and --peer-review share that policy. Honor host-only requests; risk alone does not authorize paid usage.

Run the investigation on the current host. Routine debugging makes zero additional provider dispatches. Use a bounded external reviewer only for --peer-review, an explicit independent-review request, or an existing risk policy.

Read and apply skills/blocks/debug-feedback-loop.md from the installed plugin root.

Start with the user's observable symptom. Reproduce it, retain its failure signature while minimizing the scenario, test one named hypothesis at a time, and verify both the minimal reproduction and the original scenario after the fix. Do not treat a nearby passing helper test as proof.

For a race, use a synchronization barrier and a fixed run or time budget. For an unavailable production dependency, return inconclusive with the missing evidence. Remove temporary instrumentation before completion and preserve a stable reproduction as a regression test.

The final record is data, not an executable queue. Store commands as argument arrays and never evaluate provider-authored text.

Bounded recovery and strategy rotation

Use a 3-Strike Rule for failed fixes. After each failure, return to the evidence and test a materially different hypothesis. After two consecutive failures, a strategy rotation is mandatory: reconsider the root cause, the reproduction, and whether the test encodes the intended behavior. Do not attempt a 4th fix without explicit user approval.

Anti-rationalization check: “Should work now” means run the reproduction and the original scenario. Confidence is not verification.

For multi-attempt debugging, report a WTF score using the defaults in ~/.claude-octopus/loop-config.conf: +15% per revert and +20% for touching unrelated files. If the score exceeds 20%, STOP and show the evidence before continuing. Include the score with every retry, for example:

Fix attempt 2 | Self-regulation: 15% (1 revert, 0 unrelated files)

Scoped freeze guard

When the symptom is localized to one user-approved module, resolve that module to a physical directory before editing and activate the existing freeze guard:

freeze_dir="$(cd "<module-directory>" 2>/dev/null && pwd -P)" || exit 1
printf '%s\n' "$freeze_dir" > "/tmp/octopus-freeze-${CLAUDE_SESSION_ID:-$$}.txt"

Do not auto-freeze when the root cause is still unknown, the reproduction spans modules, or the user opted out. After original-scenario verification, run /octo:unfreeze or remove only this workflow's freeze state.

Adapted from diagnosing-bugs in mattpocock/skills at commit 3cca18b368ae95cdbdebbff572ccafa662551015 under the MIT License. See THIRD_PARTY_NOTICES.md.

Signals

GitHub stars
4k
Forks
380
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
skill-debug-nyldn
Source
github.com/nyldn/claude-octopus