agentic-praxis-grimoire

APG6 RepoMap Cross-Repository Dogfooding

Outcome

• Observation date: 2026-07-19
• Additional repository: RepoMap, using the canonical public identity repo-map
• Harness class: Codex macOS application with repository-local linked skill discovery
• Observations: one read-only migration design and one accepted documentation-only implementation and closeout
• Maturity: all six APG skills remain provisional
• Superpowers state: globally installed, unchanged, and reference-only for APG

RepoMap is the second real repository in which APG skills have been used. The two observations below are distinct samples with different authority and trigger boundaries. They do not establish automatic selection, comparative or causal superiority, statistical reliability, production readiness, or stable maturity.

Authority and evidence boundary

The human maintainer accepted the completed RepoMap DOC-LAYOUT0 result and authorized APG6 to maintain the resulting APG evidence. APG6 does not reopen the archive decision, authorize another RepoMap phase, or modify RepoMap.

Application-level discovery and explicit skill-use facts are accepted observed evidence. APG6 independently reconciles the locally observable clean repository state, accepted documentation result, canonical APG leaves, linked discovery entries, and resulting APG correction. Public evidence omits private checkout identities, commits, report identifiers, local paths, usernames, and managed-report destinations.

Observation A — read-only migration design

Trigger decisions

Explicitly selected:

• designing-significant-changes, because archive placement, historical preservation, date semantics, migration, and rollback boundaries were still consequential design decisions; and
• reviewing-and-verifying-repository-work, because the bounded design artifact required evidence-backed disposition.

Material non-triggers:

• planning-repository-work, because the design had not yet been accepted;
• implementing-with-test-discipline, because implementation was not authorized;
• debugging-systematically, because no unexplained target defect required a diagnostic workflow; and
• composing-bounded-worker-assignments, because delegation was unnecessary.

Artifact and disposition

Result: pass-with-follow-up.

The design correctly rejected semantic reclassification based only on missing historical exit framing. It supplied current-state inventory, credible alternatives, introducing-committer-date and counter semantics, migration and rollback boundaries, non-goals, and maintainer decisions.

The review procedure handled the exact artifact under review, preserved the decision owner’s authority, and did not accept unresolved design implicitly. The follow-up concerned discovery precision: the review leaf’s frontmatter named changes, phases, commits, and worker results but not design records, plans, reports, or other bounded repository artifacts already covered by its procedure.

Observation B — accepted DOC-LAYOUT0 implementation

Trigger decisions

Explicitly selected:

• planning-repository-work, because the accepted migration required dependent inventory, move, reference, validation, review, and closeout units; and
• reviewing-and-verifying-repository-work, because the complete resulting repository state and evidence required acceptance or correction.

Material non-triggers:

• designing-significant-changes, because the maintainer had accepted the consequential layout and historical-preservation decisions;
• composing-bounded-worker-assignments, because implementation was not delegated and only a read-only final reviewer was needed;
• debugging-systematically, because no unexplained failure occurred; and
• implementing-with-test-discipline, because no executable behavior changed.

Completion disposition

Result: complete.

DOC-LAYOUT0 moved 41 historical ADRs and 636 historical status records to daily archive layouts while preserving every status record under docs/status/ and preserving historical numbers, basenames, H1 titles, semantic roles, and archive ownership. Accepted ADR 0042 and status exit 00637 were added. All 677 introduction-date placements matched the introducing committer timestamp’s recorded offset with zero ambiguity, target collisions were zero, and active old-layout references were zero. Review classified 266 retained old-layout occurrences as deliberate historical, remediation, plan, command or report, or synthetic evidence.

Local Markdown links and fragments passed. Independent finished-diff review accepted the documentation-only result, its commit was pushed and verified remote-equal, and no source, test, dependency, runtime, semantic archive classification, compatibility stub, symlink, or migration framework was added. The implementation observation required no APG skill correction.

Cross-repository discovery observation

A fresh Codex macOS session rooted in RepoMap discovered all six linked APG skills after a full application restart. The cross-repository links remained readable and resolved to the matching canonical APG leaves. This is an observed bootstrap fact for the sampled environment, not a universal guarantee about Codex refresh behavior. Projects adding local linked skills may need a fresh application session before testing discovery.

Superpowers was not invoked or followed in either RepoMap observation.

Bounded review-skill correction

APG6 changes only the review leaf’s frontmatter description and the matching skill-index trigger summary. The description now names a bounded repository artifact and evidence-backed disposition explicitly. The procedure already began by establishing the exact artifact or resulting state under review, so no procedure change was required.

The correction does not authorize review, create acceptance authority, turn casual design feedback into a formal disposition, accept unresolved design, or change project-owned severity and review parameters. The skill remains provisional.

Authority, privacy, ceremony, and project-policy findings

• Both observations stayed within explicit maintainer authority and kept design acceptance separate from implementation authorization.
• No private identity, local path, commit, report identifier, report destination, or raw payload entered the public record.
• Selected skills added sequencing, review, privacy, and completion discipline without imposing branches, executable tests, managed reports, delegation, or another design phase where those were not project-owned requirements.
• Material non-trigger decisions prevented premature implementation, unnecessary debugging, unnecessary delegation, and reopening an accepted design.
• No authority, privacy, destructive-action, or completion-evidence regression was observed in these samples.

Validation and independent review

A fresh read-only scenario worker returned pass across the positive design- artifact, positive resulting-state, casual-feedback non-trigger, and stale-or- changing-evidence edge cases. It confirmed that ideation remains excluded, unresolved design cannot be silently accepted, the description grants no authority, project-owned authority and severity remain external parameters, and no procedure text change is needed.

A separate fresh non-author leaf reviewer returned accept with no blocker, material finding, minor finding, or second-correction need. It confirmed valid frontmatter, exact index alignment, an otherwise unchanged procedure, preserved stop conditions and project-policy parameters, and unchanged provisional maturity.

The fresh independent complete-diff reviewer returned accept-with-follow-up with no blocker, material finding, minor finding, or optional repository finding. It independently reproduced the archive evidence, observation separation, trigger decisions, procedure stability, APG5 placement, date semantics, maturity, decommission, confidentiality, scope, and repository-state checks. The only follow-up was procedural: record the review and rerun the two Git whitespace checks after that closeout update. Both checks passed, and the same reviewer returned final disposition accept with no remaining finding.

Maturity and decommission state

Current evidence supports four components of the eventual Superpowers decommission gate: material workflow mapping, successful APG use in its own repository, successful APG use in one additional real repository, and a preserved Superpowers source and provenance snapshot.

The gate remains incomplete. Broader real-use coverage across all six skills, repeated regression evidence, an explicit uninstall rollback plan, an explicit human decommission decision, actual Superpowers removal, and post-decommission smoke validation remain unresolved. Superpowers coexistence is unchanged.

Limitations

• These are two accepted observations in one additional real repository, not a repeated or controlled comparison.
• Skill selection was explicit; automatic selection was not measured.
• Cross-repository discovery was observed after a full restart in one Codex macOS environment only.
• The implementation was documentation-only and did not exercise executable behavior, schema, or contract-change test discipline.
• Broader real-use and regression coverage across the six-skill bundle remains necessary before any maturity or decommission decision.

This site is open source. Improve this page.