Skip to content

feat(pr-pipeline): run the project's app-control skill in pr-verify after the deterministic tests #8651

Open
7 of 12 issues completed

Description

@kyle-sexton

Refs: #7580
Refs: #7555
Decision: #7555 (comment)
Blocked by: #8816 (adds the app-control plugin and /app-control:drive, whose report.json this step reads).

Agent Brief

Type: Feature
Summary: After the deterministic tests pass, the verifier runs the project's own app-control skill against the PR head and folds what it finds into its verdict. Stack-neutral: the project names its skill; Docker is used only if that skill uses it.

Problem: The verifier judges a PR from its diff and text; nothing drives the running app the way a user would. The line that asked for this ("agent browser checks run after the deterministic tests, with no write token while PR code runs") sat in Phase 9, which closed with it unbuilt.

Current behavior:

  • .github/workflows/pr-verify.yml:3-27: runs on workflow_run after pr-require-checks completes; "No pull request code runs here"; job tokens can only read (:66-68, :384-386).
  • No app-control plugin exists yet; Phase 13 creates /app-control:drive and the generated control-<app> skill shape.

Desired behavior:

  • A setting in docs/conventions/pr-pipeline.yaml names the project's app-control skill. Unset: the step is skipped and the verdict says no app check ran.
  • When set: a separate job, after the deterministic checks pass, checks out the PR head and runs /app-control:drive --unattended, which finds the project's control skill; the verdict job judges drive's report.json (drive itself never judges; D54). It holds no write token and no secret beyond the model credential the run needs, on a fresh GitHub-hosted runner, under ADR 0049's hardening.
  • Its findings go to the verdict job as data, framed as untrusted (docs/conventions/untrusted-content/README.md); a blocking finding makes the verdict FAIL.
  • Nothing in the workflow names a stack. Docker appears only if the project's skill starts it.

Key interfaces:

Done when:

  • The diff adds the setting, the new job and the verdict input, and the job's permissions: grant no write scope (read from the diff).
  • With the setting unset, the verdict notes that no app check ran (a workflow.test.mjs case). CI's test-node passes on the PR (a PR check; this run does not run it).
  • CI's check-plugins passes on the PR (a PR check; this run does not run it).
  • Security & Cost's pass on the final head is recorded on the PR (.github/workflows changes; owner rule 2026-10-20).

Out of scope:

Origin:

  • Verifier agents that run the app and click around like a user, then fix and repeat: hers (Lauren, stream ~53:10), adapted.
  • Docker app stack plus browser checks in CI: the owner's (2026-10-16); moved to Phase 13 and made stack-neutral: ours, confirmed by the owner 2026-10-20.

Review: security (runs PR code in CI).

Do not run with: other pr-verify.yml changes.

Size Large, Complexity High, Uncertainty Knowable.


Slice of #7580 (Phase 13), filed from the owner's 2026-10-20 disposition round.

Activity

  1. added
    status: confirmedA maintainer confirmed the item is real and in scope; ready to be worked.
    on Oct 20, 2026
  2. kyle-sexton commented on Oct 30, 2026

    @kyle-sexton
    MemberAuthor

    Phase text for S7 (the pr-verify capture job), from the evidence-verification plan (Brief: #11364 (comment)). Slices are sub-issues of this issue.

    Satisfies: AC1, AC2, AC5, AC9, AC10
    Size Large. Depends on: S2a (#11388), S4a (#11386), S4b (#11390), S6 (#11385), and owner answers to U1 and U2 (USER-RESERVED). Files: .github/workflows/pr-verify.yml, .github/scripts/pr-verify/prompt.md, workflow.test.mjs. Design: T8, sequence flow 2.

    • Probe step first (U1/U2 evidence): after setup-runner-isolation, a head-code step tries to read the runtime token and the credential from a hook; the job fails closed if either is readable.
    • Job capture: same if as .github/workflows/pr-verify.yml:73-77 (fork heads never run), gates, read-only token, .base plugins, --trusted-root .base, /testing:run-e2e unattended app=<merge.verify-capture.app> base=<base-sha>; uploads publishable files as pr-capture-<pr>-<sha> with retention-days; skip with note when the key is unset.
    • Gates first: the capture job runs the kill switch, resolver, trusted-trigger and agree steps as its first steps (check-intent's copies stay), so it captures nothing on a run that judges nothing (.github/workflows/pr-verify.yml:102-175). timeout-minutes: 30; the outer session runs through ci-workflows' run-claude-cli with its own role, a tool allowlist and the drive maxEntryPoints cap (S4b (feat(harness-ops): mask control-claude-code captures and launch changed plugins #11390)); a capture timeout leaves capture: blocked and never stops the judge.
    • Judge jobs: only check-intent gains needs: capture with !cancelled(); check-intent-subscription and check-intent-fallback keep their success chain (.github/workflows/pr-verify.yml:434-445) and download the same run's artifact by id. All three download into pr-verify-data/capture/; the prompt frames it untrusted and says a missing or blocked capture is never a FAIL reason, while a trusted-text manifest line Evidence loop: 3/3 cap reached, last NEEDS WORK is a FAIL (AC7's park path then runs).
    • Last step, after the probe passes on a real PR: set merge.verify-capture.app: claude-code and show one in-repo PR with a judged capture (AC1). Comment block at .github/workflows/pr-verify.yml:33-36 updated (reviewed change: jobs now take an artifact).
    • Relates to pr-merge: per-PR live smoke of changed skills before unattended merge (owner decision 16) #7529 (Claude Code UI leg of the per-PR smoke).
    • Sanity Check: node --test .github/scripts/pr-verify/workflow.test.mjs exits 0 with cases: capture job permissions contain no write; the three judge jobs carry identical capture-download steps; capture if includes the head-repository check; no job-level always() in pr-verify.yml; only check-intent needs: capture; a probe step exists, runs after setup-runner-isolation and before any head code, and has no continue-on-error; actionlint passes.

    Depends on

    🤖 Generated with Claude Code

    https://claude.ai/code/session_01UJUn7KoThXoYQoxfvdCVi7

  3. kyle-sexton commented on Oct 30, 2026

    @kyle-sexton
    MemberAuthor

    Owner decision 2026-10-30, Q11: the capture job stays in pr-verify as its own job (no write token; only its run directory reaches the judge). No verdict-forgery probe and no artifact-name filter: the only actors are the owner's own agents, and the producer and reviewer are separate agents. Q10: it uses pr-verify's existing credential legs. Remaining dependencies: #11388, #11386, #11390, #11385. Brief and decisions: #11364

  4. kyle-sexton commented on Oct 31, 2026

    @kyle-sexton
    MemberAuthor

    Decisions 2026-10-31 (audit of Q10-Q12, resolved from the owner's standing direction s1, s5, s10):

    • The capture job sets CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 so hooks and Bash in the driven session get no credentials (https://code.claude.com/docs/en/env-vars). No separate OS user and no /proc defense: reading another process's environment on purpose needs a malicious actor, which the threat model excludes.
    • No process-group kill before upload (a guard against our own actors).
    • Evidence waits between capture and review in workflow artifacts with a retention period (as pr-verify's verdict does), never on a branch.
  5. added
    host-runNeeds an unattended run on a host; the host runner takes it, not a person or the issue lane.
    on Nov 1, 2026
  6. kyle-sexton commented on Nov 1, 2026

    @kyle-sexton
    MemberAuthor

    Decision recorded.

    The capture job is built as the 2026-10-30 S7 phase text on this issue describes. S7 replaces two parts of the body: the job runs /testing:run-e2e unattended app=<merge.verify-capture.app>, not /app-control:drive --unattended, and a missing or blocked capture is never a FAIL reason (the body said a blocking finding fails the verdict); the trusted manifest line Evidence loop: 3/3 cap reached, last NEEDS WORK is a FAIL. Q10 is the owner's 2026-10-30 answer ("we already have pr-verify", #11364 issuecomment-6536099521): the job uses pr-verify's existing credential legs. Q11 was decided on 2026-10-31 from the owner's standing direction, not in the owner's words (#11364 issuecomment-6547731014, which supersedes issuecomment-6547464663): the capture job may hold that credential, and the PR's code runs in a child process whose environment and readable files hold no token. The 2026-10-31 comment here adds CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1, no process-group kill before upload, and evidence in workflow artifacts with a retention period; it drops a separate OS user and a /proc defense, not S7's probe step, so the probe step stays. S6 (#11385) already added the merge.verify-capture key, with app unset recording capture: no-app. At origin/main b9ae6bbaed this edits the pr-verify workflow, which the lanes App cannot push (.github/actions/resolve-config/patch-rules.json:75-76). The body's Security & Cost criterion stays: no later decision on this issue removes it. Of the S7 dependencies, #11388 and #11385 are closed; #11386 and #11390 are open, so the build waits on them. Agents decided this on evidence under standing direction §1.

    Acceptance criteria:

    • feat(harness-ops): let control-claude-code run unattended on sessions it launches #11386 and feat(harness-ops): mask control-claude-code captures and launch changed plugins #11390 are closed by merged pull requests before the build starts.
    • With merge.verify-capture.app unset, the capture job is skipped and the verdict's evidence records capture: no-app (a .github/scripts/pr-verify/workflow.test.mjs case).
    • A probe step runs after setup-runner-isolation and before any head code, has no continue-on-error, and fails the job if a hook can read the runtime token or the credential.
    • The step that starts the app under test passes it no credential variable, and a test asserts that the app process's environment holds none of the pr-verify credential names.
    • The capture job's permissions grant no write scope, its if keeps fork heads out, and only check-intent gains needs: capture.
    • The capture job's environment sets CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1.
    • Capture evidence reaches the judge jobs only as a workflow artifact with a retention period, framed as untrusted. A missing or blocked capture is never a FAIL reason, and the trusted line Evidence loop: 3/3 cap reached, last NEEDS WORK is a FAIL.
    • node --test .github/scripts/pr-verify/workflow.test.mjs exits 0, and CI's test-node and check-plugins pass on the PR.
    • Security & Cost's pass on the final head is recorded on the PR (owner rule 2026-10-20, for a workflow change).

    Route: builder (Automation Lanes).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    host-runNeeds an unattended run on a host; the host runner takes it, not a person or the issue lane.status: confirmedA maintainer confirmed the item is real and in scope; ready to be worked.

    Fields

    No fields configured for Feature.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions