Cross Model Plan Review

Cross-Model Plan Review

> Adversarial multi-model review for design and implementation plans. > For code review of existing files, use /supervised-review.

---

Supported Models

| Role | Supported Models | CLI / Invocation | |---|---|---| | Lead Engineer | Claude, Codex (GPT-4o), Kimi | See references/agent-invocation.md | | Skeptic Agents | Claude, Codex (GPT-4o), Kimi | See references/agent-invocation.md |

> Kimi is invoked via the Kimi Code CLI (kimi -p) — a first-class terminal agent, not a curl wrapper. > Auth is configured in ~/.kimi-code/config.toml or via kimi → /login. > Kimi Code CLI also supports Claude and OpenAI models as backends via its provider config. > See references/agent-invocation.md for full setup.

---

Phase 0 — Invocation

User provides:

1. Path to plan file (Markdown .md)
2. Lead model: claude | codex | kimi
3. Skeptic agents (1 or 2): claude | codex | kimi (must differ from each other; may overlap with lead)
4. Number of review rounds (default: 2)

Example invocation:

/cross-model-plan-review
  --plan ./docs/deployment-settings-plan.md
  --lead claude
  --skeptics codex,kimi
  --rounds 2

---

Phase 1 — Auth & Environment Check

Read references/agent-invocation.md before this phase.

Before any review begins, verify each selected model's CLI is available and authenticated. Run scripts/check_agents.sh with the selected model list.

Fail-fast rules:

• Any selected agent that fails its auth check → STOP. Print the agent name, the specific failure, and the exact setup steps from references/agent-invocation.md. Do not proceed until the user resolves it.
• If the lead model fails → STOP immediately (no review is possible).
• If a skeptic agent fails → STOP (partial reviews produce misleading plans).

Auth check outputs a [READY] or [FAIL: ] per agent. Only proceed when all selected agents are [READY].

---

Phase 2 — Plan Ingestion & Lead Analysis

Read references/prompts/lead-engineer.txt before this phase. Do NOT load references/prompts/skeptic.txt at this phase — persona contamination causes the lead to hedge findings instead of committing to them.

Invoke the lead model via scripts/invoke_agent.sh --role lead. Lead produces Plan v0 (structured with catalogued decision points) and Initial Observations — senior-level flags that prime the skeptics without pre-answering their challenges. Initial Observations are injected into Round 1 skeptic prompts only; strip them for Round 2.

---

Phase 3 — External Skeptic Review (Round N)

Read references/prompts/skeptic.txt and references/templates/skeptic-round-template.md before spawning skeptics. Do NOT reload references/prompts/lead-engineer.txt at this phase.

Invoke skeptics via scripts/invoke_agent.sh --role skeptic-a (and --role skeptic-b if two agents). When two agents: invoke in parallel, never show either agent the other's output — independent review is the mechanism. Await both completions before Phase 4.

Round 2 critical difference: Pass the list of all DEFERred IDs from Round 1 explicitly in the skeptic prompt:

DEFERRED FROM ROUND 1: SK[A/B]-003, SKA-007
These were not resolved. Probe them harder — look for whether the plan merely acknowledged
the concern without structurally addressing it.

Without this injection, Round 2 skeptics will review the revised plan cold and may miss papering-over of deferred items — which is the entire point of the second round.

Skeptic Output Contract

Skeptics emit challenges only. No fixes. No rewrites. No code.

SKA-### / SKB-### / SK-###
Challenge: 
Failure Mode: 
Evidence: 
Severity: CRITICAL | HIGH | MEDIUM | LOW
Questions: 

Minimum viable skeptic output: Fewer than 3 substantive challenges on a plan >500 words is a skeptic failure — re-invoke with explicit instruction to challenge severity assignments, omissions, and unstated assumptions. A thin challenge set is not a clean bill of health.

---

Phase 4 — Lead Synthesis & Debate

Lead model is invoked again with the full skeptic output(s).

Read references/prompts/lead-engineer.txt (already loaded — do not reload skeptic.txt).

Lead receives:

• Plan v0 (Round 1) or current revised plan (Round 2)
• SKA-### challenge set (if Agent A ran)
• SKB-### challenge set (if Agent B ran)

Lead produces a Synthesis & Debate document:

For each SKA-### / SKB-### ID:
  Decision: ACCEPT | REJECT | DEFER
  Rationale: 
  Plan Delta: 
  Cross-Agent Note: 

Convergence rule: If both agents independently raised the same challenge (same root concern, even under different IDs), the lead must treat it as at minimum HIGH severity regardless of individual agent ratings. Convergent challenges cannot be REJECT without extraordinary rationale.

---

Phases 5 & 6 — Plan Revision & Round Gate

Apply all ACCEPTed and DEFERred deltas. Produce Patch Notes (v[N-1] → vN) — every change maps to exactly one skeptic ID. No ID = change is rejected, not just flagged.

Before gating to Round 2: compile the DEFER list (all IDs with Decision: DEFER). This list is injected into Round 2 skeptic prompts per Phase 3. Do not proceed to Round 2 without it — without explicit DEFER injection, the second round cannot fulfill its core purpose of catching unresolved items.

---

Phase 7 — Final Output

MANDATORY — READ ENTIRE FILE: Before assembling the final document, read references/templates/final-plan-template.md in full. That file is the authoritative output contract for Sections A–K.

Do NOT load final-plan-template.md at any earlier phase — it is only needed here and burns context prematurely if loaded at trigger time.

Section I (Final Hardened Plan) is the deliverable. It must be clean — no review IDs, no debate artifacts, no [OPEN: ...] markers. An implementer should be able to act on it without reading the rest of the document.

---

Non-Negotiable Rules

• Never skip Phase 1. A failed auth check that isn't caught produces a partial plan with no indication of what was missed.
• Never show one skeptic agent the other's output. Independent review is the point. Contamination collapses it to a single perspective.
• Never accept a patch-note change without an originating skeptic ID. Untraceable mutations undermine the entire audit chain.
• Never let the lead invent new findings during synthesis. Synthesis responds to skeptic challenges; it does not authorize new scope without a triggering ID.
• Never skip Round 2. The second round is specifically designed to catch challenges that were nominally resolved in Round 1 but not genuinely addressed.
• Convergent challenges cannot be silently rejected. Both agents independently flagging the same concern is a strong signal; it requires explicit, detailed rationale to reject.
• Final Hardened Plan (Section I) must be clean. No review IDs, no debate artifacts. It is the deliverable — ready to hand to an implementer.

---

NEVER

• NEVER let a skeptic agent produce fixes, code, or revised plan sections. Skeptics challenge only. Any fix output from a skeptic is discarded and the agent is re-prompted.
• NEVER proceed past Phase 1 with an unauthenticated agent. Partial external review is worse than no external review — it gives false confidence.
• NEVER merge SKA and SKB outputs before the lead sees them. The lead synthesizes from two independent streams; pre-merging destroys provenance.
• NEVER omit Section K (Open Items). DEFERred challenges are unresolved design decisions. Hiding them from the final output leaves the implementer uninformed.
• NEVER run the same model as both lead and the only skeptic. A model challenging its own output is not adversarial review. Require at least one skeptic from a different model family than the lead.

---

Inline Fallbacks

If subagents are unavailable: Run lead and skeptic invocations sequentially via bash_tool. Output shape is identical; label each section clearly.

If an external agent call times out or returns under 200 words on a plan >500 words: Treat as a skeptic failure — not a clean result. Re-invoke once. If it fails again, run that skeptic slot inline in the main thread and note [INLINE FALLBACK: timed out / truncated] in the transcript. Do not paper over the failure silently.

If a skeptic agent returns fixes, rewrites, or code: Strip those sections, note [SKEPTIC OVERREACH — fixes discarded], re-prompt with explicit challenge-only instruction.

If a skeptic agent returns fewer than 3 substantive challenges on a plan >500 words: Re-invoke with explicit instruction to challenge severity assignments, omissions, and unstated assumptions. Thin output is a skeptic failure, not a clean bill of health.

If the plan file is not found at the provided path: STOP. Print the path that was tried and ask the user to confirm. Do not infer or search.

---

Reference Files

Load only at the phase indicated. The Do NOT Load constraints are correctness requirements, not suggestions — wrong-phase loads cause persona contamination or premature context burn.

| File | Load at phase | Do NOT load at | Purpose | |---|---|---|---| | references/agent-invocation.md | Phase 1, Phase 2 | — | CLI setup, auth checks, invocation per model | | references/prompts/lead-engineer.txt | Phase 2, Phase 4 | Phase 3 | Lead persona — plan analysis, synthesis, debate | | references/prompts/skeptic.txt | Phase 3 | Phase 2, Phase 4 | Skeptic persona — challenge-only, output contract | | references/templates/skeptic-round-template.md | Phase 3 | — | Per-round output shape | | references/templates/final-plan-template.md | Phase 7 only | Phases 1–6 | Final document assembly — Sections A–K | | scripts/check_agents.sh | Phase 1 | — | Auth verification for all selected agents | | scripts/invoke_agent.sh | Phase 3, Phase 4 | — | Shell wrapper — claude -p, openai, or kimi -p with timeout |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: RyanMakesAndBreaksStuff
• Source: RyanMakesAndBreaksStuff/Custom-Codex-Claude-Plugins-and-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.