Accepted
2026-07-19
The human maintainer’s APG9 assignment accepts this decision. It authorizes v0.1 closeout, current-document reconciliation, and a bounded v0.2 roadmap. It does not authorize APG10, skill edits or promotion, public-repository changes, Superpowers restoration, or v0.2 publication.
APG v0.1.0 has been published through one intentionally squashed public commit. The public checkout supplies the maintainer’s user-global Codex integration. The private development history remains distinct and has advanced through APG-TEST0, which established the project test layout and the 22-family Bats and 28-family Python integration suites.
The maintainer has decommissioned Superpowers. A fresh post-decommission RepoMap session discovered all six APG skills, explicitly selected and applied the review skill, passed the default and explicit managed checks, preserved a clean RepoMap worktree, and did not use Superpowers as workflow authority. The preserved Superpowers source remains historical external evidence.
The public v0.1.0 projection also exposes a release-process defect: it documents
the project-local projection command and contains its implementation modules,
but omits the tracked bin/apg-project-skills wrapper. APG9 records that
limitation without modifying or republishing v0.1.0. A validated public
projection and release process must prevent this class of omission in v0.2.
The six current skills remain provisional. Their accumulated scenario,
real-use, non-trigger, correction, regression, rollback, and post-decommission
evidence warrants an individual stability review. It does not authorize
promotion during APG9. Clean baseline-versus-treatment superiority evidence is
absent, but that absence is not itself a promotion blocker.
The current roadmap also contains unnumbered themes whose original questions have been answered by v0.1 artifacts, remain project-owned, need a concrete v0.2 owner, or require an explicit reconsideration condition. Leaving that section open would obscure completion of the v0.1 epic.
v0.2 must:
stable, leaving only a skill with an
unresolved material defect non-stable under a recorded correction condition;stable requires repeated real use, representative non-triggers, no unresolved
material authority, privacy, safety, or procedure defect, an explicit
post-Superpowers review, and a supported removal or rollback path. Publication
and global installation are useful integration evidence but do not independently
establish maturity.
This decision refines ADR 0003’s bootstrap minimum for v0.2: cross-repository breadth remains material evidence to inspect and report, but the absence of a positive use in a second repository is not independently a material defect. Repeated positive use may be established within APG, in additional repositories, or through both. Representative non-triggers remain a separate requirement and do not count as positive use. This refinement follows the human mandate to target all six for stability while permitting only a concrete material defect to block an individual skill.
Each skill receives an individual APG13 disposition. One bounded correction per
skill may be made and its affected scenarios and regression gates rerun. A
skill that still has a material defect remains non-stable with a concrete
correction condition. stable does not mean universal applicability,
production warranty, automatic selection, or comparative superiority. Clean
A/B evidence may be pursued under the existing isolation conditions, but it is
not a prerequisite for promotion.
The tracked Karpathy Guidelines source remains experimental evidence until APG10. APG will synthesize rather than copy by default. The source’s declared MIT frontmatter does not replace exact upstream provenance, license, and notice review before any copied or adapted use.
The likely destination is a targeted refinement to an existing owner or a small cross-cutting project standard. Source packaging does not justify a seventh broad coding-guidelines skill. APG10 owns the final adopt, defer, or reject disposition, including the tension between mandatory escalation for every ambiguity and APG’s proportional progress under reversible uncertainty.
Compare the experimental source with current APG policy and the six skills, test actual gaps, and make only supported targeted corrections. At most one correction per skill is allowed. A seventh skill requires a separate consequential decision. APG10 owns final adopt, defer, and reject dispositions.
Evaluate and formalize the existing lifecycle, provenance, scenario, maintenance, and rollback practices as a minimal reusable procedure, justified by their repeated use across APG4 through APG9. This does not require a new skill. Add deterministic checks only for stable mechanical invariants demonstrated by repeated work. Record a terminal disposition for every legacy theme and close the candidate-theme section without building a generalized evaluator or prompt-scoring framework.
Formalize public-sourced global integration and its update, check, uninstall, rollback, restart, and duplicate-scope behavior. Implement a public-projection validator and reproducible squashed-release process. The validator must detect the v0.1 omitted-wrapper class and verify licensing, notices, tests, release metadata, public/private separation, and the intended tracked surface.
Review each skill separately with positive, non-trigger, edge or stop, and
real-repository evidence. APG10 owns Karpathy-derived deltas; APG13 may correct
only a residual material defect exposed by maturity review. Permit one bounded
material correction per skill,
promote passing skills to stable, record a correction condition for any
blocked skill, and produce the release-candidate maturity matrix.
Run the complete gates, verify public global integration against the candidate, create and verify the squashed public v0.2.0 projection, and record release closeout. APG14 does not begin until separately authorized.
No phase begins automatically. This sequence is not a preapproved roadmap envelope: APG10, APG11, APG12, APG13, and APG14 each require a separate human assignment before work begins.
| Theme | Preliminary terminal disposition |
|---|---|
| Publication-surface validation | Implement in APG12. |
| Skill authoring and bounded mechanical maintenance checks | Formalize and implement in APG11. |
| Generalized prompt scoring | Defer until a supported isolated evaluation can freeze treatment and scoring before results. |
| Debugging, verification, planning, design, implementation discipline, and evidence-first review | Satisfied by the six current skills; APG13 reviews maturity rather than reimplementing them. |
| General test runner, universal coverage threshold, and duplicated language policy | Reject as unnecessary or project-owned. |
| Generic delivery workflow and autonomous worker infrastructure | Reject as native or project-owned; no additional execution-controller skill is justified. |
| Language or project-standard profiles | Defer until multiple validated skills show repository instructions are repeatedly insufficient. |
| Repository-local Codex projection and project-local lifecycle | Satisfied by APG4A and ADR 0004. |
| Public-sourced maintainer global integration | Formalize in APG12. |
| Additional non-Codex harness adapters or packages | Defer until a named harness demonstrates a separate need and rollback contract. |
| Licensing and contribution governance | Satisfied by ADR 0005 and the aligned governance files. |
| Public repository creation | Satisfied by v0.1.0 publication. |
| Public v0.2.0 projection | Implement in APG14. |
| Release cadence | Reject as a maintainer-owned publication choice. |
| Hosted services | Defer pending a concrete service, safety model, operator, and explicit authority. |
| Telemetry | Defer pending a decision deficit, privacy-safe contract, retention rules, and explicit authority. |
| Wholesale source migration | Reject; APG selectively evaluates and synthesizes practices. |
| Broader dogfooding and individual maturity dispositions | Implement in APG13. |
| Clean baseline-versus-treatment comparison | Conditionally defer; it is not a stability prerequisite. |
| Superpowers decommission, rollback plan, source preservation, and bounded smoke | Satisfied by the human action and existing records; APG9 reconciles current state. |
| Post-Superpowers skill review | Implement in APG13. |
Rejected. Current evidence makes all six credible candidates but APG13 still owes individual review, final correction handling, rollback confirmation, and explicit dispositions.
Rejected as a mandatory gate. The supported harness has not yet supplied the necessary isolation and invocation evidence, and absence of that comparison is not a material authority, privacy, safety, or procedure defect.
Rejected. It overlaps current owners, includes absolute guidance that conflicts with proportional escalation, and would require complete provenance and notice review for copied or adapted expression.
Accepted. APG10 can preserve useful principles while placing each gap in the artifact that owns it and avoiding duplicated workflow policy.
Rejected as the complete v0.2 distribution model because the maintainer now uses a public-sourced global integration. Project-local projection remains supported for repository-specific discovery and rollback.
Accepted for APG12. The current practice must become reproducible and
verifiable without turning apg-project-skills into a global installer.
Rejected. The v0.1 wrapper omission demonstrates that documentation review and an otherwise valid public surface do not prove projection completeness.
Accepted for APG12 and APG14. It must remain narrow, deterministic, and owned by demonstrated projection and release invariants rather than becoming a general framework.
Rejected. Its broad headings now mix satisfied questions, project-owned policy, conditional deferrals, and concrete v0.2 work.
Accepted. APG10 through APG14 give every remaining theme an owner or terminal disposition while preserving separate phase authorization.
provisional until APG13.