shannon

Read-only audit with severity classification. Scopes: screen, app, session, drift. No edits.

Refusal-driven retry loop around /shannon:cook. Preserves Iron Rule + refusal-discipline pattern.

Execute a plan or brief end-to-end with iron-rule validation + evidence gates. Wraps planning, execution, validation, completion gate.

Subagent dispatch with mode flag. Sequential | parallel | competitive. Each subagent gets IRON-RULE inject; judge confirms output.

Health check. Validates plugin manifest, hooks.json registration, settings.json sanity, log dirs writable, marketplace declarations, build state.

Severity-rated review producing BLOCKING/HIGH/MEDIUM/LOW verdicts

Deep multi-wave planner for complex brownfield initiatives — decomposes objectives into dependency-grouped waves with per-wave synthesis gates

Generic task executor for plan steps — the most-spawned agent type

Pillar 4: rubric YAML generator before any judge/oracle/critic runs

Plan synthesis: intake → research → design → tasks → executable plan

Full-lifecycle autopilot orchestrator for /shannon:autopilot. ALWAYS use when the user says "autopilot", "run autonomously", "Spec→Plan→Execute→QA→Validate→Cleanup", "keep retrying until it works", or invokes /shannon:autopilot. Drives a six-phase lifecycle with stall detection, multi-perspective validation, SDK harness instrumentation, resume semantics, and transcript-evidence Iron Rule enforcement.

Parallelized repo-wide context survey via the sciomc scientists pattern (3-7 parallel agents with [FINDING:][EVIDENCE:][CONFIDENCE:] tags), with cross-validation, proving-commands extraction, and incremental caching. ALWAYS use when the user says "analyze codebase", "repo context", "module map", "hot path identification", "survey the repo", or starts /shannon:scope. Produces structured machine-extractable findings under e2e-evidence/<run-id>/codebase-analysis/.

Final completion gate evaluator for Shannon runs. ALWAYS use when the user says "completion gate", "ship gate", "evaluate completion", "final gate", or is about to claim a task is done. Reads the entire evidence tree, checks every Mandatory Success Criterion (MSC) against cited evidence, requires three-reviewer consensus PASS plus Oracle quorum APPROVED, and emits report.json. No override flag, no force-complete. Outer mechanical layer above evidence-gate (per-claim) and verification-before-completion (per-commit).

Multi-validator agreement gate with 5-state synthesis and multi-round debate iteration. ALWAYS use when the user says "consensus validation", "validate with N reviewers", "agreement gate", "high-confidence validation", "consensus gate", or needs confidence-scored verdicts. Spawns ≥2 (default 3) independent validators in isolated evidence dirs, applies the 5-state synthesis table (UNANIMOUS_PASS / UNANIMOUS_FAIL / MAJORITY_PASS / MAJORITY_FAIL / SPLIT), and escalates SPLIT or borderline MAJORITY to up to 3 rounds of filesystem-mediated debate.

Create optimized prompts for Claude-to-Claude pipelines with research, planning, and execution stages. Use when building prompts that produce outputs for other prompts to consume, or when running multi-stage workflows (research -> plan -> implement).

9 hooks across 5 events