codex_run
Delegate a coding task to the Codex CLI, specify model and effort per call, and receive back git-verified file changes along with AgentBridge's independent verification output.
Instructions
Delegate a task to the Codex CLI running as an external worker, using your existing Codex login (no API key). Model and reasoning effort are chosen per call and never inherited from global Codex config. Returns the worker's final result plus git-verified files_changed and scope_violations, and — unless skipped — a verification block holding the real output of the project's checks, which AgentBridge runs itself. One call = one attempt; AgentBridge never retries on its own.
Input Schema
TableJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| goal | Yes | What the worker must accomplish. Be specific and self-contained. | |
| mode | Yes | analyze/review are read-only investigations; implement makes changes and runs checks. | |
| model | Yes | Codex model to run, passed to the CLI verbatim. gpt-5.6-sol is the strongest general engineering model, gpt-5.6-terra the deep brownfield/existing-codebase specialist, gpt-5.6-luna the economical high-volume workhorse. Availability is checked against the installed CLI's own model catalogue; an unavailable model fails with requested_model_unavailable and is never substituted. | |
| tests | No | Tests or checks the worker should run. Context for the worker only — use `verify_commands` for anything AgentBridge should execute. | |
| effort | Yes | Reasoning effort: light->low, medium->medium, high->high, 'extra high'->xhigh, max->max. | |
| context | No | Background the worker needs but cannot infer from the repository. | |
| contract | No | Interfaces, types, or API shapes the worker must conform to exactly. | |
| no_touch | No | Project-relative paths the worker must not modify. | |
| allowed_paths | No | Project-relative paths the worker may write. Required in implement mode; overlapping scopes across concurrent implement jobs are rejected. | |
| relevant_files | No | Project-relative files the worker should look at first. | |
| timeout_seconds | No | Wall-clock limit for the worker process. Default 900s, clamped to 30-3600s. | |
| verify_commands | No | Commands AgentBridge runs itself in the project directory after the worker exits, e.g. ["npm run typecheck", "node --import tsx --test"]. Their real output comes back in `verification`. Run without a shell, so no pipes, redirection, or `&&`. Omitted in implement mode, AgentBridge derives them from the project's own `typecheck` and `test` scripts. | |
| skip_verification | No | Set true to skip AgentBridge's own checks for this job. The result then carries only the worker's account of whether the change works. | |
| acceptance_criteria | No | Conditions that must hold for the job to be considered done. | |
| verify_timeout_seconds | No | Wall-clock limit for each verification command. Default 300s. |