[rocklambros]

What changed

This replaces the intake bot's two bash steps with one Python gate, tools/pr_intake.py, that queues a pull request no maintainer asked for, reading that from a lead's review or hand clear, the author's merged record, an assignment, a claim on a help wanted issue, or an editorial tick on the documentation lane. The bot owns the status:needs-triage label, keeps one comment per pull request edited in place, turns auto-merge off where the design names the harm, and lists the queue on the health issue, where the roadmap monitor raises the lead alarm when the nightly sweep fails. Enforcement stays off until two repository variables arm it, so a merge changes nothing beyond today's behavior, and the design, its premortem, and the plan are committed under design/ with the design's Amendments section recording every change made since version 3.0.

Which issue does this implement

Part of #362

• This is an editorial correction on the documentation lane (typo, grammar, link, formatting) with no change in meaning, so it needs no issue

Base branch

• integration, because this touches the specification, schemas, a reference
  implementation, an adapter, tests, or CI
• main, because every changed path is on the documentation lane allowlist in
  CONTRIBUTING.md

It changes workflows, tools, and tests, so it lands on integration and reaches main by promotion.

Type of change

• Specification change (schema, hooks, events, AgBOM)
• Reference implementation or adapter
• Documentation
• Tooling or CI
• Governance (licensing, security policy, contributor docs)

I tested this

• I synced my branch with the base branch before opening this
• uv run pytest -v passes on my machine
• uv run mkdocs build --strict passes on my machine

• uv run pytest -q: 1222 passed, 1 skipped.
• uv run mkdocs build --strict: built with no warnings.
• uv run --locked --only-group zizmor zizmor .github/workflows/: exit 14, 2 high findings (github-app), both in .github/workflows/board-reconcile.yml, which this branch does not change and which is identical on integration. The three workflows this branch changes (pr-intake.yml, roadmap-sync.yml, monitor-roadmap.yml) report no findings. The zizmor workflow runs with --no-exit-codes, so CI does not fail on these.

Checklist

• Commits are signed off with git commit -s (required by the DCO)
• Prose follows STYLE.md
• No secrets, tokens, or internal URLs in the diff

Security

• This change has no security impact

The workflow runs on pull_request_target, so it holds the base repository's token on fork pull requests. These are the controls around that token.

• permissions: {} sits at the top of pr-intake.yml. The one job holds contents: read, pull-requests: write, and issues: read, each commented with its use.
• The job checks out the pull request's base ref with persist-credentials: false, never the head. Its if admits pull request events only against main, integration, or release/*, the ruleset-protected bases, so only code a lead approved runs beside the token.
• Every gh call goes through tools/gh_runner.py: absolute /usr/bin/gh, a minimal environment, and an allowlist with one argv shape per operation. No title, body, path, or login from a pull request reaches argv, a shell, or a workflow command. The allowlist refuses -f, -F, --jq, -q, -t, and --template in every position.
• Every comment and label body goes to gh on stdin as JSON.
• The bot finds its own comment by its exact login plus a marker, and parses state only from that comment's first line with a strict pattern, so a contributor's comment cannot forge state.
• No ${{ }} appears inside any run:, and zizmor reports no findings on the three workflows this branch changes.

Rollout

The owner accepted the deviation on 2026-10-30. The line is in the design under "The owner writes one line here".

The checklist below is posted on issue #362 after merge and ticked there. Undo order at any step: unset PR_INTAKE_ENFORCED, then dispatch mode=retract.

• Precondition. A project lead: the GOVERNANCE/CODEOWNERS reconciliation pull request is merged, or a project lead writes on issue PR intake: one gate for work no maintainer asked for #362 that the three roster entries without repository access stay, and why.
• (a) Rock Lambros dispatches mode=dryrun with the variables unset. Expected: a summary with zero writes and arm with PR_INTAKE_TOOL_SHA=<value>. He posts the would-queue numbers and reasons on issue PR intake: one gate for work no maintainer asked for #362, naming PR docs: conformance is a technical measure — it cannot attest human intent #24, PR Batch Guardian log writes and drain them on shutdown #112, PR Record OpenCode host decision-honoring evidence for #92 #113, PR docs: acs-verdict-evidence-v1 predicate (draft) #297, PR Match toolCallRequest summary to schema #253, PR Tighten forward compat for 0.x versions #255, and PR Run AGT pin check and upstream watch in CI #259, and the three actions that clear them: approve, a fresh request for changes, or assign.
• (b) Rock Lambros posts the same in #team-genai-asi-acs-general and on the weekly call agenda.
• (c) Project leads clear the collateral. Done when a re-dispatched dryrun lists none of the named numbers, or a project lead writes "proceed with the remainder" on issue PR intake: one gate for work no maintainer asked for #362.
• (d) A project lead opens a throwaway pull request with auto-merge enabled, dispatches evaluate on it, and comments on issue PR intake: one gate for work no maintainer asked for #362 which write the summary reports and the value of autoMergeRequest afterwards. Expected: null.
• (e) A project lead promotes. Expected: the next roadmap sweep renders "Queued pull requests" and the next monitor run logs the intake line.
• (f) A project lead with repository admin edits the live help wanted label description, dispatches retract with authors=bots to clear the seven inherited Dependabot labels, then in one sitting outside 03:00 to 04:30 UTC sets PR_INTAKE_TOOL_SHA from the dry run's line, sets PR_INTAKE_ENFORCED=true, and dispatches evaluate. Expected: mode=enforced in the summary.
• (g) A project lead reads the next morning's health issue, sets PR_INTAKE_BASELINES from its "Queued pull requests" counts, and comments the values and the review date on issue PR intake: one gate for work no maintainer asked for #362.
• (h) A project lead marks issue PR intake: one gate for work no maintainer asked for #362 completed by hand, which GOVERNANCE counts as delivered.

Open items

• The roster names three leads without repository access: Eva Benn, Richard Bird, and Aruneesh Salhotra. Reconciling them, or a lead's note on issue PR intake: one gate for work no maintainer asked for #362 that they stay, is the precondition for the rollout above.
• The review's outside-this-pull-request items (collection split, partial-apply write counts, event-run cost, the pre-arming dry-sweep diff, the renderer tolerance note) are tracked in issue Intake gate follow-ups after PR #368 #369.

[rocklambros]

Replaced by PR #406, a much smaller check with no rollout.