Research Review via External Reviewer Backend (ultra reasoning)
Unverified●38/40Claude Code●WorksValid SKILL.md that declares allowed-tools
Cursor·UnknownWe have not crawled the repo tree, so we will not guess
Codex·UnknownWe have not crawled the repo tree, so we will not guess
Gemini CLI·UnknownThe spec defines no detection rule for Gemini
Copilot·UnknownWe have not crawled the repo tree, so we will not guess
npx agentalley add research-reviewWho is stuck, and on what
Get a deep critical review of research from an external reviewer backend (Codex or manual). Use when user says "review my research", "help me review", "get external review", or wants critical feedback on research ideas, papers, or experimental results.
The whole source
Frontmatter — 4 properties
| name | research-review |
|---|---|
| description | Get a deep critical review of research from an external reviewer backend (Codex or manual). Use when user says "review my research", "help me review", "get external review", or wants critical feedback on research ideas, papers, or experimental results. |
| argument-hint | [topic-or-scope] |
| allowed-tools | Bash(*), Read, Grep, Glob, Write, Edit, mcp__codex__codex, mcp__codex__codex-reply, mcp__manual_review__review, mcp__manual_review__review_reply |
| 1 | --- |
| 2 | name: research-review |
| 3 | description: Get a deep critical review of research from an external reviewer backend (Codex or manual). Use when user says "review my research", "help me review", "get external review", or wants critical feedback on research ideas, papers, or experimental results. |
| 4 | argument-hint: "[topic-or-scope]" |
| 5 | allowed-tools: Bash(*), Read, Grep, Glob, Write, Edit, mcp__codex__codex, mcp__codex__codex-reply, mcp__manual_review__review, mcp__manual_review__review_reply |
| 6 | --- |
| 7 | |
| 8 | # Research Review via External Reviewer Backend (ultra reasoning) |
| 9 | |
| 10 | > 🔒 **Do not wrap this skill in `/loop`, `/schedule`, or `CronCreate`.** It is |
| 11 | > verdict-bearing — it produces a cross-model review verdict, multi-round with |
| 12 | > reviewer thread continuity. An external timer re-fires the verdict on |
| 13 | > wall-clock time and breaks the reviewer's round-to-round memory: zero new |
| 14 | > signal, full token cost. Schedule the *external wait that precedes it* (work |
| 15 | > ready → then review once), not the verdict. See |
| 16 | > [`shared-references/external-cadence.md`](../shared-references/external-cadence.md). |
| 17 | |
| 18 | Get a multi-round critical review of research work from the selected external reviewer backend with maximum reasoning depth. |
| 19 | |
| 20 | ## Constants |
| 21 | |
| 22 | - 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-5.5`, `o3`). 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). |
| 23 | - **REVIEWER_BACKEND = `codex`** — Default: Codex MCP (ultra). Override with `— reviewer: oracle-pro` for Oracle MCP, or `— reviewer: manual` for Manual Review MCP. If manual-review MCP is unavailable, stop and print the install command; do not fall back to Codex. See `shared-references/reviewer-routing.md`. |
| 24 | |
| 25 | ## Reviewer Calling Convention |
| 26 | |
| 27 | When calling the reviewer, branch on REVIEWER_BACKEND: |
| 28 | |
| 29 | **If REVIEWER_BACKEND = `codex`:** |
| 30 | Use `mcp__codex__codex` for new review threads. |
| 31 | Use `mcp__codex__codex-reply` for follow-up rounds (reuse threadId). |
| 32 | |
| 33 | **If REVIEWER_BACKEND = `manual`:** |
| 34 | Use `mcp__manual_review__review` for new review threads with: |
| 35 | prompt: [exact same prompt that would go to Codex] |
| 36 | config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true} |
| 37 | Save the returned `threadId`. |
| 38 | Use `mcp__manual_review__review_reply` for follow-up rounds with: |
| 39 | threadId: [saved manual-review threadId] |
| 40 | prompt: [follow-up prompt] |
| 41 | config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true} |
| 42 | |
| 43 | Content fidelity: the manual reviewer should see the same substantive review |
| 44 | brief Codex would read. If the manual UI supports file upload / attachment, |
| 45 | reuse the same brief file; otherwise paste the brief contents inline because |
| 46 | remote web UIs cannot read your local filesystem paths. Review tracing applies |
| 47 | equally to both backends. |
| 48 | |
| 49 | ## Context: $ARGUMENTS |
| 50 | |
| 51 | ## Prerequisites |
| 52 | |
| 53 | - **Codex MCP Server** configured in Claude Code: |
| 54 | ```bash |
| 55 | claude mcp add codex -s user -- python3 "$HOME/aris_repo/mcp-servers/codex-exec/server.py" # your ARIS clone's path |
| 56 | ``` |
| 57 | - This gives Claude Code access to `mcp__codex__codex` and `mcp__codex__codex-reply` tools |
| 58 | |
| 59 | ## Workflow |
| 60 | |
| 61 | ### Step 1: Gather Research Context |
| 62 | Before calling the external reviewer, compile a comprehensive briefing: |
| 63 | 1. Read project narrative documents (e.g., STORY.md, README.md, paper drafts) |
| 64 | 2. Read any memory/notes files for key findings and experiment history |
| 65 | 3. Identify: core claims, methodology, key results, known weaknesses |
| 66 | |
| 67 | ### Step 2: Initial Review (Round 1) |
| 68 | Send a detailed prompt with ultra reasoning, using the selected backend. For |
| 69 | the `codex` backend, keep the MCP payload short: write the full briefing to |
| 70 | `RESEARCH_REVIEW_REQUEST.md`, then point Codex at that file. |
| 71 | |
| 72 | *For codex backend:* |
| 73 | |
| 74 | ``` |
| 75 | mcp__codex__codex: |
| 76 | model: gpt-6-astra |
| 77 | config: {"model_reasoning_effort": "ultra"} |
| 78 | prompt: | |
| 79 | Read the review brief at <absolute path to RESEARCH_REVIEW_REQUEST.md>. |
| 80 | Executor notes are not evidence beyond the files they cite, so verify the |
| 81 | referenced artifacts before judging. |
| 82 | Please act as a senior ML reviewer (NeurIPS/ICML level). Start from the |
| 83 | assumption that the work is broken somewhere — your job is to find where. |
| 84 | Be adversarial. Trust nothing the author tells you — verify everything |
| 85 | yourself. Identify: |
| 86 | 1. Logical gaps or unjustified claims |
| 87 | 2. Missing experiments that would strengthen the story |
| 88 | 3. Narrative weaknesses |
| 89 | 4. Whether the contribution is sufficient for a top venue |
| 90 | |
| 91 | === SCOPE LIMITS (these bound what you PROPOSE, never what you look for) === |
| 92 | Report anything that is actually wrong here — including a rare-looking case, if |
| 93 | this repo actually produces it. Then keep the fix in scope: |
| 94 | 1. This is a RESEARCH-WORKFLOW tool, not a security paper. Verification is |
| 95 | welcome; over-defense is not. Assume a cooperating operator on their own |
| 96 | machine — a malicious local user is NOT in the threat model. |
| 97 | 2. Do NOT propose SHA / hash / content-fingerprint / digest-binding schemes. |
| 98 | Reporting a real defect in hashing code that already exists is fine. |
| 99 | 3. NO speculative machinery: do not add feature flags, migration frameworks, |
| 100 | compat layers, wrappers, pins, or similar mechanisms unless evidence shows |
| 101 | a current repo defect they fix or an explicit existing invariant they must |
| 102 | preserve. "Load-bearing", "compatibility", and "not scaffolding" are labels, |
| 103 | not evidence. Point to the failing path/artifact or invariant, and check the |
| 104 | proposal's factual premises, such as whether a named package version exists. |
| 105 | 4. NO corner-case obsession: exotic encodings, symlink races, RTL text and |
| 106 | millisecond races are out of scope unless you can show the case arises here. |
| 107 | 5. Where a rubric or checklist is genuinely needed, do not over-mechanize |
| 108 | judgement. A clear sentence a human reads beats a scored table nobody |
| 109 | maintains. |
| 110 | Exception: code that runs remote commands, starts a network service, or installs |
| 111 | an MCP server runs on the user's machine with their credentials — trust-boundary |
| 112 | findings there are in scope and the default is strict. |
| 113 | Say plainly when something is correct. Do not manufacture findings. |
| 114 | Be brutally honest. If, after genuinely trying to break it, the work |
| 115 | holds up and is ready, say so clearly. |
| 116 | ``` |
| 117 | |
| 118 | The review brief should contain the full research context, the specific |
| 119 | questions, and the primary artifact / raw-result paths the reviewer should |
| 120 | inspect. |
| 121 | |
| 122 | *For manual backend:* use `mcp__manual_review__review` with the same brief |
| 123 | contents. If the manual-review UI supports attachments, attach |
| 124 | `RESEARCH_REVIEW_REQUEST.md`; otherwise paste the brief inline. Save the |
| 125 | returned `threadId`. |
| 126 | |
| 127 | ### Step 3: Iterative Dialogue (Rounds 2-N) |
| 128 | For `codex` backend: use `mcp__codex__codex-reply` with the returned `threadId`. |
| 129 | For `manual` backend: use `mcp__manual_review__review_reply` with the same `threadId`. |
| 130 | Use the appropriate tool to continue the conversation. For Codex follow-up |
| 131 | rounds, write an updated brief such as `RESEARCH_REVIEW_ROUND_2.md` and send |
| 132 | only the path: |
| 133 | |
| 134 | ```text |
| 135 | mcp__codex__codex-reply: |
| 136 | threadId: [saved reviewer threadId from Step 2] |
| 137 | # replies inherit the thread's model/effort (gpt-6-astra ultra) |
| 138 | prompt: | |
| 139 | Read the updated review brief at <absolute path to |
| 140 | RESEARCH_REVIEW_ROUND_2.md>. |
| 141 | Focus on unresolved weaknesses and whether the revision actually fixed them. |
| 142 | ``` |
| 143 | |
| 144 | For manual follow-up rounds, attach that same updated brief if possible; |
| 145 | otherwise paste it inline. |
| 146 | |
| 147 | For each round: |
| 148 | 1. **Respond** to criticisms with evidence/counterarguments |
| 149 | 2. **Ask targeted follow-ups** on the most actionable points |
| 150 | 3. **Request specific deliverables**: experiment designs, paper outlines, claims matrices |
| 151 | |
| 152 | Key follow-up patterns: |
| 153 | - "If we reframe X as Y, does that change your assessment?" |
| 154 | - "What's the minimum experiment to satisfy concern Z?" |
| 155 | - "Please design the minimal additional experiment package (highest acceptance lift per GPU week)" |
| 156 | - "Please write a mock NeurIPS/ICML review with scores" |
| 157 | - "Give me a results-to-claims matrix for possible experimental outcomes" |
| 158 | |
| 159 | ### Step 4: Convergence |
| 160 | Stop iterating when: |
| 161 | - Both sides agree on the core claims and their evidence requirements |
| 162 | - A concrete experiment plan is established |
| 163 | - The narrative structure is settled |
| 164 | |
| 165 | ### Step 5: Document Everything |
| 166 | Save the full interaction and conclusions to a review document in the project root: |
| 167 | - Round-by-round summary of criticisms and responses |
| 168 | - Final consensus on claims, narrative, and experiments |
| 169 | - Claims matrix (what claims are allowed under each possible outcome) |
| 170 | - Prioritized TODO list with estimated compute costs |
| 171 | - Paper outline if discussed |
| 172 | |
| 173 | Update project memory/notes with key review conclusions. |
| 174 | |
| 175 | > **Composed mode** — if invoked with `— composed: <canonical-report-path>` (an |
| 176 | > orchestrator like `/idea-discovery` passes this), do **not** write a standalone review |
| 177 | > `.md` in the project root. The raw conversation is already persisted to `.aris/traces/…` |
| 178 | > (see *Review Tracing* below — that audit copy is kept in every mode); fold the review |
| 179 | > *conclusions* (consensus, claims matrix, prioritized TODOs) into the orchestrator's |
| 180 | > canonical report and cite the trace path there. **Default (no `— composed:` directive): |
| 181 | > behave exactly as above — write the standalone review document.** Never infer composed |
| 182 | > mode from a report file merely existing. Full rules: |
| 183 | > [`shared-references/output-composition.md`](../shared-references/output-composition.md). |
| 184 | |
| 185 | ## Key Rules |
| 186 | |
| 187 | - ALWAYS pin `model: gpt-6-astra` + `config: {"model_reasoning_effort": "ultra"}` for reviews (deep-audit tier; capability fallback per `reviewer-routing.md`, never below `xhigh`) |
| 188 | - That pin is the **Codex** backend's. For `manual`, use the identity-bearing config from the Reviewer Calling Convention above; `model`, `sandbox` and `cwd` are Codex-only |
| 189 | - Put comprehensive context in the review brief. Codex can read local files |
| 190 | when you pass an absolute path; manual reviewers usually cannot, so attach or |
| 191 | paste the same brief there. |
| 192 | - Be honest about weaknesses — hiding them leads to worse feedback |
| 193 | - Push back on criticisms you disagree with, but accept valid ones |
| 194 | - Focus on ACTIONABLE feedback — "what experiment would fix this?" |
| 195 | - Document the threadId for potential future resumption |
| 196 | - The review document should be self-contained (readable without the conversation) |
| 197 | |
| 198 | ## Prompt Templates |
| 199 | |
| 200 | ### For initial review: |
| 201 | "I'm going to present a complete ML research project for your critical review. Please act as a senior ML reviewer (NeurIPS/ICML level)..." |
| 202 | |
| 203 | ### For experiment design: |
| 204 | "Please design the minimal additional experiment package that gives the highest acceptance lift per GPU week. Our compute: [describe]. Be very specific about configurations." |
| 205 | |
| 206 | ### For paper structure: |
| 207 | "Please turn this into a concrete paper outline with section-by-section claims and figure plan." |
| 208 | |
| 209 | ### For claims matrix: |
| 210 | "Please give me a results-to-claims matrix: what claim is allowed under each possible outcome of experiments X and Y?" |
| 211 | |
| 212 | ### For mock review: |
| 213 | "Please write a mock NeurIPS review with: Summary, Strengths, Weaknesses, Questions for Authors, Score, Confidence, and What Would Move Toward Accept." |
| 214 | |
| 215 | ## Review Tracing |
| 216 | |
| 217 | After each reviewer call (`mcp__codex__codex`, `mcp__codex__codex-reply`, `mcp__manual_review__review`, or `mcp__manual_review__review_reply`), save the trace following `shared-references/review-tracing.md` (Policy C — forensic; never silently skip). 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`). |
| 218 | A verdict-bearing manual response MUST begin with |
| 219 | `Reviewer-Model: <exact-model-id>` — pass the model THIS session is actually |
| 220 | running as in `executor_model`. Missing, unknown, or same-family identity |
| 221 | cannot acquit; emit `REVIEW_UNAVAILABLE` rather than guessing. If the executor |
| 222 | model cannot be named, manual review's cross-family claim is unprovable — say |
| 223 | so in the report instead of asserting it. |
| 224 |
Reviews
Installed this one?Write the first review and take the Trailblazer badge.
Alternatives
Training CheckPeriodically check WandB metrics during training to catch problems early (NaN, loss divergence, idle GPUs). Avoids wasting GPU hours on broken runs. Use when training is running and you want automated health checks.●····●37/40Changelog AutomationAutomate changelog generation from commits, PRs, and releases following Keep a Changelog format. Use when setting up release workflows, generating release notes, or standardizing commit conventions.◐····●36/40Hermes Tweet> Install and operate Hermes Tweet, a Hermes Agent plugin for X/Twitter research, timeline reading, tweet analysis, and approval-gated private or state-changing operations. Use this skill when installing Hermes Tweet, researching X/Twitter accounts, monitoring launch signals, investigating mentions, auditing giveaways, or preparing gated X operations. Use proactively when a Hermes Agent workflow needs current X/Twitter context. Requires XQUIK_API_KEY for read and action tools.◐····●36/40Social Publishing> Schedule and publish social media posts across 13 platforms (X, LinkedIn, Instagram, Facebook Pages, TikTok, Discord, Telegram, YouTube, Reddit, WordPress, Pinterest) via the SocialClaw API. Use when the user wants to publish, schedule, or manage social media content programmatically. Requires SOCIALCLAW_API_KEY.◐····●34/40