Use two distinct modes: get bounded read-only advice, or relay authorized work to an existing Claude conversation. Codex owns its decisions; a receiving Claude conversation retains its own owner instructions and permissions.
Route the request
| User intent | Action |
|---|---|
| "Ask Claude to review this approach" | Consult the read-only advisor, default Opus |
| "Send this over to Claude chat 'claude test'" | Resolve that exact name and hand off |
| "Pass the findings to this Claude URL" | Queue one message into that existing cloud session |
| "Continue that stopped Claude conversation with these findings" | Resolve it, then explicitly resume and run one turn |
| "Connect this task to Claude session X" | Bind the exact destination for later use; do not send yet |
| "What Claude sessions are running?" | List/inspect only |
Infer "this", "these findings", and similar content from the current task. Prepare the useful result before sending it. Carry forward the user's established scope and target. Do not ask for approval again when the request already authorizes the handoff. Ask only for missing content or a destination that cannot be uniquely resolved. A vague request for advice is not authorization to message unrelated existing Claude sessions.
Consult Claude
Read references/advisory.md for advisory model selection, prompt composition, persistence, report handling, and failure behavior. Resolve scripts relative to this skill's directory and run from the project being reviewed:
bash "$SKILL_DIR/scripts/consult-opus.sh" "<focused advisory question>"
# Only for an explicit Fable request or a justified frontier question:
bash "$SKILL_DIR/scripts/consult-fable.sh" "<focused advisory question>"
The existing wrapper commands and bindings remain compatible. Ordinary advice uses restricted Read/Grep/Glob tools, no MCP, and refuses inbound peer messages. Never broaden an advisor's permissions to make a handoff work. Skip model consultation for mechanical tasks that local tools can settle.
Resolve a handoff destination
Read references/handoffs.md for delivery, receipts, bindings, stopped-session history, and recovery. The Python helper requires Python 3.10+ and an authenticated Claude Code CLI. Check capabilities when needed:
python3 "$SKILL_DIR/scripts/claude_handoff.py" doctor
python3 "$SKILL_DIR/scripts/claude_handoff.py" list
python3 "$SKILL_DIR/scripts/claude_handoff.py" inspect --to 'claude test'
Discovery uses claude agents --json --all across local project folders, not the
current folder alone. Names containing spaces work. Match exact names or UUIDs;
use --project '/path/to/project' to distinguish duplicate records. Do not select
the newest, fuzzy-match silently, or confuse background job IDs with session UUIDs.
If two live sessions still have the same name, have the owner rename one:
Claude's native SendMessage lookup uses names, not full session UUIDs.
For an unlisted older session, use --history with the optional official
claude-agent-sdk package, or an explicit UUID and project directory. Local discovery
does not search other computers. For cloud sessions, accept an explicit
https://claude.ai/code/<id> URL or session_.../cse_... ID. Never substitute a
local session for a remote destination.
Send one handoff
Prepare a compact body with Outcome, Evidence, Requested action, and Work continuing
when useful, matching the shape of the reciprocal codex-handoff skill. Preserve
the user's exact wording when they supply a message. Keep secrets and unnecessary
private data out of the message; local artifact paths do not transfer files.
python3 "$SKILL_DIR/scripts/claude_handoff.py" send --to 'claude test' <<'HANDOFF_EOF'
Outcome:
<useful findings or prepared work>
Evidence:
<relevant paths and observations>
Requested action:
<bounded next step in the receiving conversation>
Work continuing: No
HANDOFF_EOF
The helper adds [Codex handoff], source provenance, a message ID, and the
receiving authority boundary. Use --file for a prepared UTF-8 file, --message
for short text, or --dry-run to inspect without sending. Pass arguments as data;
never interpolate handoff text into shell code.
- Live local: a restricted Opus courier can only send one exact message. A PreToolUse hook checks the target identity and literal content immediately before allowing SendMessage. It cannot read project files or execute the handoff itself.
- Stopped local: use
--resumeonly when continuing/running that conversation is within the user's request. This executes a turn; it is not a passive queue. - Cloud: use the explicit existing target; the helper queues and exits.
- Return once: add
--reply-to '<Codex task UUID>'only when the user requested a reply back. The receiver can usecodex-handoffif installed and permitted.
Do not stop, restart, fork, rename, change inbound settings, or grant permissions
to the destination just to deliver a message. Respect held/refused messages.
Never use concurrent --resume as a fallback for live messaging, write directly
into transcript files, or implement an undocumented socket protocol.
Report the evidence
Use the helper's receipt, not a courier's claim that it sent something:
prepared: dry run, nothing sent.sent/queued: transport acknowledged the handoff; receiver completion is unknown.held: receiving approval/settings are preventing immediate delivery.completed: a resumed headless turn finished; inspect its result and any denied tools.refused/not_sent/failed: report the actual limitation.uncertain: inspect the destination and receipt before any retry.
Receipts and bindings are private local state. Identical sends in the same source
task, project and target reuse an ID and do not automatically send again.
Use --new-message only for an intentionally new, repeated message, never to evade
an uncertain receipt. A source task UUID supplies return routing, not permission.
For verified capabilities, platform limits and test commands, read references/compatibility.md.