docs(pm-skills): retract the REST issues-list completeness certificate (cursor half stopped on the ratchet) by claude[bot] · Pull Request #14068 · objectstack-ai/objectstack

claude · 2026-09-01T02:29:45Z

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

claude · 2026-09-01T02:29:45Z

Part of #13900

Half of the adjudicated two-file repair lands here. The other half is stopped on a
measured ceiling, with byte-exact floors below — it is an audit input, not an omission.

What landed: the falsified completeness certificate is retracted in place

.claude/skills/pm-dispatch/references/platform-readings.md L147 certified the REST
issues-list endpoint's results as complete — (core 桶、结果**完整**、 ...) — inside the
dedup ladder that routes a seat to that endpoint precisely when search_issues has gone
to a session-level zero. That property is false. The line now reads:

(core 桶、 `labels` 真 AND;⛔ 完整性自证:`&page=N` + 总数核对;...)

The certificate is replaced by the obligation it was standing in for: completeness is
self-proved, by the page-number walk plus a total-count cross-check. Nothing else in
the bullet changed; the two adjacent readings it carries (the MCP branch's 226-of-100
manual-paging reading, and the qualifier tombstone) are untouched.

No causal claim is made, and none is implied. Whether the truncation is GitHub's own
cursor behaviour or the session egress proxy rewriting the Link header is UNMEASURED.
The prescription removes the dependency on that header; it does not diagnose it. The
landed text asserts nothing about cause.

I grepped the whole pm-dispatch surface for completeness claims afterwards: L147 was the
only one about this channel, so no false statement about it remains anywhere in the corpus
even with the second file stopped.

Reproduction, first-hand, this container, this repo

Not inherited from the card — re-measured here against objectstack-ai/objectstack:

how the open-issues population was walked	items
`per_page=100`, then follow `Link: rel="next"`	102 (100 + 2, then no further `rel="next"`)
`per_page=100&page=1..5` (classic page numbers)	448 (100+100+100+100+48)
`open_issues_count` from the repo endpoint	447, re-read 449 (live backlog; counts issues AND PRs)
open PRs, `pulls?state=open` page-walked	23 (page 2 and 3 empty)

The Link header does carry a cursor rather than a page number:

Link: .../repositories/1136691870/issues?state=open&per_page=1&after=Y3Vyc29yOnYyOpLP...&page=2; rel="next"

The cursor walk terminates the way an exhausted walk terminates — no error, no warning —
having returned 102 of ~448. The card measured the identical 102 on a sibling repo, so the
cutoff is not repo-specific.

Premise re-verified on origin/main before editing: both files still carried the
falsified certificate and lacked the prescriptions, and no open PR touches either file
(all 23 open PRs' file lists checked; two touch other files in the same references/
directory — decision-analysis.md and dispatch-runbook.md — neither overlaps).

What is stopped, and the byte-exact floor

The two prescriptions for references/rest-channel.md are not in this PR. The file has
no room and this card carries no ruling that would make room.

Line-ratchet verdict lines at this head, both files:

✓ check-skill-line-ratchet: .../references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .../references/rest-channel.md is 87 lines (ceiling 87; headroom 0).

Reflow-within-lines was tried first and measured, not assumed. Using the gate's own
splitPrefix / hanAsciiPunctTail / 120-byte budget, plus the break repertoire the corpus
actually uses (a space, or a segment break between two wide characters), an optimal
minimum-line packer was run over every block of both files:

file	lines now	optimal repack	reclaimable
`platform-readings.md`	314	314	0
`rest-channel.md`	87	87	0

The packer reproducing each file's current line count exactly, block for block, is the
calibration: these files were mechanically re-wrapped once already and are sitting on
their packing floor. There is no slack to reclaim in either file.

What that leaves, measured to the byte:

rest-channel.md, list-cards row (L18-19): 212 bytes over 2 lines; it can absorb
23 more bytes without a third line. The two prescriptions written faithfully — both
refusals, the &page=1,2,3…-until-a-short-page termination condition, the
open_issues_count cross-check with the open-PR subtraction, the cause-unknown boundary,
and the note that a keyword positive control structurally cannot see this class — are
731 bytes ⇒ 6 lines. Stripped to the two refusals plus the cause boundary: ~450 bytes
⇒ 4 lines. Funding 4-6 lines means deleting 4-6 lines of measured content or raising
the ceiling; the ratchet's own rule is that a raise needs a maintainer ruling quoted in
the raising PR, and this card has none. So: stopped, not forced.
platform-readings.md, certificate bullet (L141-150): 1,140 bytes over 10 lines, and
it absorbs exactly +26 bytes at 10 lines. The landed correction is +26 — at the
ceiling to the byte, which is why it says 总数核对 rather than naming
open_issues_count inline (+33, which needs an 11th line).

One candidate funding source, offered as an observation rather than acted on:
rest-channel.md L58 restates the red-window standing rule that platform-readings.md
owns in its quota section, against rest-channel.md's own header rule that a fact lives in
one place. Retiring it would free a line or two — but that is a content decision about
which file owns a rule, not a reflow, so it belongs to the maintainer, not to this card.

Gates

Derived after the last commit with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands
(the script takes the change set from the merge base itself). Union run at head
028dbc85e; exit codes captured before any pipe.

gate	exit
`pnpm check:pm-skill-ratchet`	0
`pnpm check:pm-skill-id-lint`	0
`pnpm check:skill-frame-sync`	0
`pnpm check:doc-authoring`	0
`pnpm check:agent-test-spelling`	0
`pnpm check:pm-governed-merges`	0
`pnpm --filter @objectstack/lint run check:doc-formula-expressions`	0
`pnpm check:nul-bytes`	0 (not derived; run because any edit owes it)
`node scripts/pm/check-governed-queue-guard.mjs`	1 — NOT MEASURED

Two readings that are not results, stated so they are not read as green or red:

check:doc-formula-expressions first exited 3 — PREREQUISITE NOT MET, its own words
"Nothing was measured". Building @objectstack/formula and then @objectstack/lint
cleared it; the 0 above is the re-run.
check-governed-queue-guard exits 1 locally because it reads GITHUB_EVENT_PATH and
there is no workflow event payload in a container. By design it refuses to be green when
it cannot look. It runs in CI.

Landing

.claude/** is a governed surface: draft only, human merge, approval pinned to the
current head. No changeset — the diff is .claude/**-only, so skip-changeset is applied
(additively) instead.

Generated by Claude Code

The dedup ladder's REST branch certified the list endpoint's results as complete. Measured in one container against this repo: walking the `Link: rel="next"` cursor the endpoint returns reports itself exhausted at 102 items, while the `&page=1,2,3...` walk over the identical query returns 448 and matches `open_issues_count` (449 at re-read, 23 open PRs). A dedup grep over the returned 102 answers "no duplicate" with total confidence -- the failure the ladder routes seats INTO when `search_issues` goes to a session-level zero. The certificate is replaced in place by the obligation it was standing in for: completeness is self-proved, by the page-number walk plus a total-count cross-check. Line-neutral -- the file stays at its 314-line ceiling, and the change is a correction of a measured-false line, not an addition. The cause is NOT explained here and is not claimed to be: whether the truncation is GitHub's own cursor behaviour or the session egress proxy rewriting the `Link` header is unmeasured. The prescription removes the dependency on that header; it does not diagnose it.

github-actionsBot added size/xs documentation Improvements or additions to documentation labels Sep 1, 2026

claudeBot added the skip-changeset PR has no user-facing published change; bypasses the changeset gate label Sep 1, 2026

claudeBot requested review from hotlong and os-zhuang September 1, 2026 02:30

This was referenced Sep 1, 2026
[finding] The REST issues-list Link: rel="next" cursor silently returns 102 of 287 open issues — the dedup channel references/rest-channel.md prescribes #13900
Closed
[PM seat] skills — 🟢 os-steve #7623
Open

os-zhuang approved these changes Sep 1, 2026
View reviewed changes

os-zhuang marked this pull request as ready for review September 1, 2026 06:33

os-zhuang enabled auto-merge September 1, 2026 06:33

os-zhuang added this pull request to the merge queue Sep 1, 2026

claudeBot mentioned this pull request Sep 1, 2026
[finding] The &page=N completeness walk refuses with HTTP 422 past 9,900 items — the ceiling on the remedy #14068 just landed #14133
Closed

Merged via the queue into main with commit fe5c4e2Sep 1, 2026
29 checks passed

os-zhuang deleted the claude/issue-13900-rest-cursor-underreturn branch September 1, 2026 06:54

huangyiirene mentioned this pull request Sep 1, 2026
[PM seat] director(项目总监) — 🟢 os-warren（第 14 场，决裁 + 契约复审） #12708
Open

os-justin mentioned this pull request Sep 1, 2026
[finding] "The MCP write rate limit" is not GitHub’s quota — GitHub read 14888/15000 core while MCP writes were refused, and GitHub’s reset field predicts nothing about the block #14014
Open