Consult Claude

Claude owns a bounded answer: an investigation, recommendation, caller map, API example, migration assessment, or finished text draft. Codex owns the conversation, execution, verification, and integration. Use delegate-claude when Claude should own edits or run an experiment.

When the user's request or chosen workflow includes Claude, choose a consultation whose result advances a named decision or replaces work Codex would otherwise do. Give Claude enough work to finish the answer; available capacity alone does not justify another opinion. A standalone consultation does not automatically launch the adversarial reviewer pair.

Define the result

Brief Claude with the desired result, intended consumer, source paths, genuine constraints, acceptance evidence, and stopping condition. For example:

Trace every caller affected by this rename API. Return a source-linked caller map, incompatible assumptions, and the smallest migration proposal. Read the current uncommitted source. Stop when every caller class is accounted for; request runtime evidence only where it decides compatibility.

Distinguish explicit requirements, observed behavior, and hypotheses. If the contract is unresolved, assign its discovery rather than assuming a fix. Check the deciding claims against sources; do not redo the whole investigation.

An exploration can advance a decision by discovering a better destination or an unknown blind spot; it need not resolve a question already identified. Name the opportunity the assignment could explore. Claude may propose worthwhile follow-up investigations, but Codex decides which to commission.

Make the design legible

Supply the desired outcome, genuine constraints, representative callsites, and relevant existing implementation excerpts with source paths. For a blind adversarial design pass, withhold proposed API code and your engineering reasoning until Claude has answered from the outcome and callers. Then supply the proposal, reasoning, and unresolved questions in the same session. For other consultations, include actual and proposed API code up front. Use an ASCII diagram when it clarifies ownership, lifecycle, or data flow. Distinguish observed facts, proposals, assumptions, and preferences. Give enough concrete evidence to judge the decision; Claude can read further to verify your account.

For design questions, have Claude apply adversarial-review itself, without launching another reviewer. That skill's coordinator owns reviewer selection; this consultation supplies its Claude reviewer or a standalone Claude opinion. A standalone request for Claude's opinion does not itself launch a Codex reviewer. Explicitly ask for the strongest greenfield direction: starting from the desired outcome and actual callers, what would it build if the current abstraction did not exist? Treat your reasoning as evidence, not constraints. Ask for concrete signatures and callsites, what disappears, new complexity, and requirements being questioned. Keeping the design is valid when it earns its place. A narrow question does not require a full architectural report.

Send a brief

Requires Bun, Git, and authenticated Claude Code 2.1.280 or later with access to the selected model. A new consultation requires an explicit --model. Choose claude-sonnet-5-5 with --effort medium for a bounded caller map, ordinary bug investigation, or draft. Choose claude-opus-5-5 with --effort high for an open-ended design question or consequential judgment. Adversarial architecture reviews explicitly use Opus high, including follow-ups. Use xhigh only for a focused unresolved decision, then return to high. Run from the checkout containing the evidence; supply the brief directly on stdin:

bun .agents/skills/consult-claude/scripts/consult-claude.ts \
  --model claude-opus-5-5 --effort high <<'BRIEF'
[Concrete question, code blocks, source paths, diagram, reasoning, uncertainties.]
BRIEF

On macOS, the Codex workspace sandbox can hide Claude Code Keychain login. Treat a sandboxed claude auth status result of loggedIn: false as inconclusive. Request the same status check outside the sandbox. If it reports true, run the launcher and its follow-ups outside the sandbox. Keep the launcher read-only Claude restrictions in place; do not log in again, export credentials, or switch the whole Codex task to Full access for this case.

The launcher runs one native print-mode turn in the current checkout with only Read, Glob, and Grep. Restricted mode, blocked MCP tools, disabled hooks, and outside-read restrictions enforce the boundary. It creates no replica, brief file, or checkpoint. Claude Code owns session storage. Keep the reviewed files stable during each turn. For parallel work, use a source snapshot containing the relevant committed, uncommitted, and untracked task files. A HEAD-only copy can omit the proposal. Copy prototype evidence into the selected checkout or run there; naming an outside path in the brief does not grant access.

Read the native JSON result, including result, session_id, is_error, and any permission denials. A process starting or exiting successfully is not proof of a successful consultation. Preserve the session ID for follow-ups.

For an architecture-review follow-up, repeat high effort:

bun .agents/skills/consult-claude/scripts/consult-claude.ts --effort high --resume <session_id> <<'EVIDENCE'
[Requested evidence, commands and relevant raw output, changed source, next question.]
EVIDENCE

Resume a completed consultation for the same assignment in the same checkout. Supply changed source, new evidence, and the next question explicitly. The launcher reapplies read-only tools but does not verify the native session's assignment or original workspace; verify that identity before resuming. Use a fresh session for independent exploration, an independent review, or a materially different frame. See delegate-claude for reviewing a worker's result. Resuming retains context: a challenge to another proposal or Codex's synthesis is an informed follow-up, not blind discovery.

The launcher sends one fixed brief per turn, reapplies the access boundary on every turn, and provides no mid-turn instruction channel or interactive attach. Steer at completed-turn boundaries using adversarial-review's coordination guidance for design investigations. If the shell tool yields a running process, monitor it and keep the user informed.

A new consultation requires --model. On resume the launcher supplies no model or effort default: Claude Code restores the session's model, while command-line effort must be repeated. Explicit --model overrides that session's model. --dry-run previews launch arguments without invoking Claude. Consult the native result or transcript before attributing findings to a model: access restrictions and fallback can change the model used.

Adjudicate and continue

Evaluate objections yourself. When Claude requests evidence, run useful checks within the user's existing authorization and return the commands, relevant raw results, and source state. An advice-only request does not authorize edits. Ask before materially expanding the task; prior authorization still applies.

This route remains read-only. When the assignment requires Claude to own implementation or experimental execution, use delegate-claude only with the user's explicit authorization for Claude execution. Return artifacts and evidence, including limitations. A longer report or more objections does not establish that a consultation was useful.

Return the recommendation, supporting evidence, and remaining disagreement. Stop when the bounded decision has enough evidence; consensus is not required.

Read the example and evaluation cases when evaluating this workflow or revising the skills. For launcher changes, verify claude --version, claude --help, the official CLI reference, and programmatic usage. Native Claude sessions own conversation history; the launcher owns only access and transport.