Deterministic pre-release gate — four agents run real bandit, ruff, pytest, and docstring checks against hard thresholds

Failure modes

Symptom	Cause	Fix	Severity
`RuntimeWarning: coroutine 'ReleasePrepTeam.assess_readiness' was never awaited`	`assess_readiness` (or `execute`) called without `await`	Both are coroutines — `await` them or use `asyncio.run`	high
Verdict is BLOCKED but the command exited 0	`success` reflects that the assessment ran, not the verdict	Read `metadata["approved"]` / the report — don't gate on the exit code alone	medium
Security gate shows `actual: -1` / a gate can't be measured	The agent errored (e.g. the tool raised); `critical_issues` falls back to `-1`	Inspect that agent's `findings["error"]`; confirm the tool runs in the project	medium
Coverage reads as estimated, not measured	`pytest --cov` output wasn't parseable (timeout, no TOTAL line)	The percentage is heuristic from test count (`findings["estimated"]` is `True`); run coverage directly to confirm	low
A gate reports `"bandit not available"` / `"ruff not available"`	The tool isn't installed in the run environment	Install the tool; the agent scores a neutral fallback otherwise	medium
Setting `quality_gates={"coverage": 0.9}` has no effect	Wrong key	Use `min_coverage` (a percentage like `90.0`), not `coverage`	medium

The async call is easy to get wrong. assess_readiness and execute are coroutines. Calling assess_readiness synchronously is the single most common mistake.
success is not the verdict. A BLOCKED release returns success=True. Branch on metadata["approved"].
Threshold keys are specific. They are max_critical_issues, min_coverage, min_quality_score, min_doc_coverage — coverage and doc-coverage are percentages (e.g. 80.0), not fractions.

Confirm you are awaiting: await workflow.execute(path=".") / await team.assess_readiness(codebase_path=".").
Read the verdict from result.metadata["approved"], not success.
For a blocked gate, read the report's blockers and the failing gate's actual-vs-threshold.
For an errored agent, inspect its findings["error"] in report.agent_results.
Confirm the tools (bandit, ruff, pytest) run in the project environment.