[kyle-sexton]

Closes #964

Summary

The model legs inside pr-review-cli.yml and intake-triage.yml kept no transcript, so the daily puller in melodic-software/claude-code-plugins#7676 had nothing to copy for them. After this PR, each run-claude-cli leg's execution file is scanned, sealed to the caller's public key and uploaded as a 7-day artifact. No plaintext transcript reaches a log or an artifact. Nothing changes for a caller until it passes the key.

Fix

• New action .github/actions/upload-lane-transcript. It runs ../scan-upload/scan.cjs on the plaintext, gzips and seals the file in memory, and uploads only transcript.json.gz.sealed as lane-transcript-<job-id>-<attempt>, retention-days: 7, overwrite: false. The record is claude-code-plugins' seal-lane-output format, bound to the run id and job id.
• New optional input lane-transcript-public-key on both reusables, default empty. An empty key skips the scan and the seal with a notice. In the model action the scan runs first; here it does not, because callers float on @main with no key until the follow-up below lands, and every leg would otherwise clone claude-code-plugins and could raise a scan error for a transcript nobody keeps.
• Steps. Each leg gets one step with if: always() && steps.<leg>.outputs.execution-file != '', continue-on-error: true and timeout-minutes: 3. That is 4 per job in the three pr-review-cli.yml model jobs and 1 per job in the three intake-triage.yml model jobs. A sealing failure or a malformed key fails only that step.
• Placement. The steps run after the outcome step, not directly after the leg. report-lane-outcome/action.yml:360-362 compares the clock at the outcome step with step-started-at, so upload time before it could turn a failed leg near its budget into timeout.
• Job limits. Each model job's limit grows by what its new steps can take: review 95/103 to 107/115 (code/security), intake 16 to 19. continue-on-error does not stop a job timeout from turning a lane red.
• Job ids. <lane>-<github.job> plus -2, -3, -4 for the chunk legs (code-review-key-2), and intake-<github.job> (intake-subscription). github.job must stay an expression because the parity tests deepEqual the three jobs' steps. In a reusable it is the called job's key: real artifacts read pr-review-cli-handoff-code-review-subscription-<run>-<sha>. The lane prefix keeps ids apart if one run calls both lanes, as this file's other artifact names already do. The puller's ARTIFACT regex (scripts/lane-transcripts/store.mjs) is greedy up to the last -, so lane-transcript-code-review-key-2-1 parses as job code-review-key-2, attempt 1.
• Docs. The run-claude-cli execution-file output no longer says "do not upload it" and now points at this action. Both SECURITY MODEL headers, both canonical callers and the README action list name the new step.

No README or package.json for the action, as with scan-upload and run-claude-cli; its action.yml description holds the contract. The seal code is CommonJS, since this repository has no .mjs and CI globs *.test.cjs, and it keeps only the single-record path (no bundle or chunk code).

Verification

• Compatibility with the real puller code (one-off script, not committed): sealed with this action, then opened with claude-code-plugins origin/main open-lane-output/open.mjs openRecord and named with store.mjs parseArtifactName, for all 27 job ids: opened 27 of 27 job ids, every name parsed, every plaintext byte-identical, transcriptMetrics read each one.
• node --test on upload-lane-transcript-seal.test.cjs, intake-triage-legs, pr-review-cli, lane-outcome-emitters, network-timeout-policy, github-script-bodies, ci-fanout-consolidation and run-claude-cli/lane.test.cjs: 261/261 pass. New tests assert every leg has an upload with a unique, parseable job id. The seal test lives in .github/scripts, so the existing test-contracts glob runs it with no workflow edit; that also lets the review lanes run here, since claude-code-action skips itself when pr-require-checks.yml differs from main.
• node --test on the 11 other suites that read these workflows (claude-lane-job-gates, claude-review-declared-outputs, claude-review-outcome-wiring, pr-review-cli-api-calls, report-api-budget, check-key-credit, intake-triage-apply, intake-triage-gate, claude-lane-status-check, claude-lane-fallback, claude-lane-install-isolation): all pass.
• actionlint from the repository root on the three changed workflows: no findings.
• zizmor --offline on the two reusables and the new action.yml: no findings (7 suppressed).
• composite-run-shellcheck.sh on the new action.yml: 2 run blocks checked, no findings.
• biome check on the edited tests, markdownlint-cli2 README.md, typos: clean.
• Left out: a live sealed upload, which needs a caller passing the key (see Remaining); the rest of .github/scripts/*.test.cjs, which CI runs.

Related

• melodic-software/claude-code-plugins#7676: the transcript sink and puller this feeds.

Remaining

• Runner policy, first: add lane-transcript-public-key to allowedInputs of both reusables' entries in melodic-software/standards components/runner-policy/policy.json, synced to .github/standards/runner-policy/policy.json in each repository. Neither entry lists it today, so the caller edit below would fall outside the policy's allowlist. The policy pins no job limit, so the limit changes here need no policy edit.
• Caller edit in melodic-software/claude-code-plugins, not in this PR: add lane-transcript-public-key: ${{ vars.LANE_TRANSCRIPT_PUBLIC_KEY }} to the with: of the pr-review job in pr-review-hosted.yml, the pr-review-security job in pr-review-security-hosted.yml and the intake-triage job in intake-triage.yml. Until it lands, every leg only prints the skip notice. The close of claude-code-plugins#7676 depends on it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VQb9Q1VgLEiwGgRVshK3b7

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-02T23:55:05.862913Z	3445cf7	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed 155848268805132ca3843768a26fe4cc9eb6b560: 0 finding(s), 0 as inline comments.

Incremental review of .github/scripts/upload-lane-transcript-seal.test.cjs (since 1f5acaa). The change is formatting only: it reflows the makeTemporaryDirectory import and wraps the path.join calls for the action.yml read and the seal-transcript.cjs spawn path. The paths and arguments are the same, and no logic, environment, or trust-control code changed. The file is a test, not a trust control.

No security findings.

critical-findings: 0

[github-actions]

claude-review reviewed 155848268805132ca3843768a26fe4cc9eb6b560: 0 finding(s), 0 as inline comments.

No findings. I reviewed only the change since 1f5acaa, which is in .github/scripts/upload-lane-transcript-seal.test.cjs.

The change is formatting only, in three places:

• The makeTemporaryDirectory import is collapsed to one line.
• The fs.readFileSync call that loads action.yml is wrapped across lines.
• The seal-transcript.cjs path argument to spawnSync is wrapped across lines.

Reading the diff, the arguments, the paths and the control flow are the same before and after. No test behaviour changes.

I did not run node --check or the tests: the sandbox blocked the command and I did not retry it. The conclusion rests on reading the diff.

I did not re-review the other files of the PR; they are unchanged since the earlier review.