[JacobStephens2]

Summary

A completed Triage investigation previously failed immediately when its recommendation had a shape or questionnaire error. Give it one correction turn while keeping validation strict.

completed investigation
  validate complete recommendation
  invalid -> one correction in the same worktree and launch policy
             keep a distinct Session trace and Model evidence
             validate the complete corrected recommendation
  check publication -> file/update issues or recommend only

The correction receives the original instructions, rejected output and exact error as separately identified evidence, and preserves the full investigation and intended issues/comments without investigating again. Execution and GitHub/publication failures do not trigger correction. Private Security-finding notification protections remain in force.

Closes #2019

Evidence

Execution evidence

• Before: cargo test --test triage_correction an_unnecessary_questionnaire_gets_one_correction_and_keeps_the_investigation -- --exact --nocapture

  assertion `left == right` failed:
  thirdshift: the Triage questionnaire is required only for needs-info or ready-for-human issues: Handle an empty cache
    left: Some(1)
   right: Some(0)
  test result: FAILED. 0 passed; 1 failed
  

  After: the same command passed: 1 passed; 0 failed.

• Confirmed cause: successful Harness output reached parse_recommendation, whose validation error propagated before publication or a correction could run. The minimized trigger is one otherwise complete issue with an unnecessary questionnaire.

• cargo test --test triage_correction: 9 passed, including missing/unnecessary questionnaires, misplaced comments, mixed entries, one-turn limits, complete publication, recommend-only, existing-issue cardinality/Agent Brief/readiness, post-correction validation, private notifications and real Bubblewrap confinement.

• cargo test --test triage_run --test triage_issue --test triage_spec_regressions --test triage_pass: 52 passed.

• cargo test --bin thirdshift prompts_page: 14 passed after regeneration; npm run docs:check --prefix site passed.

• PLAYWRIGHT_BROWSERS_PATH=/tmp/thirdshift-issue-2019-browsers CARGO_BUILD_JOBS=3 RUST_TEST_THREADS=3 scripts/affected-tests refs/remotes/origin/main: passed (68 site checks and the complete Rust suite, including all 100 integration test files).

The initial scripts/affected-tests main selected unrelated history because local main is checked out in another worktree at 27f547958b9e5f43bb2616f7e42286f35534fe0a; it stopped immediately on unavailable Go. The final selection uses the specified refs/remotes/origin/main base at de91156c424fa54e95845f1e10e46afaf0bdf406. That worktree and local main were left untouched.

Test seams

• Real thirdshift triage CLI in Report and existing-issue modes, with fake Harness responses and fake GitHub publication.
• Session traces and Model evidence retained by that CLI, and the real read-machine sandbox launch where available.

Visual evidence

Manual, Triage runs section, same scroll state.

Desktop, 1440×900

Mobile, 390×844

CI repair validation

Merged main at 482507a4 in 2d05b78a, including #2027's Ubuntu DNS recovery before musl dependency installation. The reconciliation also includes upstream version 0.29.0 and domain documentation; it changes no application sources, Rust tests or site pages. The feature diff against the merged base remains the same 13 files.

• Before reconciliation, PLAYWRIGHT_BROWSERS_PATH=/tmp/thirdshift-issue-2019-browsers CARGO_BUILD_JOBS=3 RUST_TEST_THREADS=3 scripts/affected-tests refs/remotes/origin/main passed: 68 site checks and 4,242 selected Rust tests. Rust wall time: 1682.52 seconds, including compilation.
• After reconciliation, node --test site/tests/docs.test.cjs passed 7 tests, CARGO_BUILD_JOBS=3 RUST_TEST_THREADS=3 cargo test --test update passed 10 tests, and the built binary's version output matched thirdshift 0.29.0.
• python3 -m unittest discover -s scripts -p test_generate_release.py passed 3 tests; python3 scripts/generate-release.py --check and bash -n scripts/release-dns.sh passed. The imported DNS helper also passed four isolated probes with mocked getent and sudo: healthy DNS, recovery through public resolvers, refusal to change DNS outside CI, and failure when recovery still cannot resolve a host. Host DNS was not changed.
• Attempted dist build --artifacts=local --target x86_64-unknown-linux-musl --print=linkage --output-format=json and the same command with --target aarch64-unknown-linux-musl. This Rocky Linux host cannot complete those artifact builds: x86_64 lacks x86_64-linux-musl-gcc; aarch64 lacks cargo-zigbuild. Both attempts exited 255. Local musl artifacts are not verified.

Review coverage

Independent Standards and Spec reviews against refs/remotes/origin/main (de91156c424fa54e95845f1e10e46afaf0bdf406) each raised 0 findings. Each reviewer read all 13 changed files, including generated output; no unread changed files or exclusions remain. Reports and counts are retained in .thirdshift-review-7L9Yo7/{standards.md,spec.md,summary.json} beside the Session evidence.

Unaddressed findings

Standards

None.

Spec

None.

Security

None.

Security review outcome

No unaddressed introduced findings.

CI notes

• At the original head 4e5d9a49, the x86_64 musl job and aarch64 musl job failed before compilation: APT could not resolve azure.archive.ubuntu.com / ports.ubuntu.com while installing musl-tools. The feature diff changes neither release setup nor dependency inputs. The upstream reconciliation above supplies the existing DNS fix without relaxing any check.
• The unrelated console job failed on the same Ubuntu DNS error during Playwright dependency installation; the site job was cancelled during dependency installation before npm test. Those setup steps are unchanged by the feature diff. No additional fixes were made for them.

Merge Danger

Door: two-way

Reverting restores immediate failure on invalid recommendations; no schema migration.

Blast Radius: Triage

An invalid recommendation can spend one additional Harness turn before publication or failure. Valid initial recommendations use no extra turn.

Built with codex codex-cli 0.167.1 · gpt-6.1-sol · xhigh