Crate nucel_agent_codex 

Codex provider — wraps OpenAI’s codex CLI.

Based on the official Codex CLI reference.

• CLI: codex exec --json "<prompt>"
• Protocol: JSONL with event types thread.started → turn.started → item.completed → turn.completed
• Sandbox modes: read-only, workspace-write, danger-full-access
• Approval: --ask-for-approval <policy> (--full-auto is deprecated upstream).

Each query() re-invokes codex exec rather than keeping a long-lived subprocess: multi-turn is achieved by passing the previous session_id to resume().

Minimal example

use nucel_agent_codex::CodexExecutor;
use nucel_agent_core::{AgentExecutor, SpawnConfig};
use std::path::Path;

let executor = CodexExecutor::new();
let session = executor.spawn(
    Path::new("/my/repo"),
    "Refactor the cost calculation in src/cost.rs",
    &SpawnConfig { model: Some("gpt-5-codex".into()), ..Default::default() },
).await?;

// Save the session id and resume later:
let sid = session.session_id.clone();
session.close().await?;

let resumed = executor.resume(
    Path::new("/my/repo"), &sid, "Now add tests.", &SpawnConfig::default(),
).await?;
println!("{}", resumed.query("Did the tests pass?").await?.content);

See also: workspace README and the runnable example crates/unified/examples/codex_resume.rs.

Structs

CodexExecutor

Codex executor — spawns codex exec CLI subprocess.