[finding] two unreleased changesets publish the /packages list door's filter set as two-of-three into 17.5.0 — the same release that carries the repair

[os-steve]

⏱️ This card has a deadline, and it is the release cut. The release that consumes a changeset
deletes the file and publishes its text into CHANGELOG.md. After that, a factual error in it is
amendable only by a dedicated docs-only PR against a released changelog entry. The cheapest moment to
repair is before the cut — ⛔ this is not a card that keeps its value if it waits.

Filed at 2026-09-21T16:10Z by the domain:spec seat 4 (session_01AmH9bKvGoLjiY86Q4Z3og2, seat post #18917),
from the boundary flags of the PR #19595 contract review — both rounds named it (round 1
5762666543, round 2 5763547517), and the dev's PR body carries it under Acceptance notes as
⛔ not-fixed-here. ⛔ Filed unassigned, ⛔ no priority:*, ⛔ no domain:*, ⛔ no type — routing and
grading are triage's. ⛔ Not a claim. ⛔ Not a ruling.

The defect

GET /api/v1/packages reads three query filters. Two unreleased changesets enumerate two of
them, and both publish into 17.5.0 — the same release that carries the repair for every other carrier
of the same sentence.

⇒ 17.5.0 would ship a changelog that contradicts the tombstone string shipping beside it in the same
release.

⏱️ Measured on origin/main 1f69917c5c, first-hand by this seat — ⛔ not adopted from the review

Leg 1 — what the door reads. handlePackagesRequest's parts.length === 0 && m === 'GET' branch,
packages/runtime/src/domains/packages.ts, reads exactly three query keys and no others:

key	read point	effect
enabled	readEnabledFilter(query?.enabled) (:828), applied at :845	packageCountsAsEnabled(p) === enabled.value
status	if (query?.status) (:837)	p.status === query.status
type	if (query?.type) (:840)	p.manifest?.type === query.type

The enabled leg is PR #19405's; it is merged, so the premise is live and settled. The same file's own
comment at :872-873 already says so in the door's words — 「this door applies the status / type /
enabled filters and then returns every remaining row」.

Leg 2 — the two carriers, and they are the only two left. git grep over origin/main, both
spellings, exit codes captured before any pipe:

carrier	reading
.changeset/17667-packages-query-contract.md:14	「the serving door filters on status / type and then returns every remaining row」 — grep exit 0, 1 line
.changeset/scoped-packages-dispatcher-door.md:13	「this door applies its status / type filters and returns every remaining row」 — grep exit 0, 1 line

Every other carrier of the sentence — packages/spec/src/api/package-api.zod.ts:202, the ADR-0087 D3
semantic entry, the retired-key record comment, the generated registry.ts mirror, and both rows of
content/docs/references/api/package-api.mdx — is repaired by PR #19595. After it lands, these two
files are the whole remainder.

Leg 3 — both publish into 17.5.0. Frontmatter read from the files themselves:
17667-packages-query-contract.md grades '@objectstack/spec': minor;
scoped-packages-dispatcher-door.md grades "@objectstack/runtime": minor. Both packages sit at
17.4.0 on origin/main, so both land as 17.5.0 — the exact version the tombstone string names.

⭐ Lit control, same instrument, same corpus. A line-wise regex for the three filters together
(status.{0,12}type.{0,12}enabled over packages/ content/ .changeset/) returns exit 0, 1
line: packages/client/src/index.ts:2524, the client's list() signature declaring
{ status?: string; type?: string; enabled?: boolean }. ⇒ the instrument reaches this corpus, and the
three-filter fact is independently present in the tree.

⭐ Dark control, same instrument. The same regex with zzznotafilter in place of enabled returns
exit 1, 0 lines.

⚠️ One honest note about the instrument. A first pass used a single-line probe for the literal
spelling `status` / `type` / `enabled` and read 0 hits, which would have said the tree nowhere
states the three-filter fact. That was false: the runtime comment wraps the spelling across two
lines, so a line-wise probe cannot see it. The lit control above is the replacement, and this note is
here so no later reader repeats the probe and trusts its zero.

Why this is a card and not a widening of PR #19595

Both files are outside #19407's declared file surface (packages/spec/src/migrations/,
packages/spec/src/api/). The dev reported them rather than widening into them, which is the correct
call and is ⛔ not a defect in that PR. The repair needs its own carrier, and the carrier needs to exist
before the cut.

What the repair is

Two one-line edits, each adding the third filter to an enumeration:

17667-packages-query-contract.md:14-15
FROM  the serving door filters on `status` / `type` and then returns every remaining row
TO    the serving door filters on `status` / `type` / `enabled` and then returns every remaining row

scoped-packages-dispatcher-door.md:13
FROM  this door applies its `status` / `type` filters and returns every remaining row
TO    this door applies its `status` / `type` / `enabled` filters and returns every remaining row

⚠️ Whoever takes it should re-run leg 2 first. The count of unconsumed changesets on origin/main
is 566, and a release cut can consume these two at any time. If the release has already run, the
sentence is in a published CHANGELOG.md and the repair is a different, larger act — ⛔ do not edit a
released changelog entry as though it were a changeset.

What this card does NOT claim

• ⛔ It does not say the sentence is false. It is incomplete: the load-bearing claim — no page
  was ever withheld and no continuation token was ever minted — stays true, because enabled filters
  rows and does not paginate. Only the enumeration is short.
• ⛔ It does not reopen the limit / cursor retirement, which is settled.
• ⛔ It does not touch ListInstalledPackagesResponseSchema.nextCursor, still declared and never
  emitted — a separate observation, deliberately left by the retirement.

Dedupe

changeset enumerates two of three packages filters · 17.5.0 changelog status type enabled ·
packages list door third filter unreleased changeset · 17667-packages-query-contract stale enumeration · scoped-packages-dispatcher-door status type filters

⛔ Not a duplicate of #19407 (that card is the spec/docs carriers and is delivered by PR #19595;
this is the two .changeset/ files that PR correctly left alone) and ⛔ not of #19405 (the PR that made
enabled live).

---

Generated by Claude Code

[os-steve]

Leg 2 re-run after PR #19595 landed — the two carriers stand, and a THIRD line now matches that is ⛔ NOT one, 2026-09-21T16:37Z

The card was filed while PR #19595 was still in the merge queue, so leg 2 was measured against a tree
where the repair had not landed. It has now: merge commit 5c5b67fc4140f76ca3158acea9e0845d9eebfad8,
card #19407 closed completed. Re-run against the new origin/main 5c5b67fc41 — exit codes captured
before any pipe — because the card tells whoever takes it to re-run leg 2 first, and that instruction
is worth more if the answer is already on the record.

The repair landed everywhere it said it would. Probe for the three-filter form: exit 0, 6
lines — package-api.zod.ts:202, the ADR-0087 D3 semantic entry, the retired-key record comment, the
generated registry.ts:13366 mirror, both rows of package-api.mdx, plus the TO side of the new
changeset's own before/after block.

Probe for the two-filter enumeration (`status` / `type` immediately followed by 「and then
returns」 / 「filters and returns」): exit 0, 3 lines.

line	verdict
.changeset/17667-packages-query-contract.md:15	✅ carrier — this card's item 1, unchanged
.changeset/scoped-packages-dispatcher-door.md:13	✅ carrier — this card's item 2, unchanged
.changeset/19407-packages-list-tombstone-third-filter.md:23	⛔ NOT a carrier — it is the FROM side of PR #19595's own before/after block, with the corrected TO at :27. Quoting the old sentence to show what changed is the point of the block. ⛔ Do not "repair" it — editing it would destroy the record of what moved

⇒ the card's claim holds — two carriers, both .changeset/ — but a naive re-run now returns three
lines, and the third is a deliberate quotation. Named here so it is not mistaken for a fourth defect or,
worse, silently "fixed".

⭐ And the landing makes the deadline concrete rather than theoretical. Both 17667-packages-query-contract.md
('@objectstack/spec': minor) and the just-landed 19407-packages-list-tombstone-third-filter.md
publish into the same @objectstack/spec 17.5.0 changelog. Unrepaired, that published changelog
carries the stale sentence stated as current in one entry, immediately beside an entry whose whole
content is the correction of that sentence. That is the shape of the harm, and it is one release away.

---

Generated by Claude Code

[os-steve]

⏱️⏱️ The deadline is not approaching — the release PR is OPEN, it is the 17.5.0 cut, and it ALREADY carries both stale sentences, 2026-09-21T16:54Z

The card said 「the cheapest moment to repair is before the cut」. Measured now: the cut exists, as an open
PR, and both sentences are already written into its generated changelogs. ⛔ Nothing below is adopted —
the release branch was fetched and read directly.

⏱️ Measured on changeset-release/main at 3b8ebe9a2dec9fa6fea7c3a131f4fa72fc005fec

reading	value
PR #17076 chore: version packages	open, not draft, 720 files, +47458/−20529
packages/spec/package.json on that branch	17.5.0 ⇒ this is the cut this card names
packages/spec/CHANGELOG.md:1987	「… status / type and then returns every remaining row, so no page was ever …」 — carrier 1, already published into the changelog
packages/runtime/CHANGELOG.md:654	「… this door applies its status / type filters and returns every remaining row …」 — carrier 2, already published into the changelog
⭐ control — the three-filter form, same instrument, same two files	exit 1, 0 lines ⇒ the corrected spelling is nowhere in the release changelogs

⇒ the generated artefact that publishes these sentences is not hypothetical and is not scheduled. It is
sitting in an open PR that anyone can merge.

What this changes, and what it does NOT

It does not change the repair. Still two one-line edits in .changeset/. The changesets bot regenerates
#17076 on every push to main, so repairing the two files on main rewrites the release PR's changelog
entries before they publish — ⛔ no edit to #17076 is needed and ⛔ none should be attempted.

It does change the window. The card's 「before the cut」 read as a future event with slack in it. The
honest reading: the window closes whenever a human merges #17076, which can be at any moment and
depends on nobody's schedule. There is no later cheap moment — after that merge the text is in a released
CHANGELOG.md and only a dedicated docs-only PR against a published entry can touch it.

ⓘ One reading that is NOT a defect, so nobody chases it

PR #17076's head is from 2026-09-21T12:21:19Z and the repair (PR #19595) landed at
2026-09-21T16:36:28Z, so the release branch predates the repair. That is why the three-filter
control reads 0 there: ⛔ not a failed repair, just a release PR that has not regenerated since. It will
pick the repair up on the next push to main — the same mechanism this card depends on.

⇒ ⚠️ and that is precisely the point: the same regeneration that will carry #19595's fix into the release
will carry these two stale sentences into it too, unless they are repaired first.

---

Generated by Claude Code

[os-steve]

Claim: PM loop round 17
Session: session_01AmH9bKvGoLjiY86Q4Z3og2
Branch: claude/issue-19605-changeset-carriers-third-filter
Worktree: objectstack-issue-19605
Domain: domain:spec
Seat: domain:spec#4
File surface: .changeset/ (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: opus build — quoting this run's --tier output: 「no path-derived mandate … floor sonnet · default opus · ceiling opus」 and 「Clause ② is NOT reachable from paths」. ⭐ The diff is two words, which argues sonnet; the tier is opus anyway and the reason is stated rather than left to taste: the card has a closing window, and two of its three real risks are judgment, not typing — whether a .changeset/ edit owes its own changeset, and ⛔ never touching the release PR. A lost round costs more than the tier.
Clause-②: no
Thread-read: 5764245871
Serial constraints cleared: none — the two carrier filenames probed across all 12 non-bot open PRs' file lists: exit 1, 0 hits; lit control .changeset/ on the same list: 12 hits; dark control zzznotapath: 0. PR #17076 is the changesets bot's release PR; it holds no seat's lock and is ⛔ not to be touched.

⚠️ Reading-time: 2026-09-21T17:52Z. The thread-read id above is this card's newest comment, written 2026-09-21T16:54:27Z.

⛔ This seat is routing its own filed card, and says so rather than letting it pass unnoticed

This card was filed by this seat today, unlabelled, with 「routing and grading are triage's」 on its face. Claiming it now routes it without waiting for triage. ⛔ That is a departure and it is deliberate: the release PR that consumes these two changesets is open right now, and after it merges the repair is only possible as a docs-only PR against a published CHANGELOG.md. ⛔ The grading is still triage's — this claim adds pm:dispatched and takes no priority:*.

Premise re-verified at head 5c5b67fc41, and ⛔ the first probe was WRONG

leg	reading
.changeset/17667-packages-query-contract.md	two-filter enumeration PRESENT, three-filter absent
.changeset/scoped-packages-dispatcher-door.md	two-filter enumeration PRESENT, three-filter absent
lit control — 「no page was ever withheld」 in the first file	true
dark control — 「zzznotasentence」 in the first file	false
PR #17076	still open, merged: false, mergeable_state: blocked

⭐ The honest part. This seat's first probe was a single-line git grep, and it returned exit 1, 0 lines for carrier 1 — which would have read as 「already repaired, close the card」. It is a false zero: the sentence wraps, 「the serving door filters on」 ends one line and 「status / type and」 begins the next. The readings above are from a whitespace-normalised whole-file probe.

⚠️ ⛔ This is the second time today the same line-wrap produced a false zero for this seat, and this very card's body warns about it in as many words. The warning was written and then not applied. Recorded here so the next reader distrusts a line-wise zero on this sentence specifically, ⛔ rather than trusting that the warning is enough.

---

Generated by Claude Code

[os-steve]

os-dev-report

{
  "issue": 19605,
  "status": "done",
  "branch": "claude/issue-19605-changeset-carriers-third-filter",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/19616",
  "session": "session_01AmH9bKvGoLjiY86Q4Z3og2",
  "premise_still_valid": true,
  "summary": "Both carriers repaired on a branch off main at 5c5b67fc41; the diff is 2 files, +3/-3, entirely inside .changeset/ as the claim declared. 17667-packages-query-contract.md now reads `status` / `type` / `enabled`, and so does the hasMore bullet in scoped-packages-dispatcher-door.md; the second line of the first hunk is a re-wrap only (that file is hard-wrapped at 80 columns), contained to the two lines the sentence occupies. PR #17076 was read but NOT touched in any way -- no branch, file, body or label write of any kind. THREE THINGS THE SEAT MUST ACT ON. (1) Check Changeset will be RED on PR #19616 and that is the designed state, not a failure: this PR changes changesets it did not add, which is pr-automation.yml's route 0 DELIBERATE CORRECTION class -- the counting step reads added=0 and the verdict step exits 1. (2) skip-changeset must NOT be applied; maintainer ruling on #18375 (ruling D, 2026-09-18) refuses the label for exactly this class, so the label scope for this card is empty and zero labels were written. (3) The route-0 remedy asks for a written confirmation on the PR naming each note and what changed under it -- it is written into the PR body, and the gate's own text says the check stays red either way and the approver merges over it (Check Changeset is not one of the seven required contexts). Assignee arrived already set to os-steve by the dispatch and was never written by this run. The claim comment 5765016512 names this exact branch.",
  "tests": "Exit codes captured before any pipe (cmd > file 2>&1; EXIT=$?). PREMISE re-verified at 5c5b67fc41 with a whitespace-normalised whole-file probe, never a line-wise grep: before the edit carrier 1 two-filter=1/three-filter=0, carrier 2 two-filter=1/three-filter=0; lit control ('no page was ever withheld') true in file 1, dark control ('zzznotasentence') false in both. After the edit: two-filter=0/three-filter=1 in each, lit control still 1, dark control still 0 -- the load-bearing claim is intact and only the enumeration moved. DESTINATION confirmed read-only on the release branch, fetched into a ref this branch owns (refs/issue-19605/release-head, no write to changeset-release/main): at 3b8ebe9a2d, spec/CHANGELOG.md carries stale sentence 1 (count 1), runtime/CHANGELOG.md carries stale sentence 2 (count 1), the three-filter form counts 0 in both, and both package.json read 17.5.0. PR #17076 re-read live: state open, merged false, draft false, head still 3b8ebe9a2d -- the window was still open at report time. GATES: dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 19 commands at 7bd01a6d88; all 19 run and reconciled via --ran (19 derived, 19 run, 0 NOT-MEASURED, 0 UNRUN, every row carrying an exit code), plus 2 beyond the derivation. 18 exit 0. The single exit 1 is the designed route-0 red: check-empty-changeset.mjs --base origin/main, printing 'present on the merge base and CHANGED by this PR' for both files and naming DELIBERATE CORRECTION vs COLLISION; this is COLLISION's opposite (the filenames were drawn by #17667 and #16781, this branch drew neither, and restoring from base would put the short enumeration back). Green includes check-changeset-no-major --base 0, check-adr-0087-registration --base 0 (the first file is declared-breaking and its disposition marker is intact), check:nul-bytes 0, check:published-files 0, check:pm-changeset-deadline-census 0, check:objectui-changeset 0, check-closing-keyword-parity 0. Beyond the derivation: check-changeset-fixed.mjs 0 (the one roster gate whose roster lives under .changeset, flagged by the deriver as a silence that is not evidence -- run rather than assumed) and check-issue-citations.mjs --base origin/main 0 (0 issue citations added; no bare card number lands on an added line). REGENERATION: pnpm --filter @objectstack/spec build exit 0 under os-verify-lock.sh (OS_VERIFY_LOCK_SLOT=issue-19605-spec-build, VERDICT command-exit 0, held 137s, waited 0s), then pnpm --filter @objectstack/spec check:generated exit 0 -- 'All 15 generated artifacts are up to date', measured against the dist this run built. Nothing in packages/spec/scripts reads .changeset/ content (the 3 path hits are prose in file headers). The only artefact these files feed is the release PR's generated CHANGELOG.md, and that regenerates itself on the next push to main. LINT, narrowed and the narrowing proven: eslint.config.mjs carries 7 files-bearing config objects and all 7 are TS/JS globs; a grep of that file for a markdown glob or for .changeset returns exit 1, zero hits; npx eslint --no-inline-config --format json over exactly the 2 changed files returns 2 results, 0 errors, each 'File ignored because no matching configuration was supplied'; the diff adds and removes no config object and touches no .ts/.js file, so no untouched file's verdict can move. The repo-wide scan is CI's and its verdict is invariant under this diff. TIERS: the diff touches no workspace package, so there is no affected-package closure and no package test/typecheck is owed; the spec build was run only to give check:generated a real dist. No ablation is owed -- this branch adds no guard whose ability to fail needs proving. CONTRACT REVIEW reading, report-only: needs:contract-review is NOT on PR #19616 (label list empty); scripts/pm/check-clause2-carriers.mjs --pair 19616 exits 4 (C6 -- domain:spec owes a contract review of record on head 7bd01a6d88 and none exists yet). Neither hung, stripped nor waited on, per the standing clause. CI convergence not awaited; gate status on the PR is in_progress at report time.",
  "mcp_calls": "0 -- no MCP GitHub tool was called. Every GitHub read and write went through the REST proxy with curl.",
  "api_writes": "2 -- POST /repos/objectstack-ai/objectstack/pulls (draft PR #19616); POST /repos/objectstack-ai/objectstack/issues/19605/comments (this report). ZERO label writes: the label scope for this card is (labels named by the dispatch = none) plus skip-changeset, and skip-changeset is refused by name for this class, so the scope is empty. No PATCH of the PR body was attempted -- the body was written once at PR-open. Git pushes (2: the empty-branch write-routing probe, then the commit) are outside the REST count. Reads used: GET issues/19605, GET issues/19605/comments, GET pulls/17076, GET pulls/19616, plus the 5 reads check-clause2-carriers.mjs issued.",
  "open_questions": [
    {
      "question": "PR #19616 corrects two PENDING release notes and adds no changeset of its own, so Check Changeset is red by design and route 0 asks for the correction to be confirmed in writing on the PR before it lands. Who writes that confirmation and merges over the red?",
      "options": [
        "A -- the seat posts the confirmation on PR #19616 (the PR body already names each note and what changed under it) and a human merges over the red Check Changeset, which is advisory, not one of the seven required contexts.",
        "B -- apply skip-changeset to clear the red.",
        "C -- add a changeset of this branch's own so the counting step reads added=1."
      ],
      "recommendation": "A. B is refused by name: maintainer ruling on #18375 (ruling D, 2026-09-18) says the label is never applied to a PR that edits an existing changeset -- it would be a false declaration, since the notes rewritten here are releases that are still pending, and it exempts the whole job, taking the refusal that names this class with it. C is route 1, declaring a release this branch does not make; it publishes nothing (check:published-files exit 0). The gate's own text is explicit that the red is the point and costs nothing. The urgency is real: PR #17076 is open right now at 17.5.0 with both stale sentences already generated into its changelogs, and it regenerates from main on the next push."
    }
  ],
  "out_of_scope_findings": [
    "carrier: this PR's own Acceptance notes and card #19605's second comment -- noted, not filed. .changeset/19407-packages-list-tombstone-third-filter.md line 23 matches the two-filter pattern and is the FROM side of that note's own before/after block, with the corrected TO two lines below. Verified in place: that file still counts 1 two-filter AND 1 three-filter occurrence after this change. Editing it would destroy the record of what moved, so it is deliberately untouched.",
    "carrier: card #19605's own body already records it -- noted, not filed, no duplicate opened. ListInstalledPackagesResponseSchema.nextCursor stays declared and is never emitted; the card names it as a separate observation deliberately left by the pagination retirement.",
    "carrier: none needed -- a recorded platform reading, reproduced, not a defect. Creating PR #19616 by raw REST with a body whose tail is the harness attribution block (not a rule line plus the session-URL footer) caused the platform to append its own rule line plus session-URL footer: stored body is the sent body plus exactly 90 bytes, sent is a strict prefix of stored. That is the cell already recorded in .claude/skills/pm-dispatch/references/platform-readings.md (bare-REST create side, 90/91 bytes, append only when the sent tail is not already that block). Read back in full and left alone; no PATCH attempted."
  ]
}

---

Generated by Claude Code

[os-steve]

⛔ Correction — this card's "live deadline" never existed. The sentence I counted on the release PR was the fix's own QUOTATION, 2026-09-22T06:42Z

domain:spec seat 4. This corrects a reading this seat made repeatedly, including to the maintainer, and it is wrong in the direction that manufactures work.

What I reported, and what is actually there

I reported that release PR #17076's head still carried 1 stale two-filter sentence in packages/spec/CHANGELOG.md, called it the last live piece of this card's deadline, and put a decision to the maintainer about spending a workflow_dispatch + refresh_version_pr to buy it back before the 6-hourly tick.

Re-measured with file-based patterns (⛔ not a shell-quoted grep — see the instrument note below), on both heads:

head	the old two-filter sentence	the corrected three-filter sentence
f7bc3f2d87 (before the tick)	1	1
7d326daeb7 (after the 06:15:30Z tick)	1	2

And the single occurrence, on both heads, sits inside a fenced block that opens FROM … the serving door filters on ``status`` / ``type`` and then returns every remaining row … — packages/spec/CHANGELOG.md:8216 on the old head, :8539 on the new one. It is the FROM/TO block of .changeset/19407-packages-list-tombstone-third-filter.md, the changeset that describes this very repair, quoting the sentence it corrected.

⇒ there was never a live stale sentence on the release PR. What moved at the tick is the corrected form going 1 → 2 as the newly-consumed changesets landed in the generated changelog. The mechanism worked exactly as designed and needed nothing from anybody.

Root cause, named rather than softened

I counted a grep hit without asking whether it was a live statement or a quotation of the defect being fixed. ⭐ That is the exact discipline this seat wrote into #18048's dispatch word and then verified in the delivered gate: «the gate's hardest requirement is not the joins: it is that a gate which cannot tell a tombstone from a quotation of a bad tombstone will red on the ruling that forbids the bug.» I held a subagent's gate to that standard in the same hours I failed to hold my own grep to it.

⚠️ A second instrument fault sat on top of it and nearly hid the first: the re-check I ran at 06:14Z spelled the pattern inside a double-quoted shell string, so its backticks were command substitution, not literal characters — the pattern that reached git grep was not the one I wrote. Two consecutive readings of the same ref disagreed (1, then 0) and that disagreement is what exposed the whole thing. ⇒ patterns containing backticks go through a file and grep -F -f, never through a quoted shell argument.

What this changes

• ⛔ Nothing is owed from the maintainer on this card. The workflow_dispatch decision I put to them was manufactured by my own misreading; they should not spend a thought on it.
• The repairs themselves are unaffected and remain correct: fix(changeset): the two unreleased /packages-door notes enumerate all three filters that door reads #19616 and docs(changeset): drop the deleted api-surface-declarations family from four pending release notes #19627 landed real corrections to real carriers, verified by reading the merged files out of origin/main. ⛔ This correction touches the release-PR claim only.
• This card stays closed. It was discharged when fix(changeset): the two unreleased /packages-door notes enumerate all three filters that door reads #19616 merged; the "residual deadline" I appended to it afterwards was not real.

_{domain:spec#4 · session_01AmH9bKvGoLjiY86Q4Z3og2 · GitHub os-steve · read at 2026-09-22T06:42Z}

---

Generated by Claude Code