[dan-petty]

Refine, stage, implement and alternatives left out each item that the Service App or github-actions[bot] opened, because GitHub gives both bots the association CONTRIBUTOR. 38 items waited on this rule. Now a job takes an item whose author has write access, or whose login is exactly service.github_app.bot_login or github-actions[bot]. A bot item reaches Ready only after a review of refine approves the issue and the proposal. The other jobs take it only while that approval holds for the same body. The models of refine and of the alternatives job read only the replies of trusted authors. Refine tells the author of any other item once.

Closes #3502

Change: openspec/changes/archive/2026-11-05-3502-jobs-trust-service-app-items/

[dan-petty]

Fail-to-pass run

Command: uv run devops ci fail-to-pass --base origin/main.

• 40 tests failed on base. 6 test files failed to collect on base, because InMemoryRoadmapStore.seed_issue() takes no author_login there: test_roadmap_explain, test_roadmap_implement_selection, test_roadmap_refine, test_roadmap_refine_trust, test_roadmap_repair and test_roadmap_trust.
• All of these tests pass on this branch.

Full gate run

Command: uv run devops ci on the final code, after the rebase onto origin/main (f6d57ec).

• 24 of 25 rows passed: lint, format, typecheck, security, docs, openspec, constructs, test-tree and the others.
• The test row ran 18,874 passing tests and coverage was 94.52% (floor 85%). It reported 4 failures and 1 teardown error. All five come from the state of this host, not from this change:
  • tests/crypto/test_ssh.py::test_get_ssh_key_prefix_resolution reads the name of the git root, and the worktree is nested in another repository.
  • tests/telemetry/test_config_env_overrides.py::test_find_project_config_path_skips_directory finds a file /tmp/config.yaml on the host.
  • tests/commands/test_devcontainer_install_tools_dry_run.py::test_mark_only_commands_under_global_dry_run finds devcontainer installed on the host.
  • tests/ai/test_tool_executor_contract.py::test_the_direct_executor_refuses_a_network_profile_before_it_starts_the_tool[direct] reaches the popeye binary of the host.
  • Open issues fix(ci): host state fails four tests and one case #3585, fix(tests): four tests fail on the devcontainer host #3590, fix(ci): host-state tests fail and the suite exceeds 5 minutes #3593, fix(tests): four host-dependent tests fail on main #3594 and fix(ci): four tests read host state and fail in a devcontainer #3597 track these host-state failures. The required checks of this pull request run on GitHub runners, where the host holds none of this state.
• The test row took 7m 44s on a host that other agents shared. The slowest test is tests/commands/test_ai_local_dry_run.py::test_local_ai_commands_run_cleanly (74 s). The files of this change hold none of the ten slowest tests.

Earlier review findings

I checked each finding against the code on this branch.

• Approval after a body edit: fixed. trust.is_approved also needs the body to hash as refine.body_hash and refine.section_hash record. Tests cover the stage, implement and alternatives readers, and the second refine round. The title is not part of the hashes, and design.md says why: refine ignores a retitle, and only a person with write access can retitle an item of a bot.
• Documents and specs: fixed. The change holds deltas for roadmap-alternatives, roadmap-stage and roadmap-implement-job, and docs/roadmap/alternatives.md and docs/roadmap/refine.md name the two bots.
• ASD-STE100: fixed. I counted the words of each new output string, and no sentence has more than 23 words. No paragraph of the new text has more than 6 sentences.
• Dry-run plan of refine: fixed. refine_requests lists the comment read and the notice comment, and tests/roadmap/test_roadmap_dry_runs.py has a test.
• Comments that reach the alternatives job: fixed in this pull request. _check_one filters them with trusted_replies, and tests/roadmap/test_roadmap_alternatives.py has a test. So SECURITY.md rule 14 is true as written.
• Test that patched get_llm_client: fixed. The test uses the model_fake fixture and patches nothing. The proposal and review doubles of the file go in through the model and verifier parameters of plan_refine, as in the other refine test files.

The owner decides at merge: the change widens the trust rule of #2478 to two bots (AGENTS.md rule 5).