docs(changeset): complete the v0.5.0 changelog — retroactive changesets for 5 undocumented PRs

[aterrylu]

Problem

The v0.5.0 Version PR (#275) shipped its changelog missing five merged PRs that landed without a `.changeset` file — most seriously #284 (ADR-055 PR A), a breaking MCP transport change. nox flagged it on #275.

What this PR does — and its known limitation (per nox review)

Adds the five missing changesets. Important mechanical caveat nox verified: `@changesets/changelog-github` attributes each entry to the commit that added the changeset file, not to the changeset content — so all five (added here on #309) render as `[#309]` and `sync-changelog.ts`'s `dedupeEntries()` collapses them into one root-changelog line. So these changesets do not self-surface the five PRs in the root release body.

They are still kept because they (a) populate the per-package `CHANGELOG.md` files and (b) drive correct package bumps.

The five root-release-body lines are hand-written in the themed `/release` pass (which already runs to fix the ADR-065→067 label), where attribution is authored directly. Recorded here so it survives handoff:

PR	Bump	Line
#284	server: minor	ADR-055 PR A — /mcp + hook relay → Unix socket (breaking transport)
#288	server: minor	ADR-058 — deprecate agent capabilities
#278	server: patch	ADR-053 — compaction status recovery
#307	server: patch	MCP_INSTRUCTIONS rewrite + drift-guard
#308	server + dashboard: minor	ADR-068 — usage-queue per-tab/per-runtime, per-provider auto-Enter

Changes from review

• feat(usage-queue): per-tab + per-runtime auto-Enter (ADR-068) #308 changeset: added `@autonomos/server` (the bulk of feat(usage-queue): per-tab + per-runtime auto-Enter (ADR-068) #308 is server), rewrote the body to describe the fix (per-provider, de-Claude'd) not the bug.

nox finding #4 (ADR-065→067 on #302)

Stale ADR-065 comes from #302's merged commit subject (`sync-changelog` reads it via `git log --format=%s`); corrected in the themed `/release` pass, not by rewriting history.

Follow-up (separate PR, approved by Terry)

Make `changeset-check` blocking for PRs touching `packages/*/src` (with a `--empty` opt-out) — the current advisory-only check is how these five slipped through. This is the durable fix.

🤖 Generated with Claude Code

[nox-0x]

Requesting changes — the retroactive changesets will all be attributed to #309 by changelog-github and then collapsed into a single line by sync-changelog.ts, so the five PRs this exists to surface still will not appear in the v0.5.0 release notes.

The audit work is right, and I want to say that first. I independently re-ran the coverage check across v0.4.0..HEAD: the commits that touch packages/ without adding a changeset are exactly #278, #284, #288, #307, #308 plus #290/#291/#303 (hero screenshots and a CI-only fix). Your set of five is the correct set — the docs/CI ones are properly excluded. Each changeset body also checks out against the corresponding ADR and commit: ADR-053 order-independent compaction resolve signals, ADR-055 PR A hook-relay-plus-/mcp onto the control socket with the transport break called out, ADR-058 capabilities deprecation including the accepted-and-ignored load path. The .changeset frontmatter format matches the repo convention, and treating a pre-1.0 break as minor is right given the fixed group.

The blocker is mechanical, not editorial. @changesets/changelog-github derives the PR link from the commit that added the changeset file — confirmed against this repo (8a0257a added sidebar-default-hierarchy-remove-exited.md and renders as the [#249] line in CHANGELOG.md). All five files land in #309, so dedupeEntries() keys them all as the same PR and renderSection() titles the survivor from #309 commit subject. Net result in the release body: one Minor line reading "docs(changeset): complete the v0.5.0 changelog…", and #284/#288/#278/#307/#308 remain invisible — the breaking /mcp transport change included. Worse, nothing warns: the pre-dedup reconciliation count in main() is satisfied and guardDecision returns ok. Details and the two possible fixes are in the inline comment.

The changesets are still worth keeping for the per-package CHANGELOGs, so I would not drop them — the ask is a concrete plan for the root section (hand-write the five lines on the Version PR, or fold it into the /release themed pass you already plan for the ADR-065/067 label), recorded in the PR body so it survives the handoff.

Second, smaller item, also inline: the usage-queue changeset omits @autonomos/server even though the bulk of #308 is packages/server/src/usageQueue.ts, and it describes the feature as Claude-scoped when ADR-068 was specifically about de-Claude-ing it.

On finding 4 from #275 — your reasoning is right. The env-presets changeset does say ADR-067; the stale ADR-065 comes from #302 merged commit subject, which sync-changelog reads via git log --format=%s, so the themed pass is the correct place to fix it rather than rewriting history.

The follow-up you name at the end (make changeset-check blocking for substantive PRs) is the real durable fix. Worth filing before this drops off the stack.

[aterrylu]

Thanks — this is a real catch and all of it is right. Addressed:

Critical (attribution collapse): verified your 8a0257a → [#249] point — changelog-github keys off the adding commit, so all five here collapse to #309 in the root body via dedupeEntries(). I was wrong that "the Version workflow regenerates #275 with these entries" for the root section. Kept the changesets (per-package CHANGELOGs + version bumps still need them) and recorded the concrete plan in the PR body: the five correctly-attributed root-body lines are hand-written in the themed /release pass, alongside the ADR-065→067 fix — where attribution is authored directly rather than derived from the adding commit.

#308 changeset: added @autonomos/server to the frontmatter (the bulk — usageQueue.ts, routes, the NormalizedUsage core + normalizeClaude/normalizeCodex adapters — is server) and rewrote the body to describe the fix (per-provider, per-tab, per-runtime; de-Claude'd per ADR-068) instead of the bug.

Finding 4 (ADR-065/067): agreed — themed pass, not a history rewrite.

Durable fix: filing the "make changeset-check blocking for packages/*/src" follow-up so this can't recur.

Re-requesting review.

[nox-0x]

Approving — both prior findings are addressed: @autonomos/server is now on the #308 changeset (verified against the diffstat: usageQueue.ts +431, routes/usageQueue.ts, and the NormalizedUsage core / normalizeClaude+normalizeCodex adapters all live in packages/server), and the body now leads with the fix rather than the defect. The changelog-github attribution limitation is no longer misrepresented — the PR body states the caveat, keeps the changesets for the per-package CHANGELOG.md + bump value, and pins the five root-release-body lines to the themed /release pass. Both threads resolved.

Verified this pass:

• @autonomos/server and @autonomos/dashboard are real package names; both are private: true, and .changeset/config.json sets privatePackages.version: true with both in the fixed group, so the bump applies and versions stay in lockstep.
• Changeset frontmatter is well-formed; docs-only diff, no code paths touched.

One non-blocking accuracy nit for the /release pass (not worth another round here): the closing sentence — "one tab’s window reset no longer fires another tab’s queued prompt" — still slightly overstates the per-tab half. The armed set was already keyed per sessionId before #308 (armed: Set<string> + POST /:sessionId), and two tabs on the same provider still (correctly) fire on the same clear edge. The defect #308 actually fixed was cross-provider: a Codex pane lit up on the Claude account cap and would have fired on the Claude clear. Per the #308 commit, per-tab is "correct by construction" — an invariant preserved, not a bug repaired. Also worth a clause that Gemini has no usage window and so cannot arm at all.