diff --git a/governance/OP-ACCESS-20261007.md b/governance/OP-ACCESS-20261007.md index 00aa5ba05..be8cc9ad6 100644 --- a/governance/OP-ACCESS-20261007.md +++ b/governance/OP-ACCESS-20261007.md @@ -265,6 +265,89 @@ decision, and until it is recorded the choice is pending. work continues. - Never use a ruleset bypass, and never simulate another identity's approval. +## Addendum, 2026-10-12 — observations after the record landed + +Live reads 2026-10-12 23:14Z to 2026-10-13 00:07Z. Nothing above is rewritten; +each item below carries its own date and scope. + +- **Landing.** This record landed as `b7deb198` (pull request + [#282](https://github.com/d6g8k5htny-coder/main/pull/282), merged 23:49Z by + a Codex integrator; landed tree equal to the tested carry; the three files + equal the reviewed blobs). The post-merge workflow run failed once at its + checkout step (a fetch timeout, attempt 1) and passed on a whole-run re-run + (attempt 2); no file changed between the two attempts. +- **Boards.** [main#229](https://github.com/d6g8k5htny-coder/main/issues/229) + stopped accepting comments on 8 October at GitHub's 2,500-comment limit; the + board continues as [main#307](https://github.com/d6g8k5htny-coder/main/issues/307). + Where owner actions 1 and 2 above say "report on main#229" or "reply on + main#229", read main#307. +- **Claude lane, a second observation.** A different Claude Code session + (`017Mi3…`), activated by the coordinator as successor custodian for the + thread and body tasks of #282 + ([6134522988](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134522988)), + resolved seventeen review threads through + `POST repos//main/pulls/282/ccr/comments//resolve` and patched the + pull-request body, each read back + ([6134573901](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134573901), + [6134585682](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134585682)), + and later, on the coordinator's recovery request + ([6134878898](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134878898)), + re-ran a failed Actions run through + `POST repos//main/actions/runs//rerun` + ([6134909547](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134909547)). + In the author session (`01UzNu9j…`) the harness had refused, on + 7 October, review-thread replies, thread resolutions and pull-request body + patches on #282 (the texts were held, as the pull request's comments + record; the table above does not list them). Those denials stayed in force + in that session and were neither retried nor routed by it; the coordinator + recorded that the successor's completed work establishes no ongoing route + for a denied action + ([6134690083](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134690083)). + What this adds to item 2, for exactly two operations — thread resolution + and pull-request body patch on #282 — is that a harness refusal in one + Claude session did not recur in another Claude session five days later. + Review-thread replies stay author-denied with no successor success on + record; the Actions re-run is a separate observation, with no earlier + refusal of a re-run recorded in the author session. Why the refusals did + not recur is still not established. +- **`AGENTS.md` pointer, denial lineage.** On 7 October the harness refused + this session's write of a one-line pointer to this record into `AGENTS.md` + (a shell heredoc that also carried the `docs/FORMAL_REQUIRED_CHECKS.md` + correction; the correction was then applied through the editor tool and + went through, the pointer was left out and recorded as refused in the #282 + description; the verdict text was not preserved). On 13 October, after a + context compaction had reduced that record to "deferred", this session + edited `AGENTS.md` through the editor tool and the edit went through; the + author treats that as an unintended retry of a refused action, withdrew + the edit from the pull request that proposed it ([#384](https://github.com/d6g8k5htny-coder/main/pull/384), + which now changes this file only). The owner can make the edit himself. + For agents, this session included, the pointer stays stopped unless the + original denial is resolved through a supported platform clearance or the + required handoff to the user, as [OP-WORKFLOW](OP-WORKFLOW-20260930.md) + requires for a permission or security denial; owner approval alone, or a + later write success, is not clearance, and the pointer is not routed to + another lane. The original verdict text being unavailable, which rule + refused the write is not known from this record. +- **Automated review.** The Codex automated reviewer answered the refresh push + with a usage-limit notice instead of a review + ([6134693329](https://github.com/d6g8k5htny-coder/main/pull/282#issuecomment-6134693329)); + the integrator recorded it as not run, not as a pass, and merged on the + unchanged reviewed blobs. A lane whose usage is exhausted is the third + situation that the workflow paragraph in the next item keeps apart from a + denial and from an unavailable tool. +- **Workflow text.** [OP-WORKFLOW](OP-WORKFLOW-20260930.md) gained, with + [#383](https://github.com/d6g8k5htny-coder/main/pull/383) at 23:31Z, a + paragraph on sessions that stop with a role unfinished: successor claims, + the predecessor's closeout kept UNKNOWN until its own release is found, + capability observed rather than inferred, a denied step never routed to + another executor, a new session not being clearance. The "For every lane" + Claude bullet above says the same for a single session. +- **Owner actions 1–4:** unchanged at these reads. Ruleset 23798639 still + lists the ChatGPT Codex Connector as an `always` bypass actor (`updated_at` + 2026-09-28T06:22Z); Math- 24045351 lists none. Whether a `cursor[bot]` + issue comment on main has succeeded since 7 October was not checked for + this addendum. + ## Unchanged Scientific acceptance conditions, `formal/` and register custody,