---
name: arc-local-code
description: Dispatch work or cross-model review from Claude Code on the ARC host to Codex or another ARC Agent, then wait in the background and review the task outcome through the local ARC CLI.
---

# ARC from local Claude Code

Use this flow only when this session can run the CLI on the ARC execution host.
A claude.ai MCP connector is shared by Code, Chat and Cowork and cannot identify this local session.
Use your own shell, files and Git for host work.

## Discover and select

Run `arc cli control guide`.
Use the returned `cliPath` if the shell has no `arc` shim.
To opt into the shared Local Code guide, use the returned `sourceId` with:

```sh
printf '%s' '{"capabilityProfile":{"hostAccess":"host","waitMode":"background-cli"}}' | arc cli source update <sourceId> --stdin
arc cli control guide
```

This preference changes discovery for all built-in CLI callers, not execution permissions.
Read `arc cli control dispatch_policy` before selecting the Agent, account, model, effort or tier.
Retain the returned `criteriaHash` after the first read in this session.
On later reads, pass it as `knownCriteriaHash` to omit unchanged criteria.
For multiple criteria blocks, pass a hash map keyed by Agent ID and `global` for top-level criteria.
Use `arc cli control dispatch_policy --stdin` with the args object, such as `{"agent":"codex","knownCriteriaHash":"<criteriaHash>"}`.
Read `arc cli control guide <op>` for each operation's live schema and retry contract.
Do not embed model names, quotas or dispatch criteria from previous sessions.
Refresh the guide when `arcVersion` changes.
Use `cli control guide --stdin` with `{"view":"full"}` for all permitted operations.

## Dispatch and wait

CLI stdin is the operation's args object, not an MCP envelope.
For ordinary work, pass `purpose:"task"` and `notify:null` explicitly because local dispatch defaults to self-test.
Callers SHOULD always provide `title` on `codex_create` and `agent_create`: a short human-readable task title of at most 120 characters, shown in task lists and notifications.
Omitted or blank titles fall back to the first non-empty prompt line, truncated to 120 characters; receipts include `titleHint`.
Rename Codex tasks with `codex_thread` using `action:"name/set"` and `params:{threadId,name}`.
Use the user's authorized scope and the current policy when constructing the dispatch.

```sh
printf '%s' '{"project":"/workspace/project","title":"Implement requested change","prompt":"Implement the requested change.","purpose":"task","notify":null}' | arc cli control codex_create --stdin
```

Retain `requestId`, `threadId` and `turnId` from the response.
If dispatch is pending, read the original receipt until submission is confirmed.
After an uncertain response, reconcile the original receipt or task rather than redispatching.
Run the following command through Claude Code's background-command facility so its completion can return control to this session:

```sh
printf '%s' '{"threadId":"<threadId>","turnId":"<turnId>","until":"terminal-or-interaction","view":"summary"}' | arc cli wait --control --op codex_wait --stdin --until '/conditionMet==true' --timeout 1800
```

A shell `&` alone does not arrange session re-invocation.
If background completion is unavailable, use bounded waits in the active session.
No notification binding or verification handshake is needed for this flow.
On `WAIT_TIMEOUT`, inspect `error.lastResult` and continue waiting on the same task.
On success, inspect `result.lastResult.state`; matched waits can contain failed or interrupted tasks.
Handle interactions through the returned `next` operation without resending the task.
For other Agents, use the live `agent_create` and `agent_wait` schemas with their returned Agent/task identifiers.

## Review

Summary results exclude inbox data and retain only the selected task's outcome.
Follow `completeEvidence` when the message or changed paths are truncated, or evidence is unavailable.
Inspect the actual local diff and relevant validation evidence before recording a verdict.
For an independent cross-model assessment, dispatch another review task through the current policy.
`codex_review` and `agent_review` record verdicts; they do not launch a reviewing model.
Read the review schema and record `pass` or `rework` for the exact terminal turn.
Keep unknown change attribution explicit rather than treating it as no changes.