netsky_core/runtime/

codex.rs

//! Codex runtime: build the `codex` CLI invocation for a resident agent.
//!
//! Codex (0.120.0+) has no `--append-system-prompt` / `--base-instructions-file`
//! flag. We pass the rendered netsky prompt as codex's positional `[PROMPT]`
//! arg instead, read at exec time from `$NETSKY_PROMPT_FILE` via the same
//! `$(cat ...)` shell trick the claude runtime uses to avoid tmux's
//! `command too long` argv limit.
//!
//! The rendered prompt includes the full base.md + per-agent identity
//! stanza + cwd addendum, so the Codex session receives the same
//! "you are agent<N>" shape Claude does. The separate sidecar path in
//! `codex_agent.rs` remains for one-off prompt / drain invocations; the
//! resident path lives here.
//!
//! Startup (`/up`) parity: codex's positional `[PROMPT]` arg is already
//! spoken for by the rendered netsky prompt — codex treats it as the
//! first user turn. The startup prompt therefore lands as a follow-on
//! user turn, pasted into the tmux pane via [`paste_startup`] once the
//! session is up. Claude passes startup inline as a second positional.

use std::path::Path;
use std::process::Command;
use std::time::Duration;

use crate::agent::AgentId;
use crate::consts::{ENV_NETSKY_PROMPT_FILE, TMUX_BIN};
use crate::error::{Error, Result};

use super::claude::shell_escape;

/// Per-agent codex-CLI configuration.
#[derive(Debug, Clone)]
pub struct CodexConfig {
    pub model: String,
    pub sandbox: String,
    pub approval: String,
}

impl CodexConfig {
    /// Defaults for a resident codex agent: `gpt-5.4`,
    /// `danger-full-access` sandbox, `never` approval policy. Matches
    /// the sidecar adapter's choices (codex_agent.rs) so swapping
    /// sidecar <-> resident for the same agent N produces comparable
    /// behavior. Overrides land via `AGENT_CODEX_MODEL`,
    /// `AGENT_CODEX_SANDBOX`, and `AGENT_CODEX_APPROVAL` env vars.
    pub fn defaults_for() -> Self {
        let model = std::env::var(ENV_AGENT_CODEX_MODEL)
            .unwrap_or_else(|_| DEFAULT_CODEX_MODEL.to_string());
        let sandbox = std::env::var(ENV_AGENT_CODEX_SANDBOX)
            .unwrap_or_else(|_| DEFAULT_CODEX_SANDBOX.to_string());
        let approval = std::env::var(ENV_AGENT_CODEX_APPROVAL)
            .unwrap_or_else(|_| DEFAULT_CODEX_APPROVAL.to_string());
        Self {
            model,
            sandbox,
            approval,
        }
    }
}

pub const CODEX_BIN: &str = "codex";
const DEFAULT_CODEX_MODEL: &str = "gpt-5.4";
const DEFAULT_CODEX_SANDBOX: &str = "danger-full-access";
const DEFAULT_CODEX_APPROVAL: &str = "never";
const ENV_AGENT_CODEX_MODEL: &str = "AGENT_CODEX_MODEL";
const ENV_AGENT_CODEX_SANDBOX: &str = "AGENT_CODEX_SANDBOX";
const ENV_AGENT_CODEX_APPROVAL: &str = "AGENT_CODEX_APPROVAL";

pub(super) fn required_deps() -> Vec<&'static str> {
    vec![CODEX_BIN, crate::consts::TMUX_BIN]
}

/// Build the shell command string tmux will run inside the detached
/// session for a codex-resident agent. `codex -m <model> -s <sandbox>
/// -a <approval> --no-alt-screen "$(cat "$NETSKY_PROMPT_FILE")"`.
///
/// The mcp_config path is accepted for signature parity with the claude
/// runtime but ignored: codex has its own `~/.codex/config.toml` + `-c`
/// override surface and the netsky agent-bus MCP source is not yet
/// wired through the codex MCP client (tracked in
/// `briefs/codex-side-by-side.md`). Resident codex reaches the bus via
/// `netsky channel drain|send` shell-outs instead.
///
/// The `startup` arg is ignored HERE because codex's positional slot
/// is already spoken for by the rendered prompt. Startup (`/up`) lands
/// as a follow-on user turn via [`paste_startup`], invoked by the
/// runtime's `post_spawn` hook after the tmux session exists. See this
/// module's top-level doc for the rationale.
pub(super) fn build_command(
    _agent: AgentId,
    cfg: &CodexConfig,
    _mcp_config: &Path,
    _startup: &str,
) -> String {
    let mut parts: Vec<String> = Vec::with_capacity(10);
    parts.push(CODEX_BIN.to_string());

    parts.push("-m".to_string());
    parts.push(shell_escape(&cfg.model));

    parts.push("-s".to_string());
    parts.push(shell_escape(&cfg.sandbox));

    parts.push("-a".to_string());
    parts.push(shell_escape(&cfg.approval));

    // Inline-mode TUI: preserves scrollback so `tmux capture-pane`
    // returns the full session content (alt-screen buffers would be
    // blank outside the alt screen). Critical for the smoke test +
    // post-mortem debugging.
    parts.push("--no-alt-screen".to_string());

    // Initial prompt read at exec time from the file the spawner wrote.
    // Same trick the claude path uses for --append-system-prompt.
    parts.push(format!("\"$(cat \"${ENV_NETSKY_PROMPT_FILE}\")\""));

    parts.join(" ")
}

/// Minimal pane-IO abstraction used by [`paste_startup`]. A live impl
/// ([`TmuxPaneIo`]) shells out to `tmux`; tests provide a mock that
/// records calls and simulates capture output.
pub trait PaneIo {
    fn send_text(&self, session: &str, text: &str) -> Result<()>;
    fn send_enter(&self, session: &str) -> Result<()>;
    fn capture(&self, session: &str, lines: Option<usize>) -> Result<String>;
}

/// Live `PaneIo` that shells out to the `tmux` CLI. Used by the
/// production `post_spawn` hook.
pub struct TmuxPaneIo;

impl PaneIo for TmuxPaneIo {
    fn send_text(&self, session: &str, text: &str) -> Result<()> {
        // No `-l` (literal) flag: codex's TUI composer treats bracketed-paste
        // input differently from keystrokes and drops Enter-to-submit when
        // bracketed. Plain keystroke-by-keystroke matches the smoke test.
        let status = Command::new(TMUX_BIN)
            .args(["send-keys", "-t", session, text])
            .status()?;
        if !status.success() {
            return Err(Error::Tmux(format!("send-keys text to '{session}' failed")));
        }
        Ok(())
    }

    fn send_enter(&self, session: &str) -> Result<()> {
        // Codex's composer treats C-m (Enter alone) as submit; Shift+Enter
        // as newline. We want submit.
        let status = Command::new(TMUX_BIN)
            .args(["send-keys", "-t", session, "C-m"])
            .status()?;
        if !status.success() {
            return Err(Error::Tmux(format!("send-keys C-m to '{session}' failed")));
        }
        Ok(())
    }

    fn capture(&self, session: &str, lines: Option<usize>) -> Result<String> {
        let start;
        let mut args: Vec<&str> = vec!["capture-pane", "-t", session, "-p"];
        if let Some(n) = lines {
            start = format!("-{n}");
            args.extend(["-S", &start]);
        }
        let out = Command::new(TMUX_BIN).args(&args).output()?;
        if !out.status.success() {
            return Err(Error::Tmux(format!("capture-pane '{session}' failed")));
        }
        Ok(String::from_utf8_lossy(&out.stdout).into_owned())
    }
}

const CODEX_TRUST_DIALOG_PROBE: &str = "Do you trust the contents";
const PASTE_ATTEMPTS: u32 = 6;

/// Paste `startup` into a live codex tmux pane as a follow-on user
/// turn. Dismisses codex's per-cwd "Do you trust" dialog if present,
/// then retries send-text + send-Enter until `capture` echoes the text
/// or `attempts` is exhausted. Matches the retry pattern in
/// `bin/test-resident-codex` (the smoke test paste of `netsky channel drain`).
///
/// Used by [`post_spawn`] to close the resident-codex /up parity gap
/// (round-1 review B1): the rendered netsky prompt already occupies
/// codex's positional `[PROMPT]`, so startup must be delivered as a
/// second user turn rather than a CLI flag.
pub fn paste_startup<I: PaneIo>(io: &I, session: &str, startup: &str, attempts: u32) -> Result<()> {
    let text = startup.trim().to_string();
    if text.is_empty() {
        return Ok(());
    }

    // Pre-check: codex shows a per-cwd "Do you trust" dialog on first
    // spawn. Send Enter to accept if present. The claude path handles
    // its own TOS dialog via `spawn::dismiss_tos`; codex's dialog is
    // handled here so the paste that follows lands on the composer, not
    // the trust prompt.
    for _ in 0..3 {
        let pane = io.capture(session, None).unwrap_or_default();
        if pane.contains(CODEX_TRUST_DIALOG_PROBE) {
            io.send_enter(session)?;
            delay(Duration::from_secs(1));
        } else {
            break;
        }
    }

    for _ in 0..attempts {
        io.send_text(session, &text)?;
        delay(Duration::from_millis(500));
        io.send_enter(session)?;
        delay(Duration::from_millis(1500));
        let pane = io.capture(session, Some(2000)).unwrap_or_default();
        if pane.contains(&text) {
            return Ok(());
        }
    }
    Err(Error::Tmux(format!(
        "codex pane '{session}' never echoed startup prompt within {attempts} paste attempts"
    )))
}

/// Runtime-dispatched post-spawn hook for codex: pastes startup into
/// the fresh tmux pane as a follow-on user turn. The caller
/// (`spawn::spawn`) invokes this immediately after
/// `tmux::new_session_detached` returns.
pub(super) fn post_spawn(session: &str, startup: &str) -> Result<()> {
    paste_startup(&TmuxPaneIo, session, startup, PASTE_ATTEMPTS)
}

// Real sleep in production; no-op in tests. Tests run the state
// machine without wall-clock delays.
#[cfg(not(test))]
fn delay(d: Duration) {
    std::thread::sleep(d);
}
#[cfg(test)]
fn delay(_d: Duration) {}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn default_config_picks_documented_defaults() {
        // Lock in against env bleed from the test runner's shell.
        unsafe {
            std::env::remove_var(ENV_AGENT_CODEX_MODEL);
            std::env::remove_var(ENV_AGENT_CODEX_SANDBOX);
            std::env::remove_var(ENV_AGENT_CODEX_APPROVAL);
        }
        let cfg = CodexConfig::defaults_for();
        assert_eq!(cfg.model, DEFAULT_CODEX_MODEL);
        assert_eq!(cfg.sandbox, DEFAULT_CODEX_SANDBOX);
        assert_eq!(cfg.approval, DEFAULT_CODEX_APPROVAL);
    }

    #[test]
    fn cmd_for_clone_invokes_codex_with_prompt_file_cat() {
        let cfg = CodexConfig {
            model: "gpt-5.4".to_string(),
            sandbox: DEFAULT_CODEX_SANDBOX.to_string(),
            approval: DEFAULT_CODEX_APPROVAL.to_string(),
        };
        let cmd = build_command(
            AgentId::Clone(42),
            &cfg,
            Path::new("/tmp/mcp-config.json"),
            "/up",
        );
        assert!(cmd.starts_with("codex "), "unexpected prefix: {cmd}");
        assert!(cmd.contains("'gpt-5.4'"));
        assert!(cmd.contains("-s 'danger-full-access'"));
        assert!(cmd.contains("-a 'never'"));
        assert!(cmd.contains("--no-alt-screen"));
        assert!(
            cmd.contains("$(cat \"$NETSKY_PROMPT_FILE\")"),
            "cmd must read prompt from NETSKY_PROMPT_FILE: {cmd}"
        );
    }

    #[test]
    fn cmd_shell_escapes_injection_attempts() {
        let cfg = CodexConfig {
            model: "gpt;touch /tmp/pwned".to_string(),
            sandbox: DEFAULT_CODEX_SANDBOX.to_string(),
            approval: DEFAULT_CODEX_APPROVAL.to_string(),
        };
        let cmd = build_command(
            AgentId::Clone(1),
            &cfg,
            Path::new("/tmp/mcp-config.json"),
            "",
        );
        assert!(
            cmd.contains("'gpt;touch /tmp/pwned'"),
            "model not shell-escaped: {cmd}"
        );
        assert!(
            !cmd.contains(" gpt;touch "),
            "model leaked unescaped: {cmd}"
        );
    }

    // ----- paste_startup tests: mock PaneIo, no real tmux touched. -----

    use std::cell::RefCell;

    /// Mock PaneIo that appends every `send_text` call to a fake pane
    /// buffer, so the first capture after the first paste contains the
    /// pasted text (happy path). Events are recorded for sequence
    /// assertions.
    struct EchoingPaneIo {
        events: RefCell<Vec<String>>,
        pane: RefCell<String>,
    }

    impl EchoingPaneIo {
        fn new() -> Self {
            Self {
                events: RefCell::new(Vec::new()),
                pane: RefCell::new(String::new()),
            }
        }
    }

    impl PaneIo for EchoingPaneIo {
        fn send_text(&self, session: &str, text: &str) -> Result<()> {
            self.events
                .borrow_mut()
                .push(format!("text:{session}:{text}"));
            self.pane.borrow_mut().push_str(text);
            Ok(())
        }
        fn send_enter(&self, session: &str) -> Result<()> {
            self.events.borrow_mut().push(format!("enter:{session}"));
            Ok(())
        }
        fn capture(&self, session: &str, _lines: Option<usize>) -> Result<String> {
            self.events.borrow_mut().push(format!("capture:{session}"));
            Ok(self.pane.borrow().clone())
        }
    }

    #[test]
    fn paste_startup_sends_text_then_enter_and_returns_on_echo() {
        let io = EchoingPaneIo::new();
        paste_startup(&io, "agent998", "/up", 3).expect("paste should succeed");
        let events = io.events.borrow();
        // First attempt should paste the text, then press Enter, then
        // the capture that follows echoes it back.
        let text_idx = events
            .iter()
            .position(|e| e == "text:agent998:/up")
            .expect("text event missing");
        let enter_idx = events
            .iter()
            .skip(text_idx)
            .position(|e| e == "enter:agent998")
            .expect("enter event after text missing");
        assert!(enter_idx > 0, "enter must follow text: {events:?}");
    }

    #[test]
    fn paste_startup_errors_when_pane_never_echoes() {
        struct SilentPaneIo;
        impl PaneIo for SilentPaneIo {
            fn send_text(&self, _: &str, _: &str) -> Result<()> {
                Ok(())
            }
            fn send_enter(&self, _: &str) -> Result<()> {
                Ok(())
            }
            fn capture(&self, _: &str, _: Option<usize>) -> Result<String> {
                Ok(String::new())
            }
        }
        let err = paste_startup(&SilentPaneIo, "agent998", "/up", 2)
            .expect_err("silent pane must yield an error");
        match err {
            Error::Tmux(msg) => {
                assert!(msg.contains("never echoed"), "unexpected tmux error: {msg}")
            }
            other => panic!("expected Error::Tmux, got {other:?}"),
        }
    }

    #[test]
    fn paste_startup_dismisses_trust_dialog_before_pasting() {
        struct TrustDialogIo {
            captures_seen: RefCell<u32>,
            events: RefCell<Vec<String>>,
        }
        impl PaneIo for TrustDialogIo {
            fn send_text(&self, _: &str, text: &str) -> Result<()> {
                self.events.borrow_mut().push(format!("text:{text}"));
                Ok(())
            }
            fn send_enter(&self, _: &str) -> Result<()> {
                self.events.borrow_mut().push("enter".to_string());
                Ok(())
            }
            fn capture(&self, _: &str, _: Option<usize>) -> Result<String> {
                let mut n = self.captures_seen.borrow_mut();
                *n += 1;
                // First capture: dialog up. Second+ captures: text
                // echoed (so the main paste loop can succeed).
                if *n == 1 {
                    Ok("Do you trust the contents of this directory?".to_string())
                } else {
                    Ok("> /up".to_string())
                }
            }
        }
        let io = TrustDialogIo {
            captures_seen: RefCell::new(0),
            events: RefCell::new(Vec::new()),
        };
        paste_startup(&io, "agent0", "/up", 3).expect("paste should succeed");
        let events = io.events.borrow();
        // First event must be an Enter (trust-dialog dismissal), before
        // any text paste.
        assert_eq!(
            events.first().map(String::as_str),
            Some("enter"),
            "first event should dismiss trust dialog: {events:?}"
        );
        assert!(
            events.iter().any(|e| e == "text:/up"),
            "startup text was never sent: {events:?}"
        );
    }

    #[test]
    fn paste_startup_noop_on_empty_startup() {
        let io = EchoingPaneIo::new();
        paste_startup(&io, "agent998", "", 3).expect("empty startup should no-op");
        assert!(
            io.events.borrow().is_empty(),
            "empty startup must touch no pane: {:?}",
            io.events.borrow()
        );
    }
}