agentic-praxis-grimoire

APG Roadmap

Roadmap policy

The human maintainer retains ultimate roadmap authority. ChatGPT may advance successive bounded phases only within a human-approved task or preapproved roadmap envelope and must stop at its defined boundary. Codex recommendations, commits, and reports are evidence rather than authorization.

APG0 through APG8 are the closed v0.1 development epic, with every historical terminal outcome preserved, including APG3’s blocked preflight. The epic established governance, the bootstrap maturity model, six provisional skills, Codex repository discovery, APG and RepoMap dogfooding, project-local projection and rollback, and RepoMap managed adoption. The APG9 human assignment supplies APG8’s previously requested final external acceptance.

Public v0.1.0 was subsequently published in one intentionally squashed public commit. The public checkout is the canonical source for the maintainer’s separately managed user-global Codex integration. Superpowers was decommissioned by explicit human decision, and a bounded fresh RepoMap smoke discovered all six skills, successfully applied the review skill, passed managed checks, preserved the tracked repository, and did not use Superpowers as workflow authority. These later facts reconcile current state without rewriting contemporaneous phase records.

APG-TEST0 is the first post-v0.1 development-foundation phase. It establishes the current test layout and the 22-family Bats and 28-family Python integration suites. APG13 promoted all six current skills to stable under ADR 0010. APG14 publishes those six stable leaves as public v0.2.0 through APG12’s exact projection and reproducible squashed-release process, correcting the v0.1.0 omitted-wrapper class without rewriting the historical release.

APG9 accepts the bounded APG10 through APG14 v0.2 sequence. It is not a preapproved roadmap envelope. Separate human assignments completed APG10’s experimental-source disposition, APG11’s skill-maintenance formalization, APG12’s public distribution and release validation, APG12A’s bounded correction, APG13’s six individual maturity dispositions, and APG14’s v0.2.0 publication sequence. No successor roadmap epic is authorized.

APG0 — Foundation and governance

• Status: complete; see the APG0 Foundation Exit.
• Objective: establish project identity, artifact ownership, provenance, manager-worker coordination, a compatible skill leaf, research evidence, and a bounded migration direction.
• Artifact: concise root instructions, core governance documents, skill index, publication-excluded research, roadmap, and exit record.
• Result: governance adopted; no production skill, taxonomy, validator, dependency, or publication.

APG1 — Public projection and documentation boundaries

• Status: complete; see ADR 0001 and the APG1 exit.
• Objective: establish the private-working/public-projection contract, two-level provenance, corrected report direction, independent record namespaces, and a publication-safe documentation surface.
• Artifact: ADR 0001, record indexes, private evidence ledger, sanitized public documents, normalized APG0 exit, and APG1 exit.
• Result at APG1 close: publication boundary adopted; repository creation, skills, validation, licensing, and release remained future work.

APG2 — Roadmap reconciliation and first-implementation proposal

• Status: complete as a proposal phase; externally accepted by APG2A. See ADR 0002, which is Accepted, and the APG2 exit.
• Objective: compare the credible first-implementation choices under common criteria, correct the public authority chain, and recommend exactly one bounded next phase.
• Artifact: authority-model corrections, publication-excluded decision evidence, proposed ADR 0002, this reconciled roadmap, and the APG2 exit.
• Result: one evaluation-first skill vertical slice was accepted as APG3. No implementation occurred during APG2 or APG2A.

APG3 — Bounded worker-assignment skill vertical slice

• Status: phase closeout complete; terminal outcome blocked at the mandatory harness preflight. See the public evaluation summary and APG3 exit.
• Objective: evaluate whether ordinary Codex assignment composition has a material deficit and, only when the frozen baseline-adequacy rule demonstrates that deficit, author and evaluate at most one narrowly triggered, reversible composing-bounded-worker-assignments skill. It would turn an already authorized objective into one proportional internal worker assignment only after delegation is separately permitted and selected.
• Artifact: a public-safe blocked evaluation record, exact publication- excluded preflight evidence, updated skill and provenance records, and one APG3 exit. No candidate leaf exists.
• Evaluation result: the accepted method required an independent evaluator to freeze calibration, sealed confirmation, bounded sampling, a rubric, a baseline-adequacy rule, and safe downstream cases before authoring. That sequence did not begin. Active overlapping skill behavior could not be removed from both conditions through a supported internal-agent profile, and the harness exposed neither an invocation event nor candidate availability and deliberate-loading controls. The experiment therefore could not preserve the candidate as the sole intended treatment difference.
• Dependency result: Accepted ADR 0002, APG2A authorization, manager-worker boundaries, and the manual public-surface gate were present. The required comparable harness was absent, which activated the accepted stop condition.
• Non-goals: delegation authorization, a manager runtime, scheduler, registry, external-prompt generator, reusable evaluation framework, publication validator, taxonomy, harness adapter, dependency, license choice, publication decision, or another roadmap theme.
• No-leaf state: the candidate was never authored. No active or empty candidate directory, rollback mutation, or rejected-candidate snapshot exists.
• Authorization boundary: APG2A authorized only this vertical slice. APG3 applied its stop boundary and is closed. Later themes below remain unauthorized.

APG4 — Bootstrap APG v0.1 under Superpowers coexistence

• Status: complete as a provisional bootstrap phase; see ADR 0003, the APG4 evaluation summary, and the APG4 exit.
• Objective: make Superpowers reference-only for this repository, accept a bootstrap-before-clean-evaluation maturity model, and author a small canonical APG skill bundle without changing global plugin state.
• Artifact: ADR 0003, the bootstrap model, six direct-child skill leaves, the Superpowers transition map, public and publication-excluded scenario and review evidence, updated indexes, and one APG4 exit.
• Skills: composing-bounded-worker-assignments, designing-significant-changes, planning-repository-work, implementing-with-test-discipline, debugging-systematically, and reviewing-and-verifying-repository-work.
• Maturity: every retained skill is provisional. No skill is called evaluated, stable, production-proven, or superior.
• Evidence: 18 deliberate public-safe scenario walkthroughs and independent per-skill and complete-diff review. This is not automatic-trigger telemetry or isolated baseline-versus-treatment evidence.
• Superpowers state at phase close: installation was unchanged and decommission had not occurred. Repository instructions suppressed workflow authority behaviorally without claiming mechanical unloading.
• Non-goals: plugin mutation, isolated profile, adapter, evaluator framework, dependency, runtime, scheduler, registry, taxonomy, publication, license decision, or another implementation epic.
• Authorization boundary: APG4 closes after bootstrap, push, and reporting. It does not authorize promotion, decommissioning, or a subsequent phase.
• Subsequent disposition: skill content and evidence accepted; correction-required for the omitted Codex repository discovery projection.

APG4A — Codex repository skill discovery integration

• Status: complete as a forward integration correction; see the APG4A exit.
• Objective: expose the six canonical APG v0.1 leaves through Codex’s repository discovery location without copying or changing their content.
• Artifact: a real .agents/skills/ directory containing six relative symbolic links to the matching canonical leaves under skills/, reconciled public documentation, and exit 00007.
• Evidence: exact relative targets, repository containment, readable matching leaves, unchanged canonical SHA-256 values, Git mode 120000, Markdown and confidentiality gates, and fresh finished-diff review.
• Maturity: every skill remains provisional; projection correctness is integration evidence, not invocation or dogfooding evidence.
• Superpowers state at phase close: installation was unchanged and behaviorally reference-only for APG.
• Authorization boundary: APG4A ends after integration, validation, push, and reporting. It does not begin dogfooding or authorize promotion, decommissioning, packaging, publication, or another implementation phase.

APG5 — First Codex dogfooding and commit-message hygiene

• Status: complete; see the APG5 evaluation and APG5 exit.
• Objective: record the accepted fresh-session discovery and first explicit APG skill use, diagnose APG4A’s literal backslash-n commit-body sequences, and correct only the repository-owned cause supported by evidence.
• Artifact: public and publication-excluded dogfooding evidence, focused commit-message hygiene in the manager-worker reporting contract, reconciled project history, and exit 00008.
• Result: six-skill discovery and explicit review-skill use passed. Exact-byte evidence localized the anomaly to caller-side message construction; the Git report preserved the commit correctly and required no executable change.
• Maturity: all six skills remain provisional; automatic invocation, comparison, stable maturity, and production readiness remain unmeasured.
• Superpowers state at phase close: installation was unchanged and reference-only for APG; no Superpowers workflow governed APG5.
• Authorization boundary: APG5 ends after validation, commit, push, and linked reporting. It does not authorize promotion, decommissioning, publication, licensing, or another implementation phase.

APG6 — RepoMap cross-repository dogfooding

• Status: complete and externally accepted; see the APG6 evaluation and APG6 exit.
• Objective: record separate RepoMap migration-design and accepted documentation-only implementation observations, correct review-skill discovery for bounded repository artifacts, and align APG exit placement with introducing-committer-date semantics.
• Artifact: public and publication-excluded cross-repository evidence, one review frontmatter and index correction, explicit status-date policy, the corrected APG5 exit placement, reconciled maturity and decommission records, and exit 00009.
• Evidence: successful use of designing-significant-changes, planning-repository-work, and reviewing-and-verifying-repository-work in RepoMap; proportional material non-triggers; fresh correction scenarios; non-author leaf review; complete finished-diff review; and public/private, sequence, link, and projection checks.
• Maturity: every skill remains provisional. The observations are neither automatic-selection nor comparative evidence.
• Superpowers state at phase close: installation was unchanged and reference-only for APG. Material mapping, APG use, one additional-repository use, and a preserved source snapshot supported the gate, while decommission still awaited later human action and smoke.
• Authorization boundary: APG6 ends after validation, one commit, push, and linked reporting. It does not authorize promotion, decommissioning, publication, licensing, another skill, or another implementation phase.

APG7 — Project-local skill projection and rollback tooling

• Status: accepted in substance and corrected forward by APG7A; see the APG7 evaluation, ADR 0004, and APG7 exit 00010.
• Objective: provide safe opt-in install, adopt, check, and uninstall operations for the six canonical APG skills without tracked target or global Codex and Superpowers mutation.
• Artifact: one Python 3.10+ standard-library executable with two non-executable helpers, 26 behavioral test families, project-local deployment guidance, a human Superpowers decommission and rollback runbook, public and publication-excluded evidence, and exit 00010.
• Evidence: accepted ownership and rollback design; intended pre-production failing evidence; 26 passing behavioral tests in temporary Git worktrees and temporary HOME; one separate passing disposable lifecycle; fresh code and safety review; fresh complete-diff review; and integrated source, projection, Markdown, confidentiality, sequence, mode, and Git checks.
• Skill use: design, planning, implementation discipline, review, and bounded reviewer-assignment composition applied within their triggers. Systematic debugging remained a material non-trigger because no unexplained failure occurred.
• Maturity: every skill remains provisional. The executable observation is neither automatic-selection, comparative, stable, nor production-readiness evidence.
• Superpowers state at phase close: installation was unchanged and reference-only for APG. The documented and tested local uninstall plus human rollback runbook supported the rollback-plan component; later human action and smoke remained pending.
• Authorization boundary: APG7 ends after validation, one commit, push, and linked reporting. It does not authorize global plugin action, promotion, publication, licensing, another skill, or a successor phase.

APG7A — Idempotent projection compliance correction

• Status: complete and externally accepted; see the APG7 evaluation’s subsequent-correction section and exit 00011.
• Objective: make repeated install and adopt success require the same semantic managed-path cleanliness that check already enforces.
• Artifact: one shared idempotent-compliance guard in the existing command helper, two focused behavioral families, strengthened clean controls, reconciled APG7 records, focused publication-excluded evidence, and exit 00011.
• Evidence: exact disposable reproduction for install and adopt; two tests failing against APG7 production for the intended exit-0 reason; four focused corrected controls; 28 passing behavioral families; a full disposable lifecycle; a semantic-override and exact-recovery control; and fresh code-and-safety and complete-diff review.
• Skill use: systematic debugging, implementation test discipline, and review apply. Repository planning and significant-change design remain material non-triggers because the accepted architecture and one-step correction remain unchanged. Bounded assignment composition applies only to the separately authorized reviewers.
• Architecture: command grammar, state format version 1, state path, exclusion grammar and markers, ownership, locking, transaction ordering, dependencies, portability boundary, and global behavior remain unchanged.
• Maturity: every skill remains provisional. The correction is neither automatic-selection, comparative, stable, nor production-readiness evidence.
• Superpowers state at phase close: installation was unchanged and reference-only for APG; decommission remained pending.
• Authorization boundary: APG7A ends after validation, one forward-only commit, push, and linked reporting. It does not authorize another command, architecture phase, skill promotion, external project mutation, publication, licensing, decommissioning, or a successor phase.

APG8 — RepoMap managed projection adoption

• Status: complete and externally accepted by the APG9 human assignment; see the APG8 evaluation and exit 00012.
• Objective: adopt RepoMap’s six existing compatible manual APG links into managed Git-local state and verify them without changing RepoMap’s tracked repository.
• Artifact: one public deployment evaluation, focused publication-excluded before/adoption/check evidence, reconciled rollout and decommission records, and exit 00012. The command and tests are unchanged.
• Evidence: synchronized clean target checkpoint; exact six-link and canonical validation; expected unmanaged pre-check; one successful default adoption; default and explicit managed checks; preserved link inodes, targets, hashes, user-owned exclusion bytes, tracked tree, index, status, global sentinels, and report state; and fresh independent review.
• Skill use: planning and review apply. Significant-change design, implementation test discipline, and systematic debugging remain material non-triggers. Bounded assignment composition applies only to the authorized reviewer.
• Maturity: every skill remains provisional. Managed deployment is not discovery-after-restart, invocation, automatic-selection, comparative, stable, or production-readiness evidence.
• Decommission effect: material workflow mapping, real APG and additional- project use, source preservation, tested local lifecycle behavior, real-project managed adoption and check, and the human rollback runbook are supported. The full gate remains incomplete.
• Subsequent disposition: the maintainer completed Superpowers decommission, and the bounded fresh RepoMap smoke passed. It discovered all six skills, explicitly applied the review skill, passed managed checks, preserved RepoMap, and did not use Superpowers as authority.
• Authorization boundary: APG8 ends after one APG commit, push, and linked reporting. It does not authorize RepoMap uninstall, another target, command change, skill promotion, global plugin action, publication, licensing, the fresh-session smoke, or another phase.

APG-TEST0 — Repository test layout and report-tool coverage

• Status: complete; see exit 00013.
• Objective: establish the project-owned unit and integration test layout, add behavior coverage for both report commands, and migrate the existing project-skill suite without changing executable behavior.
• Artifact: two Bash Bats suites under src/test/unit/bash, the migrated Python suite under src/test/int/python, documented layout and commands, and exit 00013.
• Evidence: 22 report-tool behavior families passed; all 28 project-skill integration families passed; Bash syntax, executable help, test discovery, repository links, confidentiality, sequence, and Git whitespace checks passed.
• Skill use: planning, implementation test discipline, and review apply. Significant-change design and systematic debugging remain material non-triggers. No worker assignment was composed.
• Maturity: the phase strengthens executable regression evidence but does not change the maturity of any APG workflow skill.
• Authorization boundary: APG-TEST0 changes test ownership and layout only. It does not authorize report-format changes, a new dependency manifest, coverage policy, publication, release, skill promotion, or another phase.

APG9 — v0.1 closeout and v0.2 roadmap adoption

• Status: complete; see ADR 0006, the APG9 public evaluation, and exit 00014.
• Objective: accept public v0.1.0 and its squashed-history model, reconcile user-global public integration and Superpowers decommission, close the v0.1 epic, and assign every remaining roadmap concern to a bounded v0.2 phase or terminal disposition.
• Result: APG8 receives final external acceptance; APG0 through APG8 close as the v0.1 epic with historical outcomes preserved; APG-TEST0 remains the first post-release foundation; and APG10 through APG14 are accepted as separately authorized phases.
• Maturity: all six skills remain provisional. APG13 targets stable for each skill after individual post-Superpowers review; a concrete material defect may block one skill, while absent clean A/B or second-repository positive evidence does not block by itself.
• Release limitation: public v0.1.0 omitted the documented bin/apg-project-skills wrapper. APG12 owns the validated correction; APG9 does not modify or republish public.
• Karpathy status: the tracked source remains experimental. APG10 owns adopt, defer, and reject dispositions and defaults to independent synthesis, not a seventh overlapping skill.
• Authorization boundary: APG9 ends after validation, one private development commit and push, and linked reports. It does not start APG10, edit or promote a skill, restore Superpowers, modify public/reference, or publish v0.2.

APG10 — Karpathy evaluation and selective integration

• Status: complete; see ADR 0007, the APG10 evaluation, and exit 00015.
• Objective: establish exact public provenance and reuse limits for the experimental source, test current APG behavior under a frozen concealed- source contract, and adopt only a demonstrated gap into its current owner.
• Result: assumptions and alternatives, semantic change traceability, speculative-scope avoidance, necessary consistency work, proportional reversible progress, and project-owned boundaries were already covered. Two independent current-change-cleanup positives were materially partial and supported one APG-native correction to implementing-with-test-discipline.
• Provenance: APG10-KARPATHY-SOURCE-01 identifies the unversioned public source inspected for APG10. Its mutable locator and MIT declaration do not supply a semantic revision, complete license, or notice payload, so APG10 copies or adapts no expression and adds no guessed third-party notice.
• Scope: one current leaf changes once. No root universal, other leaf, seventh skill, support directory, dependency, test policy, or maturity state changes.
• Maturity: all six skills remain provisional; the APG10 evidence becomes input to APG13 rather than a promotion decision.
• Authorization boundary: APG10 ends after one private development commit, push, and linked reports. It does not begin APG11, alter public/reference, change global integration or RepoMap, restore Superpowers, or publish v0.2.

Legacy theme dispositions

APG11 closes all 24 former candidate and explicitly deferred questions in the legacy roadmap closure ledger. The ledger preserves implemented, conditional, and rejected/project-owned dispositions. APG12 through APG14 now have completed artifacts where the ledger records them.

Accepted v0.2 sequence

APG10 — Karpathy evaluation and selective integration

Completed. The experimental source was evaluated against existing owners under frozen scenarios. APG10 adopted one current-change cleanup correction in the implementation leaf, recorded every other idea as already covered, project-owned, or rejected, and added no seventh skill.

APG11 — Skill authoring, maintenance, and legacy-roadmap closure

Completed. ADR 0008 and the maintainer guide formalize the repeated lifecycle, the dependency-free checker enforces only adopted mechanical invariants, and the terminal ledger closes 24 legacy themes. No skill leaf, maturity state, dependency, public release, or external integration changed.

APG11A subsequently corrected required-key shadowing and fenced-code closure false negatives in that checker. The correction preserves ADR 0008’s narrow parser architecture, the accepted APG11 lifecycle and ledger, all six skill leaves and provisional maturity states, and the APG12 boundary.

APG12 — Public distribution and release validation

Completed. ADR 0009, the strict public-surface policy, and apg-public-release establish exact projection of every tracked non-private path, critical-owner checks, and one deterministic local squashed commit after the accepted public base. apg-user-skills separately establishes verified public-source install, adopt, check, update, rollback, and uninstall with direct links and user-local ownership state. Disposable dogfood detects the v0.1 omitted-wrapper class and leaves public and active integration state unchanged.

APG12A subsequently corrects four bounded tooling defects without reopening the accepted architecture. Both release and user tools now share an exact public-lineage verifier, executable candidate validation uses only isolated disposable repository copies and temporary state, and user check creates nothing while using an existing shared lock. The APG12 historical test record, v0.1 omission regression, schema versions, six provisional skills, and APG13 and APG14 ownership boundaries remain unchanged.

APG13 — Six-skill post-Superpowers stability review

Completed. ADR 0010 records six individual evidence inventories, frozen current applications, zero procedure corrections, six accept-stable final reviews, and six stable catalog entries. Historical provisional records and public v0.1.0 remain unchanged. APG14 subsequently corrects two overbroad APG9 evidence labels without changing any procedure or maturity disposition.

APG14 — v0.2.0 release candidate and publication

Completed. APG14 prepares the exact public-safe source, independently verifies two deterministic candidates and the isolated user lifecycle, atomically publishes one squashed v0.2.0 release commit and annotated tag over v0.1.0, and fast-forwards the existing public-backed integration source without changing its link ownership shape. The maintainer subsequently completed the requested full restart and fresh-session discovery smoke and reported that it passed.

Accepted v0.3 architecture and bounded sequence

APG15 proposed a synthesis-oriented capability architecture rather than one-to-one Codex VC migration. The maintainer accepted APG15 and ADR 0011 through APG16. APG18 accepts ADR 0012 after bounded contract correction and one Python vertical slice. APG17 accepts ADR 0013 for bounded pre-rewrite guidance classification.

APG15 — v0.3 foundation design

Completed and externally accepted as Complete — v0.3 foundation architecture proposed. The phase designed the public workflow router, guidance-synthesis candidate, ten language profiles, shared warning proposal, root-guidance migration, private-overlay boundary, ownership matrix, risks, and successor roadmap without implementing a skill or beginning release work.

APG16 — Public APG workflow router

Completed in private development. The maintainer subsequently confirmed the requested duplicate-name discovery result: both same-name routers were visible and selectable in the APG repository, while only the private router was visible and selectable globally. One provisional agentic-praxis-grimoire-workflow router selects among the six stable process leaves for ambiguity, routing audit, or capability-health diagnosis. Twenty-one frozen scenario families passed without candidate correction initially; after semantic review required one bounded non-trigger inspection correction, 21 fresh applications reran and passed the complete contract. A checked skill-local map enforces future catalog disposition. The phase does not modify the six leaves, replace private integration, implement profiles or synthesis, change stable maturity, or publish.

APG17 — Repository-guidance synthesis

Completed in private development. Its historical terminal outcome retained a pending fresh-session application smoke. One provisional synthesizing-repository-guidance leaf owns decomposition and bounded owner, provenance, privacy, migration, and rollback dispositions before rewrite. Sixteen frozen families passed with zero candidate corrections, and a 34-unit dogfood ledger exercised every authorized source class without changing or removing source guidance. The router map now covers seven routable leaves; the six-skill v0.2 distribution contracts remain unchanged. No language profile, root migration, private cutover, manager-prompt skill, release, or APG18 work occurred.

APG17A subsequently corrected two publication-excluded public-release identity fields without changing public state or any APG17 result. The maintainer deferred APG17 and APG18 application smoke to the aggregate APG23 readiness and APG24 release-preparation gates; that deferral does not reverse APG17 repository acceptance.

APG18 — Language-profile foundation and Python vertical slice

Completed in private development. ADR 0012 and the normative language-profile contract establish accessible Green/Yellow/Orange/Red responses, project-policy precedence, artifact classification, exception evidence, and rollback. One provisional python-language-profile passed twenty-four frozen scenarios with one bounded candidate correction and a seven-case read-only dogfood ledger. Private development contains six stable process skills and three provisional skills; the router map contains eight routable non-router capabilities. Public v0.2.0 and its project, user, and release lifecycle remain six-skill contracts. Application smoke remains deferred to APG23/APG24. Retention of one profile does not authorize the remaining nine or APG19.

APG19 — Shell and shell-test profiles

Completed in private development. ADR 0014 accepts separate language and test-harness ownership. Bash, Bats, and Zsh are retained-provisional; ZUnit is deferred-source-or-version. The corrected candidates use concrete monotonic thresholds, combined-signal crisis handling, frozen scenarios, and read-only dogfood. Private development contains six stable and six provisional skills, twelve canonical leaves and projections, and eleven routable non-router map entries. Public v0.2.0 and its six-skill lifecycle remain unchanged.

APG19A — Semantic phase identity and APG19 reconciliation

Completed under the maintainer’s two-phase authorization. ADR 0015 adopts globally unique semantic phase IDs, independent ADR and exit sequences, semantic durable references, precommit record finalization, and deterministic identity validation. APG19’s substantive dispositions remain accepted. One reproduced Bats fallback-count defect is corrected for the supported comment function form; Bash and Zsh remain byte-identical, ZUnit remains deferred, and catalog, maturity, distribution, and application-smoke boundaries are unchanged.

APG20 — Go and Ruby profiles

Completed under the maintainer’s APG19A-to-APG20 assignment. Independent Go and Ruby candidates completed current-source research, concrete threshold design, frozen scenarios, and one bounded design correction each. Fresh review found additional material semantic, measurement, source-license, dogfood, and classification defects. The one-correction limit required both candidates to be deferred-material-defect; no candidate leaf or integration is retained. Private development remains six stable and six provisional skills, twelve canonical leaves and projections, and eleven routable non-router map entries. Public v0.2.0 and its six-skill lifecycle remain unchanged.

APG20A — Go and Ruby profile corrections

Completed under the maintainer’s APG20A-to-APG21 assignment. APG20A uses the complete APG20 defect ledger as its corrected baseline, reproduces and fixes the report append-lock release race, and retains corrected Go and Ruby profiles as provisional development skills. Both profiles pass frozen and defect- specific scenarios, read-only calibration, and fresh non-author review. Private development contains six stable and eight provisional skills, fourteen canonical leaves and projections, and thirteen routable non-router map entries. Public v0.2.0, schema version 1, and active integration remain unchanged.

APG21 — Nix and relational-engine profiles

Completed under the maintainer’s APG20A-to-APG21 assignment. APG21 accepts ADR 0016, retains separate provisional PostgreSQL and SQLite profiles, and adopts no generic SQL owner. Nix is deferred-material-defect after one bounded correction and a second material scenario contradiction. PostgreSQL and SQLite pass frozen scenarios, read-only dogfood, and fresh non-author review. Private development contains six stable and ten provisional skills, sixteen canonical leaves and projections, and fifteen routable non-router map entries. No profile grants evaluation, build, live database, migration, activation, or destructive authority. Public v0.2.0, schema version 1, and active integration remain unchanged.

APG21A — Nix profile correction

Completed under the maintainer’s APG21A-to-APG22 assignment. APG21A uses the complete APG21 Nix defect ledger as its corrected baseline, separates structural merge-family breadth from semantic merge behavior, and retains the corrected Nix profile provisionally. One newly discovered PostgreSQL false-escalation contradiction is corrected with failing evidence and focused re-review; SQLite and router behavior remain accepted. Private development contains six stable and eleven provisional skills, seventeen canonical leaves and projections, and sixteen routable non-router map entries. Public v0.2.0, schema version 1, and active integration remain unchanged.

APG22 — Cross-repository dogfood and guidance migration proposal

Completed under the maintainer’s APG21A-to-APG22 assignment. APG22 applies the router, synthesis procedure, and nine retained profiles to 35 frozen read-only APG, RepoMap, private-classification, and public-safe synthetic cases. All expected dispositions match, with no behavior-bearing APG defect. Application discovery remains unavailable and is not inferred from explicit reading. The phase records bounded root and private transition proposals without performing a cutover, decommission, public release, active-integration mutation, or application smoke.

APG22A — Approved-roadmap manager assignments

Completed under the maintainer’s APG22A-to-APG22B assignment. APG22A retains composing-approved-roadmap-assignments provisionally after an exhaustive historic-corpus inventory, current native-capability review, strong ordinary- prompting baseline, and 30/30 frozen candidate cases. The leaf translates approved authority without approving, dispatching, executing, accepting, or continuing work. Development becomes 18/18/18 with 17 routes; the six-skill public v0.2 boundary remains unchanged.

APG22B — Version-bounded ZUnit profile

Completed under the maintainer’s APG22A-to-APG22B assignment. APG22B retains a provisional zunit-test-profile for exact ZUnit v0.8.2 with Zsh 5.9.2 in the tested environment. The exact ZUnit v0.8.2 with Zsh 5.3.1 pair is unsupported because its dependency probe does not complete. The retained leaf owns ZUnit runner, discovery, assertion, output, fixture, hook, isolation, and lifecycle judgment while pairing with the Zsh profile for shell semantics. It passes 48/48 frozen cases with zero candidate corrections. Development becomes 19/19/19 with 18 routes; the six-skill public v0.2 boundary remains unchanged.

APG22C — ZUnit startup-isolation evidence correction

Completed under the maintainer’s APG22C-to-APG23 assignment. APG22C reproduces that the APG22B startup sentinel was outside the path selected by ZDOTDIR, then corrects the disposable harness to require a loading positive control, a suppressing negative control, and sentinel absence inside the focused ZUnit test process. The corrected exact matrix retains ZUnit v0.8.2 with Zsh 5.9.2, preserves Zsh 5.3.1 as unsupported because its runner dependency probe times out, and leaves the ZUnit skill bytes and zero-correction count unchanged. Development remains 19/19/19 with 18 routes; public v0.2 remains 6/6/6. Application smoke and APG23 do not occur in the APG22C session.

APG23 — Individual readiness and maturity review

Completed after the required full restart in a new APG-rooted session. All nineteen repository skills were directly discoverable and selectable, the duplicate workflow router was source-qualified with client path evidence, and all required router, synthesis, profile, and process/domain applications passed. Eight v0.3 rows are promoted to stable, five remain provisional, all thirteen are include-v0.3, and no behavior correction is required.

APG24 — v0.3 release candidate and publication

Completed under the maintainer’s explicit APG24 publication authority. ADR 0019 expands the release, user, and project distribution contracts to the exact nineteen-skill v0.3.0 set while retaining schema version 1, source-specific user rollback, and existing project-subset ownership. Two deterministic candidates precede one appended public commit and annotated tag. The active public-backed source fast-forwards without changing aggregate-link ownership. The personal router remains installed, and source-qualified fresh-session shadow smoke remains the external terminal observation.

APG24A — v0.3 external smoke and router transition closeout

Completed under the maintainer’s explicit APG24A authority. The human-reported fresh-session public-v0.3 shadow passed, and the personal same-name router was subsequently decommissioned under separate authority. Focused verification confirms unchanged public and active v0.3.0 state, nineteen aggregate skills, preserved personal transition targets, and an exact private restoration source. No behavior-bearing skill, release, active source, or target repository changes.

APG25 — v0.4 structured project work foundation

Completed under the maintainer’s APG24A-to-APG26 assignment. ADRs 0020-0022 accept structured formal/non-phase defaults, prompt compression, scoped tests, coverage remediation and mock boundaries, Python-first reporting, pytest/xdist coverage architecture, nested ChatGPT-manager ownership, and evidence-gated personal hygiene transitions. One bounded correction makes the approved-roadmap assignment skill consume repository defaults without dropping authority or stop boundaries. Development remains 19/19/19 with fourteen stable and five provisional rows; no dependency, executable conversion, test migration, canonical move, private-skill change, or release occurs.

The complete dependency order and unallocated slices are maintained in the v0.4 roadmap rather than duplicated here.

APG26 — pytest and Bash-to-Python capabilities

Completed under the maintainer’s bounded sequence. APG26 retains provisional pytest-test-profile and converting-bash-scripts-to-python after current- source calibration, sixty frozen scenario families, failing-first focused contracts, read-only report-tool dogfood, and fresh non-author review. Pytest uses no behavior-bearing correction; conversion uses one bounded option- injection correction. Development becomes 21/21/21 with fourteen stable rows, seven provisional rows, and twenty routes. Public and active v0.3.0 remain unchanged; no report executable or test is migrated.

APG26A — formal-phase commit-message enforcement

APG26A records that APG26’s formal-phase commit has only its subject. APG26 is not rewritten, and its substantive capability, scenario, routing, and maturity decisions remain accepted. A dependency-free Python checker enforces the canonical phase subject, one blank separator, and ordered nonempty Scope, Result, Verification, and Not run sections before and after future APG formal-phase commits. The correction changes no skill, report format or executable, public or active v0.3.0 state, target repository, or personal skill.

APG27 — Python agent-reporting core and Git-record association

APG27 accepts ADR 0023 and produces an uncommitted candidate for the first report slice. The candidate implements the importable standard-library core, fixed-vector Git access, exact Git-show compatibility, temporary-index Git-diff evidence, operational association, and safety corrections. It is not adopted.

Git-show format version 2 and compatible standalone operational format version 1 remain byte-compatible on the characterized POSIX platform. Git-diff format version 1 records deterministic state evidence without mutating the real index or worktree. Operational records name an existing complete show or diff record in the same canonical phase report whenever Git evidence exists. APG27 stops partial after the release-policy component needs a second material correction: historical v0.2 configured validation is repaired, but immutable v0.3.0 policy lacks a version-bounded historical surface. The candidate also retains a Red Python branch-count stop. No commit or push occurs.

APG27A — Python reporting correction and adoption

APG27A preserves APG27’s partial records and stopped worktree, then corrects both recorded acceptance defects. Immutable v0.3.0 policy is independently frozen from the current v0.4 development inventory, malformed and unsupported identities fail closed, and the Red source-path validator is decomposed without changing diagnostics. The resulting dependency-free Python report core, Git-diff evidence, and Git/operational association are adopted after focused parity, safety, historical-release, and independent review gates.

APG28 — pytest, xdist, coverage, and mirrored paths

APG28 closes Partial. Its uncommitted candidate includes an exact dependency stack, eight-worker runner, mirrored paths, strict source inventory, and coverage-data union. Fresh review rejected the first narrowed denominator; the corrected full-source gates then exposed material unit, integration, and union branch deficits, four release-policy failures, incomplete process accounting, and missing unconditional Git-show Bats equivalence. ADR 0024 remains Proposed. Public and active v0.3.0 remain unchanged.

APG28A — pytest migration correction and adoption

APG28A preserves APG28’s stopped worktree and Partial result, corrects runner aggregation, release policy, process completeness, artifact lifecycle, and coverage quality, and satisfies exact 80/80 component and 85/85 union gates. Both Bats owners remain. ADR 0024 is Accepted. Public and active v0.3.0 remain unchanged.

APG29 — process-skill alignment

APG29 aligns four existing owners with the adopted structured-project and test defaults. implementing-with-test-discipline owns bounded useful-contract coverage remediation; planning-repository-work owns scoped evidence and justified expansion; reviewing-and-verifying-repository-work checks honest coverage, mock and real-boundary claims, and Git/operational evidence; and composing-approved-roadmap-assignments never assumes a broad suite while compressing ordinary procedure. All four receive one bounded correction after fifty-one frozen cases. No trigger, maturity row, route, catalog description, release object, ChatGPT path, personal skill, or target repository changes.

APG30 — ChatGPT-manager topology and subrouter

APG30 implements ADR 0022’s accepted actor-qualified topology. Canonical discovery now supports direct leaves and exactly one skills/chatgpt/<name> class, while flat Codex projections remain globally named. The new provisional chatgpt-manager-workflow owns one local manager-leaf route; the general router owns the subrouter route and no manager leaf. Project, user, and current-development release owners resolve source-declared canonical paths, while immutable v0.1.0 through v0.3.0 policy remains direct-child.

Development is 22/22/22 with fourteen stable rows, eight provisional rows, twenty general-map entries, one ChatGPT-local entry, and twenty-one route edges. Public and active v0.3.0 remain 19/19/19. No personal skill is changed and no post-restart application-discovery evidence is claimed.

APG31 — personal-hygiene shadow and conditional transition

APG31 passes APG30’s fresh-session source-qualified topology, routing, semantic-identity, and managed-report gates without correction. Current APG, private, and RepoMap owners replace the three targeted general procedures, but the independent transition gates produce a partial result.

The personal docs-only capability is decommissioned after source-qualified shadow, exact restoration, and non-author review. Git-history scope reduction remains deferred because current private routing and destructive-stop ownership do not yet satisfy the narrowed contract. RepoMap phase-hygiene decommissioning remains deferred because a current private caller route is outside APG31 write authority. APG development stays 22/22/22, public and active v0.3.0 remain 19/19/19, and no target repository changes.

APG31 subsequent gate

At APG31 exit, the maintainer still had to restart Codex and run the required post-transition discovery smoke. Any later assignment had to record or correct that result before a remaining v0.4 profile slice could begin.

APG31A later supplies the bounded smoke-closeout and caller-correction authority without altering that historical stop.

APG31A — personal-hygiene transition completion

APG31A records the maintainer-reported APG31 external smoke as passed while preserving APG31’s historical Partial result. Publication-excluded review supports two independent results: git-history-hygiene is scope-reduced after generalized behavior returns to APG and repository owners, and repomap-phase-hygiene is decommissioned after generalized behavior returns to current RepoMap owners.

APG development remains 22/22/22 with fourteen stable and eight provisional rows. The general and ChatGPT-local maps remain twenty and one. Public and active v0.3.0 remain 19/19/19, and no target repository changes.

APG32 — Minitest test profile

APG32 retains minitest-test-profile provisionally after current official Minitest 6.0.6, minitest-mock 5.27.0, and Ruby 4.0.6 calibration; source and rights review; ownership analysis; thirty-six frozen scenarios; a failing-first mirrored contract; structural classification; integration; and fresh non-author review. The candidate uses one bounded trigger and ownership correction.

Development becomes 23/23/23 with fourteen stable and nine provisional rows. The general map becomes twenty-one edges, the ChatGPT-local map remains one edge, and public and active v0.3.0 remain 19/19/19. APG32 adds no dependency, selects no framework or project command, runs no readiness or smoke gate, and does not construct or publish a release.

APG33 — Dockerfile profile

APG33 retains dockerfile-profile provisionally after current official Docker documentation, stable Dockerfile frontend 1.25.0, BuildKit 0.31.2, and OCI Image Spec 1.1.1 calibration; source and rights review; ownership analysis; forty frozen scenarios; a failing-first mirrored contract; structural classification; integration; and fresh non-author review.

Development becomes 24/24/24 with fourteen stable and ten provisional rows. The general map becomes twenty-two edges, the ChatGPT-local map remains one edge, and public and active v0.3.0 remain 19/19/19. APG33 adds no dependency, selects no image, builder, platform, project command, or runtime policy, performs no Docker operation, runs no readiness or smoke gate, and does not construct or publish a release.

APG34 — Vagrantfile profile

APG34 retains vagrantfile-profile provisionally after current official Vagrant 2.4.9 and development-source calibration, current Vagrant documentation, Vagrant’s declared Ruby compatibility boundary, source and rights review, ownership analysis, forty frozen scenarios, a failing-first mirrored contract, structural classification, integration, and fresh non-author review. One bounded source-semantics and machine-measurement correction scopes forwarded-port behavior by provider and prevents an implicit default machine from being added to named multi-machine counts.

Development becomes 25/25/25 with fourteen stable and eleven provisional rows. The general map becomes twenty-three edges, the ChatGPT-local map remains one edge, and public and active v0.3.0 remain 19/19/19. APG34 adds no dependency, selects no provider, box, plugin, host platform, network, synced folder, provisioner, project command, or lifecycle action, performs no Vagrantfile evaluation or Vagrant operation, runs no readiness or smoke gate, and does not construct or publish a release.

APG35 — v0.4 remaining-skill authoring

APG35 authors, on the dedicated branch claude/apg35-v0.4-remaining-skill-authoring, the five remaining v0.4 candidates: go-test-profile, matryer-is-test-profile, go-cmp-test-profile, go-testing-stack, and nix-test-profile. It also proposes ADR 0025 and two specifications, freezes one hundred fifty-four scenario families, and produces a complete Codex integration handoff.

APG35 integrates nothing. It adds no flat projection, catalog row, router entry, release-policy entry, strict inventory entry, executable fixture, or test, and it changes no Python or shell code. It runs no project test, no go test, and no Nix evaluation, build, flake check, or virtual-machine or container test. Development therefore remains 25/25/25 with fourteen stable and eleven provisional rows, the general map remains twenty-three edges, the ChatGPT-local map remains one edge, and public and active v0.3.0 remain 19/19/19. ADR 0025 remains Proposed.

The phase’s fresh non-author review requirement was not satisfied as specified: independent reviewers could not be spawned, so the eight lanes were executed by the author as separate adversarial passes with primary-source re-checks. That limitation is recorded in the evaluation, the exit, and the handoff, and Codex must re-run the affected lanes independently before accepting ADR 0025.

APG36 — Claude-authored skill integration

APG36 preserves and adopts the exact APG35 authoring commit, reruns six fresh independent review lanes, converts all 154 frozen families into transient public-safe fixtures, establishes failing-first mirrored contracts, and runs a disposable pinned Go compatibility harness.

Independent review finds more than one material behavior correction in go-test-profile, matryer-is-test-profile, go-cmp-test-profile, and nix-test-profile; each is deferred-material-defect. go-testing-stack is rejected-no-independent-value because its required native owner is not retained and its trigger, duplicate-owner response, thinness, and removal semantics need multiple corrections. ADR 0025 is Rejected.

All five candidates and both proposed specifications are removed through the APG36 forward commit. Development remains 25/25/25 with fourteen stable and eleven provisional rows, the general map remains twenty-three entries, the ChatGPT-local map remains one entry, and checked route edges remain twenty-four. Public and active v0.3.0 remain 19/19/19.

The Claude/Codex division-of-labor trial is supported for another bounded trial only with immutable author history, independent Codex review, executable fixtures, a one-correction allowance, forward-only corrections, and Codex-owned final status, tests, reports, publication, and deployment.

APG37 — Go and Nix test-profile redesign

APG37 redesigns the four APG36-deferred candidates on the authoring branch claude/apg37-v0.4-go-nix-redesign, working from the APG36 defect dossier and reverified current primary sources. It authors four replacement leaves, proposes ADR 0026 and two specifications, freezes 130 scenario families mapped to their APG35 predecessors, measures structural calibration against the complete Go 1.25.10 standard-library and toolchain test corpus, and produces a publication-excluded Codex integration handoff.

All twenty-four APG36 material findings received a terminal authoring disposition: twenty-two closed in redesign and two removed as invalid requirements. Measurement falsified the APG35 numeric structural bands, which escalate 9.7% of maintained upstream Go test files to crisis on line count alone, so every numeric crisis cutoff was removed in favour of categorical conditions.

ADR 0026 proposes three independent component owners and no composition owner. go-testing-stack remains rejected-no-independent-value and absent, and ADR 0025 remains Rejected and unreopened.

APG37 integrates nothing. It adds no flat projection, catalog row, router entry, maturity row, release policy, inventory entry, executable fixture, or test, and runs no project test, go test, or Nix operation. Its review was an author self-review, not an independent review, and it makes no retention prediction. Development remains 25/25/25 with fourteen stable and eleven provisional rows, twenty-three general-map entries, one ChatGPT-local entry, and twenty-four checked route edges. Public and active v0.3.0 remain 19/19/19.

APG38 — APG37 Go and Nix integration

APG38 preserves the exact APG37 authoring object and performs independent source, rights, privacy, owner-graph, structural, fixture, and compatibility review. It retains provisional go-test-profile and go-cmp-test-profile; accepts ADR 0026’s two-component no-stack architecture; and keeps ADR 0025 Rejected.

matryer-is-test-profile and nix-test-profile are deferred-material-defect after fresh corrected-state review finds new attribution/false-escalation and FreeBSD sandbox-default behavior defects. Each candidate has already used its coherent correction cycle, so its leaf and integration surfaces are forward removed rather than corrected again.

Development becomes 27/27/27 with fourteen stable and thirteen provisional rows, twenty-five general-map entries, one ChatGPT-local entry, and twenty-six checked route edges. Public and active v0.3.0 remain 19/19/19.

APG39 — matryer/is and Nix final redesign

APG39 authors final replacement candidates for the two APG38-deferred profiles on a Claude authoring branch. It reverifies matryer/is v1.4.1, Nix 2.35.1, and Nixpkgs/NixOS 26.05 as exact current sources; closes the registered-wrapper attribution, relaxed-mode count-escalation, and sandbox platform-default defects; freezes 24 and 40 replacement scenario families with complete predecessor mapping; proposes ADR 0027 conditionally for a version-bounded third Go component; and hands independent review, fixtures, compatibility probes, the ADR decision, and atomic integration to a later separately authorized Codex phase.

Both candidates are authored-pending-independent-review. Development remains 27/27/27 with unchanged maturity, router, and edge counts. Public and active v0.3.0 remain 19/19/19.

APG40 — APG39 matryer/is and Nix integration

APG40 preserves and adopts the exact APG39 authoring object, converts both frozen scenario sets into corrected manager evidence, runs exact-version matryer/is probes and mandatory Nix source/corpus review, and applies one coherent correction pass per candidate. Corrected-state review retains nix-test-profile provisionally and defers matryer-is-test-profile after its equality correction remains materially inaccurate. ADR 0027 is Rejected; ADR 0026 remains Accepted; ADR 0025 remains Rejected; go-testing-stack remains absent.

Development becomes 28/28/28 with fourteen stable and fourteen provisional rows, twenty-six general-map entries, one ChatGPT-local entry, and twenty-seven checked route edges. Public and active v0.3.0 remain 19/19/19. No Nix execution, target-repository testing, readiness, smoke, release, publication, deployment, or successor phase is included.

APG41 — v0.4 readiness and pre-release smoke

APG41 retains all fourteen provisional rows without maturity promotion after complete trigger, non-trigger, owner-boundary, adverse, limitation, and removal review. Cross-profile dogfood preserves direct selection and no mandatory chain. One bounded wording correction makes Minitest, Dockerfile, and Vagrantfile removal candidate-independent; fresh corrected-state review passes.

Complete repository gates, two repository-owned disposable v0.4.0 candidates, and isolated user/project lifecycle smoke pass on the current host. The terminal result is ready-for-publication-with-provisional-limitations. Public and active v0.3.0 remain 19/19/19. APG41 publishes and deploys nothing and authorizes no successor phase.

APG42 — v0.4 publication and active deployment

APG42 reproduces the exact APG41 candidate, retains the canonical Agentic Praxis Grimoire NOTICE bytes, freezes the release source, builds and independently reviews two deterministic final candidates, and atomically publishes only public main and the annotated v0.4.0 tag. A live remote and fresh clone verify the append-only lineage before the aggregate-owned active public source advances by exact fast-forward.

Public and active v0.4.0 contain 28/28/28 with fourteen stable and fourteen provisional rows. ADR 0025 and ADR 0027 remain Rejected; ADR 0026 remains Accepted; matryer/is and the stack remain absent. No procedure or maturity changes, GitHub Release, signing, announcement, plugin distribution, target mutation, or successor phase are included.

Next action

No phase after APG42 is authorized. Signing, announcement, GitHub Release, plugin publication, target mutation, rollback, and any successor work require separate explicit maintainer authority. APG42 does not allocate a successor phase.

This site is open source. Improve this page.