Repository navigation
[finding] platform-readings: the REST proxy's write identity changed mid-shift with no seat action — content writes read back claude[bot] (installation) until 2026-09-15T20:04Z and os-zhuang / User via app claude (user-to-server) from 2026-09-16T01:53Z — the register has no row saying the token class is re-read, not assumed #18350
Description
Activity
Consequence measured — skills seat, session
session_01HZfg2AwVX191qCizp88gQr, 2026-09-16T02:33Z. With the proxy writing asos-zhuang, the draft PRs this seat's devs open are AUTHORED by os-zhuang (PR #18351 read backuser.login=os-zhuang), and GitHub then refuses to request a review from the author:POST /pulls/18351/requested_reviewerswith["os-zhuang","hotlong"]→ 422 「Review cannot be requested from pull request author」; the same call with["hotlong"]alone → 201. GitHub also does not let a PR's author submit an APPROVED review on it, so while this identity stands, os-zhuang cannot be the authorized approver of any governed PR this seat opens — only hotlong can, and ruling C depends on that approval. This turns the identity change from an audit concern into a throughput one: every rules-layer PR from this seat now waits on hotlong alone. Remedy stays the one in the card body: a dedicated low-privilege GitHub user behind seat sessions, or the App installation token (claude[bot]), which authored every seat PR before 20:04Z and was requestable / approvable by both approvers.
Generated by Claude Code
Closed as a duplicate of #18334 (skills-lane self-triage): the triage seat measured the same write-identity flip first (2026-09-15T22:55Z,
claude[bot]→os-samon #6015), and this card's table is the second measurement of it on this seat's writes (claude[bot]at 2026-09-15T20:04Z →os-zhuangat 2026-09-16T01:53Z). Nothing here is dropped — the table, theGET /userreading, the maintainer's account of the mechanism and the reviewer-request consequence (422 on PR #18351) are carried onto #18334 in one comment, and #18334's grading names the three rows this card asked for (:129 / :130 re-keyed, one consequence row), folded with #18320 into one platform-readings PR. Skills seat, sessionsession_01HZfg2AwVX191qCizp88gQr, 2026-09-16T03:19Z.
Generated by Claude Code
- added a commit that references this issue
on Sep 17, 2026
Filed by the
domain:skillsexecution PM seat, sessionsession_01HZfg2AwVX191qCizp88gQr, at 2026-09-16T02:11Z, from its own readings this shift.Measured (comment authors read back through
GET /issues/comments/{id}, same session, same toolscripts/pm/post-stamped.mjs, same REST proxy)user.loginuser.typeperformed_via_github_appclaude[bot]claudeos-zhuangclaudeos-zhuangclaudeGET /useron the same credential readsos-zhuang/Userat both instants (it did at seating, 02:03Z the day before, when the writes read backclaude[bot]).os-zhuang) where it injected the installation token before;performed_via_github_appreadsclaudein both classes, so it is NOT a discriminator (already on the register);user.login+user.typeon a read-back IS.merged_by/ comment authors alone cannot tell the seat from the approver any more.What is asked (⛔ not asserted — the seat grades)
One or two rows in
references/platform-readings.md(write side): the proxy's token class can change between two writes of one session with no seat action; the class is read on every write's read-back (user.login+user.type), never carried forward from the round-open marker; a change is recorded on the seat post at the next refresh with both instants. Optionally one line in the seat-post protocol: the round-open marker's identity reading is dated, not standing. References tier (in-seat contract-tier review). Fold candidate with #18320 (same file); the ceiling is 466 with 463 in use.Grading (lane
findingself-triage; class (c), a platform fact change measured with a control): p3 · Task ·pm:queue·domain:skills.查重词
write identity flipped mid-shift·installation token to user-to-server·claude[bot] os-zhuang read-back·performed_via_github_app not a discriminator·token class re-read every writeGenerated by Claude Code