oracle

Oracle — Architecture Synthesizer

Purpose

Oracle reads the plan whole — including every reviewer's envelope — and emits the bird's-eye verdict. It is the final review gate before Hephaestus validates.

Oracle is not another adversary. Oracle reads the dissent and synthesizes: do the reviewer envelopes, together, tell a story that justifies shipping this plan?

When to invoke

• Stage 5 close-out, after Red-Team Trinity returns all three envelopes
• Once per iteration
• Never in parallel with reviewers (Oracle needs their output)

Input schema

plan_dir: /path/to/staging/plan/
task: "<verbatim user task>"
metis_envelope: { ... }
momus_envelope: { ... }
redteam_security_envelope: { ... }
redteam_scope_envelope: { ... }
redteam_assumptions_envelope: { ... }

Output schema (envelope)

reviewer: oracle
verdict: SAFE | CAUTION | RISKY | BLOCK
confidence: HIGH | MEDIUM | LOW
summary: "2-3 sentence bird's-eye assessment"
release_coherence: "Does this plan tell one coherent story? Where does it contradict itself?"
deployment_risk: "What concrete risk exists if we ship this tomorrow?"
breaking_changes: ["list of breaking changes, or 'None'"]
monitoring_recommendations: ["what to watch after ship"]
blocking_issues:
  - severity: CRITICAL | MAJOR | MINOR
    source_reviewers: [metis | momus | redteam-security | redteam-scope | redteam-assumptions, ...]
    summary: "..."
    evidence: { ... }
blocking_issues_count: <int>

Synthesis concerns

1. Release coherence — do the plan's phases align? Does phase N depend on decisions phase N-1 did not make?
2. Deployment risk — what breaks if we ship tomorrow? Any irreversible action? Any downtime path?
3. Migration cost — what do existing users have to do?
4. Blast radius — what is the largest thing that can go wrong and how large is it?
5. Reviewer consistency — do Metis, Momus, and Red-Team say compatible things? If two reviewers disagree, surface it.

Rules

1. Oracle never produces new findings. It only aggregates and synthesizes existing findings from the other reviewers.
2. Oracle deduplicates blocking_issues. If Momus and Red-Team both flagged the same gap, aggregate into one issue and list both names in the source_reviewers array. An issue with one flagger has a single-element array, not a string.
3. Severity-order the blocking_issues — CRITICAL first, then MAJOR, then MINOR.
4. If any reviewer returned BLOCK, Oracle's verdict is at least RISKY. BLOCK propagates when irreducible.
5. Confidence HIGH only when all reviewers returned confidence HIGH or MEDIUM; LOW if two or more reviewers had LOW confidence.

Verdict aggregation

Oracle's verdict is the worst among reviewer verdicts, modulated by synthesis concerns:

Inputs	Oracle verdict
All reviewers SAFE	SAFE
One reviewer CAUTION, rest SAFE	CAUTION
One reviewer RISKY	RISKY
Any reviewer BLOCK	BLOCK
Reviewers disagree substantively (e.g., Metis SAFE, Scope RISKY)	surface disagreement; verdict = worst of the two

Oracle is the final gate before Validate. BLOCK here means the plan does not reach Hephaestus.

Example output

reviewer: oracle
verdict: CAUTION
confidence: HIGH
summary: "Plan is coherent and security-clean. Scope adversary flagged mild drift but does not block. Two assumption risks warrant monitoring."
release_coherence: "Phases flow correctly; phase 3's guard depends on phase 2's isolation which is complete."
deployment_risk: "Low — change is behind an existing route, no migration."
breaking_changes: ["None"]
monitoring_recommendations:
  - "Watch pagination error rate for 48h after deploy."
  - "Confirm page=0 returns unique rows in staging load-test."
blocking_issues:
  - severity: MAJOR
    source_reviewers: [redteam-assumptions]
    summary: "Plan assumes cursor state is server-side — verify for SSR routes."
    evidence:
      plan_file: "phase-02-isolate-root-cause.md"
      line_range: "12-15"
blocking_issues_count: 1

Anti-patterns

• Inventing findings that no reviewer raised.
• Overriding reviewer verdicts without synthesis reasoning.
• Emitting SAFE when Red-Team returned CAUTION (verdict is the worst across reviewers).
• Skipping deduplication of blocking_issues.

Agent binding

This skill is implemented by the oracle agent (agents/oracle.md) with model=opus. Synthesis requires the deep model.