APG17 accepts ADR 0013 and implements one provisional
synthesizing-repository-guidance leaf. The skill owns decomposition and
bounded ownership, provenance, privacy, migration, rejection, rollback, and
source-preservation dispositions for mixed guidance before any rewrite. Its
conceptual boundary is classify and disposition; do not rewrite or migrate.
The retained capability is narrower than a writing skill, standards catalog, migration engine, or root-modularization command. It grants no implementation, cutover, deletion, publication, maturity, or successor-phase authority.
The maintainer confirmed that both same-name workflow routers were visible and selectable in the APG private development repository and that only the personal/private router was visible and selectable globally. APG17 records that supplied result as the subsequent external disposition that closes APG16’s duplicate-name discovery and coexistence observation. It infers no routing-case transcript, selector metadata, user-interface detail, or precedence rule.
Current official Codex skill documentation and the available native technical authoring capability were compared with APG’s authoring guide, project model, and provenance policy.
A separate public skill-writing leaf is not justified.
The source analysis used bounded maintainer-authored private guidance, one general engineering source, a representative historic manager-prompt corpus, three representative private skills, current official Codex guidance, native authoring capability, and APG’s current owners. Public expression is independently written. Private ownership does not establish publication eligibility, and source packaging does not establish a public capability.
Raw medical, cognitive, diagnostic, accessibility, deeply personal,
productivity, machine-specific, and private-integration detail was neither
retained nor published. keep-private is a valid terminal result. Missing
source identity or reuse rights blocks copied or adapted work; bounded
independent analysis may continue only where allowed.
retain-root
route-existing-owner
synthesize-existing-owner-correction
candidate-new-skill
candidate-language-profile
core-documentation
deterministic-tool-candidate
project-owned
keep-private
evidence-only
defer
reject
split-required
blocked-source-or-rights
Each coherent unit receives a stable local identifier, one proposed owner, one disposition, rights and precedence results, acceptance evidence, migration dependency, rollback or restoration boundary, and public/private treatment. A disposition is not implementation authority.
The sixteen-family contract was frozen after four read-only source and baseline reviews and before candidate wording:
Ordinary reasoning reached correct owners and safe dispositions for all sixteen. Its material limitation was not correctness or table formatting; it was lower repeatability for coherent splitting, coverage, stable records, migration dependencies, rollback, and explicit source preservation in the mixed cases. This is baseline and retention evidence, not a claim of comparative superiority.
Two fresh evaluators applied disjoint A-H and I-P halves to the authored candidate. All positives produced complete dispositions, every non-trigger remained a non-trigger, rights and authority faults stopped safely, personal and machine guidance remained private, rewrite and mirroring demands were rejected, and sources remained unchanged. Candidate correction count: zero.
The retained candidate classified one bounded 34-unit inventory across every authorized source class. The selected sample was large enough to exercise mixed ownership, existing owners, project policy, private retention, future capabilities, profile candidacy, evidence-only treatment, rights, migration, and rollback. It did not attempt wholesale private-guidance migration.
| Disposition | Count |
|---|---|
retain-root |
1 |
route-existing-owner |
16 |
synthesize-existing-owner-correction |
1 |
candidate-new-skill |
2 |
candidate-language-profile |
2 |
core-documentation |
1 |
project-owned |
5 |
keep-private |
5 |
evidence-only |
1 |
| Total | 34 |
No selected unit ended at a rights blocker; the frozen missing-rights scenario separately verifies that stop. Each private-skill route records explicit shadow behavior, discovery and non-trigger checks, project/private overrides, cutover acceptance, and byte-restorable rollback prerequisites.
The application demonstrates distinct procedural value by retaining stable unit identifiers, one proposed owner and disposition per coherent unit, coverage, rights, precedence, migration, restoration, public/private treatment, and source preservation. It is not a comparative experiment.
No inventoried source was rewritten, removed, retargeted, migrated, published, or decommissioned.
The historic manager-prompt corpus supports one future-capability recommendation for constructing a top-level Codex assignment from an already human-approved roadmap. A single-project sample does not establish ordinary prompting as inadequate. APG17 creates no final name, skill, prompt, or implementation and grants no future authority.
Two bounded language units received candidate-language-profile. ADR 0012
remains Proposed; its profile contract and warning levels are not accepted or
implemented. No language profile exists.
Current private development contains eight canonical leaves, eight relative
checked-in projections, six stable process rows, and two provisional rows.
The router’s sorted capability map contains seven routable non-router entries
and routes mixed guidance to synthesizing-repository-guidance. A focused
catalog-driven exact-set test and one mixed-guidance scenario test cover that
contract. The router procedure remains byte-identical.
The catalog-growth test failed first while the map was intentionally stale, then passed after one map entry was added. One current-development checker integration expectation was corrected from the prior seven-entry summary to eight. The project command recognizes the router and synthesis leaf as known unmanaged development skills while continuing to manage exactly the six stable v0.2 leaves.
Public v0.2.0, active integration, project/user state schemas, managed default sets, install/update/rollback behavior, public critical policy, and release architecture remain six-skill contracts. No public or active integration was modified.
Rollback removes the synthesis leaf, projection, catalog row, map entry, focused test, and second known-unmanaged name, and restores every APG17 current-state count and compatibility statement to the prior seven-canonical, six-routable development shape. Historical phase records remain as evidence.
agentic-praxis-grimoire-workflow routed the ambiguous phase to the relevant
current leaves and later advertises the retained synthesis owner.designing-significant-changes owned the capability boundary, source/privacy
model, native-authoring distinction, output contract, and rollback.planning-repository-work owned the dependent baseline, freeze, candidate,
dogfood, integration, review, and closeout sequence.composing-bounded-worker-assignments owned the already-authorized source,
baseline, scenario, and independent-review assignments.implementing-with-test-discipline owned only deterministic catalog-map and
current-development compatibility tests.debugging-systematically applied only to a temporary baseline-fixture
inconsistency whose cause was established before correction.reviewing-and-verifying-repository-work governed worker disposition,
candidate retention, ledger correction, integration, and completion claims.Automatic delegation, a mandatory skill chain, native technical authoring as the classification owner, a language profile, root rewrite, private cutover, release, and APG18 were material non-triggers.
The fresh synthesis semantic reviewer returned accept with no blocker or
material finding. The dogfood-ledger reviewer initially returned
correction-required for bundled multi-owner units, one stale native-source
identity, and insufficient private-skill migration gates. The ledger was
corrected by splitting each procedure into one-owner units, recording the
resulting 34-unit count and current source identity, and making shadow,
discovery, override, cutover, restoration, and rollback evidence explicit. The
reviewer then returned accept with no remaining blocker or material finding.
The router and integration reviewer initially required correction of one stale
7/7/7 current-checker expectation and then required documentation rollback to
include current-state count and compatibility surfaces. Both corrections were
applied without changing the candidate. The reviewer then returned accept
with no remaining blocker or material finding. Complete-diff review follows the
final exit and validation records.
The fresh complete-diff reviewer returned accept with no blocker or material
finding after inspecting every tracked and untracked APG17 artifact. It
independently verified APG16 wording, ADRs 0012 and 0013, one new skill,
34-unit arithmetic, catalog and distribution boundaries, public privacy and
links, external read-only state, rollback, validation claims, non-goals, and
the pending-smoke terminal outcome.
Resulting-state evidence includes 17 passing skill-library unit tests, 58 passing checker integration tests with the exact v0.1 fixture, 8 release unit and 32 release integration tests, 8 user unit and 39 user integration tests with the exact v0.1 fixture, 22 report-tool Bats families, and 28 project-lifecycle integration families. Development checker text and JSON pass at 8/8/8; public v0.2.0 passes at 6/6/6. Python byte compilation, applicable Bash syntax, all six command-help checks, and Git whitespace checking pass. Link, confidentiality, schema, mode, counter, and public-read-only boundaries are covered by the checker, release, user, report, and project suites and the final diff review.
The public v0.2.0 checkout, reference corpus, RepoMap checkout, active public-backed integration, private source guidance, private skills, and retired Superpowers state remain unchanged. APG17 creates no dependency, plugin, runtime, scheduler, registry, daemon, state-schema change, release candidate, tag, or public push.
File, checker, scenario, and review evidence cannot prove that a newly launched Codex session discovers and explicitly applies the repository-scoped synthesis leaf. That observation remains external.
Terminal repository outcome:
synthesis-skill-implemented-pending-fresh-session-smoke
APG18 remains a separately authorized future decision and has not started.
APG17A corrected two publication-excluded public-release identity fields after fresh remote, checkout, lineage, and release-record verification. The public release itself was unchanged. No skill, architecture, scenario, dogfood disposition, maturity state, distribution contract, or external state changed.
The maintainer intentionally deferred APG17 application discovery and explicit use smoke to the aggregate v0.3 readiness and release-preparation gates. This deferral does not reverse APG17 repository acceptance. APG18 is separately authorized by the APG17A-to-APG18 assignment.