Provider for ChatGPT-plan-backed Codex execution via codex exec.
This provider treats Codex as a structured completion backend rather than a full agent runtime. Alloy remains responsible for the tool loop, while Codex receives the current transcript and available tool definitions, then returns JSON describing either a final assistant response or one or more tool calls.
Config
Required:
:model- Codex model name (for example"gpt-5.4")
Optional:
:codex_bin- Executable path (default:"codex"):workdir- Directory passed tocodex exec(defaults to a temp dir):profile- Optional Codex config profile:codex_home- OverrideCODEX_HOMEinstead of creating an isolated temp home:command_runner- Test hook matchingSystem.cmd/3:system_prompt- System prompt string
Notes
- Authentication is handled by the local
codexCLI login state. - By default the provider creates a minimal temporary
CODEX_HOMEcontaining onlyauth.json, which avoids pulling in the user's full MCP/plugin config. - Usage accounting is not exposed by
codex execin a structured form yet, so this provider currently reports zero token counts. - Streaming is emulated by running a normal completion and replaying the final text to the provided callback.