debug-mission

# Debug Mission Start `mission.inspect` with template `debug` and the exact bug report as the objective. Follow the returned evidence gates rather than blindly dispatching every stage. - Keep one `diagnostician` session until it reports a minimized red feedback loop and confirmed cause. - Declare clear tasks before their prerequisites finish if useful; a declaration executes nothing. Admit `implementer` only after dependency-connected diagnosis. Require behavioral red/green evidence and the original loop green. - Then delegate `review-standards` and `review-spec` to two fresh distinct native sessions in parallel over the same latest live implementation. Neither reviewer may specify a continuation target or `reuseFromTaskKey`; a fresh native child satisfies freshness. - For native `resolver`, set `reuseFromTaskKey` to the exact dependency-connected completed implementer task and pass both review reports in the brief. Reuse only its actual bound child after an exact current native-ended/returned observation and authoritative idle check. Never infer idleness from completion or fall back to an unrelated sole implementer. A resolver without native mode reuses its dependency-connected implementer's root. - Use a fresh `validator` session, read-only, for configured typecheck, lint, test, build, and focused regression checks. - Pass each ready native declaration's canonical `assignmentPrompt` into its ordinary native call, preserving its declared execution profile and continuation checks. Native children return actual evidence and the required structured `artifact`; the coordinator supplies them in `mission.report` with `taskKey` and no `contract`. Do not request child duplicate reports or forward task generations/bindings to helpers. Independent-root actors submit their own assigned reports. The plugin rejects completed reports that do not prove their role-specific gate. - Finish green only when applicable diagnosis, both review axes, resolution, and all validation categories prove the latest live implementation; count only reviews of that implementation. Never stage, commit, push, publish, or mutate issue trackers as part of this mission. Run independent ready work in parallel when useful, including bounded research/checks inside each stage; preserve the actual evidence gates and both fresh review axes. Each native child owns its assignment and may recursively decompose useful independent subtasks. Pass relevant scope, role/read-only constraints, safety boundaries and this delegation policy to each helper yourself. Integrate actual returned evidence before returning to the immediate parent. Helpers cannot create mission topology or submit the task actor's report. Respect native permissions and the user's configured runtime depth: do not change configuration, force delegation levels, or convert a denied/depth-limited helper into an undeclared independent root. Finish within scope or return the limitation. Avoid conflicting edits/shared mutable checks, and never treat background launch as completion. Independent roots require an explained location, lifetime, existing-root, or playbook exception. Use only the attached server's supported contract; never convert real missions or fabricate native bindings. For UI bugs, an actor whose session belongs to a project open in a local CodeNomad desktop window may use `codenomad.inspect`/`codenomad.act` as an optional feedback loop (targeting rules in `dev-docs/BROWSER_AUTOMATION.md`); otherwise use repository-native checks.