agentic-praxis-grimoire

APG5 First Codex Dogfooding

Observation

The human maintainer supplied and accepted the application observation. APG5 independently reproduced the locally observable repository facts, including the clean expected project state, canonical skill files, and repository discovery links. Shell inspection cannot independently reproduce the application UI or its explicit invocation event.

This is APG’s first real explicit-use dogfooding observation. It does not measure automatic selection and does not promote the skill beyond provisional.

Evidence categories

The accepted observation covered the active instruction chain, clean repository state, six-skill availability, canonical-to-projection identity, exact relative symlink targets, target containment, readable canonical leaves, matching frontmatter names, absence of duplicates or broken links, and an explicit review-skill result.

Superpowers remained globally installed and visible in the broader registry. It was not invoked or followed and did not obtain effective workflow authority. APG’s repository-local behavioral policy remained effective; no mechanical plugin disablement is claimed.

Commit-message investigation

APG4A’s structured commit body contains literal backslash-n sequences in the commit object. Normal Git rendering preserves them, and the managed Git report preserves the same commit-message payload byte for byte. The report executable did not introduce or escape the sequences.

The cause was caller-side construction: a multiline value was serialized and its escaped representation was interpolated into a shell command. Shell quoting then passed the backslash characters literally to Git. Historical evidence was left unchanged. The narrow correction adds commit-message construction and self-verification guidance to the manager-worker reporting contract; no report executable changed.

APG5 skill-use decisions

Positive use:

Material non-triggers:

No skill procedure required correction. The phase makes no automatic-selection, comparative, stable-maturity, production-readiness, or decommission-readiness claim.

Limitations and next evidence

One explicit observation in APG does not establish universal trigger quality, automatic invocation, causal advantage, or repeated real-project reliability. Serena was not exposed in the accepted fresh-session smoke, although that did not prevent the bounded application observation.

The next evidence should be natural, bounded use of APG skills in another real repository, including proportionate non-trigger decisions and review of authority, privacy, ceremony, and project-owned parameters. Another synthetic activation gate is not required.