Land on the organization's repository as the merging account; stale-review rule and local reviews for the Codex gate - #1061

Scope

• What this PR changes:
  • Ports three operational landing fixes, each with fail-before tests:
    • nas_land.sh takes the repository from LANDING_REPO, default native-agent-stack/native-agent-stack;
    • the no-reply author is the merging account, LG_AUTHOR_LOGIN (CC correction Wire the order-contract boundary into every Alpaca paper submit (local refusal before any network call) #198);
    • a review that stays shown running for 90 minutes is "unavailable" (5) (CC correction Reconcile trading records with merged evidence (runtime-target, snapshot gate, dashboard, new-host EdgarTools receipt) #200).
  • Adds a reviewer-agnostic local review input for the Codex gate: LOCAL_REVIEWS, with the alias CODEX_LOCAL_REVIEW.
  • Adds tools/landing/review_normalize.py, which writes the normalized file from a codex review rollout or from the official /code-review plugin's comment.
  • The second commit, the CC's follow-up:
    • Codex's verdict is carried, and "patch is incorrect" blocks;
    • the plugin's arm is advisory.
  • The third commit fixes the reviews at c999eee, and the fourth the reviews at 974090c (below).
• Base commit: c63c83a822f2687c3e7aed2d3b3ede0413119c9b
• Lane: lane:foundation
• Owned paths touched:
  • tools/landing/ (README.md, codex_gate.py, nas_land.sh, uet_land.sh, new review_normalize.py);
  • tests/test_landing_scripts.py;
  • manifests/evidence.json (the six records of those files).

Per item

1. Repository slug.
   • R=${LANDING_REPO:-native-agent-stack/native-agent-stack}. The old slug is redirected, but GitHub deletes those redirects if a repository is created at the old location, so the default is the new slug.
   • Since the third commit, R must have GitHub's owner/name shape (exit 2 before any call), and every -R "$R" is quoted. Since the fourth, the rule cites the REST API's owner and repo path parameters and the docs' repository-name rule.
2. No-reply author (Wire the order-contract boundary into every Alpaca paper submit (local refusal before any network call) #198).
   • The shared lg_author_email uses ${LG_AUTHOR_LOGIN:-<R's owner>}, and both scripts set LG_AUTHOR_LOGIN=${LG_AUTHOR_LOGIN:-seathatflowsinourveins}.
   • A login that is not GitHub's shape is refused before any call. Since the third commit, only a users/<login> reply of type User gives an id (exit 11 otherwise). Since the fourth, the filter cites the API: type is a free string, "User" in its example and "Organization" for an organization's account.
   • The address form <id>+<login>@users.noreply.github.com is unchanged.
3. Stale running review (Reconcile trading records with merged evidence (runtime-target, snapshot gate, dashboard, new-host EdgarTools receipt) #200).
   • The operational block, byte for byte: a row shown running with no connector activity for 90 minutes gives 5.
   • Since the third commit, the age runs from the later of the last activity and the latest trigger, and a trigger after a stall gets the 10-minute NOSTART grace.
   • Since the fourth commit:
     • the latest review of our head since the trigger decides: one that ended other than Completed gives 4, before the stale 5 and before a 0, whatever completed before it, while a later completion still supersedes it;
     • the 👀 reaction with no Running row waits (3) with no bound again, as at c999eee.
   • OpenAI's Codex docs state no review time limit. The value's measured basis is nas#1056, and it lives in one constant.
4. Local reviews. The gate decides from the files alone and calls nothing. codex review is the required arm; the /code-review plugin is advisory until the review-replay addendum measures it.
   • A codex-review file carries Codex's verdict (patch is incorrect blocks with 4; any other value than the two stops with 2) and, since the third commit, the merge base its review request named.
   • lg_gate passes the gate <head>:<git merge-base <head> origin/main> for the read head and each pure rebase of it (the scripts' lg_same_change). A review against another base stops (2).
   • Since the third commit, a set but empty variable stops (2) instead of falling back to the connector.
   • Since the fourth commit, nas_land.sh requires local reviews. With neither variable set it stops with 2 before any call. The one exception is LOCAL_REVIEW_WAIVER, a file whose last Waiver: line reads Waiver: local reviews for #<pr> at <head8>. It is checked in verdict_gate and again by final_gate. A waived landing asks the connector as before, and its squash body records the waiver after GitHub's default body. uet_land.sh is unchanged.

The third commit: CHANGES_REQUESTED at c999eee, all three reviews

• The CC's Claude read.
  • P1 base: the review's base is checked.
  • P1 empty variable: a set but empty variable stops.
  • P2 order: our head's failed review gave 4 before the stale 5 only when no review of our head had completed. Astra's read at 974090c showed that an earlier completion still hid it, and the fourth commit fixes this (below).
  • The four P3s: the LANDING_REPO shape and quoting, type == "User", the Legacy rollout refused (fail closed, documented), and the bounded 👀. The fourth commit reverts the bound.
• The CC's /code-review probe (advisory).
  • Finding 1: the trigger grace.
  • Finding 4: the docs search, with the measured basis recorded.
  • Finding 8: codex_wait's message on a local 2.
  • Finding 9: the plugin comment is bound to --head by its full-SHA links.
• Astra.
  • P1-1: a set but empty variable stops (2). The residual Astra reproduced was an unset variable that proceeded on the connector's 5; it stood as policy at that commit. The fourth commit closes it on nas, where local reviews are now required.
  • P1-2: covered by the base check, with Astra's regressions: missing branch, another branch, base = head, base = head^, and the correct base.
  • P1-3: every digest and prefix of a host-only file (the CC's live copies) is removed from the README and this body, and a test refuses any that returns.
  • Follow-up 1 (fixed): a refused retry removes the earlier --out file.
  • Follow-up 2 (fixed): the gate refuses a non-finite confidence.
  • Follow-up 3: covered by P1-3.

The fourth commit: CHANGES_REQUESTED at 974090c, the CC's Claude delta and Astra

• The CC's Claude delta.
  • P1-1 closed on nas, by the CC's ruling: local reviews are required.
    • With neither variable set, nas stops with 2 before any call: before the lock, update-branch and ready.
    • The only exception is a waiver bound to the pull request and the read head's first 8 hex digits. It is read in verdict_gate, again by final_gate, and recorded in the squash body.
    • The check is in neither the shared block nor codex_gate.py, and uet_land.sh keeps the connector path.
  • The 👀 path waits (3) again, so this PR opens nothing new.
  • P3 sources: github/rest-api-description@58b1e0c0 for the users API's type and the owner/repo parameters, with the docs' repository-name rule.
  • P3 digest guard:
    • it reads any letter case and bare prefixes, and a commit id must be a named public commit;
    • it reads the four scripts, and the body when LANDING_PR_BODY names it;
    • its known digests are only the files git tracks.
• Astra's P2. An earlier completion after the trigger hid a later Cancelled review of the same head, so with another stale Running row the gate gave 5 instead of 4.
  • The latest review of our head since the trigger decides now, before the stale branch and before a 0.
  • Astra's sequence gives 4; it gave 5 at 974090c.
  • The same two reviews with no row running give 4; they gave 0 at 974090c. The base's test, which pinned "any completion evidence passes", is flipped.
  • The false claims in this body's earlier text and at the README's former :69 are corrected.
• Astra's follow-ups, the one-line fixes.
  • 2: the README states that a clean plugin comment binds no head.
  • 3: the probe's findings 6 and 7 are recorded below, with triggers.
  • 4: sources are cited for each mechanism: the set-or-unset tests (bash(1), cpython's os.environ), the users API's type, the repository name, os.replace and os.unlink, and json with math.isfinite.
  • 6: the date is on the test comment.
  • The other follow-ups are listed under Follow-ups.

SOTA sources

• Codex CLI, review mode: openai/[email protected] (installed: codex-cli 0.162.1), https://github.com/openai/codex/tree/rust-v0.162.1
  • codex-rs/protocol/src/items.rs:173-183: EnteredReviewModeItem.target, and ExitedReviewModeItem.review_output: Option<ReviewOutputEvent>.
  • codex-rs/protocol/src/protocol.rs:3503-3545: ReviewOutputEvent and ReviewFinding.priority: i32. :3507 gives overall_correctness: String, a free String, and :3509 the confidence.
  • codex-rs/prompts/templates/review/rubric.md:85: the two verdict values. :58 and :78: a priority may be omitted or null.
  • codex-rs/core/src/tasks/review.rs:94-95: an abort. :188-208: the parse fallback.
  • codex-rs/history/src/rollout_payload.rs:32-33: how a rollout line is tagged.
• The review's base (third commit):
  • codex-rs/prompts/src/review_request.rs:21: "The merge base commit for this comparison is {{merge_base_sha}}". :20: the fallback prompt with no SHA. :62-70: when each is used (sha256 f4896d67…).
  • codex-rs/git-utils/src/branch.rs:13-15: merge_base_with_head "mirrors git merge-base HEAD <branch>" (sha256 86774c34…).
  • Measured on the 6 local baseBranch rollouts: 3 named the merge base with origin/main, and 3 were delta reviews against a commit of the pull request's own branch.
• Legacy rollouts (third commit):
  • codex-rs/exec/src/lib.rs:1425-1432: falls back from paginated history; :1461 is the paginated default (sha256 51b62cef…).
  • codex-rs/rollout/src/policy.rs:141-155: only Legacy rollouts persist EnteredReviewMode/ExitedReviewMode (sha256 65ce50e8…).
• No upstream review time limit (Reconcile trading records with merged evidence (runtime-target, snapshot gate, dashboard, new-host EdgarTools receipt) #200's value):
  • https://developers.openai.com/codex/llms.txt (sha256 72459eb4…).
  • https://learn.chatgpt.com/docs/code-review.md (:129; sha256 c18b2fe1…).
  • https://learn.chatgpt.com/docs/third-party/github.md (:52, :170, :196-205; sha256 1ab50213…).
  • All fetched 2026-10-13; none states a limit. The basis is measured with the gate's own query on nas#1056: Running since 2026-10-12T22:56:31Z, its last connector activity 22:56:34Z, silent for 250 minutes at 03:06Z.
• GitHub repository-transfer redirects: https://docs.github.com/en/repositories/creating-and-managing-repositories/transferring-a-repository (github/docs@be38ec5d78e24172587e61b3c6ff40ace1865c71 content/repositories/creating-and-managing-repositories/transferring-a-repository.md:62, :68-69, sha256 ed316b3f…).
• GitHub no-reply address: https://docs.github.com/en/account-and-profile/reference/email-addresses-reference (github/docs@be38ec5d content/account-and-profile/reference/email-addresses-reference.md:47, "[email protected]", sha256 fbbf1c5f…).
• The login shape: github/docs@be38ec5d username-considerations-for-external-authentication.md:94, :103, :126-127 (sha256 591df6bd…).
• Claude Code's official /code-review plugin, run as shipped: https://github.com/anthropics/claude-plugins-official/blob/f359d43b406057f1a8cdbc4ea93b827642c162e6/plugins/code-review/commands/code-review.md (sha256 7d5a0bc9…).
  • :2 is the allowed-tools line, :11 the skip of an already reviewed PR, :26 "do not proceed" below 80, :50-92 the comment format, :80-84 the clean comment, which has no link, and :86-92 the full-SHA links.
  • Its README.md:157-167 covers "No review comment posted".
  • Installed client: Claude Code 2.1.296.
• GitHub GraphQL schema: https://docs.github.com/public/fpt/schema.docs.graphql (sha256 3ed09565…). It gives IssueComment.author and lastEditedAt (nullable) at :21215 and :21287-21290, and PullRequest.headRefOid at :44787. For the fourth commit: :45597-45605 viewerMergeBodyText(mergeType: PullRequestMergeMethod): String!, and :26079-26082 MergePullRequestInput.commitBody, "if omitted, a default message will be used".
• The users API and the repository name (fourth commit): github/rest-api-description@58b1e0c00b39 descriptions/api.github.com/api.github.com.json (sha256 da3d22e4…).
  • :88612-88614 GET /users/{username}.
  • :126248-126250 and :160055-160058: type is a free string, "User" in the example. :311038: an organization's "Organization".
  • :344898-344915: the owner and repo path parameters.
  • With github/docs@be38ec5d data/reusables/repositories/repo-name.md:1 (sha256 5a1343a0…): at most 100 characters, ASCII letters, digits, ., - and _.
• A waived landing's squash body (fourth commit): cli/[email protected] (the commit fc4b137c).
  • pkg/cmd/pr/merge/merge.go:146-147 and :176: --body sets the body. :721-735: the "edit the commit message" step starts from viewerMergeBodyText (sha256 ca3849b4…).
  • pkg/cmd/pr/merge/http.go:72-75 and :121-158: CommitBody, and the query (sha256 17eb5418…).
  • Measured on 2026-10-13 with the script's query on this pull request: the reply is the commit messages, the repository's squashMergeCommitMessage is COMMIT_MESSAGES, and --jq prints the string raw.
• Set or unset, and Python's file and number semantics (fourth commit):
  • bash(1) of the installed bash 5.3, Parameter Expansion: "Omitting the colon tests only for a parameter that is unset".
  • python/[email protected]: Doc/library/os.rst:184-186 (os.environ, a mapping), Lib/os.py:704, and Lib/_collections_abc.py:808-813 (get gives None only for a missing key).
  • The same tag: os.rst:2754-2760 (os.replace), :3671-3676 and :2661-2665 (os.unlink), Doc/library/json.rst:308-311 (float(num_str)), and Doc/library/math.rst:370-373 (isfinite).

Evidence-class table

Local commands run

$ python3 -m unittest tests.test_landing_scripts      (LANDING_PROHIBITED_NAMES from the CC's list)
Ran 156 tests in 93.126s
OK (skipped=1)
$ CI=true python3 -m unittest tests.test_landing_scripts
Ran 156 tests in 91.708s
OK (skipped=1)
$ python3 -m unittest tests.test_landing_scripts      (against 974090cf's tools/landing)
Ran 156 tests in 100.928s
FAILED (failures=31, skipped=1)
$ LANDING_PR_BODY=<this body> python3 -m unittest tests.test_landing_scripts.ScriptTextTests.test_the_pull_request_body_binds_published_bytes
Ran 1 test in 0.021s
OK
$ python3 -m unittest tests.test_validate_shards
Ran 13 tests in 3.658s
OK
$ python3 -m unittest tests.test_published_digest_privacy      (with the names)
Ran 15 tests in 12.728s
OK
$ python3 -m unittest tests.test_evidence_manifest
Ran 5 tests in 0.005s
OK
$ python3 scripts/validate.py
{"components": 71, "hashed_files": 11861, "profiles": 4, "receipts": 240, "status": "passed"}
$ python3 scripts/evidence_manifest.py --check
{"files": 11861, "status": "passed"}
$ gitleaks git --staged --no-banner --redact .        (and --log-opts="c63c83a8..HEAD")
INF no leaks found (exit 0)
$ privacy_scan.sh <the added lines> <the commit message> <this body>
clean (exit 0)

shellcheck was not run: CI does not run it on these scripts, and it is not installed on this host. bash -n passes for both.

Decision record

None separate. The README's "Local reviews" section and its 2026-10-13 sources tables record the schema, the required and advisory arms, the base rule, nas's requirement and its waiver, the pins, the measured facts and the untested boundaries.

Host evidence

Not applicable: no file under evidence/hosts/ changes.

Checklist

• No GitHub Actions or workflows changed.
• No workflow permissions changed.
• No secrets are printed, logged or committed; no new secret.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

Risks

• --commit reviews are refused. The normalizer accepts only a baseBranch review, the pull request's whole change against its base. A review the co-op wrapper runs with --commit covers one commit and is refused, so the required arm must run in the wrapper's default mode, against the head's merge base with origin/main.
• The latest-review rule also applies on uet's connector path. One trigger can be followed by a completion and then a Cancelled review of the same head. That now stops both scripts (4, exit 9) where it passed, so a person decides. A re-run that completes after the failure lands as before.

Follow-ups

• Astra's follow-up 1 at 974090c. A usage error in review_normalize.py (exit 2, such as --head bad) leaves an earlier --out file. Trigger: retry automation, or a claim that every failed attempt invalidates the output.
• Astra's follow-up 2. A clean plugin comment binds no head; the README's Known gaps state it. Trigger: before the plugin arm becomes required.
• Astra's follow-up 3, the probe's findings 6 and 7.
  • The owner fallback of lg_author_email is reachable only through the shared helper, since both scripts set a login. Trigger: the helper's calling convention changes.
  • lg_gate repeats per-file git work at every gate call. Trigger: large local review sets or repositories; measure it first.
• Astra's follow-up 4. The aggregate fail-before counts and the compatibility controls are this body's figures, not packaged receipts.
• Astra's follow-up 5. PUBLIC_DIGESTS names public artifacts but does not verify their bytes against a mirror. Trigger: the guard's next change.
• Astra's follow-up 6 (inherited). Host-name occurrences on 18 unchanged lines of manifests/evidence.json; this PR changes none of them.
• Astra's follow-up 7. The cited captures are not packaged: the rollout and live-copy measurements, nas#1056's activity capture, the client-version captures and the fail-before receipts. Packaging them is a records task.

🤖 Generated with Claude Code