fix(pm): H64 keys on the session id in the text, not on the author account by claude[bot] · Pull Request #18257 · objectstack-ai/objectstack

claude · 2026-09-15T04:30:02Z

H64 keyed on the author account. The rule it was built on had already been superseded: write identity follows the CHANNEL, the REST actor is claude[bot] or a user per session, and user.login records the token rather than the seat — attribution is the session_ id in the text (.claude/skills/pm-dispatch/references/rest-channel.md :54, .claude/skills/pm-dispatch/SKILL.md :98). So the row fired on compliant REST writes from user-token sessions, told their authors the write had gone through the MCP GitHub tool, and prescribed a re-post through the REST proxy that lands under the same login and fires again.

This re-keys the row onto the attribution the landed rule actually requires. One file, scripts/pm/check-half-states.mjs (non-governed PM tooling; check-governed-merges.mjs --test on this file list reads 0 of 1 path(s) hit the register).

The measurement first — the discriminator the card doubted does not exist

The PM's mechanism assumption P1 was that performed_via_github_app.slug tells the MCP tool from the REST proxy. Measured with GET-only reads on live rows, 2026-09-15:

row	`user.login`	`user.type`	`performed_via_github_app`
comment 5673548571 on #7623 — REST-proxy write, App token	`claude[bot]`	`Bot`	`{ id: 1236702, slug: 'claude', name: 'Claude' }`
comment 5673265919 on #17076 — REST-proxy write, user token	`os-warren`	`User`	`{ id: 1236702, slug: 'claude', name: 'Claude' }`
comment 5673413139 on #18237 — REST-proxy write, user token	`os-warren`	`User`	`{ id: 1236702, slug: 'claude', name: 'Claude' }`

And the four specimens H64 was built from, all read the same slug: #18045 (os-project-manager / User), objectui#9404 (os-project-manager / User), PR #18051 (os-project-manager / User), comment 5652138683 on objectui#9370 (os-tesla / User); with the clean control comment 5654046782 (claude[bot] / Bot).

P1 is falsified. The field names the App whose credential signed the write, never the tool. The first two rows above are both REST-proxy writes differing only in the token class the session was handed, and they are byte-identical in this field. So no channel is inferred anywhere in this family any more, and artefactChannel's docblock now says what it does read: an App credential versus none. Following P1's own instruction for this outcome, the MCP half of the seat direction (candidate (i)) collapses — it has no readable population — and candidate (ii) is the whole firing row.

What fires now

A seat/dev-signed artefact — a claim, a report, a contract review, a filing header — carrying no session_ id anywhere in its text. The text asserts a seat or a dev wrote it; nothing says which session, and the author field records only the token. Remedy: the owner gives the artefact its id (an edit in place is enough). The account is explicitly stated not to be the repair.

A structural consequence worth naming: two of the six signature forms are session ids (the bare id in the head window, and the attribution footer's claude.ai/code/session_ URL), so a text matching either carries its attribution by construction and can never fire. The firing population is the other four forms with no id anywhere. The id test reads the whole body while the signature's head window stays narrow — pinned as a case.

The exposure is not relaxed — it moved to where it costs nothing

The addendum stands: a suspended user account 404s everything it authored, measured on this board. Nothing here relaxes it. It is now reported as an informational count plus a login roster inside the H64 summary clause, on every run, with no remedy named and no row filed.

That is P5's first option, chosen for the reason P5 gives: no act available to a user-token session moves its content to the App, because the token class is handed to a session at start rather than chosen at write time. A row naming no remedy re-files every sweep at full weight against an anchor that trimmed 423 of 441 rows. The clause states the same fact for free, unconditionally, and is strictly more visible than the old rows were — those competed for the cap and were trimmed; a clause is not.

Board measurement, before and after (live sweep, GET only)

reading	before (`ecf91cb4`)	after (`c509364b`)
texts read / signed	1697 / 873	1703 / 875
finding population	671 user-authored	116 naming no session id
judged (on/after the pin)	70	26
rows filed (cap 10)	10	10
census behind the pin	601, back to 2026-08-05T19:32:57Z	90, back to 2026-08-08T13:49:21Z
informational exposure	not reported	671 texts, 17 logins, 3 PAT, 3 unread channel

The two sweeps ran ~40 min apart, so the corpus moved slightly on its own (1697 to 1703 texts).

Reverse verification on the live board. Every row the card named as a false positive is gone: comments 5673315903 (#17396), 5673859871 (#17502), 5673413139 (#18237), 5670954004 (#13801) and 5671601791 (#16695) each appear in 0 rows after, having appeared in 1 before. #18237's own two rows are gone.

And the retired test was blind in the other direction too. Of the 10 rows filed now, 7 are authored claude[bot] — for example comment 5674562475 on #16166 and comment 5674534458 on #18174, both os-dev-report payloads whose only footer is the platform's bare one, and card #18167's Filed by the ... seat body. Every one of those is exactly as unattributable as the ones the old row shouted about, and the old row could not see any of them. Spot-checked by GET: none carries a session id anywhere.

Fixtures — one kept its role, three were re-classified

Each with its reason, as P4 asks:

fixture	before	after	why
comment 5652138683 (objectui#9370, `os-tesla`)	fires	fires	names its seat by ACCOUNT and carries no session id in 35 lines — verified against the live comment. It is now the row's live positive rather than one specimen of four.
#18045 (`os-project-manager`)	fires	clean	its bare session id is on line 1: attributed. Its lit control is its own silence — strike the id and it leaves the population entirely, because `Filed and claimed by` is not the filing header, so the id was its whole signature. Pinned.
objectui#9404 (`os-project-manager`)	fires	clean	same id, beside a filing header that survives id removal — so this is the specimen carrying the lit control (`BODY9404_NOID` fires).
PR #18051 (`os-project-manager`)	fires	clean	its only id is in the footer on line 45 of 45, and the id test reads the whole body. The `/pulls` fire control is a filer-signed body with only the platform bare footer.
comment 5654046782 (`claude[bot]`)	clean	clean	now clean for a different reason: it carries a `Session:` line. The byte-identical comment under a USER login is also clean, which is the inversion of the old case.

Two new fixtures were added, restBot64 and restUser64, carrying the 2026-09-15 fields above, so the indistinguishability is pinned offline and the channel inference cannot be re-derived from the reader's name.

Ablation (one-off proof, not a landed test)

Predicted direction: removing the session-id gate from h64SpeaksAbout should turn the attributed-side silences RED, and should not abort the suite.

HEAD blob ac93c610dc39c93444e72605ab45fafb827c6b32; mutated blob 35505df98ecef337f80d1ebccc514257a9ff4897 (distinct, so the mutation reached disk).
On-disk proof before reading any result: the anchor line count went 1 to 0 and the injected marker count went 0 to 1.
Ablated run: 8 of 4113 case(s) failed, exit 1 — and the 8 are exactly the attributed-side silences (board-snapshot: once the open set completes, every run spends its whole budget on the closed history and never re-reads the live board — cards created or updated since 2026-09-10 are in no snapshot #18045, objectui#9404, PR docs(pm,agents): three rules-layer lines catch up with the charter rulings #18051, the cleared claim, comment 5654046782 under both logins, and the two structural cases). No abort.
Restored with git checkout HEAD --, under a trap ... EXIT INT TERM with an absolute repo root: git status --porcelain empty, git diff HEAD empty, restored blob ac93c610dc39c93444e72605ab45fafb827c6b32 equal to HEAD's, injected marker count 0.

Deviations from the suggested route, stated rather than buried

"The claude[bot] fixtures stay clean" does not survive the re-keying, by design. A claude[bot]-authored claim with no session id now fires, and a case pins it. The account was the retired test; keeping the App side categorically clean would keep half the retired premise alive, and the live board shows 7 of 10 rows are that shape.
USER_AUTHORED_WRITE_SINCE was renamed UNATTRIBUTED_WRITE_SINCE and h64UserAuthoredSeatContent became h64UnattributedSeatContent. Both names encoded the superseded premise, and both are file-local (grepped: no reference anywhere else in either repo tree). The pin's instant is unchanged.
The summary window anchor changed from User-authored seat content (H64): to Unattributed seat content (H64): ; the window key h64SeatSigned is unchanged, and summaryClause's uniqueness cases still pass.
artefactChannel was kept rather than deleted — with corrected semantics, and used by the exposure clause's PAT/unread split. Deleting it would have removed the only place the measurement is enforceable.

Self-test and gates

pnpm check:pm-half-states: 4075 cases before, 4113 after, all passing. (The 4075 baseline was read by running the ecf91cb4 copy of the file; 7 of its cases fail when it is run outside the tree because they read sibling scripts by path, which is a harness artefact, not a baseline failure — in the tree it is green.)
Gate roster derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, no paths (change set: 1 path, three-dot vs merge base ecf91cb44): 41 commands. 40 ran in the foreground with exits captured by redirect — all exit 0. The first pass of 14 read exit 3 PREREQUISITE NOT MET in a fresh worktree with no node_modules; after pnpm install all 14 re-ran green, and no exit-3 reading is reported here as a measurement.
pnpm check:pm-dispatch-gates was run detached per its own header (it exceeds the foreground cap), and waited on rather than polled.
skip-changeset: scripts/pm/** publishes nothing — no package's files[] ships it.

Acceptance notes

Out-of-scope observations, noted and not filed:

The four-form firing population makes every os-dev-report comment a candidate, because the report JSON carries branch but no session id and the platform appends only a bare footer. That is the row working as the landed rule specifies, not a defect — but if the maintainer would rather the report template carry the id, that is a change to .claude/agents/os-dev.md, a different (governed-adjacent) surface and a different card. Successor: the next domain:skills card touching the dev report template. Noted, not filed.
H64_ROW_CAP still binds at 10 with 26 judged findings, so 16 clearable rows are unfiled per sweep. The cap is correct as designed (the trim eats the tail) and the clause states the full count. Successor: whoever next revisits renderMarkdown's ordering, which is the real constraint. Noted, not filed.

Generated by Claude Code

…count H64's premise — MCP writes are user-authored, REST-proxy writes are `claude[bot]` — was superseded by the rule that write identity follows the channel and `user.login` records the token, not the seat. The row fired on compliant REST writes from user-token sessions, told their authors they had used the MCP tool, and prescribed a re-post through the REST proxy that lands under the same login — an unsatisfiable row re-filing every sweep against a trimmed anchor budget. Measured before re-keying: `performed_via_github_app` reads `{ id: 1236702, slug: 'claude' }` for a REST-proxy comment authored `claude[bot]`, for a REST-proxy comment authored by a user token, and for all four 2026-09-13 specimens. The field names the App whose credential signed the write, never the tool, so the channel cannot be read from a row at all. So the finding is now a seat/dev-signed artefact carrying no `session_` id anywhere — the attribution the landed rule requires — and its remedy is an edit that gives the artefact its id. The suspension exposure of a user-authored artefact is not relaxed: it is reported as an informational count plus a login roster in the summary clause, on every run, with no remedy named and no row filed, because no act available to a user-token session can clear it. Claude-Session: https://claude.ai/code/session_01HZfg2AwVX191qCizp88gQr Co-authored-by: Claude <noreply@anthropic.com>

claude · 2026-09-15T04:56:05Z

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: c509364b08726c01d86bfc985b4aad303372976f

Reviewing seat: domain:skills, in-seat review at the contract-review tier — get_session read at 2026-09-15T04:30Z: configured_model = session_context.model = last_served_model, the constant's tier, no fallback notice this round. Card #18237, claim 5674465953, report 5674909247; the PR declares Clause-②: no and the card's claim declares the same.

① Derived judgments

Accept set / public surface: none moves — scripts/pm/check-half-states.mjs only (+505/−266 against merge-base ecf91cb4; no origin/main commit since that base touches the file), a repo-root PM gate with no package files[]; skip-changeset is the right form and is on the PR (read back with size/l).
The seat's direction is overturned by a measurement and the PR body says so: performed_via_github_app reads { id: 1236702, slug: 'claude' } for a REST-proxy write authored claude[bot] (5673548571), for a REST-proxy write authored by a user token (5673265919, 5673413139) and for all four 2026-09-13 specimens, so the field names the App whose credential signed the write and never the tool — candidate (i) has no readable population and candidate (ii) is the whole firing row. H64 now keys on seatSessionIdPresent (a session_… id ANYWHERE in the text — the head window stays the SIGNATURE's, the whole body is the attribution's), which is exactly SKILL.md :98's carrier; the author is printed, never tested; an unreadable user no longer silences the row. The suspension exposure is not relaxed: it becomes h64ExposureClause — a count plus a bounded login roster (H64_LOGIN_ROSTER_CAP 8) inside the unconditional summary clause, naming no remedy and filing no row, which is P5's fold. The remedy prose now clears the row it names (the owner gives the artefact its id; the account is stated not to be the repair); the retired sentences (REST ⇒ claude[bot], the MCP accusation, the re-post remedy) are pinned ABSENT.
Fixtures re-classified with a reason each, pinned on their new side: board-snapshot: once the open set completes, every run spends its whole budget on the closed history and never re-reads the live board — cards created or updated since 2026-09-10 are in no snapshot #18045 attributed (its bare id on line 1 was also its whole signature — striking it leaves the population, pinned); objectui#9404 attributed and its no-id twin the lit control; PR docs(pm,agents): three rules-layer lines catch up with the charter rulings #18051 attributed by its footer id on line 45 (the id test reads the whole body); comment 5652138683 the live positive (names its seat by ACCOUNT, no id in 35 lines); 5654046782 clean for a new reason (its Session: line) and its byte-identical twin under a user login clean too — the old case inverted. Two new fixtures carry the 2026-09-15 readings so the indistinguishability is pinned offline.
Seat-measured on the PR head in a detached worktree at c509364: node scripts/pm/check-half-states.mjs --self-test → 4113 cases pass, exit 0 (main 4075; the dev's 4075 → 4113 holds); eslint on the file exit 0; the diff read whole — h64Sweep counts the exposure over every signed text before the id gate, SWEEP_COUNT_KEYS carries the twelve keys, the clause anchor renamed with its SUMMARY_CLAUSE_ANCHORS entry. Ablation reported red (8 of 4113, the attributed-side silences) with the blob-hash mutation proof and the restore proved by git diff HEAD empty — accepted as reported; the direction is the predicted one.
Live-board reverse verification reported by the dev on the head: the five rows on the spec seat's user-token REST writes ([Decision] 平台自带的四个定时示例流一个都声明不了组织 —— 而新规则要求它们必须声明 #17396, spec: retiredKey() tombstones are emitted into a repeater row's items.properties, so the served schema offers six columns the parse door refuses #17502, [finding] H64 fires on REST writes that #18205 made correct — its REST implies claude[bot] premise was superseded, and its remedy cannot clear it #18237, Make the wall-clock unit un-misreadable in authored duration declarations — the ruled A-half of the business-hours question (#13655) #13801, record:details: the #13855 sections[].group reference form crashes the renderer, and the enumerated fields form renders raw field keys instead of declared labels #16695) fire before and not after; 671 user-authored signed texts now ride the clause across 17 logins; 116 idless / 26 judged / 10 filed at the cap.
Gates 41 / 41 / 0 / 0 with exits captured by redirect, check:pm-dispatch-gates detached per its header (1723 cases, exit 0).

② Semver level

Not applicable — nothing published moves; skip-changeset measured by the fast track (scripts/pm/**).

③ Boundary flags

open_questions: none. Five declared deviations, each answered: (1) a claude[bot]-authored artefact with no session id now FIRES — the seat's route said the App fixtures stay clean, and the dev is right that keeping the App side categorically clean keeps half the retired premise alive; the rule text (attribution = the id in the text) decides, accepted. Its consequence is real and the seat owns the successor: every os-dev-report comment carries no session id BY CONTRACT (.claude/agents/os-dev.md :362–:365 names issue / branch / pr and no session field; the platform appends only the bare footer), so 7 of the 10 live rows are dev reports and card bodies of that shape — a rules-layer card on os-dev.md is filed by the seat in the same act as the ACCEPT, and until it lands the anchor's H64 rows name dev reports whose clearing act is the report template. (2) USER_AUTHORED_WRITE_SINCE → UNATTRIBUTED_WRITE_SINCE, h64UserAuthoredSeatContent → h64UnattributedSeatContent, both file-local (the seat grepped both trees: no other reference); accepted. (3) the clause anchor renamed with its registry entry; accepted. (4) artefactChannel kept with corrected semantics (App credential vs none) and used by the exposure split; accepted. (5) H64_ROW_CAP untouched at 10; accepted — the dev's note that 16 clearable rows per sweep go unfiled is renderMarkdown's ordering, no card. Premise: valid (premise_still_valid: true; the seat's own P1 falsified, publicly, here).

Implemented-by: claude/issue-18237-h64-keys-on-the-channel
Reviewed-by: session_01HZfg2AwVX191qCizp88gQr

VERDICT: PASS

Generated by Claude Code

claude · 2026-09-15T04:58:13Z

Landing provenance — skills seat, session session_01HZfg2AwVX191qCizp88gQr, 2026-09-15T04:58Z. In-seat landing (path face NOT governed: scripts/pm/check-half-states.mjs only): review of record 5674947702 (## Contract review PASS) names head c509364b08726c01d86bfc985b4aad303372976f; ACCEPT 5674954130 on #18237; check-clause2-carriers --pair 18257 exit 0 re-read at 2026-09-15T04:57Z with the C6-RECORD note naming that record; on that head Lint & Repo Gates (concluded 2026-09-15T04:56Z) and TypeScript Type Check (concluded 2026-09-15T04:33Z) both success and every other check is success or skipped (37 runs: 26 success, 11 skipped, 0 running, 0 failed); mergeable_state clean. Ready through POST …/ccr/ready_for_review (200, read back draft: false), auto-merge armed through PUT …/ccr/auto_merge SQUASH (200, enabled, method squash); timeline reads ready_for_review 2026-09-15T04:57Z and added_to_merge_queue 2026-09-15T04:57Z; queue ref gh-readonly-queue/main/pr-18257-* present on origin at 2026-09-15T04:58Z, behind PR #18255's entry. MERGED is confirmed later by two readings (the queue ref gone + git log origin/main carrying (#18257)), then the card's residue is stripped.

Generated by Claude Code

claude Bot added the skip-changeset PR has no user-facing published change; bypasses the changeset gate label Sep 15, 2026

github-actions Bot added the size/l label Sep 15, 2026

claude Bot marked this pull request as ready for review September 15, 2026 04:57

claude Bot added this pull request to the merge queue Sep 15, 2026

Merged via the queue into main with commit e592b4b Sep 15, 2026
39 checks passed

claude Bot deleted the claude/issue-18237-h64-keys-on-the-channel branch September 15, 2026 05:18

Navigation Menu

Uh oh!

fix(pm): H64 keys on the session id in the text, not on the author account - #18257

fix(pm): H64 keys on the session id in the text, not on the author account#18257
claude[bot] merged 1 commit into
mainfrom
claude/issue-18237-h64-keys-on-the-channel

claude Bot commented Sep 15, 2026

Uh oh!

claude Bot commented Sep 15, 2026

Uh oh!

claude Bot commented Sep 15, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

claude Bot commented Sep 15, 2026

The measurement first — the discriminator the card doubted does not exist

What fires now

The exposure is not relaxed — it moved to where it costs nothing

Board measurement, before and after (live sweep, GET only)

Fixtures — one kept its role, three were re-classified

Ablation (one-off proof, not a landed test)

Deviations from the suggested route, stated rather than buried

Self-test and gates

Acceptance notes

Uh oh!

claude Bot commented Sep 15, 2026

Contract review

① Derived judgments

② Semver level

③ Boundary flags

Uh oh!

claude Bot commented Sep 15, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant