Tracker: `AGENTS.md > Where the Rules Live`

[ptr727]

Tracker for AGENTS.md > Where the Rules Live. Filed under #1315's roll-up of the open prose backlog: every finding a bundle issue enumerated against this unit is listed here once, one line each, naming the issue it came from, and the bundle is closed as a duplicate of this tracker so the finding is not lost and the bundle stops being a unit of work. A new pre-existing finding on this unit gathers here per pr-review-conduct's Merge Gate, numbering continuing this list. This tracker closes when the unit is rewritten and each line below is settled or superseded.

1. AGENTS.md:108 says the skills path is hub-local and not carried into every fleet repo, while spec/files.json carries .github/skills to every repository. (Carried Text Says the Skills Are Not Carried, While the Manifest Carries Them to Every Repo #1184)

[ptr727]

Appended from the class 6 whole-unit read of this section (#1388), which changed the "Editing a workflow YAML file" row, numbering continuing the list above.

2. The "Editing a workflow YAML file" row's "surfaced with the full WORKFLOW.md contract by the workflow-ci-contract Skill" sends a reader to a skill whose own "Why This Exists" says it carries the summary, section 4 whole as an include, and an index of section 5, so the rest of WORKFLOW.md is still read from the document. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)
3. The closing paragraph says Supported Development Platforms, Devcontainer, and Editor and Tasks are "each already routed to by the procedures and Skills that need it", and no heading reference or anchor to those three exists in STANDUP.md, AUDIT.md, RESYNC.md, the spec files' intent refs, or any Skill, so the reason given for leaving those rows Skill-less is a claim the tree does not hold. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)
4. The "Reviewing a pull request" row says the code-review Skill "routes to the applicable general, language, documentation, and workflow skills", and that skill's "Establish the Contract" routes to the three language skills, comment-and-doc-style, and workflow-ci-contract with no general skill existing, so the row names a target a reader cannot find. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)

[ptr727]

Five more from the class 6 whole-unit read of this section on #1388, run at the opus tier after the change added a connector to the "Editing a workflow YAML file" row, numbering continuing the list in the comment above.

5. "refreshing this machine's installed Skills first when the repo is this hub" states an order merge-and-release inverts: that skill puts the refresh at step 7, after the merge, the release dispatch, and the bounded run watch, and runs it regardless of whether the dispatch happened, so an agent following this sentence literally refreshes before dispatching. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)
6. "invoked separately from drive-pr so the promotion merge and the release dispatch each keep their own explicit go-ahead" is false of the skill it describes, whose "What Invoking This Skill Authorizes" grants both in one invocation and whose "How Far to Go" says to act on "merge and release" without asking again. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)
7. "hand-authored at .agents/skills/ in the hub", closing with "the generated plugin tree is never hand-edited", is stale since include regions landed: build_dist.py calls that tree the one place a skill is hand-edited "its include regions excepted", three regions in agent-conduct are generated, and --check fails a pull request that edits one, so the source-hand-authored versus plugin-generated dichotomy tells a reader the whole source tree is safe to edit. This file's own opening paragraph already states the include mechanism. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)
8. Hub-Hosted Tooling is listed among "short reference sections a task reads at the moment it touches their subject", and it is neither short (911 words against 45 to 334 for the other five, and more than double the Foundational Principles the same paragraph sets aside) nor occasional, being always-on law about what a repository carries versus reaches that RESYNC.md, AUDIT.md, and three skills each cite. Distinct from the routing claim in the same sentence, filed as item 3. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)
9. carried-instruction-file-guard is named nowhere in this unit and no row routes to it, while RESYNC.md step 1 mandates its probe before touching AGENTS.md, GOVERNANCE.md, CODESTYLE.md, or WORKFLOW.md every time without exception, so an agent reaching this map before re-vendoring a carried instruction file is never routed to the skill written for that destructive incident. It is the only skill under .agents/skills/ this file mentions nowhere. [AGENTS.md > Where the Rules Live] (Cut WORKFLOW.md Section 2 to a Pointer at Workflow YAML Conventions #1388)

[ptr727]

Five more from a #1311 second-round pass (complete-brief whole-unit read, opus tier, no PR — read-only) over this section, numbering continuing the list above. The reviewer confirmed items 2-9 all still present as worded without re-deriving any of them, and reports the brief worked as a suppressant: every candidate finding adjacent to a filed item's own sentence was checked against the brief and discarded where it restated one.

10. The "Reviewing a pull request" row promises routing to "the applicable general, language, documentation, and workflow skills," but no row in this table names a language skill, mentions CODESTYLE.md, or mentions .NET, Python, or Shell code style anywhere — a reader consulting this map for "what governs the file I am writing" in a language context finds no route at all, distinct from item 4 (which is that "general" names a skill that does not exist; this is that "language" names skills that exist but are never listed here). [AGENTS.md > Where the Rules Live] (Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311 second round)
11. "run the installer ... once per machine ... installs them for every repo touched from that machine" is contradicted two sentences later by "--report ... says whether this machine is current" and by merge-and-release's per-release refresh (the subject of item 5) — a one-time install cannot go stale, need a currency report, or need refreshing at each release, and nothing in this paragraph reconciles "once" with "current"/"stale". [AGENTS.md > Where the Rules Live] (Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311 second round)
12. The closing paragraph's own promise ("A row above naming no Skill, or naming one only for part of its section, is doc-only by decision rather than by omission, and the reason differs by row") gives reasons for five cases and omits three: the "Durable Knowledge and Self-Improvement", "Verification Discipline", and "Communicating with the User" rows each read "surfaced at its decision moment by the agent-conduct Skill, and the section keeps the full rules" — textbook partial naming — with no reason given for any of them, while the same paragraph frames partial naming as peculiar to the boundaries row alone. [AGENTS.md > Where the Rules Live] (Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311 second round)
13. carried-instruction-file-guard is confirmed absent from this unit by direct search — named nowhere, routed to by no row — while this is the map an agent consults before touching a carried instruction file, and (per the tracker) RESYNC.md step 1 mandates that skill's probe before every such edit. Sharpens item 9 with an explicit verification rather than a new claim. [AGENTS.md > Where the Rules Live] (Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311 second round)
14. Item 6 is independently falsifiable from this unit alone, without fetching merge-and-release: "invoked separately from drive-pr so the promotion merge and the release dispatch each keep their own explicit go-ahead" is a non-sequitur on its face, since separating merge-and-release from drive-pr produces exactly one separation and cannot give the merge and the release each their own go-ahead when the unit's own preceding clause already places both inside the single merge-and-release package ("Merging a ready promotion PR and dispatching the release it unblocks ... is merge-and-release"). [AGENTS.md > Where the Rules Live] (Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311 second round)

New supporting evidence for item 5, not a new number: item 11 above shows the "refresh first" order is now also contradicted from inside this same unit's own "once per machine" paragraph, not only from merge-and-release's own text.