Consult Claude
Claude owns a bounded answer and the strategy for finding it: an investigation, recommendation, caller map, API example, migration assessment, or finished text draft. The coordinating Codex owns overall continuity, acceptance of the answer, and any execution it leads to. Use delegate-claude when Claude should own edits or run an experiment. If the user already authorizes a delivered change, assign the investigation and implementation together through that route; do not add an initial consultation merely because the approach is unresolved.
Use Claude early when the result depends on reasoning, design judgment, taste, or choosing between competing interpretations. Codex need not solve the question before briefing Claude or wait until it is stuck. Give Claude the desired outcome and raw source evidence, and enough ownership to finish the answer. A short but consequential question can warrant consultation; routine mechanical work stays in Codex. A standalone consultation does not automatically launch an adversarial review.
Choose Claude for the quality of its reasoning and judgment, including on a small question when the choice matters. Use Codex assistance for supporting evidence that Claude can inspect without giving up its understanding of the problem. Judge the answer by its reasoning, evidence, and quality of judgment. Elapsed time and usage are context, not the objective.
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.
Make the design legible
Supply the desired outcome, genuine constraints, representative callsites, and relevant existing implementation excerpts with source paths. Include actual and proposed API code up front unless an explicitly blind review is selected under adversarial-review. That skill owns the briefing method. 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 a requested adversarial design review, have Claude apply adversarial-review itself without commissioning another reviewer. Its coordinator owns staffing, the starting questions, and relaying exchanges between reviewers; this consultation supplies the review's Claude seat or a standalone opinion. For other design questions, answer the bounded question from source evidence; a complete architectural review is not implied. Treat the implementer's reasoning as contestable evidence. Ask for a concrete smaller alternative when that would decide the question; keeping the design is valid when its guarantees earn their cost.
Send a brief
Requires Bun, Git, and authenticated Claude Code 2.1.280 or later with access to
the selected model. New and resumed consultations default to claude-opus-5-5
at high effort. Sonnet is an explicit choice for routine bounded work, not the
default reasoning lead. Adversarial architecture reviews 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 \
--account personal --account-email claude@example.com \
--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
Read, Glob, and Grep. Without --codex, MCP tools are blocked. With --codex,
only the host's read-only evidence tool is configured and approved; Claude can
choose bounded evidence requests while retaining the investigation and judgment.
Read Codex assistance before enabling it for
setup, limits, result handling, and the difference from /codex:rescue.
Restricted mode, disabled hooks, and outside-read restrictions retain the
Claude boundary. The launcher creates no source replica or checkpoint; native
Claude and Codex sessions own conversation history. 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 assignment that may outlive the current turn or context, record a compact locator in existing task notes: native session ID, account and verified email, absolute checkout, mode, actual model and chosen effort. For a delivery, add the workspace choice, source baseline, allowed or transferred paths, completed checks, blocker, and next step. If Codex assistance is enabled, retain that launch choice, helper model, limits and write roots, and returned helper IDs. The coordinating Codex owns these notes for every Claude assignment. Native history owns the full conversation; do not build another session database. Update the same record as work proceeds.
For an architecture-review follow-up, repeat high effort:
bun .agents/skills/consult-claude/scripts/consult-claude.ts \
--account personal --account-email claude@example.com \
--effort high --resume <session_id> <<'EVIDENCE'
[Requested evidence, commands and relevant raw output, changed source, next question.]
EVIDENCE
Resume the same session for the same assignment in the same checkout and account. New evidence or a correction does not need a new conversation; state source changes and new evidence in the follow-up. Use a fresh session for an independent review of Claude-authored work or a materially different frame. The launcher reapplies read-only tools but does not verify the native session's assignment or original workspace; verify that identity before resuming. For review of a delivered result, first supply the desired behavior, source baseline, actual callers, and cumulative diff from the idle assigned checkout or a snapshot containing that evidence. For a clone delivery, invoke the launcher by its absolute path in the coordinating checkout, with the clone as the working directory, so code under review does not set its own access boundary. Withhold the worker's explanation until the reviewer has reconstructed the system. Resuming retains context and does not restore independence. The launcher reapplies the access boundary on every turn. If the shell tool yields a running process, keep monitoring it and provide progress updates. There is no interactive attach step.
The launcher sends one fixed brief per turn and provides no mid-turn instruction channel. Steer at completed-turn boundaries: state what remains accepted, what changed, and the question the next turn should answer. As a review-group seat, Claude receives other reviewers' findings, questions, and objections the same way: the coordinator resumes the session with them attributed to their authors, and Claude replies, revises its position, or holds its objection.
The shared launcher owns launch defaults. New consultations, delegations and
resumes use Opus 5.5 high unless explicit --model or --effort flags override
them. Codex helpers use medium reasoning independently. Repeat any deliberate
nondefault model or effort on each follow-up. A resume retains the same native
conversation history even when the model changes; an older Sonnet conversation
can continue with Opus without starting over. Native command-line effort does
not persist, so the launcher reapplies high on every turn.
--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.
Choose a subscription account
Every launch, including resumes, requires --account personal|work and
--account-email <verified Claude email>.
personal uses the existing native default login; work uses ~/.claude-work.
The email assertion blocks a profile signed into the wrong identity. Codex carries
the verified email and original account into follow-ups; no alias registry is
stored. Do not infer a Claude login from a ChatGPT billing email. For example:
bun .agents/skills/consult-claude/scripts/consult-claude.ts \
--account personal --account-email claude@example.com \
--model claude-opus-5-5 --effort high <<'BRIEF'
[Bounded consultation.]
BRIEF
The launcher clears inherited Anthropic credentials, endpoints, provider flags,
and secure-storage overrides in the child environment. Native auth status
must identify the expected paid claude.ai account and configuration directory
before the brief is sent. Restricted settings apply to both preflight and prompt;
managed policy still applies. Claude resolves --resume in the selected profile
and fails if it cannot find the session. The launcher does not inspect transcript
files. Dry-run previews the selection without checking login and prints removed
variable names, never credential values.
Claude owns login, Keychain storage, refresh, and history. For a new work login, remove conflicting authentication overrides and run the native login with the user's work email:
CLAUDE_CONFIG_DIR="$HOME/.claude-work" claude auth login --claudeai --email "work@example.com"
The user completes the browser login and verifies its Claude email. Keep profile
login changes out of active launches: preflight does not lock the login while
Claude runs. Do not copy credentials between profiles or log out of the default
account. Explicitly setting CLAUDE_CONFIG_DIR to
~/.claude selects a different Keychain entry from leaving it unset, so preserve
the unset default for personal. This follows the official
multiple-account guidance.
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 within the user's authorized scope. Do not turn an advice-only request into execution. Return artifacts and evidence, including limitations. A longer report or more objections does not establish that a consultation was useful.
For a standalone consultation, return the recommendation, supporting evidence, and remaining disagreement. Stop when the bounded decision has enough evidence; consensus is not required. As a seat in an adversarial review group, Claude keeps answering relayed findings and objections until it can explicitly support the group's direction or names the objection that still stands. Adversarial-review owns closure.
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.