Research Review via External Reviewer Backend (ultra reasoning)

🔒 Do not wrap this skill in /loop, /schedule, or CronCreate. It is verdict-bearing — it produces a cross-model review verdict, multi-round with reviewer thread continuity. An external timer re-fires the verdict on wall-clock time and breaks the reviewer's round-to-round memory: zero new signal, full token cost. Schedule the external wait that precedes it (work ready → then review once), not the verdict. See shared-references/external-cadence.md.

Get a multi-round critical review of research work from the selected external reviewer backend with maximum reasoning depth.

Constants

• REVIEWER_MODEL = gpt-6-astra — Default model for the Codex backend, reasoning effort ultra (deep-audit tier). Must be an OpenAI model (e.g., gpt-6-astra, gpt-6.1-sol). Manual backend uses a model the user chooses — it must be a recognized model from a different family (OpenAI, Anthropic, Google, DeepSeek, Moonshot/Kimi, Qwen).
• REVIEWER_BACKEND = codex — Default: Codex MCP (ultra). Resolve — reviewer:, project .aris/reviewer.json, then this default using reviewer_route.py (--host claude-code --tier deep), per reviewer-backends.md. Gemini/Antigravity, Grok, and Claude Code use that contract at every call below. Legacy Oracle/manual contracts remain in reviewer-routing.md; unavailable selected routes do not silently fall back.

Reviewer Calling Convention

When calling the reviewer, branch on REVIEWER_BACKEND:

If REVIEWER_BACKEND ∈ {antigravity, grok, claude-code}: Use the selected first/reply protocol in reviewer-backends.md, with the same review prompt, explicit cwd containing the primary files, and resolved provider-specific model/effort. Save route, cwd and session id for continuation. A Claude Code bridge may run DeepSeek; derive family from its reported model.

If REVIEWER_BACKEND = codex: Use mcp__codex__codex for new review threads. Use mcp__codex__codex-reply for follow-up rounds (reuse threadId).

If REVIEWER_BACKEND = manual: Use mcp__manual_review__review for new review threads with: prompt: [exact same prompt that would go to Codex] config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true} Save the returned threadId. Use mcp__manual_review__review_reply for follow-up rounds with: threadId: [saved manual-review threadId] prompt: [follow-up prompt] config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true}

Content fidelity: the manual reviewer should see the same substantive review brief Codex would read. If the manual UI supports file upload / attachment, reuse the same brief file; otherwise paste the brief contents inline because remote web UIs cannot read your local filesystem paths. Review tracing applies equally to both backends.

Context: $ARGUMENTS

Prerequisites

• The selected MCP must be available. For the default Codex MCP Server:

  claude mcp add codex -s user -- python3 "$HOME/aris_repo/mcp-servers/codex-exec/server.py"   # your ARIS clone's path
  

  Copy
• This gives Claude Code access to mcp__codex__codex and mcp__codex__codex-reply tools

Workflow

Step 1: Gather Research Context

Before calling the external reviewer, compile a comprehensive briefing:

1. Read project narrative documents (e.g., STORY.md, README.md, paper drafts)
2. Read any memory/notes files for key findings and experiment history
3. Identify: core claims, methodology, key results, known weaknesses

Step 2: Initial Review (Round 1)

Send a detailed prompt with ultra reasoning, using the selected backend. For the codex backend, keep the MCP payload short: write the full briefing to RESEARCH_REVIEW_REQUEST.md, then point Codex at that file.

For codex backend:

mcp__codex__codex:
  model: gpt-6-astra
  config: {"model_reasoning_effort": "ultra"}
  prompt: |
    Read the review brief at <absolute path to RESEARCH_REVIEW_REQUEST.md>.
    Executor notes are not evidence beyond the files they cite, so verify the
    referenced artifacts before judging.
    Please act as a senior ML reviewer (NeurIPS/ICML level). Start from the
    assumption that the work is broken somewhere — your job is to find where.
    Be adversarial. Trust nothing the author tells you — verify everything
    yourself. Identify:
    1. Logical gaps or unjustified claims
    2. Missing experiments that would strengthen the story
    3. Narrative weaknesses
    4. Whether the contribution is sufficient for a top venue

    === SCOPE LIMITS (these bound what you PROPOSE, never what you look for) ===
    Report anything that is actually wrong here — including a rare-looking case, if
    this repo actually produces it. Then keep the fix in scope:
    1. This is a RESEARCH-WORKFLOW tool, not a security paper. Verification is
       welcome; over-defense is not. Assume a cooperating operator on their own
       machine — a malicious local user is NOT in the threat model.
    2. Do NOT propose SHA / hash / content-fingerprint / digest-binding schemes.
       Reporting a real defect in hashing code that already exists is fine.
    3. NO speculative machinery: do not add feature flags, migration frameworks,
       compat layers, wrappers, pins, or similar mechanisms unless evidence shows
       a current repo defect they fix or an explicit existing invariant they must
       preserve. "Load-bearing", "compatibility", and "not scaffolding" are labels,
       not evidence. Point to the failing path/artifact or invariant, and check the
       proposal's factual premises, such as whether a named package version exists.
    4. NO corner-case obsession: exotic encodings, symlink races, RTL text and
       millisecond races are out of scope unless you can show the case arises here.
    5. Where a rubric or checklist is genuinely needed, do not over-mechanize
       judgement. A clear sentence a human reads beats a scored table nobody
       maintains.
    Exception: code that runs remote commands, starts a network service, or installs
    an MCP server runs on the user's machine with their credentials — trust-boundary
    findings there are in scope and the default is strict.
    Say plainly when something is correct. Do not manufacture findings.
    Be brutally honest. If, after genuinely trying to break it, the work
    holds up and is ready, say so clearly.

For Antigravity, Grok, or Claude Code, send the same prompt through its selected first-call tool; do not copy Codex's config into another tool. Check successful completion and actual file coverage before interpreting the response.

The review brief should contain the full research context, the specific questions, and the primary artifact / raw-result paths the reviewer should inspect.

For manual backend: use mcp__manual_review__review with the same brief contents. If the manual-review UI supports attachments, attach RESEARCH_REVIEW_REQUEST.md; otherwise paste the brief inline. Save the returned threadId.

Step 3: Iterative Dialogue (Rounds 2-N)

For codex backend: use mcp__codex__codex-reply with the returned threadId. For manual backend: use mcp__manual_review__review_reply with the same threadId. For Antigravity, Grok, or Claude Code: use the saved route's continuation protocol in reviewer-backends.md; keep its model/cwd and repeat Claude cwd/tools. Use the appropriate tool to continue the conversation. For Codex follow-up rounds, write an updated brief such as RESEARCH_REVIEW_ROUND_2.md and send only the path:

mcp__codex__codex-reply:
  threadId: [saved reviewer threadId from Step 2]
  # replies inherit the thread's model/effort (gpt-6-astra ultra)
  prompt: |
    Read the updated review brief at <absolute path to
    RESEARCH_REVIEW_ROUND_2.md>.
    Focus on unresolved weaknesses and whether the revision actually fixed them.

For manual follow-up rounds, attach that same updated brief if possible; otherwise paste it inline.

For each round:

1. Respond to criticisms with evidence/counterarguments
2. Ask targeted follow-ups on the most actionable points
3. Request specific deliverables: experiment designs, paper outlines, claims matrices

Key follow-up patterns:

• "If we reframe X as Y, does that change your assessment?"
• "What's the minimum experiment to satisfy concern Z?"
• "Please design the minimal additional experiment package (highest acceptance lift per GPU week)"
• "Please write a mock NeurIPS/ICML review with scores"
• "Give me a results-to-claims matrix for possible experimental outcomes"

Step 4: Convergence

Stop iterating when:

• Both sides agree on the core claims and their evidence requirements
• A concrete experiment plan is established
• The narrative structure is settled

Step 5: Document Everything

Save the full interaction and conclusions to a review document in the project root:

• Round-by-round summary of criticisms and responses
• Final consensus on claims, narrative, and experiments
• Claims matrix (what claims are allowed under each possible outcome)
• Prioritized TODO list with estimated compute costs
• Paper outline if discussed

Update project memory/notes with key review conclusions.

Composed mode — if invoked with — composed: <canonical-report-path> (an orchestrator like /idea-discovery passes this), do not write a standalone review .md in the project root. The raw conversation is already persisted to .aris/traces/… (see Review Tracing below — that audit copy is kept in every mode); fold the review conclusions (consensus, claims matrix, prioritized TODOs) into the orchestrator's canonical report and cite the trace path there. Default (no — composed: directive): behave exactly as above — write the standalone review document. Never infer composed mode from a report file merely existing. Full rules: shared-references/output-composition.md.

Key Rules

• For the default Codex route, pin model: gpt-6-astra + config: {"model_reasoning_effort": "ultra"} (deep-audit tier; capability fallback per reviewer-routing.md, never below xhigh). Other selected routes use their resolved model/effort.
• For manual, use the identity-bearing config above. Tool fields belong to their own provider contracts; do not copy a payload between bridges.
• Put comprehensive context in the review brief. Codex can read local files when you pass an absolute path; manual reviewers usually cannot, so attach or paste the same brief there.
• Be honest about weaknesses — hiding them leads to worse feedback
• Push back on criticisms you disagree with, but accept valid ones
• Focus on ACTIONABLE feedback — "what experiment would fix this?"
• Document the threadId for potential future resumption
• The review document should be self-contained (readable without the conversation)

Prompt Templates

For initial review:

"I'm going to present a complete ML research project for your critical review. Please act as a senior ML reviewer (NeurIPS/ICML level)..."

For experiment design:

"Please design the minimal additional experiment package that gives the highest acceptance lift per GPU week. Our compute: [describe]. Be very specific about configurations."

For paper structure:

"Please turn this into a concrete paper outline with section-by-section claims and figure plan."

For claims matrix:

"Please give me a results-to-claims matrix: what claim is allowed under each possible outcome of experiments X and Y?"

For mock review:

"Please write a mock NeurIPS review with: Summary, Strengths, Weaknesses, Questions for Authors, Score, Confidence, and What Would Move Toward Accept."

Review Tracing

After each reviewer call, save the trace following shared-references/review-tracing.md (Policy C — forensic; never silently skip). Record the selected route/session/cwd, requested and reported models, actual effort, raw response, completion and scope, verdict, independence and acceptance per reviewer-backends.md. Use save_trace.sh (resolved per the chain in shared-references/integration-contract.md §2) or write files directly to .aris/traces/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full). A verdict-bearing manual response MUST begin with Reviewer-Model: <exact-model-id> — pass the model THIS session is actually running as in executor_model. Missing, unknown, or same-family identity cannot acquit; emit REVIEW_UNAVAILABLE rather than guessing. If the executor model cannot be named, manual review's cross-family claim is unprovable — say so in the report instead of asserting it.