Phase ID: APG22A
Complete — approved-roadmap manager-assignment skill implemented.
APG22A retains composing-approved-roadmap-assignments provisionally. The
skill translates one human-approved roadmap phase, or an explicitly approved
bounded sequence, into one reviewable top-level manager assignment. It does
not approve, revise, dispatch, execute, review, accept, or continue the work.
The complete authorized historic corpus contained 63 tracked prompt artifacts: 18 control or orchestration prompts, 8 runner prompts, and 37 development prompts. It supports one reusable translation boundary but also contains project mechanics, private detail, inconsistent forms, and no established public copying grant. The candidate was therefore written independently and has no public dependency on that corpus.
Current official Codex documentation confirms native skill authoring, repository instructions, worker delegation, task continuation, and ordinary prompt generation. Two fresh ordinary-reasoning workers produced strong, proportional assignments for all 14 positive scenarios before candidate wording. Ordinary prompting remains the supported fallback; APG22A makes no comparative-superiority claim.
The contract froze 14 positive, 6 non-trigger, and 10 fault or authority cases before candidate wording. Thirty fresh clean-context workers each received one case, the candidate leaf, and only directly named public APG owners.
| Scenario class | Result |
|---|---|
| Positive assignments | 14/14 reviewable and proportional |
| Non-triggers | 6/6 routed to direct work or the existing owner |
| Fault and authority stops | 10/10 stopped safely |
| Material candidate corrections | 0 |
Missing project-owned IDs, paths, commands, report mechanics, and acceptance actors were surfaced as blockers rather than fabricated. Planning, worker assignment, routing, review, and acceptance remained separate owners. Every case preserved the no-successor boundary.
The phase adds one canonical leaf, one relative checked-in projection, one provisional catalog row, one capability-map entry, focused semantic and route tests, and one development-only known-unmanaged name. Resulting private development is 18/18/18 with 6 stable rows, 12 provisional rows, and 17 routable non-router entries.
Resulting-state review also corrected the live catalog heading from the
historical APG v0.1 catalog label to Current development catalog. The
checker accepts that current heading and the legacy heading retained by older
public releases; a focused compatibility test protects both forms.
Project, user, and public-release lifecycle defaults remain exactly the six stable v0.2 skills under schema version 1. Public v0.2.0 and active integration are unchanged.
ADR 0017 accepts the bounded translation owner. The retained name is
composing-approved-roadmap-assignments. No supporting template is added;
the leaf’s proportional core and risk-dependent modules are sufficient and
avoid a mandatory giant prompt form.
Focused failing-first evidence showed the absent leaf and map route. The resulting 41-test skill-library unit suite and 18/18/18 checker pass. All 248 configured tests pass with two expected fixture skips. Development and active public-backed text/JSON checks, record identity, Python compilation, applicable Bash syntax, wrapper help, changed-Markdown links, privacy, durable-reference, whitespace, and three fresh non-author review lanes pass. Final staged checks remain the last precommit gate.
APG22A changes no stable skill, dependency, schema, public release, active integration, private prompt, private skill, target repository, application state, or maturity. It performs no application smoke and does not begin APG22B until the separate APG22A commit, push, remote-parity, and report gate passes.