[arkaccess]

Summary

Most of the fix proposed in #13 is already on master: agent.py exits when all features pass, and the orchestrator breaks out of its loop on get_all_complete(). feature_get_next no longer exists, so the prompt change proposed in the issue doesn't apply anymore.

One version of the bug was still there. If every remaining feature was waiting for human input, had permanently failed, or was blocked by dependencies that can't be met (for example a dependency on a failed or missing feature), get_all_complete() returned False. The orchestrator then sat in its "No ready features available" branch. Every time a testing agent finished, _maintain_testing_agents() started another one, so the same passing features were regression-tested over and over. That is the same token drain the issue describes.

Changes

• parallel_orchestrator.py
  • New has_coding_work(): returns True if a coding agent is running or a feature can be resumed or started.
  • _maintain_testing_agents() no longer starts testing agents when there's no coding work.
  • In the stalled branch (no ready features, no coding agents):
    • re-checks a fresh DB snapshot for features that just became ready;
    • waits for any testing agents still running, since a regression they find makes that feature ready again;
    • if features are waiting for human input, it keeps polling quietly with no agents. It prints the notice once instead of every 10s. The user can answer from the UI and the loop picks the feature up;
    • otherwise it exits and lists the failed and blocked features, with a hint to fix them and restart.
• agent.py: the completion check now runs before every session, not only the first. Retries after an error or rate limit no longer start a session on a finished project.
• test_orchestrator_completion.py: 10 new tests for has_coding_work, testing-agent spawning, and how run_loop exits or waits. Agent spawns are stubbed.

Verification

• python -m pytest test_orchestrator_completion.py: 10 passed
• python -m pytest test_client.py test_dependency_resolver.py test_rate_limit_utils.py: 55 passed
• python test_security.py: all passed
• ruff check .: clean
• mypy on the changed files: no new errors. The errors it reports are already on master, in temp_cleanup.py and client.py.
• No UI changes, so no browser check.

Assumptions

• Exiting is the right move when nothing can progress without outside changes. Failure counts live in memory, so restarting the agent retries failed features.
• Waiting for human input is the one stalled state where the orchestrator stays alive, because the UI can unblock it.
• No --continuous-regression flag was added. Regression testing still runs alongside active coding exactly as before.

Closes #13

🤖 Generated with Claude Code

[arkaccess]

🔍 QA report: ✅ Passed

Tester: Ada (cubefarm agent, in a fresh session) · Round: 1 · Author: Ken

The PR fixes the part of #13 that was still broken: when every remaining feature is waiting for human input, has permanently failed, or is blocked by dependencies that can't be met, the orchestrator no longer starts regression testing agents over and over. It now exits with a list of failed and blocked features, or waits quietly without agents when a human can unblock it. agent.py also checks for completion before every session, so retries after an error no longer start sessions on a finished project. The new tests fail on master and pass on this branch, extra edge-case probes pass, and lint, type check and the test suite show no new failures.

	Check	Details
✅	Bug reproduced on master, fixed on PR	Ran the PR's test_orchestrator_completion.py against origin/master's code: 3 failed there (testing agents keep starting when the only remaining feature waits for human input; the loop never exits when the only remaining feature depends on a missing feature; the human-input wait starts agents). The same tests pass on the PR.
✅	Exit when the rest is failed or blocked	Probe test: feature #2 at MAX_FEATURE_RETRIES and #3 depending on #2. The loop exits, prints 'Failed after N attempts: #2' and 'Blocked by unmet dependencies: #3', and starts no testing agents.
✅	Dependency cycle	Probe test: #2 and #3 depend on each other. The loop exits and lists both as blocked.
✅	Waiting for human input, then answered	Probe test: the loop prints the waiting notice once and starts no agents. After needs_human_input is cleared in the DB from another thread, it starts coding feature #2 and testing agents start again.
✅	Waits for a running testing agent	With a testing agent registered as running and the remaining feature blocked, the loop waits instead of exiting, since a regression the agent finds could make a feature ready again.
✅	agent.py completion check on retries	Stubbed client whose session marks the last feature passing and returns an error. On master the agent ran 4 sessions (capped with max_iterations=4). On the PR it ran 1 session, printed 'ALL FEATURES COMPLETE!' and stopped.
✅	Regression testing during active coding unchanged	Code review: has_coding_work returns True while coding agents run or features are resumable or ready, so _maintain_testing_agents behaves as before. The test test_spawns_while_coding_work_exists passes.
✅	Security and schema	security.py and the localhost-only middleware are untouched, there is no DB schema change, and nothing new is written to logs beyond feature IDs.
✅	Backend boot and REST smoke test	uvicorn on 127.0.0.1:5571: /api/health 200, /api/projects 200, path-traversal project name returns 404, project creation and the features endpoint work.
✅	UI renders a stalled project	Loaded a project with 2 passing features, 1 waiting for input and 1 blocked. The dashboard, Kanban and graph views rendered with 0 console errors. The PR doesn't change the UI.
⏭️	UI Playwright e2e	The PR doesn't touch the UI.

🧪 Commands run

Command	Result
python -m pytest test_orchestrator_completion.py	10 passed
python -m pytest -q (full suite)	154 passed, 21 failed; the 21 are in test_devserver_security.py and master has the identical set
python -m pytest test_orchestrator_completion.py (on an origin/master checkout)	3 failed, 3 passed (bug reproduced)
python -m pytest qa probe tests (4 edge cases, not committed)	4 passed
python probe_agent.py (stubbed agent.py retry scenario, PR vs master)	PR: 1 session then stop; master: 4 sessions (capped)
ruff check .	All checks passed
mypy .	25 errors, the same set as on master
python test_security.py	ALL TESTS PASSED
python test_security_integration.py	ALL INTEGRATION TESTS PASSED
npm --prefix ui run build	built successfully, no type errors
npm --prefix ui run lint	0 errors, 6 warnings
uvicorn server.main:app --host 127.0.0.1 --port 5571 + curl smoke tests	health 200, projects 200, traversal 404; server stopped afterwards

No browser screenshots were taken in this round.

_{Posted by cubefarm · merges automatically once GitHub's checks pass}