Run Codex¶

Requires the OpenAI Codex CLI on PATH and its credentials (OPENAI_API_KEY or a completed codex login; CODEX_HOME is forwarded).

harnesslab doctor                        # shows "codex cli ok <version>"
harnesslab run demo --variants codex-default

What the adapter runs¶

codex exec --json --full-auto --sandbox workspace-write --skip-git-repo-check
      --color never -C <worktree> -c sandbox_workspace_write.network_access=false
      -o <last-message-file> -

with the prompt on stdin. The JSONL stream is parsed incrementally into command executions (with exit codes and output), file changes, MCP tool calls, web searches, todo lists, assistant messages, per-turn usage and errors. Reasoning items are counted as reasoning_event; their text is discarded before anything is written. A legacy {"id", "msg": {...}} stream shape is also understood.

Two honest gaps:

Codex does not report cost, so reported_cost_usd stays null. Give it a pricing table for estimates.
Codex's stream does not expose model invocations, so llm_calls is null.

Variant options¶

Option	Default	Effect
`model`	CLI default	`-m`
`sandbox`	`workspace-write`	`workspace-write`, `read-only`, or the explicit opt-in `danger-full-access`
`full_auto`	`true`	`--full-auto` (only with `workspace-write`)
`network_access`	`false`	`-c sandbox_workspace_write.network_access=`
`reasoning_effort`	unset	`-c model_reasoning_effort=`
`profile`	unset	`--profile`
`config_overrides`	`{}`	`-c key=value` for each entry, e.g. a compaction limit
`action_policy`	unset	`batched`, `fine` or free text prepended to the prompt
`harness`	unset	a harness bundle directory
`skip_git_repo_check`	`true`	`--skip-git-repo-check`
`extra_args`	`[]`	appended verbatim
`env_passthrough`	`[]`	extra environment variables forwarded
`executable`	`codex`	the binary to run

danger-full-access is never selected implicitly.

With a harness bundle¶

Codex has no plugin mechanism, so the bundle's system_prompt.md and each skill's text become a prompt prefix placed before the action policy and the task prompt. hooks.json and agents/ are unsupported; the adapter records one harness_components_ignored system event naming them rather than dropping them silently.

Example variant¶

variants:
  - id: codex-high-effort
    runner: codex
    model: gpt-5-codex
    reasoning_effort: high
    config_overrides: { model_context_window: 200000 }
    network_access: false

The adapter is verified against recorded JSONL fixtures and a stand-in executable that replays them through the real adapter code. To exercise the real CLI: HARNESSLAB_INTEGRATION=1 uv run pytest tests/test_integration_real.py from a clone.