lifecycle-audit

Mycelium Lifecycle Audit

Resolve bundled paths relative to the Mycelium plugin root, two directories above this SKILL.md. Read references/audit-protocol.md completely before launching a host process.

This is a black-box integration audit. A manually invoked hook can diagnose hook logic, but it cannot prove that Claude Code or Codex discovered, trusted, and dispatched the hook.

1. Fix the audit boundary

• Record the host, target repository, source commit, installed plugin version, and expected plugin root.
• Use a disposable initialized repository when possible. If using a real analytical project, capture a baseline and prohibit scientific code, data, output, report, and pre-existing .living edits.
• Confirm the requested CLI is installed. Inspect its current help rather than guessing flags.
• Freeze the source checkout and installed artifact for the entire host run. Do not edit, reinstall, or replace an exercised hook while its process may be reading it.
• If the user did not explicitly request a real host run, confirm before sending repository context to an external agent service.

For Codex local development, verify codex plugin list --json, use the supported cachebuster/reinstall flow, and start a new task. Hash exercised installed files against the source candidate. For Claude, identify the plugin source actually loaded by the CLI and verify its files the same way. A source/cache mismatch is an audit failure even if hooks otherwise work.

2. Capture a pre-run baseline

Record:

• git status and relevant file hashes;
• existing session log, registry, handoff, lineage, and runtime markers;
• scientific files and output state;
• expected next log or session identifier when determinable.

Do not delete suspicious state merely to make the test pass. If a clean-room audit is needed, create a fresh disposable repository.

3. Launch one genuinely fresh host task

Run the host CLI from the exact target root. Do not invoke any Mycelium hook script in the task or immediately afterward. Use the canonical child-task prompt in references/child-task-prompt.md — every probe it carries is required, including the nested-subagent probe, which may be marked N/A only when the host exposes no native subagent mechanism. Ask the agent to expose the exact automatic lifecycle context it receives and perform only the probes in the audit protocol. The launcher, not the child agent, owns artifact verification and baseline capture. Where the CLI supports it, disable skills in the child task so the lifecycle-audit skill does not recursively launch or expand itself. Keep ordinary file-reading and editing tools available: the child needs them to observe state and satisfy the expected Stop-compliance retry.

Exercise:

1. SessionStart injection and active-state creation.
2. When the host exposes a native subagent/task mechanism, one read-only child start and Stop while the primary remains active.
3. A known-positive harmless analysis command.
4. A near-neighbor negative command that must remain silent.
5. One disposable successful edit using the host's normal edit tool. Remove it with that tool when deletion is supported; otherwise, after edit-tracking evidence is captured, use the protocol's single explicit shell cleanup.
6. Natural Stop, including the expected enforcement block, compliant retry, and a complete five-section handoff using Mycelium's exact headings once each, in order, with nonblank section bodies.

Use python3 unless the target explicitly provides another interpreter; do not assume python exists. Preserve the real command exit status—an optional dependency failure may still be a valid lineage probe.

4. Inspect only after natural Stop

Compare post-run state with the baseline. Verify the first Stop blocked until the disposable repository's living layer was updated, then verify exact-once finalization, registry and lineage effects, handoff structure, cleanup, and absence of the disposable file. Separate lifecycle-owned changes from agent-authored scientific changes.

Classify every unexpected result as one of:

• Artifact identity: wrong source, version, cache, or packaged file.
• Host dispatch: registration, trust, discovery, matcher, or host payload never reached the hook.
• Hook logic: dispatch occurred, but state or context violated the contract.
• Environment: missing interpreter/dependency or CLI limitation unrelated to Mycelium behavior.
• Agent compliance: hook instructed the agent correctly, but it ignored or rewrote required state.
• Inconclusive: evidence was overwritten, manually stimulated, or never captured.

Never relabel an environment or compliance failure as a hook pass without showing the hook evidence that did pass.

5. Report and clean up

Return the evidence table from the audit protocol, exact errors, plugin identity, and a scientific-state diff. Remove only disposable probe files you created; retain failing lifecycle state when it is needed to reproduce a retry bug.

Do not repair failures during the audit. Diagnose first so the observed state is not contaminated. If a fix is requested afterward, switch to $mycelium:develop, add a red regression test, and repeat the black-box audit on the new installed artifact.