Nice, tight migration: one logical change (generators emit per-page metadata as YAML front matter, drop the leading H1) across both doc generators plus the mechanical regeneration of 166 CLI + 31 API pages. The AI-authorship disclosure and manual verification are appreciated, and several reviewers independently confirmed the load-bearing claims: manifest.json/swagger.json/docs.go are byte-identical in the diff, MD041 is disabled so the H1 removal breaks no lint rule, and the front matter is well-formed (colons, backticks, and DEPRECATED: prefixes are correctly JSON-quoted, state renders as a YAML sequence). Hisoka's verdict on the code itself: "Boring code, and I mean that as a compliment."

Severity count: 0 P0/P1, 5 P3, 3 Nit, 7 Note. No blocking defects; the P3s are correctness/maintainability gaps worth resolving before merge. Because this panel assumes no follow-up (agent-authored, no backlog), CRF-5, CRF-6, and CRF-9 each need a decision now, not "later."

Inline findings CRF-1 through CRF-11 below. Four items belong here in the body:

CRF-12 (P3, process, Mafu-san). The PR's "no body changes" proof is a vacuous command: git diff docs/reference/ | grep '^-' | grep -v '^---' with no ref compares the working tree against HEAD, which on a committed tree is empty regardless of what the PR changed. The underlying claim is in fact true (all 197 base..HEAD removals are # H1 lines), but the command shown would print (empty) even if a real body removal had slipped in, and it inspects removals while the comment describes added lines. Show the command that actually tests the claim: git diff <base>..HEAD -- docs/reference/ | grep '^-' | grep -v '^---'.

CRF-13 (Nit, commit hygiene, Leorio). Commit e54fdada uses scope feat(docs):, but its source changes are under scripts/, and AGENTS.md requires a scope to be a real path containing every changed file. This is the cross-cutting case that should be scope-less (feat:), as the later commit 7791007f correctly is. Moot under squash-merge on the (correct) PR title.

CRF-14 (Note, cross-repo gate, Hisoka + Mafu-san + Melody). The PR gates its merge on coder.com#968. Prior deep-review knowledge (vault, unverifiable from this repo) suggests #968 was re-scoped and the llms + Algolia front-matter awareness actually shipped in #964/#974, and that a second raw-markdown surface, the .md proxy twin advertised via <link rel="alternate" type="text/markdown">, keeps serving raw front matter until #974 deploys. Please confirm which coder.com PR(s) actually strip/consume the front matter so the manual merge-ordering decision rests on the right surface(s). This is an inference from prior review context, not something the panel could execute here.

CRF-15 (Note, pre-existing, Komugi). sort.Slice at main.go:220 uses slices.IsSorted([]string{a, b}) as its Less, which returns true for equal keys in both directions, undefined behavior per sort.Slice. Pre-existing (blame ed908ed019), safe today only because API section names are unique. Flagging for the record; not this PR's job to fix.

docs/support/support-bundle.md:70

P3 [CRF-6] Dropping the leading # {name} H1 orphans the page-title anchor #login, and this doc still links to it. (Hisoka)

docs/support/support-bundle.md:70 points at ../reference/cli/login.md#login, and base login.md opened with # login (anchor #login); the regenerated page's first heading is now ## Usage, so #login resolves to nothing.

Confirmed: login.md now starts with ## Usage; front matter does not mint a heading anchor. A full-repo sweep found this is the only live instance (every other reference/*.md#... link keys on a ##/### sub-heading such as #--flag or #get-users, which are unaffected). pnpm check-docs is markdownlint + table-formatter only and does not validate cross-file fragments, so this ships green and silent. Consequence is cosmetic (the reader still lands at the top of login.md), hence P3. Fix: drop the fragment (../reference/cli/login.md).

🤖

docs/reference/api/chat.md:1

Note [CRF-9] A stale orphan page survives the API-docs regeneration and now diverges from the migration's "every generated page is self-describing" invariant. (Kite)

docs/reference/api/chat.md contains only # Chat. It is not in docs/manifest.json, is not linked from any doc, and is not one of the 31 generated sections (chats.md is the real page).

Confirmed: not in the manifest, content is just # Chat. The API pipeline never wipes docs/reference/api/ (the copy-back is per-file), so files with no generated counterpart are never removed. This PR converts every generated API page to front matter, but chat.md keeps a bare H1 and no front matter, so after Phase 3 it becomes a permanent front-matter-less page the generator cannot touch. A one-line git rm docs/reference/api/chat.md closes the gap while this diff is already in the directory; otherwise it needs a ticket, not a silent drop.

🤖

🤖 This review was automatically generated with Coder Agents.

Strong round-2 turnaround. Every one of the 15 round-1 findings was engaged with, and the panel verified the fixes hold rather than taking the replies on faith: the escaper and the manifest schema/traversal are genuinely de-duplicated into one scripts/docgenenv package (CRF-1/CRF-4), the reserved-scalar quoting is proven by a real yaml.v3 round-trip test that also survived timestamp/hex/underscore-number probes (CRF-2), both index pages now mirror their manifest route (CRF-5), the #login anchor class is fully closed with no siblings (CRF-6), and prependFrontMatter/extractSectionName now share an enforced heading contract (CRF-11). CRF-12's vacuous verification command was fixed at the reasoning level, not just reworded. Nice work. Hisoka's read of the code: "You came looking for a fight and found the code had already won it."

Round-2 severity count: 1 P1, 1 P3, 2 Nit, 4 Note. The one blocker:

CRF-16 (P1) is why CI lint (job 88677231209) is red: three emdash U+2014 characters in the new package's doc comments. It is mechanical to fix but it blocks merge, and it is the inline comment below. This connects to a process point worth raising here in the body:

Mafu-san (P2, process). The PR's "Verification & evidence" section lists go build/go vet/go test and pnpm check-docs as passing but omits make lint, which is the one CI gate that is red. The listed checks are truthful, but the verification set was assembled from a subset that skipped the mandatory gate, so a reader trusting it believes the PR is merge-ready when it is not. AGENTS.md lists lint as "always run after changes." The deeper pattern: the emdash ban lives in AGENTS.md, the docs style guide, make lint, and the agent's own prompt, yet three fresh emdashes landed in comments written during this fix cycle. Before presenting a PR as verified, run make lint, and describe the verification set by what the CI gates require, not by what happened to pass. Fixing CRF-16 clears the gate; wiring make lint into the pre-submit habit clears the class.

One structural item (CRF-18) and four low-severity items are inline. Because this panel assumes no follow-up, the CRF-18 divergence trap is worth closing now while the abstraction boundary is fresh, but it is not a merge blocker on its own.

🤖 This review was automatically generated with Coder Agents.

This round closes the book on the round-1/round-2 findings. All 24 prior findings are addressed or accepted, and the panel verified the fixes at the code rather than trusting the replies: front-matter emission is now a single shared docgenenv.FrontMatter(Route) reached by the API postprocessor directly and by command.tpl via a frontMatter func, so the two emitters can no longer drift on field set or escaping (CRF-17/18); the CLI index carries its state arm; slices.SortFunc replaced the index comparator; the prependFrontMatter comment now names extractSectionName as the load-bearing guard; icon_path and trailing-space scalars have real coverage; and the emdash gate is green. Netero found nothing, and six panelists returned clean. The tests are honest: TestYAMLScalarRoundTrip parses the emitted YAML back and asserts string identity rather than pinning golden bytes. Ryosuke's read: "When the architecture is clean, say so: this is clean."

I'm dismissing the prior REQUEST_CHANGES: the P1 that blocked it (CRF-16 emdash / red lint) is fixed and CI is green, so this is no longer blocked. Round-3 severity count: 0 P0/P1/P2, 1 P3, 2 Nit, 2 Note. None block merge; the P3 is a maintainability/regression-trap worth closing while the code is fresh, but it guards a future field, not a present bug.

One correction to a panel finding, recorded for honesty: a reviewer flagged that make gen is not manifest-idempotent (a fresh clidocgen run adds 7 Command Line nav children). Verified against the Makefile: make gen runs clidocgen in a tmpdir and copies back only the .md pages, discarding its rebuilt manifest; the committed docs/manifest.json comes from the apidocgen staging manifest, which leaves the Command Line children untouched. So make gen is idempotent for what it writes. The real, pre-existing kernel (7 ai-gateway/server_fix-oidc-links pages absent from nav) is CRF-29 below, out of scope for this PR.

🤖 This review was automatically generated with Coder Agents.

Round 4 is a clean polish pass. All four addressed round-3 findings verify at the code, not just in the replies: CRF-25 is closed at the root by making the CLI index a struct copy (idx := *cmdLine; idx.Children = nil) that mirrors the API index and extracting the per-command mapping into one cliCommandRoute(cmd) used by both the template func and the manifest rebuild, with TestCLICommandRoute pinning it; the leaked (CRF-22) review IDs are gone; the FrontMatter comment leads with the why; and the double-wrapped manifest-load error is dropped in both generators. CRF-29 is deferred with a real ticket (DOCS-594). Netero and six of the ten panelists returned clean. Ryosuke's read: "the abstractions earn their indirection."

Retrying the dismissal of the stale round-2 REQUEST_CHANGES (the round-3 attempt hit a token-permission 403). If it fails again, the marker is stale, not active: the P1 that raised it (CRF-16) has been fixed for two rounds and CI is green.

Round-4 severity count: 0 P0/P1/P2, 1 P3, 3 Note. None block merge. The P3 is a test-authenticity gap, not a runtime bug, and the Notes are minor or pre-existing. This PR is in good shape to merge once the P3 test assertion is tightened (optional but cheap).

🤖 This review was automatically generated with Coder Agents.

Round 5 is clean test-hardening. The CRF-30 fix is verified, not just claimed: five reviewers independently reran the mutations and confirmed both die (collapsing the isBareScalar reserved-word switch and forcing the number check to return true now fail TestYAMLScalarBareWhenSafe), and the round-trip docstring is honest about what it can't see. CRF-31 (wrap guarantee documented on LoadManifest) and CRF-32 (route-not-found errors name the manifest file) both verify, and the new TestExtractSectionName / TestCLIIndexRouteMirrorsManifest cover the previously-unguarded arms. Every prior finding is resolved, accepted, or deferred with a ticket. Kite's read: "the right way to retire a test-exists-but-proves-nothing finding."

Round-5 severity count: 0 P0/P1/P2, 1 P3, 2 Nit. Nothing blocks merge. One inline P3 (non-blocking, not reachable today) and one inline Nit. Two body items:

CRF-36 (Nit, commit message, Leorio). The commit body tags six bullets with bare CRF-N review IDs ((CRF-2), (CRF-3), ...) that resolve to nothing a future git log/git blame reader can open. This is the same class as CRF-26 (leaked review IDs in test comments, since fixed), and stronger here because a scopeless single commit's message is permanent history and the first thing a bisecting reader sees. The prose already tells the story; drop the (CRF-...) parentheticals and keep the Linear: DOCS-483 reference, which is the one a reader can act on.

Process note (Gon): the duplicated-rationale-comment pattern (CRF-24, CRF-31, and CRF-35 below) keeps recurring as the refactor cycles re-mint the same "mirror the manifest route via the shared emitter" rationale on a declaration and its callers. When a helper carries the why, its callers shouldn't repeat it.

Retrying the dismissal of the stale round-2 REQUEST_CHANGES (the round-3 and round-4 attempts hit a token-permission 403). If it fails again, the marker is stale, not active: the P1 that raised it was fixed three rounds ago and CI is green.

🤖 This review was automatically generated with Coder Agents.

A few comments, but they look good to me. Please get another code review on Go Code.