Multi llm review

Multi-LLM Code Review

Use this skill for a real review gate, not a popularity vote or a summary-only second opinion. Dispatch review requests through the current agent's local stdio gateway MCP surface, never a direct provider CLI, shadow connector, or remote gateway when the requested validation path is stdio.

Completion contract

A required reviewer is complete only when it inspected the target and returned an evidence-backed APPROVED_UNCONDITIONALLY. A qualified approval, a conditional result, missing evidence, a malformed response, a timeout, or a provider error is not approval.

Do not set a review-round, turn, token, price, budget, or wallclock cap. Do not use route_request, select:"cheapest", maxCostUsd, maxOutputTokens, maxTurns, maxPrice, or a provider print deadline to reduce an exhaustive review. Continue evidence-based repair and re-review until every required healthy reviewer returns APPROVED_UNCONDITIONALLY, the user explicitly cancels, or it returns BLOCKED_EXTERNAL with a concrete external error. A blocked reviewer never becomes an implied approval.

Discover the actual roster

The gateway has seven canonical CLI provider types: Claude, Codex, Gemini (Antigravity), Grok, Mistral Vibe, Devin, and Cursor Agent. Start a full CLI review with all seven in its intended roster. Before dispatching, inspect what this gateway can actually run:

cli_versions()
list_models()
provider_tool_capabilities({cli:"claude"})
provider_tool_capabilities({cli:"codex"})
provider_tool_capabilities({cli:"gemini"})
provider_tool_capabilities({cli:"grok"})
provider_tool_capabilities({cli:"mistral"})
provider_tool_capabilities({cli:"devin"})
provider_tool_capabilities({cli:"cursor"})

Use provider-tools://{provider} as a cached read-only alternative. Optional API providers are configured availability discovered through list_models and their reported capabilities. They have no local CLI checkout/worktree or native ACP guarantee, so they are not silently interchangeable with a required source-inspecting CLI reviewer.

Record a roster before sending prompts:

Provider	Available/authenticated	Target checkout verified	Required controls verified	Result
Claude	yes/no	yes/no	yes/no	required or blocked
Codex	yes/no	yes/no	yes/no	required or blocked
Gemini	yes/no	yes/no	yes/no	required or blocked
Grok	yes/no	yes/no	yes/no	required or blocked
Mistral	yes/no	yes/no	yes/no	required or blocked
Devin	yes/no	yes/no	yes/no	required or blocked
Cursor	yes/no	yes/no	yes/no	required or blocked

If the user explicitly requests a narrower review, name the excluded providers, the reason, and the resulting limitation before dispatch. Otherwise repair an unavailable required provider, its authentication, its workspace routing, or its request shape. Do not skip it and call the result comprehensive.

Verify the target checkout

Never allow a provider to inspect a configured default workspace by accident. Capture the repository identity, exact change set, dirty-file list, and the working directory that each provider will see.

Provider	Exact target selection
Claude	Pass local workingDir:"<repo>". With Claude managed approval this is a high-risk posture input and needs an approval decision plus the operator's managed-bypass setting.
Codex	Pass local workingDir:"<repo>"; start with sandboxMode:"read-only" for inspection, or use workspace-write only for write-producing checks.
Gemini	Pass local workingDir:"<repo>". includeDirs is an extra read path, not cwd selection.
Grok	Pass local workingDir:"<repo>".
Mistral	Pass local workingDir:"<repo>".
Devin	Pass local workingDir:"<repo>" or a verified registered workspace. Use gateway worktree only with an explicit provider-native sessionId not overridden by createNewSession.
Cursor	Pass local workingDir:"<repo>". workspace is Cursor's own selector; an absolute workspace path disagreeing with workingDir is rejected, not ranked.

Do not request a fresh gateway worktree for Grok, Mistral, or Devin. For these providers, gateway worktree admission requires an explicit provider-native sessionId that is not overridden by createNewSession; fresh, createNewSession, and resumeLatest-only worktree requests fail closed. For a fresh isolated review, prepare the target checkout separately and select it with workingDir or a verified registered workspace.

Do not use workspace_* administration tools as an ad hoc workaround for a stdio review path. Correct the gateway launch configuration or use a verified registered workspace.

An unscoped local CLI child runs in a fresh private neutral directory, not the gateway process repository. Do not rely on process cwd as implicit target routing. A cwd-scoped resumeLatest fails closed unless workingDir, workspace, or a configured default workspace provides a stable target.

First-class safe repository review

Prefer review_changes for an ordinary read-only Git review when durable SQLite/PostgreSQL validation storage is available. It captures committed, staged, unstaged, and regular non-ignored untracked evidence without truncation, fences the artifact as untrusted data, returns exact byte and SHA-256 identities, and starts provider-native read-only jobs in the selected repository.

Pass an absolute local workingDir or a registered workspace, then choose the scope, providers, optional judge, stance, and optional literal path filters. Use each returned validation job_status/job_result reference for progress and human visibility. If a judge was requested, wait for every result to become terminal, then call synthesize_validation with the validationId and same repository selector. For review_changes, synthesis ignores caller question/providerResults, reloads exact owned durable linked terminal jobs, reconstructs unavailable requested seats as skipped, and atomically claims the stored judge once. Stored repository, owner, judge, and consent are authoritative. General validation still requires caller question/results. The exact fenced prompt is retained in expiry-bound job payload_json; persisted argv contains a hash marker and the flight recorder receives no repository-review prompt.

HTTP/API reviewer seats require allowApiUpload:true, and remote workspace reviews reject API upload. The tool is absent in Kit mode and when a durable validation-run store is unavailable. It is intentionally read-only, so it does not replace the explicit user-authorized full-access protocol below.

Provider controls and prompt shape

• approvalStrategy:"mcp_managed" and approvalPolicy work only for Claude. Every other CLI must use approvalStrategy:"legacy"; managed approval is rejected before launch because its ambient MCP configuration cannot be isolated.
• Claude managed requests use only provisioned gateway-owned local MCP definitions. Non-empty tool selectors, workingDir, native resume, and other posture changes need a managed approval decision and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1. For a routine source review, do not pass Claude allowedTools/tools selectors. Use legacy only when the operator deliberately accepts provider-native posture.
• Query capability data before applying a provider-specific tool list, output format, MCP setting, session field, sandbox, or native ACP transport. Do not copy Claude tool names into another provider.
• Claude, Codex, Gemini, Grok, and Mistral accept either prompt or promptParts, never both. Devin and Cursor accept only prompt. Preserve a canonical flat evidence packet for all seven and use promptParts only on the five supporting providers. Prefix hashes show gateway-side stable-prefix tracking, not proof of equivalent provider cache hits.
• Use *_request_async when async tools are registered. SQLite/PostgreSQL results are durable, acknowledged memory is ephemeral, and none registers no async/job tools. Poll on a non-blocking cadence without turning cadence into a deadline. A failed job must be diagnosed and retried or reported blocked, never counted as a review.
• Personal Agent Config Kit is not a seven-provider review mode: it is local only, supports Claude/Codex/Mistral only, requires healthy SQLite/PostgreSQL durable admission, and disables validation and least-cost routing. Do not disable it or reduce the roster without the user's explicit choice of security boundary. The normal workingDir rule is not available in Kit mode: Claude and Mistral Kit requests reject caller-supplied workingDir before context compilation. Use explain_effective_config to inspect a candidate folder, then select the target through an already configured registered workspace alias or the configured default workspace. It never inherits the gateway process cwd.

Explicit user-authorized full-access review protocol

Keep the normal inspection defaults above for ordinary review work. Replace them only when the user explicitly authorizes full provider permissions and native MCP access for a review. This is a review capability override, not an authorization to modify the target: tell each reviewer not to edit, stage, commit, reset, or otherwise mutate the repository unless the user separately asks for that work.

1. Build the exact target checkout, then start a fresh local stdio gateway from that checkout with node dist/index.js --transport=stdio. Do not dispatch the review through a globally installed or already-running gateway process whose source revision might differ from the reviewed tree.
2. Before every iteration, use that fresh process to inspect live provider capability and availability. Reapply the full-access control on every new provider job. Do not assume a previous job, a resumed session, or a provider configuration reload retained the grant. The gateway drops sandboxMode on a Codex resume, so that field cannot select a posture there, and the resumed posture is not guaranteed to match the original either (configOverrides still passes through and Codex re-resolves configuration). Start a new Codex session when the access posture must be known.
3. Preserve each provider's ambient native MCP configuration. Do not pass an allowedTools, disallowedTools, tools, allow, deny, mcpServers, or strictMcpConfig list as a purported full-access setting. The gateway cannot create a general full native-MCP allowlist for the non-Claude providers; their native configuration is the actual MCP surface. Require the reviewer to use available MCP tools when useful, and record a missing required native MCP tool as an access finding or external blocker rather than pretending it was available.
4. Do not set caller-imposed maxTurns, maxBudgetUsd, maxPrice, maxTokens, maxCostUsd, maxOutputTokens, printTimeout, or review-loop limits. Do not use a provider-specific deadline to replace a complete review. A gateway's configured safety defaults are not evidence of a caller review cap.

Use these native controls for a new full-access CLI review. All use legacy approval because managed approval is Claude-only. Omit every restriction named in the final column.

Provider	Full-access request controls	Do not combine with the full-access review
Claude	approvalStrategy:"legacy", permissionMode:"bypassPermissions"	allowedTools, disallowedTools, tools, mcpServers, strictMcpConfig
Codex	approvalStrategy:"legacy", sandboxMode:"danger-full-access", dangerouslyBypassApprovalsAndSandbox:true	fullAuto, a resume that needs a new sandbox, MCP restriction fields
Gemini (Antigravity)	approvalStrategy:"legacy", yolo:true	sandbox, skipTrust, non-empty allowedTools, policy/attachment fields, mcpServers
Grok	approvalStrategy:"legacy", alwaysApprove:true	permissionMode (the alwaysApprove branch wins), sandbox, allow, deny, tool lists, mcpServers
Mistral Vibe	approvalStrategy:"legacy", permissionMode:"auto-approve", trust:true	allowedTools, disallowedTools, maxTurns, maxPrice, maxTokens, mcpServers
Devin	approvalStrategy:"legacy", permissionMode:"dangerous", respectWorkspaceTrust:false	sandbox (a true value adds a sandbox)
Cursor	approvalStrategy:"legacy", force:true, sandbox:"disabled", trust:true	mode, autoReview

These are current gateway request controls, not claims that every provider has identical tool behavior. Check the live capability response and the provider's native configuration on each review iteration. If a required control is rejected or a provider cannot access the target or its native MCP tools, retain the exact error as a concrete blocker and repair it before treating that provider as reviewed.

Full-access evidence packet and verdict

For every iteration, give every reviewer a corrective-program specification that contains all of the following:

• The exact base SHA and current HEAD SHA, or an explicit dirty-worktree statement.
• The exact reviewable change identity: a git diff --binary <base> artifact and digest, or an exhaustive git diff --name-status <base> list together with every product-relevant untracked file.
• The verification report used as the corrective-program specification: requirements, invariants, commands actually run, results, remaining risks, prior findings, and each FIXED, DISAGREE, or BLOCKED_EXTERNAL response.
• Persistent evidence locations: job IDs, correlation IDs, raw terminal outputs, test/build logs, and any durable-store or artifact references.

The verification report is a set of corrective claims, not evidence that the claims are true. In the prompt, require the reviewer to inspect the current source, changed and neighboring files, docs, tests, command results, and any needed provider/MCP evidence independently. It must not approve from an intent, plan, summary, green test claim, or "should be fixed" statement. A disagreement with a finding requires exact code, documentation, test, or command evidence, not assertion.

Require this terminal JSON contract in every full-access review prompt:

{
  "verdict": "APPROVED_UNCONDITIONALLY | CHANGES_REQUIRED | BLOCKED_EXTERNAL",
  "findings": [
    {
      "severity": "blocker | major | minor | nit",
      "file_or_doc": "path:line",
      "issue": "concrete defect or verification gap",
      "evidence": "code/doc/test/command evidence actually inspected",
      "required_action": "concrete correction or external dependency"
    }
  ],
  "inspected": ["code, docs, tests, commands, MCP evidence actually inspected"],
  "reviewed_base": "sha or dirty-worktree statement",
  "reviewed_change_identity": "diff artifact digest or exhaustive file list"
}

APPROVED_UNCONDITIONALLY is valid only when the reviewer has directly inspected the final exact change identity and has no finding or verification gap. CHANGES_REQUIRED must cite a resolvable finding. BLOCKED_EXTERNAL must name a concrete unavailable dependency or access failure and include its exact error. A qualified approval, a response based only on the packet, or an output without inspected evidence is not an approval.

When the user specifies a 90-second progress cadence, make a non-blocking wait and do not call llm_job_status or otherwise check progress more often than once every 90 seconds. That cadence is not a deadline. Keep the raw durable review evidence, refresh the packet after every material change, and re-run the full required roster until every reachable reviewer returns APPROVED_UNCONDITIONALLY or a concrete external blocker remains.

Normal safe-review packet and dispatch

Every required reviewer receives the same substantive packet. It is a set of claims to verify, not a substitute for opening the target. This section keeps the normal safe provider controls, but it uses the same strict terminal verdict contract above. It is not an alternative approval grammar for the explicit user-authorized full-access protocol.

• Exact base and head commit IDs or an explicit dirty-worktree statement.
• git diff --no-ext-diff <base>..<head>, changed-file list, and relevant neighboring files.
• Commands actually run, result summaries, and raw-output/digest locations.
• Intended behavior, invariants, security constraints, migration/deployment implications, and documentation contract.
• Earlier findings with a per-finding FIXED, DISAGREE, or BLOCKED_EXTERNAL response backed by file:line, command, test, or upstream-doc evidence.

Require every reviewer to inspect changed files and surrounding code directly, run or inspect relevant verification, cite evidence for every finding, and say what it could not verify. Do not tell reviewers to rely only on a supplied summary or to avoid tools.

Dispatch the full CLI roster

Substitute the verified packet and target. The Gemini and Devin calls below are valid only after their target-routing conditions above are satisfied.

claude_request_async({
  prompt:"Review <packet>. Inspect <repo> directly. Return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  approvalStrategy:"legacy",
  correlationId:"review-claude"
})

codex_request_async({
  prompt:"Review <packet>. Inspect <repo> directly for correctness, races, error paths, and tests. Return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  sandboxMode:"read-only",
  approvalStrategy:"legacy",
  correlationId:"review-codex"
})

gemini_request_async({
  prompt:"Review <packet> in the target checkout. Inspect security, edge cases, and documentation. Return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  approvalStrategy:"legacy",
  correlationId:"review-gemini"
})

grok_request_async({
  prompt:"Independently review <packet> and <repo>. Verify claims directly and report blind spots with evidence. Return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  approvalStrategy:"legacy",
  correlationId:"review-grok"
})

mistral_request_async({
  prompt:"Independently review <packet> and <repo>. Verify correctness, maintainability, tests, and security directly. Return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  approvalStrategy:"legacy",
  correlationId:"review-mistral"
})

devin_request_async({
  prompt:"Independently review <packet> and <repo>. Inspect directly and return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  approvalStrategy:"legacy",
  correlationId:"review-devin"
})

cursor_request_async({
  prompt:"Independently review <packet> and <repo>. Verify claims directly and return the required terminal JSON verdict: APPROVED_UNCONDITIONALLY, CHANGES_REQUIRED, or BLOCKED_EXTERNAL.",
  workingDir:"<repo>",
  approvalStrategy:"legacy",
  correlationId:"review-cursor"
})

For a Claude managed review, replace only the Claude request after the operator has approved the required target/posture. Do not add an allowedTools list just to make it look restrictive: a reviewer needs real inspection access.

Triage and re-review

1. Poll each async job until its terminal result is available. Keep job IDs, provider, target identity, and raw result evidence.
2. Validate each verdict. A response lacking direct-inspection evidence is incomplete even if it says APPROVED_UNCONDITIONALLY.
3. Consolidate duplicate findings, but verify unique findings rather than discarding them because only one provider noticed them.
4. Fix correct findings and run relevant local verification. For disputed findings, send the reviewer the exact counter-evidence and require an evidence-backed withdrawal or revised finding.
5. Refresh the evidence packet and exact change set. Re-dispatch every reviewer whose prior result was not unconditional, plus any reviewer whose evidence became stale after the fix.
6. Finish only when every required healthy reviewer returns APPROVED_UNCONDITIONALLY and the final evidence packet matches the reviewed change set.

If a provider repeatedly reaches a terminal external failure, preserve its error and attempted repair evidence, report BLOCKED_EXTERNAL, and ask the user how to proceed. Do not proceed with a smaller set by default.

Anti-patterns

• Calling a provider directly instead of through the local stdio gateway.
• Reviewing a default or wrong workspace because target routing was not proved.
• Passing unsupported provider controls or Claude tool names to another CLI.
• Treating a gateway bookkeeping session ID as a provider-native resume handle.
• Selecting a cheapest reviewer or applying a cost/time cap to an exhaustive review.
• Accepting summaries, intent, plan compliance, or "should be fixed" as proof.
• Skipping an unavailable provider, treating a timeout as a pass, or ending after an arbitrary number of rounds.