Oracle review

Skill krzemienski/shannon/skills/oracle-review

Quorum-gated plan review (pre-execution) AND evidence audit (post-execution) for /shannon:forge. ALWAYS use at forge Phase 4 (plan review) and Phase 9 (evidence audit), or when the user says "oracle review", "plan review", "evidence audit", "quorum review", or "final gate audit". Spawns 3 structurally-independent auditor agents (critic + reviewer + validator) in isolated contexts; requires ≥2 APPROVE and 0 unresolved critical blockers, else invokes refusal-discipline and halts.From its SKILL.md

Install

npx -y skills add krzemienski/shannon --skill oracle-review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.9 KB, as published. Nobody here has run it

oracle-review

Quorum gate. Invoked at two points: BEFORE execution (plan review) AND AFTER execution (evidence audit).

Agent mapping note (target adaptation): Shannon's v1 line absorbed the dedicated oracle role into the critic / reviewer / validator triad (see docs/migration-map.md). The quorum therefore spawns those three distinct agent types — structural independence satisfies no-self-review (RL-3) without a separate oracle agent. The Phase-8 reviewer consensus (3× reviewer) and this Phase-9 quorum (critic + reviewer + validator) remain separate gates.

Behavior contract

Pre-execution (plan-review)

Spawn 3 auditor agents in parallel, isolated contexts: critic, reviewer, validator. Each receives the plan + codebase-analysis + documentation-research artifacts.

Each independently emits:

verdict: APPROVE | REFUSE
cited_blocker: <if REFUSE, file:line evidence>
summary: <one paragraph>

Aggregate: ≥2 APPROVE AND 0 unresolved critical blockers → PASS; pipeline proceeds.
Otherwise: REFUSE → invoke refusal-discipline; halt.

Post-execution (evidence-audit)

Spawn the same 3 auditor agents in parallel, isolated contexts. Each receives the full evidence tree.
Each reads cited evidence per MSC, votes APPROVE / REFUSE.
Aggregate via the same quorum rule.

File ownership

Coordinator: writes nothing.
Auditor-N: writes ONLY evidence/oracle-quorum/<run-id>/auditor-<role>/verdict.md.
Synthesizer: writes ONLY evidence/oracle-quorum/<run-id>/report.md.

When to use

Phase 4 of /shannon:forge (pre-execution plan review).
Phase 9 of /shannon:forge (post-execution evidence audit).
Standalone audit via /shannon:audit.

When NOT to use

Routine /shannon:cook runs (lightweight) — only forge invokes the oracle quorum.

Iron rules

Distinct from reviewer consensus. Quorum = pre-execution plan gate OR final evidence audit (critic+reviewer+validator). Reviewer consensus = post-implementation review (3× reviewer). Both can run for the same forge run.
No self-review. An agent that produced an artifact may not audit it.
Quorum is hard. ≥2/3 APPROVE OR refusal — no --force.
Citations specific. Directory references do not count; file path + line range where relevant.