agent_abstraction/run.rs
1//! Spawning an agent and turning its output into events and an outcome.
2//!
3//! Two entry points over the same machinery:
4//! - [`run`] waits and hands back the finished [`Outcome`].
5//! - [`stream`] hands back a [`Run`] that yields [`Event`]s as they arrive, for
6//! a UI that shows work in progress.
7//!
8//! Both read stdout and stderr concurrently. Draining only one would deadlock
9//! the moment the other filled its pipe buffer, which for a chatty agent is a
10//! matter of seconds.
11
12use std::collections::{HashMap, VecDeque};
13use std::process::Stdio;
14use std::time::Duration;
15
16use tokio::io::{AsyncReadExt, AsyncWriteExt, BufReader};
17use tokio::process::{Child, Command};
18use tokio::sync::mpsc;
19
20use crate::agent::{Continue, EnvPolicy};
21use crate::error::{Error, Result};
22use crate::event::{Event, MAX_LINE, Parser, Terminal, append_capped};
23use crate::outcome::{Outcome, Stop};
24use crate::proc::{kill_group_by_pid, kill_process_group};
25use crate::request::Request;
26
27/// Read one line, giving up on a line that never ends.
28///
29/// `AsyncBufReadExt::lines` buffers until a newline arrives, so a stream that
30/// emits megabytes without one exhausts memory before any total cap applies.
31/// This reads a bounded amount and, past the limit, returns what it has and
32/// discards the remainder of that line. Returns `None` at end of input.
33async fn read_bounded_line<R>(reader: &mut R, buf: &mut String) -> std::io::Result<Option<bool>>
34where
35 R: tokio::io::AsyncBufRead + Unpin,
36{
37 buf.clear();
38 let mut bytes = Vec::new();
39 let mut truncated = false;
40 loop {
41 let mut byte = [0u8; 1];
42 match reader.read(&mut byte).await? {
43 // End of input: a trailing fragment still counts as a line.
44 0 => {
45 if bytes.is_empty() {
46 return Ok(None);
47 }
48 break;
49 }
50 _ if byte[0] == b'\n' => break,
51 _ => {
52 if bytes.len() < MAX_LINE {
53 bytes.push(byte[0]);
54 } else {
55 // Keep draining to the newline so the pipe does not block,
56 // but stop accumulating.
57 truncated = true;
58 }
59 }
60 }
61 }
62 // Output is not guaranteed to be valid UTF-8, and one bad byte should not
63 // end a run.
64 buf.push_str(&String::from_utf8_lossy(&bytes));
65 Ok(Some(truncated))
66}
67
68/// Aborts a task when dropped.
69///
70/// The decision forwarder holds the child's stdin, so leaving it running past
71/// the run would keep a pipe open to a process that is gone.
72struct AbortOnDrop(tokio::task::JoinHandle<()>);
73
74impl Drop for AbortOnDrop {
75 fn drop(&mut self) {
76 self.0.abort();
77 }
78}
79
80/// How many decisions may queue on the way back to the agent.
81///
82/// Small on purpose: the agent asks one question at a time and waits, so a deep
83/// queue here would only mean answers piling up for questions nobody asked.
84const APPROVAL_BUFFER: usize = 8;
85
86/// How many events may queue before the producer waits for the consumer. Deep
87/// enough that a burst of tool events does not stall the agent, shallow enough
88/// that a consumer which stops reading does not grow without bound.
89const EVENT_BUFFER: usize = 256;
90
91/// A host action travelling back to an interactive agent.
92///
93/// Claude and Codex encode these differently, so the public handle preserves
94/// the intent and lets the selected transport serialize it at the boundary.
95#[derive(Debug)]
96enum Control {
97 Message {
98 body: String,
99 receipt: tokio::sync::oneshot::Sender<Result<()>>,
100 },
101 Approval {
102 id: String,
103 decision: crate::Decision,
104 },
105}
106
107/// A run in progress.
108///
109/// Yields events through [`Run::recv`] and settles into an [`Outcome`] through
110/// [`Run::finish`].
111///
112/// **Dropping a `Run` kills the agent.** That is the safe default for the hosts
113/// this crate targets: closing a window or cancelling a request should stop the
114/// work, not leave an agent running invisibly, spending quota and touching
115/// files with nobody watching. Call [`Run::detach`] when background execution is
116/// genuinely what you want.
117///
118/// On Unix, dropping **synchronously signals** the run's process group and then
119/// aborts the driver task. What it cannot do is *wait*: `Drop` cannot await, so
120/// it does not block until the child has exited or its readers have been
121/// joined. Use [`Run::cancel`] when you need to know the tree has actually gone
122/// before continuing, such as before touching the files it was working on. On
123/// Windows only the direct child is signalled.
124#[derive(Debug)]
125pub struct Run {
126 events: mpsc::Receiver<Event>,
127 /// A cloneable route back into the run. Kept separate from the event
128 /// receiver so a host can drain output while a delivery receipt is pending.
129 control: RunControl,
130 /// The typed command line, kept so both the plain and redacted views come
131 /// from the same source.
132 typed: Vec<crate::agent::Arg>,
133 /// The child's pid, so `Drop` can tear the group down itself rather than
134 /// depending on an aborted task being polled.
135 pid: Option<u32>,
136 /// Set by the driver once the child has been reaped, so `Drop` never
137 /// signals a pid the OS may since have handed to someone else.
138 reaped: std::sync::Arc<std::sync::atomic::AtomicBool>,
139 /// Dropping or firing this asks the driver to tear down in order. Held as
140 /// an `Option` so `detach` can discard it without signalling.
141 cancel: Option<tokio::sync::oneshot::Sender<()>>,
142 /// `None` only after [`Run::finish`], [`Run::cancel`] or [`Run::detach`]
143 /// has taken ownership, which is what stops `Drop` from aborting a run that
144 /// was already settled deliberately.
145 task: Option<tokio::task::JoinHandle<Result<Outcome>>>,
146 argv: Vec<String>,
147}
148
149/// A cloneable route for sending input to a live [`Run`].
150///
151/// Keep this beside the task that drains [`Run::recv`]. A delivery receipt may
152/// arrive after a burst of agent events, so awaiting [`Self::send`] on the same
153/// task that owns the mutable `Run` can apply backpressure to the event stream
154/// and prevent the receipt itself from being read. A separate control handle
155/// lets both directions make progress concurrently.
156#[derive(Clone, Debug)]
157pub struct RunControl {
158 agent: crate::Agent,
159 to_agent: Option<mpsc::Sender<Control>>,
160 bin: String,
161}
162
163/// Interactive transports acknowledge steer requests immediately. A missing
164/// receipt means the live process can no longer be trusted to accept input.
165const CONTROL_RECEIPT_TIMEOUT: Duration = Duration::from_secs(15);
166
167impl RunControl {
168 /// Send another message while the agent is still working.
169 ///
170 /// The future resolves only after the transport accepts the message. It is
171 /// safe to await from a separate task while the owner continues draining
172 /// [`Run::recv`].
173 ///
174 /// # Errors
175 ///
176 /// Returns [`Error::Unsupported`] when the run is not interactive, or
177 /// [`Error::Cancelled`] when the run finishes before delivery is confirmed.
178 /// Returns [`Error::ControlTimeout`] when the live transport stops
179 /// acknowledging input. The caller should stop that run before retrying on
180 /// a resumed session.
181 pub async fn send(&self, message: &str) -> Result<()> {
182 self.send_with_timeout(message, CONTROL_RECEIPT_TIMEOUT)
183 .await
184 }
185
186 async fn send_with_timeout(&self, message: &str, timeout: Duration) -> Result<()> {
187 let Some(channel) = &self.to_agent else {
188 return Err(Error::Unsupported {
189 agent: self.agent,
190 what: "sending a follow-up on a run that is not interactive",
191 });
192 };
193 let cancelled = || Error::Cancelled {
194 bin: self.bin.clone(),
195 };
196 let (receipt, delivered) = tokio::sync::oneshot::channel();
197 channel
198 .send(Control::Message {
199 body: message.to_string(),
200 receipt,
201 })
202 .await
203 .map_err(|_| cancelled())?;
204 match tokio::time::timeout(timeout, delivered).await {
205 Ok(receipt) => receipt.map_err(|_| cancelled())?,
206 Err(_) => Err(Error::ControlTimeout {
207 bin: self.bin.clone(),
208 timeout,
209 }),
210 }
211 }
212
213 /// Answer an [`Event::ApprovalRequest`] without borrowing the event stream.
214 ///
215 /// # Errors
216 ///
217 /// Returns [`Error::Unsupported`] when the run does not accept approvals,
218 /// or [`Error::Cancelled`] when the run finishes before the response is sent.
219 pub async fn respond(&self, id: &str, decision: &crate::Decision) -> Result<()> {
220 let Some(channel) = &self.to_agent else {
221 return Err(Error::Unsupported {
222 agent: self.agent,
223 what: "answering an approval on a run that did not request them",
224 });
225 };
226 channel
227 .send(Control::Approval {
228 id: id.to_string(),
229 decision: decision.clone(),
230 })
231 .await
232 .map_err(|_| Error::Cancelled {
233 bin: self.bin.clone(),
234 })
235 }
236}
237
238impl Run {
239 /// The next event, or `None` once the agent has finished producing them.
240 pub async fn recv(&mut self) -> Option<Event> {
241 self.events.recv().await
242 }
243
244 /// Clone the route used for follow-up messages and approval decisions.
245 ///
246 /// Use this when input and output must progress concurrently. The handle
247 /// does not keep the process alive after the `Run` settles or is dropped.
248 #[must_use]
249 pub fn control(&self) -> RunControl {
250 self.control.clone()
251 }
252
253 /// Send another message while the agent is still working.
254 ///
255 /// The whole point of [`crate::Request::interactive`]: a user who types a
256 /// correction mid-turn should not have to wait for the turn to finish.
257 ///
258 /// The agent takes it at its **next step boundary**, not mid-token, so an
259 /// answer already being written finishes first and a long tool-using task
260 /// changes course at its next step. Verified against claude 2.1.212 and
261 /// codex-cli 0.146.0.
262 ///
263 /// # Ordering and delivery acknowledgement
264 ///
265 /// The caller already knows what it sent, so the intended pattern is to
266 /// append the message to the transcript immediately, below the user's
267 /// previous one, and carry on. This deliberately does not ask the agent to
268 /// echo the message back for sequencing. The future resolves only after
269 /// the transport accepts the message: Codex has acknowledged `turn/steer`,
270 /// or Claude's input was written and flushed successfully. Rendering stays
271 /// immediate while a settle-race becomes an error the host can recover.
272 ///
273 /// # Errors
274 /// [`Error::Unsupported`] on a run that did not open the channel with
275 /// [`crate::Request::interactive`]. [`Error::Cancelled`] once the channel
276 /// has closed, which happens when the turn settles or the run is torn down:
277 /// **a message sent after the turn ends is too late** and belongs in a new
278 /// run resuming the session, so this reports it rather than dropping it.
279 pub async fn send(&self, message: &str) -> Result<()> {
280 self.control.send(message).await
281 }
282
283 /// Answer an [`Event::ApprovalRequest`].
284 ///
285 /// The agent is blocked until this is called, so a consumer that receives an
286 /// approval request and never responds stalls the run until its timeout.
287 ///
288 /// The id must be the one from the request. The agent ignores an answer
289 /// carrying any other id and keeps waiting, so a mismatch presents as a
290 /// hang rather than an error; this passes the id straight through and does
291 /// not invent one.
292 ///
293 /// # Errors
294 /// [`Error::Unsupported`] on a run that did not ask for approvals, since
295 /// there is no channel to answer on. [`Error::Cancelled`] if the run has
296 /// already finished or been torn down, which is the same reason a decision
297 /// can no longer be delivered.
298 pub async fn respond(&self, id: &str, decision: &crate::Decision) -> Result<()> {
299 self.control.respond(id, decision).await
300 }
301
302 /// The exact command line that was spawned.
303 ///
304 /// **This contains the prompt and any session id.** Treat it as sensitive:
305 /// logging it verbatim puts user content into your logs. Use
306 /// [`Run::redacted_argv`] for diagnostics.
307 #[must_use]
308 pub fn argv(&self) -> &[String] {
309 &self.argv
310 }
311
312 /// The command line with every non-public value replaced by a placeholder.
313 ///
314 /// Prompts, system prompts, session ids and anything from
315 /// [`crate::Request::unchecked_args`] are removed; flag names are kept so
316 /// the command stays recognisable. Sensitivity is recorded where each
317 /// argument is built rather than inferred from the finished line, so a
318 /// bare positional prompt or an opaque raw argument is covered too.
319 #[must_use]
320 pub fn redacted_argv(&self) -> Vec<String> {
321 redact(&self.typed)
322 }
323
324 /// Wait for the run to finish.
325 ///
326 /// Drains any events still queued, so a caller that only wants the result
327 /// can call this without having consumed the stream.
328 ///
329 /// # Errors
330 /// Whatever the run failed with. See [`Error`].
331 pub async fn finish(mut self) -> Result<Outcome> {
332 // The driver owns teardown from here; `Drop` must not also fire.
333 self.pid = None;
334 while self.events.recv().await.is_some() {}
335 // Taking the handle disarms the `Drop` guard: this run is settling
336 // normally, not being abandoned.
337 let Some(task) = self.task.take() else {
338 unreachable!("the handle is only taken by a consuming method")
339 };
340 match task.await {
341 Ok(result) => result,
342 // The driver task panicked or was cancelled. The process itself
343 // started fine, so this is not a spawn failure and must not claim
344 // to be one.
345 Err(join) => Err(Error::Interrupted {
346 bin: self.argv.first().cloned().unwrap_or_default(),
347 detail: if join.is_panic() {
348 "the driver task panicked".into()
349 } else {
350 "the driver task was cancelled".into()
351 },
352 }),
353 }
354 }
355
356 /// Stop the run and wait until the agent is actually gone.
357 ///
358 /// Cooperative rather than an abort: the driver is asked to stop, signals
359 /// the process group, reaps the child and joins its readers, and only then
360 /// does this return. So when it returns the tree really has exited, which
361 /// matters if the next thing you do touches the files it was working on.
362 ///
363 /// Returns the partial [`Outcome`] if the run happened to finish first,
364 /// otherwise [`Error::Cancelled`].
365 ///
366 /// # Errors
367 /// [`Error::Cancelled`] in the normal case, or whatever the run failed with
368 /// if it failed before the request arrived.
369 pub async fn cancel(mut self) -> Result<Outcome> {
370 // The driver tears down cooperatively and this awaits it, so `Drop`
371 // must not race that with a kill of its own.
372 self.pid = None;
373 // Dropping the sender is itself the signal, so this cannot fail in a
374 // way that leaves the driver waiting.
375 drop(self.cancel.take());
376 let Some(task) = self.task.take() else {
377 unreachable!("the handle is only taken by a consuming method")
378 };
379 match task.await {
380 Ok(result) => result,
381 Err(join) => Err(Error::Interrupted {
382 bin: self.argv.first().cloned().unwrap_or_default(),
383 detail: if join.is_panic() {
384 "the driver task panicked".into()
385 } else {
386 "the driver task was cancelled".into()
387 },
388 }),
389 }
390 }
391
392 /// Let the run continue after this handle goes away.
393 ///
394 /// The opposite of the default. Nothing can observe or stop the agent
395 /// afterwards, so reach for this only when an unsupervised background run
396 /// is genuinely intended.
397 pub fn detach(mut self) {
398 // Disarm `Drop` before it runs, or detaching would immediately kill the
399 // run it exists to keep alive.
400 self.pid = None;
401 // Leak the cancel signal rather than dropping it: a dropped sender is
402 // read by the driver as "stop", which is the opposite of detaching.
403 if let Some(cancel) = self.cancel.take() {
404 std::mem::forget(cancel);
405 }
406 // Dropping the handle without aborting is what detaches a tokio task.
407 drop(self.task.take());
408 }
409}
410
411impl Drop for Run {
412 fn drop(&mut self) {
413 // Abandoned rather than finished, cancelled or detached.
414 //
415 // Kill the group here, directly. Signalling the driver and aborting it
416 // is not enough on its own: that leaves teardown waiting on the runtime
417 // to poll the aborted task so its guard runs, and a dropped `Run` was
418 // observed leaving grandchildren alive and sleeping on Linux while the
419 // same teardown worked from `cancel`. `Drop` cannot await, so it does
420 // the one thing it can do synchronously.
421 if let Some(pid) = self.pid
422 && !self.reaped.load(std::sync::atomic::Ordering::SeqCst)
423 {
424 kill_group_by_pid(pid);
425 }
426 drop(self.cancel.take());
427 if let Some(task) = self.task.take() {
428 task.abort();
429 }
430 }
431}
432
433/// Placeholder substituted for a sensitive argv value.
434const REDACTED: &str = "<redacted>";
435
436/// Render a typed command line for logging, keeping flag names and replacing
437/// every value that is not `Public`.
438///
439/// Derived from the sensitivity recorded where each argument was built, so it
440/// cannot miss a case the way matching on flag names and positions can.
441fn redact(argv: &[crate::agent::Arg]) -> Vec<String> {
442 use crate::agent::Sensitivity;
443
444 argv.iter()
445 .map(|arg| match arg.sensitivity {
446 Sensitivity::Public => arg.value.clone(),
447 _ => REDACTED.to_string(),
448 })
449 .collect()
450}
451
452/// Run `request` to completion, discarding the intermediate events.
453///
454/// # Errors
455/// See [`Error`]; notably [`Error::NotInstalled`], [`Error::Timeout`],
456/// [`Error::RateLimited`] and [`Error::Failed`].
457///
458/// [`Error::Unsupported`] for a request that asked for approvals: this entry
459/// point discards events, so an approval request would reach nobody and the run
460/// would sit blocked until its timeout. Use [`stream`] instead.
461pub async fn run(request: &Request) -> Result<Outcome> {
462 if request.plan().approvals {
463 return Err(Error::Unsupported {
464 agent: request.agent,
465 what: "approvals on a run whose events are discarded; use `stream`",
466 });
467 }
468 stream(request)?.finish().await
469}
470
471/// Start `request`, returning a handle that streams its events.
472///
473/// Returns as soon as the child is spawned; the work proceeds on a task.
474///
475/// # Errors
476/// [`Error::NotInstalled`] if the binary is missing, [`Error::Unsupported`] if
477/// the agent cannot honour the request, or [`Error::Spawn`] on an OS failure.
478#[allow(
479 clippy::too_many_lines,
480 reason = "one spawn boundary keeps command posture, pipes, process group, and driver selection together"
481)]
482pub fn stream(request: &Request) -> Result<Run> {
483 // `tokio::spawn` panics outside a runtime. A fallible signature must not
484 // hide that, so the context is checked and reported as an ordinary error.
485 let runtime = tokio::runtime::Handle::try_current().map_err(|_| Error::NoRuntime)?;
486
487 let mut request = request.clone();
488 let session_lease = if let Some(binding) = &request.binding {
489 let lease = binding.store.lease(&binding.project, &binding.name)?;
490 // `Request::session` may have been built while a prior run still held
491 // the lease. Re-plan now, under the lease, so retrying that same request
492 // resumes the binding the prior run committed rather than starting from
493 // the stale token it observed earlier.
494 let (phase, cont) =
495 binding
496 .store
497 .plan(request.agent, &binding.project, &binding.name, binding.fork)?;
498 request.cont = cont;
499 if let Some(binding) = &mut request.binding {
500 binding.phase = phase;
501 }
502 Some(lease)
503 } else {
504 None
505 };
506
507 let initial_plan = request.plan();
508 let codex_app_server =
509 request.agent == crate::Agent::Codex && (initial_plan.duplex || initial_plan.approvals);
510
511 // Written before the argv is built, because the argv has to name it. The
512 // app-server protocol accepts the schema inline instead.
513 let schema_file = match (&request.schema, request.agent.caps().schema) {
514 (Some(_), _) if codex_app_server => None,
515 (Some(schema), crate::agent::SchemaSupport::File) => {
516 Some(SchemaFile::write(schema).map_err(|source| Error::Spawn {
517 bin: request.agent.bin().to_string(),
518 source,
519 })?)
520 }
521 _ => None,
522 };
523 if let Some(file) = &schema_file {
524 request.schema_file = Some(file.0.display().to_string());
525 }
526 let request = &request;
527
528 let plan = request.plan();
529 let typed = request.typed_argv()?;
530 let argv: Vec<String> = typed.iter().map(|a| a.value.clone()).collect();
531
532 let mut command = Command::new(&argv[0]);
533 command
534 .args(&argv[1..])
535 .stdin(if plan.stdin_prompt || plan.duplex || plan.approvals {
536 // An interactive run needs stdin for the whole turn, not just to
537 // deliver a prompt: it is the channel follow-up messages and
538 // approval decisions travel back on.
539 Stdio::piped()
540 } else {
541 // Close stdin so an agent that would otherwise wait on it exits
542 // instead of hanging forever with nothing to read.
543 Stdio::null()
544 })
545 .stdout(Stdio::piped())
546 .stderr(Stdio::piped())
547 // Without this a killed run can leave the child alive holding the pipes.
548 .kill_on_drop(true);
549 if let Some(cwd) = &request.cwd {
550 command.current_dir(cwd);
551 }
552 // Narrow the environment first, then apply explicit variables, so an
553 // explicit `env()` always wins over the policy.
554 match &request.env_policy {
555 EnvPolicy::Inherit => {}
556 EnvPolicy::Minimal => {
557 command.env_clear();
558 inherit_named(&mut command, &request.agent.essential_env());
559 }
560 EnvPolicy::Only(names) => {
561 command.env_clear();
562 inherit_named(&mut command, names);
563 }
564 }
565 for (key, value) in &request.env {
566 command.env(key, value);
567 }
568
569 // Applied last so the dedicated `thinking(false)` switch is authoritative
570 // over the general env map. Only Claude has a lever, delivered as
571 // `MAX_THINKING_TOKENS=0`. See `Agent::thinking_env`.
572 if let Some((key, value)) = request.agent.thinking_env(request.thinking) {
573 command.env(key, value);
574 }
575
576 // Put the agent in its own process group so the whole tree can be signalled
577 // together. Killing only the CLI leaves the commands *it* spawned running:
578 // a build, a test run, a server, still holding files and credentials after
579 // the run is supposedly over.
580 // 0 means "make this child its own group leader". `tokio::process::Command`
581 // exposes this directly on unix.
582 #[cfg(unix)]
583 command.process_group(0);
584
585 // Reserve an assigned session id before the child exists. Doing it inside
586 // the driver leaves a window where a spawn that half-succeeds loses the
587 // binding, and this is the id the caller may already be showing in a UI.
588 if let Some(token) = preassigned_token(request) {
589 persist_session(request, &token)?;
590 }
591
592 let child = command.spawn().map_err(|source| {
593 // A missing binary is the common case and deserves an actionable error
594 // with an install hint. Reading it off the spawn avoids resolving PATH
595 // twice, and with it the window where the resolved path is replaced
596 // between the check and the exec.
597 if source.kind() == std::io::ErrorKind::NotFound {
598 Error::NotInstalled {
599 agent: request.agent,
600 bin: plan.bin.clone(),
601 hint: request.agent.install_hint(),
602 }
603 } else {
604 Error::Spawn {
605 bin: plan.bin.clone(),
606 source,
607 }
608 }
609 })?;
610
611 let request_agent = request.agent;
612 let pid = child.id();
613 let (tx, rx) = mpsc::channel(EVENT_BUFFER);
614 // Only created for an approvals run, so `respond` can tell "no channel" from
615 // "channel closed" and refuse the first rather than hanging on it.
616 let (decisions_tx, decisions_rx) = if plan.duplex || plan.approvals {
617 let (tx, rx) = mpsc::channel::<Control>(APPROVAL_BUFFER);
618 (Some(tx), Some(rx))
619 } else {
620 (None, None)
621 };
622 let (cancel_tx, cancel_rx) = tokio::sync::oneshot::channel();
623 let reaped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
624 let reaped_for_task = std::sync::Arc::clone(&reaped);
625 let request = request.clone();
626 let task = runtime.spawn(async move {
627 // Held across the whole read-run-commit cycle. It deliberately lives in
628 // the driver task rather than `Run`, so `detach` keeps the session
629 // reserved until the background agent actually exits.
630 let _session_lease = session_lease;
631 // Moved in so the file outlives the run and is removed with it.
632 let _schema_file = schema_file;
633 if codex_app_server {
634 drive_codex_app_server(child, request, tx, cancel_rx, reaped_for_task, decisions_rx)
635 .await
636 } else {
637 drive(child, request, tx, cancel_rx, reaped_for_task, decisions_rx).await
638 }
639 });
640 Ok(Run {
641 events: rx,
642 control: RunControl {
643 agent: request_agent,
644 to_agent: decisions_tx,
645 bin: argv.first().cloned().unwrap_or_default(),
646 },
647 typed,
648 pid,
649 reaped,
650 cancel: Some(cancel_tx),
651 task: Some(task),
652 argv,
653 })
654}
655
656/// Copy the named variables from this process into `command`, skipping any that
657/// are unset so nothing is invented.
658fn inherit_named<S: AsRef<str>>(command: &mut Command, names: &[S]) {
659 for name in names {
660 if let Some(value) = std::env::var_os(name.as_ref()) {
661 command.env(name.as_ref(), value);
662 }
663 }
664}
665
666/// A schema file written for one run, removed when the run ends.
667///
668/// Codex reads its schema from disk, so the file has to outlive the spawn and
669/// not outlive the process. Tying it to a guard means every exit path removes
670/// it, including a cancel or a timeout, without each one remembering.
671struct SchemaFile(std::path::PathBuf);
672
673impl SchemaFile {
674 /// Write `schema` somewhere the agent can read it.
675 fn write(schema: &str) -> std::io::Result<SchemaFile> {
676 use std::io::Write as _;
677 use std::sync::atomic::{AtomicU64, Ordering};
678 static COUNTER: AtomicU64 = AtomicU64::new(0);
679
680 let path = std::env::temp_dir().join(format!(
681 "agent-abstraction-schema-{}-{}.json",
682 std::process::id(),
683 COUNTER.fetch_add(1, Ordering::Relaxed)
684 ));
685 let mut options = std::fs::OpenOptions::new();
686 options.write(true).create_new(true);
687 // A schema can encode what a caller is looking for, so it is no more
688 // public than the prompt.
689 #[cfg(unix)]
690 {
691 use std::os::unix::fs::OpenOptionsExt as _;
692 options.mode(0o600);
693 }
694 options.open(&path)?.write_all(schema.as_bytes())?;
695 Ok(SchemaFile(path))
696 }
697}
698
699impl Drop for SchemaFile {
700 fn drop(&mut self) {
701 let _ = std::fs::remove_file(&self.0);
702 }
703}
704
705/// Owns the child and tears down its whole process group when dropped.
706///
707/// `kill_on_drop` alone is not enough: it kills the CLI, leaving the commands
708/// *it* spawned running. Since aborting the driver task drops this guard, the
709/// same teardown covers cancellation, a dropped [`Run`] and a timeout, without
710/// each path having to remember to do it.
711struct ChildGuard {
712 child: Child,
713 /// Cleared once the child has been reaped, so a pid the OS may since have
714 /// recycled is never signalled.
715 armed: bool,
716}
717
718impl Drop for ChildGuard {
719 fn drop(&mut self) {
720 if self.armed {
721 kill_process_group(&self.child);
722 }
723 }
724}
725
726/// Feed the child, read both its pipes, and assemble the outcome.
727#[allow(
728 clippy::too_many_lines,
729 reason = "one linear lifecycle: feed, read, wait, classify. Splitting it \
730 would thread the child, parser, buffers and cancellation state \
731 through helpers and obscure the ordering that matters, such as \
732 killing the group before reaping."
733)]
734async fn drive(
735 child: Child,
736 request: Request,
737 events: mpsc::Sender<Event>,
738 cancel: tokio::sync::oneshot::Receiver<()>,
739 reaped: std::sync::Arc<std::sync::atomic::AtomicBool>,
740 decisions: Option<mpsc::Receiver<Control>>,
741) -> Result<Outcome> {
742 // From here on the child is owned by a guard, so every exit path from this
743 // task, including an abort, takes the process group with it.
744 let mut child = ChildGuard { child, armed: true };
745 let plan = request.plan();
746 let bin = plan.bin.clone();
747
748 // An approvals run owns stdin for the whole turn: the handshake and the
749 // prompt go out first, then it stays open carrying decisions until the run
750 // ends. Closing it after the prompt, as the plain piped path does, would
751 // take the answer channel with it.
752 let mut decision_task = None;
753 let mut close_stdin = None;
754 if plan.duplex || plan.approvals {
755 let Some(mut stdin) = child.child.stdin.take() else {
756 return Err(Error::Spawn {
757 bin: bin.clone(),
758 source: std::io::Error::other("stdin was not piped for an interactive run"),
759 });
760 };
761 let opening = format!(
762 "{}{}",
763 crate::approval::handshake(),
764 crate::approval::user_message(&request.agent.effective_prompt(&plan)),
765 );
766 stdin
767 .write_all(opening.as_bytes())
768 .await
769 .map_err(|source| Error::Spawn {
770 bin: bin.clone(),
771 source,
772 })?;
773 let _ = stdin.flush().await;
774 // Forwarding runs on its own task so a decision can be written while
775 // stdout is being read. It ends on whichever comes first: the channel
776 // closing, or the turn settling.
777 let (close_tx, mut close_rx) = tokio::sync::oneshot::channel::<()>();
778 close_stdin = Some(close_tx);
779 let decision_bin = bin.clone();
780 decision_task = decisions.map(|mut rx| {
781 tokio::spawn(async move {
782 loop {
783 tokio::select! {
784 biased;
785 // Once the terminal record has arrived there is no
786 // turn left to receive another message. Prioritizing
787 // closure makes a simultaneous late send fail instead
788 // of being reported as delivered to a finished turn.
789 _ = &mut close_rx => break,
790 reply = rx.recv() => {
791 let Some(control) = reply else { break };
792 match control {
793 Control::Message { body, receipt } => {
794 let wire = crate::approval::user_message(&body);
795 let result = async {
796 stdin.write_all(wire.as_bytes()).await?;
797 stdin.flush().await
798 }
799 .await
800 .map_err(|source| Error::Spawn {
801 bin: decision_bin.clone(),
802 source,
803 });
804 let failed = result.is_err();
805 let _ = receipt.send(result);
806 if failed {
807 break;
808 }
809 }
810 Control::Approval { id, decision } => {
811 let wire = decision.wire(&id);
812 if stdin.write_all(wire.as_bytes()).await.is_err() {
813 break;
814 }
815 let _ = stdin.flush().await;
816 }
817 }
818 }
819 }
820 }
821 drop(stdin);
822 })
823 });
824 }
825
826 // Deliver a piped prompt and close the pipe, or the agent waits on EOF.
827 if plan.stdin_prompt {
828 if let Some(mut stdin) = child.child.stdin.take() {
829 let prompt = request.agent.effective_prompt(&plan);
830 stdin
831 .write_all(prompt.as_bytes())
832 .await
833 .map_err(|source| Error::Spawn {
834 bin: bin.clone(),
835 source,
836 })?;
837 drop(stdin);
838 }
839 }
840
841 // Drain stderr on its own task: a full stderr pipe blocks the child even
842 // while stdout still has room.
843 // Aborted on every exit path from here, so a forwarder never survives the
844 // run it belongs to.
845 let _decision_guard = decision_task.map(AbortOnDrop);
846
847 let stderr = child.child.stderr.take();
848 let stderr_task = tokio::spawn(async move {
849 let mut buf = String::new();
850 if let Some(handle) = stderr {
851 let mut reader = BufReader::new(handle);
852 let mut line = String::new();
853 // Keep draining after the cap is hit: an undrained pipe blocks the
854 // child even though we no longer want the bytes.
855 while let Ok(Some(_)) = read_bounded_line(&mut reader, &mut line).await {
856 append_capped(&mut buf, &line);
857 }
858 }
859 buf
860 });
861
862 let stdout = child.child.stdout.take();
863 let mut parser = Parser::new(request.agent, plan.format);
864 // Raw stdout is retained only as a fallback answer for a run that exited
865 // cleanly without producing a structured one, and as evidence when
866 // classifying a failure. It is capped for the same reason as everything
867 // else here: an agent can stream for hours.
868 let mut raw = String::new();
869 // Tracks the first `Started`, so the binding is written once, and carries a
870 // store failure back out instead of discarding it.
871 let mut bound = false;
872 let mut persist_result: Result<()> = Ok(());
873
874 let read_stdout = async {
875 if let Some(handle) = stdout {
876 let mut reader = BufReader::new(handle);
877 let mut line = String::new();
878 while read_bounded_line(&mut reader, &mut line).await?.is_some() {
879 append_capped(&mut raw, &line);
880 let parsed = parser.push(&line);
881 // Close stdin as soon as the turn settles. Under stream-json
882 // input claude waits for another message otherwise, so the run
883 // would only end at its timeout even though the answer already
884 // arrived.
885 if parser.saw_terminal()
886 && let Some(close) = close_stdin.take()
887 {
888 let _ = close.send(());
889 }
890 for event in parsed {
891 // Bind a printed id the moment it appears rather than at the
892 // end. Codex announces its thread before answering, so a
893 // turn killed mid-answer stays resumable.
894 if let Event::Started { session, .. } = &event
895 && !bound
896 {
897 bound = true;
898 persist_result = persist_session(&request, session);
899 }
900 // A receiver that went away is not a failure: the run should
901 // still finish and produce its outcome.
902 if events.send(event).await.is_err() {
903 break;
904 }
905 }
906 }
907 }
908 Ok::<_, std::io::Error>(())
909 };
910
911 // Race three outcomes: the run finishing, the deadline, and a cancellation
912 // request. Reading and waiting are one future so a child that produces
913 // output forever is still bounded by the timeout.
914 let work = async {
915 read_stdout.await?;
916 child.child.wait().await
917 };
918 // A timeout is optional; `pending()` makes the un-timed case the same shape
919 // rather than duplicating the whole select.
920 let deadline = async {
921 match request.timeout {
922 Some(limit) => tokio::time::sleep(limit).await,
923 None => std::future::pending().await,
924 }
925 };
926
927 let status = tokio::select! {
928 // Biased so a finished run is reported as finished even if a deadline
929 // or cancellation lands in the same tick.
930 biased;
931 result = work => result,
932 () = deadline => {
933 // Order matters: signal the group *before* reaping. Reaping clears
934 // the child's pid, and the group kill needs that pid to target the
935 // group, so the other order silently leaves grandchildren running.
936 let partial = shut_down(&mut child, stderr_task).await;
937 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
938 return Err(Error::Timeout {
939 bin,
940 timeout: request.timeout.unwrap_or_default(),
941 partial: parser.finish().text,
942 })
943 .inspect_err(|_| drop(partial));
944 }
945 _ = cancel => {
946 // Cooperative teardown: the caller is waiting on this, so the tree
947 // is signalled, reaped and joined before returning.
948 shut_down(&mut child, stderr_task).await;
949 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
950 return Err(Error::Cancelled { bin });
951 }
952 }
953 .map_err(|source| Error::Spawn {
954 bin: bin.clone(),
955 source,
956 })?;
957
958 // The child has been reaped, so its pid must not be signalled again, by the
959 // guard here or by `Run::drop` racing this.
960 child.armed = false;
961 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
962
963 drop(events);
964 let stderr = stderr_task.await.unwrap_or_default();
965 let saw_structured = parser.saw_structured_record();
966 let saw_terminal = parser.saw_terminal_record();
967 let terminal = parser.finish();
968 let exit_code = status.code().unwrap_or(-1);
969
970 // Under a structured format, silently handing back raw stdout would turn a
971 // protocol failure into a plausible-looking answer. A run that recognized
972 // nothing, or never reached its terminal record, did not produce a result
973 // this crate can vouch for, so it is reported rather than papered over.
974 let structured = plan.format != crate::Format::Text;
975 if structured && exit_code == 0 {
976 if !saw_structured {
977 return Err(Error::Parse {
978 agent: request.agent,
979 detail: format!(
980 "no recognizable {} records in {} lines of output; the CLI's output shape has probably changed",
981 request.agent,
982 raw.lines().count()
983 ),
984 });
985 }
986 if !saw_terminal {
987 return Err(Error::Parse {
988 agent: request.agent,
989 detail: "the stream ended without its terminal record, so the turn did not complete"
990 .into(),
991 });
992 }
993 }
994
995 // Plain text has no structure to validate: the stream is the answer.
996 let mut terminal = terminal;
997 if terminal.text.is_empty() && !structured {
998 terminal.text = raw.trim().to_string();
999 }
1000
1001 // A provider refusal is not always an exit code. Claude can report a
1002 // blocking `rate_limit_event` and still exit 0, and the crate promises that
1003 // quota refusals surface as `Error::RateLimited`, so the terminal state is
1004 // checked regardless of how the process exited.
1005 let quota_blocked = terminal
1006 .rate_limit
1007 .as_ref()
1008 .is_some_and(crate::outcome::RateLimit::is_blocking);
1009 // An unauthenticated Claude run exits 0 and reports the problem in its
1010 // result text, so checking only the exit code would hand back a successful
1011 // Outcome whose answer is "Please run /login".
1012 //
1013 // Read from stderr and the agent's own prose rather than the raw stream, for
1014 // the reason `classify` does the same with quota wording: a phrase hunted
1015 // through structured output matches ids and field names, not statements.
1016 //
1017 // The answer is read at its opening only, by the same rule and the same
1018 // helper `classify_run` uses. This gate used to read `terminal.text` whole
1019 // while the classifier it guards read three lines, so the two could
1020 // disagree: a healthy answer that merely discussed logging in opened the
1021 // error path, no classifier would name it, and the turn fell out the far
1022 // end as `Error::Failed` with exit code 0 and nothing to report. One rule,
1023 // one helper, so that disagreement cannot exist.
1024 let unauthenticated =
1025 answer_reports_no_credentials(&terminal.text) || looks_unauthenticated(&stderr);
1026 // The agent saying its turn failed is as much a failure as a non-zero exit,
1027 // and Claude reports an unknown model exactly this way: exit 0, `is_error`
1028 // true, and the explanation where the answer would be.
1029 let turn_failed = terminal.stop == Stop::Error;
1030 if exit_code != 0 || quota_blocked || unauthenticated || turn_failed {
1031 let error = classify_run(request.agent, &bin, exit_code, &stderr, &raw, &terminal);
1032 /*
1033 * The backstop, and the reason a false positive can no longer cost an
1034 * answer.
1035 *
1036 * Everything above is a heuristic reading of text the agent wrote, and
1037 * a heuristic will be wrong eventually: these phrases are ordinary
1038 * English, and an agent asked about rate limits or logging in answers
1039 * in exactly the vocabulary that describes being rate limited or
1040 * logged out. What must never follow from being wrong is discarding a
1041 * finished answer.
1042 *
1043 * So a run whose process exited cleanly, whose terminal record says
1044 * the turn completed, and which carries no parsed quota block is only
1045 * ever failed by a classifier that can *name* the failure. A generic
1046 * `Failed` on that run is the classifiers disagreeing with the gate,
1047 * not evidence, and the answer stands.
1048 */
1049 let exited_clean = exit_code == 0 && !quota_blocked && !turn_failed;
1050 if !exited_clean || names_a_failure(&error) {
1051 return Err(error);
1052 }
1053 }
1054
1055 // A fork lands on a *new* id the agent only reveals at the end, so the name
1056 // has to be repointed once the run settles. Everything else was bound above.
1057 persist_result?;
1058 // Resolved before the terminal is consumed by the Outcome below.
1059 let structured = terminal.structured.clone().or_else(|| {
1060 request
1061 .schema
1062 .as_ref()
1063 .and_then(|_| serde_json::from_str(&terminal.text).ok())
1064 });
1065 if let Some(token) = &terminal.session
1066 && !bound
1067 {
1068 persist_session(&request, token)?;
1069 }
1070 Ok(Outcome {
1071 agent: request.agent,
1072 session: terminal.session,
1073 text: terminal.text,
1074 usage: terminal.usage,
1075 stop: terminal.stop,
1076 rate_limit: terminal.rate_limit,
1077 exit_code,
1078 stderr,
1079 unparsed: terminal.unparsed,
1080 first_unparsed: terminal.first_unparsed,
1081 // Claude reports the conforming value separately; Codex returns it as
1082 // the answer text, so that is parsed only when a schema was asked for.
1083 // Prose is never reinterpreted as data.
1084 structured,
1085 })
1086}
1087
1088/// Drive one interactive Codex turn over app-server's JSON-RPC transport.
1089///
1090/// Unlike `codex exec`, app-server remains alive after a turn completes. This
1091/// driver therefore treats `turn/completed` as the terminal record, closes the
1092/// protocol pipe, and reaps the service process itself.
1093#[allow(
1094 clippy::too_many_lines,
1095 reason = "one select loop owns the protocol, control channel, deadline, and child lifecycle"
1096)]
1097async fn drive_codex_app_server(
1098 child: Child,
1099 request: Request,
1100 events: mpsc::Sender<Event>,
1101 cancel: tokio::sync::oneshot::Receiver<()>,
1102 reaped: std::sync::Arc<std::sync::atomic::AtomicBool>,
1103 controls: Option<mpsc::Receiver<Control>>,
1104) -> Result<Outcome> {
1105 let mut child = ChildGuard { child, armed: true };
1106 let plan = request.plan();
1107 let bin = plan.bin.clone();
1108 let Some(mut stdin) = child.child.stdin.take() else {
1109 return Err(Error::Spawn {
1110 bin,
1111 source: std::io::Error::other("stdin was not piped for Codex app-server"),
1112 });
1113 };
1114 let Some(stdout) = child.child.stdout.take() else {
1115 return Err(Error::Spawn {
1116 bin,
1117 source: std::io::Error::other("stdout was not piped for Codex app-server"),
1118 });
1119 };
1120 let Some(mut controls) = controls else {
1121 return Err(Error::Spawn {
1122 bin,
1123 source: std::io::Error::other("Codex app-server has no host control channel"),
1124 });
1125 };
1126
1127 let stderr = child.child.stderr.take();
1128 let stderr_task = tokio::spawn(async move {
1129 let mut buf = String::new();
1130 if let Some(handle) = stderr {
1131 let mut reader = BufReader::new(handle);
1132 let mut line = String::new();
1133 while let Ok(Some(_)) = read_bounded_line(&mut reader, &mut line).await {
1134 append_capped(&mut buf, &line);
1135 }
1136 }
1137 buf
1138 });
1139
1140 let mut protocol = crate::codex_app_server::Protocol::new(request.clone());
1141 for opening in protocol.opening() {
1142 stdin
1143 .write_all(opening.as_bytes())
1144 .await
1145 .map_err(|source| Error::Spawn {
1146 bin: bin.clone(),
1147 source,
1148 })?;
1149 }
1150 stdin.flush().await.map_err(|source| Error::Spawn {
1151 bin: bin.clone(),
1152 source,
1153 })?;
1154
1155 let mut reader = BufReader::new(stdout);
1156 let mut line = String::new();
1157 let mut raw = String::new();
1158 let mut pending = VecDeque::new();
1159 let mut steer_receipts = HashMap::new();
1160 let mut bound = false;
1161 let mut persist_result: Result<()> = Ok(());
1162 let deadline = async {
1163 match request.timeout {
1164 Some(limit) => tokio::time::sleep(limit).await,
1165 None => std::future::pending().await,
1166 }
1167 };
1168 tokio::pin!(deadline);
1169 tokio::pin!(cancel);
1170
1171 while !protocol.finished {
1172 tokio::select! {
1173 biased;
1174 // User steering and approval answers outrank the agent's output.
1175 // app-server can keep stdout continuously ready with reasoning and
1176 // text deltas; reading it first in a biased select could starve a
1177 // correction precisely while Codex was busiest.
1178 control = controls.recv() => {
1179 let Some(control) = control else {
1180 continue;
1181 };
1182 pending.push_back(control);
1183 flush_codex_controls(
1184 &mut protocol,
1185 &mut pending,
1186 &mut steer_receipts,
1187 &mut stdin,
1188 &bin,
1189 ).await?;
1190 stdin.flush().await.map_err(|source| Error::Spawn {
1191 bin: bin.clone(), source
1192 })?;
1193 }
1194 record = read_bounded_line(&mut reader, &mut line) => {
1195 if record.map_err(|source| Error::Spawn { bin: bin.clone(), source })?.is_some() {
1196 append_capped(&mut raw, &line);
1197 if let Ok(value) = serde_json::from_str::<serde_json::Value>(&line) {
1198 let step = protocol.push(&value);
1199 settle_codex_steers(&mut steer_receipts, step.steer_responses, &bin);
1200 for event in step.events {
1201 if let Event::Started { session, .. } = &event
1202 && !bound
1203 {
1204 bound = true;
1205 persist_result = persist_session(&request, session);
1206 }
1207 let _ = events.send(event).await;
1208 }
1209 for write in step.writes {
1210 stdin.write_all(write.as_bytes()).await.map_err(|source| {
1211 Error::Spawn { bin: bin.clone(), source }
1212 })?;
1213 }
1214 flush_codex_controls(
1215 &mut protocol,
1216 &mut pending,
1217 &mut steer_receipts,
1218 &mut stdin,
1219 &bin,
1220 ).await?;
1221 stdin.flush().await.map_err(|source| Error::Spawn {
1222 bin: bin.clone(), source
1223 })?;
1224 } else {
1225 protocol.terminal.unparsed += 1;
1226 if protocol.terminal.first_unparsed.is_none() {
1227 protocol.terminal.first_unparsed = Some(line.clone());
1228 }
1229 }
1230 } else {
1231 protocol.failure.get_or_insert_with(|| {
1232 "app-server closed stdout before turn/completed".to_string()
1233 });
1234 protocol.finished = true;
1235 }
1236 }
1237 () = &mut deadline => {
1238 let partial = protocol.terminal.text.clone();
1239 shut_down(&mut child, stderr_task).await;
1240 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
1241 return Err(Error::Timeout {
1242 bin,
1243 timeout: request.timeout.unwrap_or_default(),
1244 partial,
1245 });
1246 }
1247 _ = &mut cancel => {
1248 shut_down(&mut child, stderr_task).await;
1249 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
1250 return Err(Error::Cancelled { bin });
1251 }
1252 }
1253 }
1254
1255 // app-server is a service rather than a one-shot process. EOF asks it to
1256 // stop cleanly; the short fallback prevents a completed turn from hanging
1257 // because a future CLI release keeps serving after its input closes.
1258 drop(stdin);
1259 if tokio::time::timeout(std::time::Duration::from_secs(2), child.child.wait())
1260 .await
1261 .is_err()
1262 {
1263 kill_process_group(&child.child);
1264 let _ = child.child.kill().await;
1265 }
1266 child.armed = false;
1267 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
1268 drop(events);
1269 let stderr = stderr_task.await.unwrap_or_default();
1270
1271 persist_result?;
1272 if let Some(detail) = protocol.failure {
1273 return Err(Error::Parse {
1274 agent: request.agent,
1275 detail,
1276 });
1277 }
1278
1279 let terminal = protocol.terminal;
1280 if terminal.stop == Stop::Error {
1281 return Err(classify_run(
1282 request.agent,
1283 &bin,
1284 0,
1285 &stderr,
1286 &raw,
1287 &terminal,
1288 ));
1289 }
1290 let structured = terminal.structured.clone().or_else(|| {
1291 request
1292 .schema
1293 .as_ref()
1294 .and_then(|_| serde_json::from_str(&terminal.text).ok())
1295 });
1296 Ok(Outcome {
1297 agent: request.agent,
1298 session: terminal.session,
1299 text: terminal.text,
1300 usage: terminal.usage,
1301 stop: terminal.stop,
1302 rate_limit: terminal.rate_limit,
1303 exit_code: 0,
1304 stderr,
1305 unparsed: terminal.unparsed,
1306 first_unparsed: terminal.first_unparsed,
1307 structured,
1308 })
1309}
1310
1311/// Write every control whose protocol ids are available, preserving earlier
1312/// messages until thread and turn startup have both completed.
1313async fn flush_codex_controls(
1314 protocol: &mut crate::codex_app_server::Protocol,
1315 pending: &mut VecDeque<Control>,
1316 steer_receipts: &mut HashMap<u64, tokio::sync::oneshot::Sender<Result<()>>>,
1317 stdin: &mut tokio::process::ChildStdin,
1318 bin: &str,
1319) -> Result<()> {
1320 let mut waiting = VecDeque::new();
1321 while let Some(control) = pending.pop_front() {
1322 let encoded = match control {
1323 Control::Message { body, receipt } => {
1324 if let Some(request) = protocol.steer(&body) {
1325 steer_receipts.insert(request.id, receipt);
1326 Some(request.wire)
1327 } else {
1328 waiting.push_back(Control::Message { body, receipt });
1329 None
1330 }
1331 }
1332 Control::Approval { id, decision } => {
1333 if let Some(encoded) = protocol.respond(&id, &decision) {
1334 Some(encoded)
1335 } else {
1336 waiting.push_back(Control::Approval { id, decision });
1337 None
1338 }
1339 }
1340 };
1341 if let Some(encoded) = encoded {
1342 stdin
1343 .write_all(encoded.as_bytes())
1344 .await
1345 .map_err(|source| Error::Spawn {
1346 bin: bin.to_string(),
1347 source,
1348 })?;
1349 }
1350 }
1351 pending.append(&mut waiting);
1352 Ok(())
1353}
1354
1355/// Resolve the public `Run::send` future only after app-server has answered
1356/// the matching `turn/steer` request. Dropping an unresolved sender when the
1357/// turn completes is deliberate: the caller receives `Error::Cancelled` and
1358/// can put the message into a fresh resumed turn.
1359fn settle_codex_steers(
1360 receipts: &mut HashMap<u64, tokio::sync::oneshot::Sender<Result<()>>>,
1361 responses: Vec<crate::codex_app_server::SteerResponse>,
1362 bin: &str,
1363) {
1364 for response in responses {
1365 let Some(receipt) = receipts.remove(&response.id) else {
1366 continue;
1367 };
1368 let result = response
1369 .result
1370 .map(|_| ())
1371 .map_err(|message| Error::AgentError {
1372 agent: crate::Agent::Codex,
1373 bin: bin.to_string(),
1374 status: None,
1375 message,
1376 });
1377 let _ = receipt.send(result);
1378 }
1379}
1380
1381/// Kill the process group, reap the child, and join the stderr reader.
1382///
1383/// The orderly teardown both cancellation and timeout share. Returns whatever
1384/// stderr had been captured, so a caller can still report why a run was stopped.
1385async fn shut_down(child: &mut ChildGuard, stderr_task: tokio::task::JoinHandle<String>) -> String {
1386 kill_process_group(&child.child);
1387 // Reap, so the caller is not left with a zombie once this returns.
1388 let _ = child.child.kill().await;
1389 child.armed = false;
1390 // The pipes are closed now that the child is gone, so this finishes
1391 // promptly rather than hanging the cancellation.
1392 stderr_task.await.unwrap_or_default()
1393}
1394
1395/// Turn a failure into the most specific error available, agent included so an
1396/// auth failure can carry the right login command.
1397fn classify_run(
1398 agent: crate::Agent,
1399 bin: &str,
1400 code: i32,
1401 stderr: &str,
1402 stdout: &str,
1403 terminal: &Terminal,
1404) -> Error {
1405 // Checked before quota and before a plain failure: a login problem is the
1406 // most specific reading of the output, and the only one a user can act on
1407 // directly.
1408 let named = |source: &str| Error::NotAuthenticated {
1409 agent,
1410 bin: bin.to_string(),
1411 message: first_meaningful_line(source).unwrap_or_default(),
1412 hint: agent.login_hint(),
1413 };
1414
1415 // The CLI's own channel, read whole: Copilot's notice runs to five lines.
1416 if looks_unauthenticated(stderr) {
1417 return named(stderr);
1418 }
1419
1420 /*
1421 * The agent's own answer, read only at the top.
1422 *
1423 * An agent with no credentials has nothing to say but the notice, so the
1424 * phrase is in its opening lines and the whole answer is those lines.
1425 * An agent that *writes about* logging in buries the same words in
1426 * paragraphs, and reading the whole answer counted that as a login
1427 * failure: a reply explaining why a publish had been refused mentioned
1428 * not being authenticated, so the run was reported as an auth error, the
1429 * hint told the user to run `/login`, and the answer itself was replaced
1430 * by the report. An agent's prose is not a diagnosis of the agent.
1431 */
1432 for source in [terminal.text.as_str(), stdout] {
1433 if answer_reports_no_credentials(source) {
1434 return named(&opening_lines(source, OPENING_LINES));
1435 }
1436 }
1437 classify(agent, bin, code, stderr, stdout, terminal)
1438}
1439
1440/// The longest an agent's answer may be and still be read as a notice.
1441///
1442/// A CLI that has been stopped says so briefly: Claude's is one sentence and a
1443/// reset time. An answer that *discusses* limits runs to paragraphs and uses
1444/// exactly the same words, so length is the only thing separating them.
1445const NOTICE_MAX: usize = 240;
1446
1447/// How far into an agent's own output a diagnosis may be read from.
1448///
1449/// Three rather than one, because a CLI is entitled to a banner line before it
1450/// says what is wrong, and three rather than more, because past that an agent
1451/// is answering the question it was asked.
1452const OPENING_LINES: usize = 3;
1453
1454/// The first `count` non-blank lines, trimmed and rejoined.
1455fn opening_lines(text: &str, count: usize) -> String {
1456 text.lines()
1457 .map(str::trim)
1458 .filter(|line| !line.is_empty())
1459 .take(count)
1460 .collect::<Vec<_>>()
1461 .join("\n")
1462}
1463
1464/// Whether an agent's own answer is a credentials notice rather than an answer
1465/// that happens to discuss credentials.
1466///
1467/// The single rule for reading an answer as a diagnosis of the run, shared by
1468/// the gate in `run` and by `classify_run`. They read the same text for the
1469/// same phrases and used to apply different rules to it: whole text at the
1470/// gate, opening lines in the classifier. A healthy answer about logging in
1471/// satisfied one and not the other, which opened the error path for a run no
1472/// classifier would then name.
1473///
1474/// Short *and* at the top, which is the same rule the quota branch applies,
1475/// and it takes both halves. Lines alone were not enough: asked to explain the
1476/// difference between a rate limit and an auth failure, a live Claude answered
1477/// in one 900-character paragraph, so "the first three lines" was the entire
1478/// essay and the phrase inside it convicted the run. Prose wraps at the
1479/// window, not at a newline, so length is what distinguishes a notice from an
1480/// answer. A real notice is a sentence: `Not logged in, please run /login`.
1481fn answer_reports_no_credentials(text: &str) -> bool {
1482 let opening = opening_lines(text, OPENING_LINES);
1483 opening.len() <= NOTICE_MAX && looks_unauthenticated(&opening)
1484}
1485
1486/// Whether a classifier named the failure rather than falling through to the
1487/// generic one.
1488///
1489/// `Error::Failed` is what `classify` returns when nothing more specific fits.
1490/// On a run that exited cleanly that is not a diagnosis, it is the absence of
1491/// one, and an answer must not be discarded for it.
1492fn names_a_failure(error: &Error) -> bool {
1493 !matches!(error, Error::Failed { .. })
1494}
1495
1496/// Whether text is an agent saying it has no usable credentials.
1497///
1498/// Narrow on purpose. Mislabelling an ordinary failure as an auth problem sends
1499/// someone to re-login over something unrelated, so these are phrases the CLIs
1500/// actually emit rather than every string containing "auth".
1501fn looks_unauthenticated(text: &str) -> bool {
1502 const PHRASES: &[&str] = &[
1503 // Claude, verified: an unauthenticated run answers exactly this.
1504 "not logged in",
1505 "please run /login",
1506 // Copilot, verified: it exits 1 with plain text, and none of the other
1507 // phrases here appear in it. Its wording shares no vocabulary with the
1508 // other two, which is why this had to be observed rather than guessed.
1509 "no authentication information",
1510 "invalid api key",
1511 "authentication_error",
1512 "unauthorized",
1513 "not authenticated",
1514 "no credentials",
1515 "credentials not found",
1516 "please log in",
1517 ];
1518 let lower = text.to_ascii_lowercase();
1519 PHRASES.iter().any(|needle| lower.contains(needle)) || mentions_status(&lower, "401")
1520}
1521
1522/// Whether `code` appears as a standalone token rather than inside a longer run
1523/// of characters.
1524///
1525/// `401` was previously matched as a bare substring, which made any Copilot
1526/// failure an auth failure whenever one of the UUIDs it prints happened to
1527/// contain those three digits: `"id":"1b0b1401-cb86-..."` was enough. That is
1528/// not rare, since a run emits several ids, so the misdiagnosis was
1529/// intermittent and told someone to re-login over an unrelated failure.
1530///
1531/// A status code is a word. Requiring non-alphanumeric neighbours keeps
1532/// `HTTP 401` and `(status 401)` while rejecting every hex blob, and a UUID
1533/// cannot produce a standalone `401` at all because its groups are four, eight
1534/// or twelve characters long.
1535fn mentions_status(haystack: &str, code: &str) -> bool {
1536 haystack.match_indices(code).any(|(at, _)| {
1537 let before = haystack[..at].chars().next_back();
1538 let after = haystack[at + code.len()..].chars().next();
1539 let free = |c: Option<char>| c.is_none_or(|c| !c.is_alphanumeric());
1540 free(before) && free(after)
1541 })
1542}
1543
1544/// Turn a non-zero exit into the most specific error available.
1545fn classify(
1546 agent: crate::Agent,
1547 bin: &str,
1548 code: i32,
1549 stderr: &str,
1550 stdout: &str,
1551 terminal: &Terminal,
1552) -> Error {
1553 let quota_signalled = terminal
1554 .rate_limit
1555 .as_ref()
1556 .is_some_and(crate::outcome::RateLimit::is_blocking);
1557 // Scanning the *raw* stream for quota wording is a false-positive machine:
1558 // under `stream-json` Claude prints a `rate_limit_event` record on every
1559 // run, including one whose status is `allowed`, so the substring
1560 // `rate_limit` is present in perfectly healthy output. Where the stream
1561 // parsed, the parsed signal and the agent's own prose decide; the raw scan
1562 // is only the fallback for output that produced neither.
1563 /*
1564 * The agent's own answer is evidence about the *topic*, not about the run.
1565 *
1566 * `terminal.text` is what the agent said. A turn that discusses quotas at
1567 * any length contains the vocabulary this function searches for, so a
1568 * finished, successful answer on that subject classified its own run as
1569 * blocked and replaced itself with a banner quoting one of its own
1570 * sentences. The same shape as the auth misclassification fixed in 0.4.2,
1571 * one branch further down the same function.
1572 *
1573 * So the run's own channels stay authoritative. `error_message` is the
1574 * CLI's own field rather than the model's words, and is read whole. The
1575 * answer is read only when it is short enough to *be* a notice: a run that
1576 * was really stopped has the notice and nothing else to say, in a couple
1577 * of lines, while an answer that discusses the subject runs to paragraphs.
1578 * Length is the one thing that separates them, because the vocabulary is
1579 * identical by definition.
1580 */
1581 let reported = terminal.error_message.clone().unwrap_or_default();
1582 let answered = if terminal.text.len() <= NOTICE_MAX {
1583 opening_lines(&terminal.text, OPENING_LINES)
1584 } else {
1585 String::new()
1586 };
1587 let prose = if terminal.text.is_empty() {
1588 format!("{reported}\n{}", opening_lines(stdout, OPENING_LINES))
1589 } else {
1590 format!("{reported}\n{answered}")
1591 };
1592 if quota_signalled || looks_rate_limited(stderr) || looks_rate_limited(&prose) {
1593 return Error::RateLimited {
1594 bin: bin.to_string(),
1595 message: first_meaningful_line(stderr)
1596 .or_else(|| first_meaningful_line(&prose))
1597 .unwrap_or_else(|| "usage limit reached".to_string()),
1598 };
1599 }
1600 // A rejected flag is not a failed request, it is this crate and the CLI
1601 // disagreeing about what the CLI accepts. Naming that is the difference
1602 // between "the run failed" and "your codex is a different version".
1603 if let Some(detail) = rejected_flag(stderr).or_else(|| rejected_flag(stdout)) {
1604 return Error::FlagRejected {
1605 bin: bin.to_string(),
1606 detail,
1607 };
1608 }
1609 // Checked before the generic failure but after quota and a rejected flag,
1610 // which are more specific readings of the same output.
1611 if terminal.stop == Stop::Error {
1612 return Error::AgentError {
1613 agent,
1614 bin: bin.to_string(),
1615 status: terminal.error_status,
1616 // Codex reports the reason apart from the answer; Claude puts it
1617 // where the answer would be.
1618 message: terminal
1619 .error_message
1620 .clone()
1621 .or_else(|| first_meaningful_line(&terminal.text))
1622 .or_else(|| first_meaningful_line(stderr))
1623 .unwrap_or_else(|| "the agent reported a failure without explaining it".into()),
1624 };
1625 }
1626
1627 Error::Failed {
1628 bin: bin.to_string(),
1629 code,
1630 // Fall back to stdout when stderr explains nothing. Codex reports a
1631 // rejected schema as an `{"type":"error"}` event on *stdout* while
1632 // stderr carries only "Reading additional input from stdin...", so
1633 // reporting stderr alone describes the failure as a status message.
1634 stderr: first_meaningful_line(stderr)
1635 .filter(|line| looks_explanatory(line))
1636 .or_else(|| first_meaningful_line(stdout))
1637 .or_else(|| first_meaningful_line(stderr))
1638 .unwrap_or_default(),
1639 }
1640}
1641
1642/// Whether a line plausibly explains a failure rather than narrating progress.
1643fn looks_explanatory(line: &str) -> bool {
1644 const NOISE: &[&str] = &[
1645 "reading additional input",
1646 "reading prompt",
1647 "waiting",
1648 "connecting",
1649 "loading",
1650 ];
1651 let lower = line.to_ascii_lowercase();
1652 !NOISE.iter().any(|noise| lower.contains(noise))
1653}
1654
1655/// The CLI's complaint, if it refused an argument.
1656///
1657/// The phrasings are clap's and commander's, which is what all three CLIs are
1658/// built on. Matched narrowly: a false positive would relabel a genuine failure
1659/// as a version problem and send someone chasing the wrong thing.
1660fn rejected_flag(text: &str) -> Option<String> {
1661 const REJECTIONS: &[&str] = &[
1662 "unexpected argument",
1663 "unknown option",
1664 "unrecognized option",
1665 "unknown flag",
1666 "invalid option",
1667 "unexpected option",
1668 ];
1669 let lower = text.to_ascii_lowercase();
1670 REJECTIONS
1671 .iter()
1672 .any(|needle| lower.contains(needle))
1673 .then(|| first_meaningful_line(text).unwrap_or_default())
1674}
1675
1676/// Whether text carries a provider quota refusal.
1677///
1678/// Deliberately a small set of unambiguous phrases: a false positive here would
1679/// relabel an ordinary failure as a quota problem and send a caller into a
1680/// pointless backoff.
1681fn looks_rate_limited(text: &str) -> bool {
1682 let lower = text.to_ascii_lowercase();
1683 [
1684 "rate limit",
1685 "rate_limit",
1686 "usage limit",
1687 "quota exceeded",
1688 "too many requests",
1689 ]
1690 .iter()
1691 .any(|needle| lower.contains(needle))
1692 // A status code is a word, and `429` as a bare substring is in every
1693 // line number, byte count, sha fragment and identifier that happens to
1694 // contain those digits. `401` was already given this treatment after it
1695 // matched inside a UUID and sent someone to re-login; this is the same
1696 // rule, applied to the code that had been left as a substring.
1697 || mentions_status(&lower, "429")
1698}
1699
1700/// The most useful line of a CLI's output for an error message.
1701///
1702/// Not simply the first non-blank one. CLIs open with progress and status
1703/// chatter, so the first line is often "Reading additional input from stdin..."
1704/// while the actual cause is further down. That turns a report into a
1705/// misdirection: it looks like an explanation and is not one.
1706///
1707/// So a line that looks like an error wins, and the first non-blank line is the
1708/// fallback when nothing does.
1709fn first_meaningful_line(text: &str) -> Option<String> {
1710 const ERROR_MARKERS: &[&str] = &[
1711 "error",
1712 "failed",
1713 "fatal",
1714 "panic",
1715 "denied",
1716 "invalid",
1717 "unexpected",
1718 "cannot",
1719 "unable",
1720 ];
1721 let lines: Vec<&str> = text
1722 .lines()
1723 .map(str::trim)
1724 .filter(|line| !line.is_empty())
1725 .collect();
1726
1727 lines
1728 .iter()
1729 .find(|line| {
1730 let lower = line.to_ascii_lowercase();
1731 ERROR_MARKERS.iter().any(|marker| lower.contains(marker))
1732 })
1733 .or_else(|| lines.first())
1734 .map(|line| (*line).to_string())
1735}
1736
1737/// Write the session binding back, reporting any store failure.
1738///
1739/// Called as soon as an id is known rather than only on a clean exit. Waiting
1740/// for success would lose the binding for exactly the runs where continuity
1741/// matters most: a timeout, a crash, or a cancelled turn.
1742fn persist_session(request: &Request, token: &str) -> Result<()> {
1743 let Some(binding) = &request.binding else {
1744 return Ok(());
1745 };
1746 binding
1747 .store
1748 .bind(request.agent, &binding.project, &binding.name, token)
1749 .map(|_| ())
1750}
1751
1752/// The id this run is already known by before it starts, if any.
1753///
1754/// Only a caller-assigned id qualifies: a printed id does not exist yet. This
1755/// is what makes an assigned session survive a run that never finishes.
1756fn preassigned_token(request: &Request) -> Option<String> {
1757 match &request.plan().cont {
1758 Continue::NewWith(id) => Some(id.clone()),
1759 _ => None,
1760 }
1761}
1762
1763/// Reported by an agent that exited cleanly but said nothing useful.
1764impl Outcome {
1765 /// Whether the agent produced any answer at all.
1766 #[must_use]
1767 pub fn is_empty(&self) -> bool {
1768 self.text.trim().is_empty() && self.stop == Stop::Completed
1769 }
1770}
1771
1772#[cfg(test)]
1773mod tests {
1774 use super::*;
1775 use crate::agent::Agent;
1776
1777 #[test]
1778 fn quota_phrases_are_recognized_and_ordinary_errors_are_not() {
1779 assert!(looks_rate_limited("Error: rate limit exceeded"));
1780 assert!(looks_rate_limited("HTTP 429 Too Many Requests"));
1781 assert!(looks_rate_limited("You have hit your usage limit"));
1782 // A plain failure must not be mistaken for a quota problem.
1783 assert!(!looks_rate_limited("error: no such file or directory"));
1784 assert!(!looks_rate_limited("model not found"));
1785 }
1786
1787 #[test]
1788 fn a_blocking_rate_limit_event_classifies_as_rate_limited() {
1789 let terminal = Terminal {
1790 rate_limit: Some(crate::outcome::RateLimit {
1791 status: "rejected".into(),
1792 window: Some("five_hour".into()),
1793 resets_at: None,
1794 overage_status: None,
1795 is_using_overage: None,
1796 }),
1797 ..Terminal::default()
1798 };
1799 assert!(matches!(
1800 classify(Agent::Claude, "claude", 1, "", "", &terminal),
1801 Error::RateLimited { .. }
1802 ));
1803 }
1804
1805 #[test]
1806 fn an_allowed_rate_limit_event_is_not_a_failure_cause() {
1807 let terminal = Terminal {
1808 rate_limit: Some(crate::outcome::RateLimit {
1809 status: "allowed".into(),
1810 window: None,
1811 resets_at: None,
1812 overage_status: None,
1813 is_using_overage: None,
1814 }),
1815 ..Terminal::default()
1816 };
1817 assert!(matches!(
1818 classify(Agent::Claude, "claude", 1, "boom", "", &terminal),
1819 Error::Failed { .. }
1820 ));
1821 }
1822
1823 /// The exact shape that made a Copilot run look unauthenticated: a UUID
1824 /// carrying the digits 401. Copilot prints several ids per run, so this
1825 /// misfired intermittently and told the user to re-login over a failure
1826 /// that had nothing to do with credentials.
1827 #[test]
1828 fn an_id_containing_401_is_not_an_auth_failure() {
1829 let line = r#"{"type":"session.mcp_server_status_changed","id":"1b0b1401-cb86-4276-9874-e84b94c96499"}"#;
1830 assert!(
1831 !looks_unauthenticated(line),
1832 "a hex blob is not a status code"
1833 );
1834 }
1835
1836 /// The needle still has to work where it was meant to. A status code is a
1837 /// word, and these are the forms an agent actually prints.
1838 #[test]
1839 fn a_real_401_is_still_recognized() {
1840 for text in [
1841 "HTTP 401",
1842 "request failed (status 401)",
1843 "401: unauthorized",
1844 "got a 401 from the API",
1845 ] {
1846 assert!(looks_unauthenticated(text), "should match: {text}");
1847 }
1848 }
1849
1850 /// Neighbouring digits mean it is part of some longer number, not a status.
1851 #[test]
1852 fn digits_around_401_keep_it_from_matching() {
1853 for text in ["error 4010", "code 1401", "seq 24019"] {
1854 assert!(!looks_unauthenticated(text), "should not match: {text}");
1855 }
1856 }
1857
1858 /// Verbatim from a healthy claude 2.1.205 run. Every `stream-json` run
1859 /// carries this record, and its status is `allowed`: nothing is refused.
1860 /// Scanning the raw stream for `rate_limit` matched it anyway, so any
1861 /// Claude failure was reported as a quota refusal, sending a caller to back
1862 /// off when the real cause was something they could fix.
1863 #[test]
1864 fn a_healthy_rate_limit_heartbeat_is_not_a_refusal() {
1865 let stdout = r#"{"type":"rate_limit_event","rate_limit_info":{"status":"allowed","resetsAt":1785331800,"rateLimitType":"five_hour","overageStatus":"rejected","isUsingOverage":false}}"#;
1866 let terminal = Terminal {
1867 stop: Stop::Error,
1868 error_status: Some(404),
1869 text: "There's an issue with the selected model (bogus-model-xyz).".into(),
1870 rate_limit: Some(crate::outcome::RateLimit {
1871 status: "allowed".into(),
1872 window: Some("five_hour".into()),
1873 resets_at: Some(1_785_331_800),
1874 overage_status: None,
1875 is_using_overage: None,
1876 }),
1877 ..Terminal::default()
1878 };
1879 let err = classify_run(Agent::Claude, "claude", 0, "", stdout, &terminal);
1880 assert!(
1881 matches!(err, Error::AgentError { .. }),
1882 "the heartbeat must not mask the real cause: {err:?}"
1883 );
1884 }
1885
1886 /// The counterpart: a refusal the parser did read must still be one, even
1887 /// though it arrives with the same zero exit code.
1888 #[test]
1889 fn a_rejected_quota_signal_is_still_a_refusal() {
1890 let terminal = Terminal {
1891 rate_limit: Some(crate::outcome::RateLimit {
1892 status: "rejected".into(),
1893 window: Some("five_hour".into()),
1894 resets_at: None,
1895 overage_status: None,
1896 is_using_overage: None,
1897 }),
1898 ..Terminal::default()
1899 };
1900 assert!(matches!(
1901 classify_run(Agent::Claude, "claude", 0, "", "", &terminal),
1902 Error::RateLimited { .. }
1903 ));
1904 }
1905
1906 /// Verbatim from a real run with an unknown model. Claude exits **0** with
1907 /// `subtype: "success"` while `is_error` is true and the explanation sits
1908 /// where the answer would be, so a caller checking only `Result::is_ok`
1909 /// renders "There's an issue with the selected model" as the answer.
1910 #[test]
1911 fn a_failed_turn_is_an_error_even_though_the_process_exited_cleanly() {
1912 let terminal = Terminal {
1913 stop: Stop::Error,
1914 error_status: Some(404),
1915 text: "There's an issue with the selected model (bogus-model-xyz). \
1916 It may not exist or you may not have access to it."
1917 .into(),
1918 ..Terminal::default()
1919 };
1920 let err = classify_run(Agent::Claude, "claude", 0, "", "", &terminal);
1921 let Error::AgentError {
1922 agent,
1923 status,
1924 message,
1925 ..
1926 } = &err
1927 else {
1928 panic!("expected AgentError, got {err:?}")
1929 };
1930 assert_eq!(*agent, Agent::Claude);
1931 assert_eq!(*status, Some(404), "the provider status must survive");
1932 assert!(message.contains("selected model"), "{message}");
1933 }
1934
1935 /// A quota refusal and a missing login are more specific readings of the
1936 /// same shape, so they must not be swallowed by the general case.
1937 #[test]
1938 fn a_failed_turn_does_not_mask_a_more_specific_cause() {
1939 let auth = Terminal {
1940 stop: Stop::Error,
1941 text: "Not logged in · Please run /login".into(),
1942 ..Terminal::default()
1943 };
1944 assert!(
1945 classify_run(Agent::Claude, "claude", 0, "", "", &auth).is_auth_failure(),
1946 "an unauthenticated failed turn must stay an auth failure"
1947 );
1948
1949 let quota = Terminal {
1950 stop: Stop::Error,
1951 rate_limit: Some(crate::outcome::RateLimit {
1952 status: "rejected".into(),
1953 window: None,
1954 resets_at: None,
1955 overage_status: None,
1956 is_using_overage: None,
1957 }),
1958 ..Terminal::default()
1959 };
1960 assert!(
1961 matches!(
1962 classify_run(Agent::Claude, "claude", 0, "", "", "a),
1963 Error::RateLimited { .. }
1964 ),
1965 "a quota-blocked failed turn must stay a rate limit"
1966 );
1967 }
1968
1969 /// Verified against the real CLI: with `USER` withheld, claude answers
1970 /// "Not logged in · Please run /login" and exits **0**. Checking only the
1971 /// exit code hands back a successful Outcome whose answer is a login
1972 /// prompt.
1973 #[test]
1974 fn an_unauthenticated_run_is_named_even_though_it_exits_zero() {
1975 let terminal = Terminal {
1976 text: "Not logged in · Please run /login".into(),
1977 ..Terminal::default()
1978 };
1979 let err = classify_run(Agent::Claude, "claude", 0, "", "", &terminal);
1980 let Error::NotAuthenticated { agent, hint, .. } = &err else {
1981 panic!("expected NotAuthenticated, got {err:?}")
1982 };
1983 assert_eq!(*agent, Agent::Claude);
1984 assert!(hint.contains("/login"), "{hint}");
1985 assert!(err.is_auth_failure());
1986 }
1987
1988 /// Verbatim from an unauthenticated Copilot run, captured by pointing it at
1989 /// an empty HOME. Its wording shares no phrase with Claude's or Codex's, so
1990 /// before this was observed the phrase list did not match it at all and a
1991 /// missing Copilot login was reported as a generic failure.
1992 #[test]
1993 fn copilots_own_unauthenticated_wording_is_recognized() {
1994 let stderr = "Error: No authentication information found.\n\n\
1995 Copilot can be authenticated with GitHub using an OAuth Token or a \
1996 Fine-Grained Personal Access Token.\n\n\
1997 To authenticate, you can use any of the following methods:\n\
1998 \u{2022} Start 'copilot' and run the '/login' command\n\
1999 \u{2022} Set the COPILOT_GITHUB_TOKEN, GH_TOKEN, or GITHUB_TOKEN \
2000 environment variable";
2001 let err = classify_run(
2002 Agent::Copilot,
2003 "copilot",
2004 1,
2005 stderr,
2006 "",
2007 &Terminal::default(),
2008 );
2009 let Error::NotAuthenticated { agent, hint, .. } = &err else {
2010 panic!("expected NotAuthenticated, got {err:?}")
2011 };
2012 assert_eq!(*agent, Agent::Copilot);
2013 assert!(hint.contains("copilot login"), "{hint}");
2014 }
2015
2016 /// Each agent's hint has to name its own login route, since they differ:
2017 /// Codex and Copilot have `login` subcommands, Claude does not.
2018 #[test]
2019 fn every_agent_offers_its_own_login_route() {
2020 for (agent, expected) in [
2021 (Agent::Claude, "setup-token"),
2022 (Agent::Codex, "codex login"),
2023 (Agent::Copilot, "copilot login"),
2024 ] {
2025 let err = classify_run(
2026 agent,
2027 agent.bin(),
2028 1,
2029 "error: unauthorized",
2030 "",
2031 &Terminal::default(),
2032 );
2033 let Error::NotAuthenticated { hint, .. } = &err else {
2034 panic!("{agent}: expected NotAuthenticated, got {err:?}")
2035 };
2036 assert!(hint.contains(expected), "{agent}: {hint}");
2037 }
2038 }
2039
2040 /// Reported from the field: a run was stopped, the user was told `claude`
2041 /// was not authenticated, and the answer was replaced by a login hint. The
2042 /// agent had been explaining why a `cargo publish` was refused, and its own
2043 /// prose contained the phrases this classifier looks for. An answer is not
2044 /// a diagnosis of the thing that produced it.
2045 #[test]
2046 fn an_agent_writing_about_authentication_is_not_an_auth_failure() {
2047 let answer = "The publish was refused before it ran.\n\n\
2048 What denied it was the auto mode classifier, not a missing \
2049 credential.\n\
2050 In auto mode there is no human to receive the prompt, so an \
2051 ask collapses into a refusal.\n\
2052 The message said the CLI was not authenticated, which is \
2053 unrelated: an unauthorized upload is exactly what the rule \
2054 is there to stop.";
2055 let terminal = Terminal {
2056 text: answer.into(),
2057 ..Terminal::default()
2058 };
2059 let err = classify_run(Agent::Claude, "claude", 1, "", answer, &terminal);
2060 assert!(
2061 !err.is_auth_failure(),
2062 "an answer that discusses auth was read as an auth failure: {err:?}"
2063 );
2064 }
2065
2066 /// The other half of the same rule: the notice itself still has to be
2067 /// caught, and it arrives as the agent's entire answer.
2068 #[test]
2069 fn the_notice_is_still_caught_when_it_is_the_whole_answer() {
2070 for text in [
2071 "Not logged in · Please run /login",
2072 // A banner first, which is why the opening is three lines deep.
2073 "claude 2.1.212\n\nNot logged in · Please run /login",
2074 ] {
2075 let terminal = Terminal {
2076 text: text.into(),
2077 ..Terminal::default()
2078 };
2079 let err = classify_run(Agent::Claude, "claude", 0, "", "", &terminal);
2080 assert!(
2081 err.is_auth_failure(),
2082 "{text:?} was not read as auth: {err:?}"
2083 );
2084 }
2085 }
2086
2087 /// The gate into the error path and the classifier behind it must read an
2088 /// answer by the same rule.
2089 ///
2090 /// Shaped after the report from the field: a healthy, finished turn about
2091 /// a stalled crate release, which mentions an auth error in a later
2092 /// paragraph because that was the subject. Read whole, as the gate used
2093 /// to, the phrase convicts the run; read at its opening, as everything
2094 /// now does, the answer is an answer. An agent with no credentials leads
2095 /// with the notice, which is what makes the opening the honest place to
2096 /// look.
2097 #[test]
2098 fn an_answer_mentioning_auth_late_does_not_open_the_error_path() {
2099 let answer = "So the duplicate pastes cost nothing.\n\
2100 The publish chain is done: 0.4.1 and 0.4.2 are both on the \
2101 registry and tagged.\n\
2102 Everything downstream already consumes them.\n\
2103 The only thing still open anywhere is the crate PR, \
2104 `pathscale/RustAgentAbstraction#18`, which is the auth error \
2105 that ate your reply: the run was reported as `not \
2106 authenticated` and the hint sent you to /login, while the \
2107 credentials were fine the whole time.";
2108 assert!(
2109 !answer_reports_no_credentials(answer),
2110 "an answer discussing auth opened the error path"
2111 );
2112 // And the notice itself, which is what the rule exists to catch.
2113 assert!(answer_reports_no_credentials(
2114 "Not logged in \u{b7} Please run /login"
2115 ));
2116 }
2117
2118 /// The backstop, which is what makes a false positive survivable at all.
2119 ///
2120 /// Every phrase check here is a heuristic over ordinary English and will
2121 /// be wrong eventually. When it is, the run reaches a classifier that
2122 /// cannot name any failure and returns the generic one. On a process that
2123 /// exited cleanly with a completed turn, that verdict is the absence of
2124 /// evidence rather than evidence, and the finished answer must stand.
2125 #[test]
2126 fn a_generic_failure_does_not_name_a_failure() {
2127 let unnamed = Error::Failed {
2128 bin: "claude".into(),
2129 code: 0,
2130 stderr: String::new(),
2131 };
2132 assert!(
2133 !names_a_failure(&unnamed),
2134 "a generic failure was treated as a diagnosis, which discards the answer"
2135 );
2136 // Everything a classifier can actually name still stands on its own.
2137 assert!(names_a_failure(&Error::RateLimited {
2138 bin: "claude".into(),
2139 message: "usage limit reached".into(),
2140 }));
2141 assert!(names_a_failure(&Error::NotAuthenticated {
2142 agent: Agent::Claude,
2143 bin: "claude".into(),
2144 message: "Not logged in".into(),
2145 hint: Agent::Claude.login_hint(),
2146 }));
2147 }
2148
2149 /// Reported from the field: a finished, successful turn that happened to be
2150 /// *about* usage limits classified its own run as blocked, and the banner
2151 /// quoted one of the answer's own sentences back as the provider's message.
2152 /// An answer is evidence about its topic, not about the run that produced it.
2153 #[test]
2154 fn an_agent_writing_about_limits_is_not_a_limit() {
2155 let answer = "The three retries cost nothing extra, so that is not where it \
2156 came from.\n\n\
2157 Providers do not rate limit on repetition: the limit is on \
2158 tokens per window, and a 429 is what you would see if one had \
2159 actually been reached. Each retry does re-send the whole \
2160 conversation, which is real spend, but spend is not the same \
2161 thing as a block and the run reported no blocking signal at \
2162 all.\n\
2163 Nothing here suggests the usage limit was reached, and the \
2164 banner quoted a sentence of this answer back as though a \
2165 provider had written it.";
2166 let terminal = Terminal {
2167 text: answer.into(),
2168 ..Terminal::default()
2169 };
2170 let err = classify(Agent::Claude, "claude", 1, "", answer, &terminal);
2171 assert!(
2172 !matches!(err, Error::RateLimited { .. }),
2173 "an answer discussing limits was read as one: {err:?}"
2174 );
2175 }
2176
2177 /// A status code is a word. These are the shapes that used to trip it.
2178 #[test]
2179 fn a_bare_429_in_prose_is_not_a_status_code() {
2180 assert!(!looks_rate_limited("see run.rs:4291 for the caller"));
2181 assert!(!looks_rate_limited("sha 8f429ac"));
2182 assert!(looks_rate_limited("HTTP 429"));
2183 assert!(looks_rate_limited("(status 429)"));
2184 assert!(looks_rate_limited("Error: too many requests"));
2185 }
2186
2187 /// Auth is the most specific reading, so it wins over a generic failure,
2188 /// but must not swallow unrelated errors.
2189 #[test]
2190 fn ordinary_failures_are_not_mistaken_for_auth_problems() {
2191 for stderr in [
2192 "error: no such file or directory",
2193 "model not found",
2194 "rate limit exceeded",
2195 "error: unexpected argument '--sandbox' found",
2196 ] {
2197 let err = classify_run(Agent::Codex, "codex", 1, stderr, "", &Terminal::default());
2198 assert!(
2199 !err.is_auth_failure(),
2200 "{stderr:?} was misread as an auth failure: {err:?}"
2201 );
2202 }
2203 }
2204
2205 /// The exact failure that cost a round of debugging: `codex exec resume`
2206 /// rejects `--sandbox`, which `Error::Failed` reported as a generic
2207 /// non-zero exit naming a flag rather than a version mismatch.
2208 #[test]
2209 fn a_rejected_flag_is_named_as_a_version_mismatch() {
2210 let err = classify(
2211 Agent::Codex,
2212 "codex",
2213 2,
2214 "error: unexpected argument '--sandbox' found",
2215 "",
2216 &Terminal::default(),
2217 );
2218 let Error::FlagRejected { bin, detail } = err else {
2219 panic!("expected FlagRejected, got {err:?}")
2220 };
2221 assert_eq!(bin, "codex");
2222 assert!(detail.contains("--sandbox"), "{detail}");
2223 }
2224
2225 #[test]
2226 fn ordinary_failures_are_not_mistaken_for_version_drift() {
2227 for stderr in [
2228 "error: no such file or directory",
2229 "model not found",
2230 "permission denied",
2231 ] {
2232 assert!(
2233 matches!(
2234 classify(Agent::Codex, "codex", 1, stderr, "", &Terminal::default()),
2235 Error::Failed { .. }
2236 ),
2237 "{stderr:?} should stay a plain failure"
2238 );
2239 }
2240 }
2241
2242 /// Real output from a failing codex run: the first line is status, the
2243 /// cause is below it. Reporting the first line looks like an explanation
2244 /// while pointing at the wrong thing.
2245 #[test]
2246 fn a_status_line_does_not_masquerade_as_the_cause() {
2247 let stderr = "Reading additional input from stdin...\n\
2248 error: invalid value 'nope' for '--sandbox <SANDBOX_MODE>'";
2249 let err = classify_run(Agent::Codex, "codex", 1, stderr, "", &Terminal::default());
2250 let Error::Failed {
2251 stderr: reported, ..
2252 } = err
2253 else {
2254 panic!("expected Failed, got {err:?}")
2255 };
2256 assert!(reported.contains("invalid value"), "reported {reported:?}");
2257 }
2258
2259 /// Codex reports a rejected schema as a JSON error event on **stdout**
2260 /// while stderr carries only a status line. Reporting stderr alone
2261 /// described the failure as "Reading additional input from stdin...",
2262 /// which is not what went wrong.
2263 #[test]
2264 fn a_cause_on_stdout_is_reported_when_stderr_only_narrates() {
2265 let stdout = r#"{"type":"error","message":"invalid_json_schema: 'additionalProperties' is required to be supplied and to be false."}"#;
2266 let err = classify_run(
2267 Agent::Codex,
2268 "codex",
2269 1,
2270 "Reading additional input from stdin...",
2271 stdout,
2272 &Terminal::default(),
2273 );
2274 let Error::Failed {
2275 stderr: reported, ..
2276 } = err
2277 else {
2278 panic!("expected Failed, got {err:?}")
2279 };
2280 assert!(
2281 reported.contains("additionalProperties"),
2282 "reported {reported:?}, which explains nothing"
2283 );
2284 }
2285
2286 #[test]
2287 fn failures_report_the_first_useful_line() {
2288 let err = classify(
2289 Agent::Claude,
2290 "claude",
2291 2,
2292 "\n\n real problem \nstack",
2293 "",
2294 &Terminal::default(),
2295 );
2296 let Error::Failed { code, stderr, .. } = err else {
2297 panic!("expected a plain failure")
2298 };
2299 assert_eq!(code, 2);
2300 assert_eq!(stderr, "real problem");
2301 }
2302
2303 /// Prompts and session ids ride the argv, and `Run::argv` invites logging
2304 /// it. The redacted form must keep the shape while dropping the content.
2305 #[test]
2306 fn redaction_removes_prompts_and_session_ids_but_keeps_flags() {
2307 let request = crate::Request::new(Agent::Claude, "my secret prompt")
2308 .system("secret system")
2309 .session_id("11111111-2222-3333-4444-555555555555");
2310 let safe = redact(&request.typed_argv().unwrap());
2311
2312 for secret in [
2313 "my secret prompt",
2314 "secret system",
2315 "11111111-2222-3333-4444-555555555555",
2316 ] {
2317 assert!(
2318 !safe.iter().any(|a| a.contains(secret)),
2319 "{secret:?} survived redaction: {safe:?}"
2320 );
2321 }
2322 // Still recognisable as the same command.
2323 assert_eq!(safe[0], "claude");
2324 assert!(safe.contains(&"--permission-mode".to_string()));
2325 assert!(safe.contains(&"--session-id".to_string()));
2326 }
2327
2328 #[test]
2329 fn codex_trailing_prompt_is_redacted_even_without_a_flag() {
2330 let request = crate::Request::new(Agent::Codex, "my secret prompt");
2331 let safe = redact(&request.typed_argv().unwrap());
2332 assert_eq!(safe.last().unwrap(), REDACTED);
2333 assert_eq!(safe[1], "exec", "the subcommand must survive");
2334 }
2335
2336 /// Redaction must cover the two shapes positional guesswork misses: Codex's
2337 /// bare trailing prompt, and raw arguments whose contents are unknowable.
2338 #[test]
2339 fn redaction_covers_positional_prompts_and_unchecked_arguments() {
2340 let request = crate::Request::new(Agent::Codex, "my secret prompt")
2341 .unchecked_args(["-c", "api_key=hunter2"]);
2342 let safe = redact(&request.typed_argv().unwrap());
2343 assert!(!safe.iter().any(|a| a.contains("my secret prompt")));
2344 assert!(
2345 !safe.iter().any(|a| a.contains("hunter2")),
2346 "unchecked arguments may hold secrets: {safe:?}"
2347 );
2348 assert_eq!(safe[1], "exec", "the subcommand must survive");
2349 }
2350
2351 /// A resume id is a capability: it continues someone's conversation.
2352 #[test]
2353 fn redaction_covers_the_codex_positional_resume_id() {
2354 let request = crate::Request::new(Agent::Codex, "hi").resume("thread-secret-9");
2355 let safe = redact(&request.typed_argv().unwrap());
2356 assert!(
2357 !safe.iter().any(|a| a.contains("thread-secret-9")),
2358 "{safe:?}"
2359 );
2360 assert!(safe.contains(&"resume".to_string()));
2361 }
2362
2363 /// `stream` is synchronous but spawns a task. Outside a runtime that would
2364 /// panic, which a `Result`-returning function must not do.
2365 #[test]
2366 fn stream_outside_a_runtime_errors_instead_of_panicking() {
2367 let err = stream(&crate::Request::new(Agent::Claude, "hi")).unwrap_err();
2368 assert!(matches!(err, Error::NoRuntime), "got {err:?}");
2369 }
2370
2371 #[tokio::test]
2372 async fn a_missing_binary_names_the_install_command() {
2373 let request = Request::new(Agent::Claude, "hi").bin("definitely-not-a-real-binary-xyz");
2374 let err = run(&request).await.unwrap_err();
2375 let Error::NotInstalled { hint, agent, .. } = err else {
2376 panic!("expected NotInstalled, got {err:?}")
2377 };
2378 assert_eq!(agent, Agent::Claude);
2379 assert!(hint.contains("claude-code"));
2380 }
2381
2382 /// Regression for the `AgencyZero` hang: Codex can put its steer receipt
2383 /// behind more events than fit in the bounded channel. Input must wait on a
2384 /// handle independent from the mutable event receiver, or the producer and
2385 /// consumer block each other permanently.
2386 #[tokio::test]
2387 async fn a_control_receipt_can_wait_behind_a_full_event_buffer() {
2388 let (events_tx, events_rx) = mpsc::channel(EVENT_BUFFER);
2389 let (controls_tx, mut controls_rx) = mpsc::channel(1);
2390 let (cancel_tx, cancel_rx) = tokio::sync::oneshot::channel();
2391 let task = tokio::spawn(async move {
2392 let _cancel_rx = cancel_rx;
2393 let Some(Control::Message { receipt, .. }) = controls_rx.recv().await else {
2394 panic!("follow-up message")
2395 };
2396 for index in 0..=EVENT_BUFFER {
2397 events_tx
2398 .send(Event::Thinking(index.to_string()))
2399 .await
2400 .expect("the host keeps draining events");
2401 }
2402 let _ = receipt.send(Ok(()));
2403 Ok(Outcome {
2404 agent: Agent::Codex,
2405 session: None,
2406 text: String::new(),
2407 usage: crate::Usage::default(),
2408 stop: Stop::Completed,
2409 rate_limit: None,
2410 exit_code: 0,
2411 stderr: String::new(),
2412 unparsed: 0,
2413 first_unparsed: None,
2414 structured: None,
2415 })
2416 });
2417 let reaped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(true));
2418 let mut run = Run {
2419 events: events_rx,
2420 control: RunControl {
2421 agent: Agent::Codex,
2422 to_agent: Some(controls_tx),
2423 bin: "codex".into(),
2424 },
2425 typed: Vec::new(),
2426 pid: None,
2427 reaped,
2428 cancel: Some(cancel_tx),
2429 task: Some(task),
2430 argv: vec!["codex".into()],
2431 };
2432
2433 let control = run.control();
2434 let delivery = tokio::spawn(async move { control.send("change course").await });
2435 let mut seen = 0;
2436 while run.recv().await.is_some() {
2437 seen += 1;
2438 }
2439
2440 assert_eq!(seen, EVENT_BUFFER + 1);
2441 delivery
2442 .await
2443 .expect("delivery task")
2444 .expect("transport receipt");
2445 run.finish().await.expect("run outcome");
2446 }
2447
2448 /// A live app-server can keep its pipes open after it stops processing
2449 /// requests. Input delivery needs a deadline so the host can reap that run
2450 /// and retry the visible message on the resumed session.
2451 #[tokio::test]
2452 async fn a_control_receipt_that_never_arrives_times_out() {
2453 let (controls_tx, mut controls_rx) = mpsc::channel(1);
2454 let control = RunControl {
2455 agent: Agent::Codex,
2456 to_agent: Some(controls_tx),
2457 bin: "codex".into(),
2458 };
2459 let receiver = tokio::spawn(async move {
2460 let Some(Control::Message { receipt, .. }) = controls_rx.recv().await else {
2461 panic!("follow-up message")
2462 };
2463 tokio::time::sleep(Duration::from_secs(1)).await;
2464 drop(receipt);
2465 });
2466
2467 let err = control
2468 .send_with_timeout("are you there?", Duration::from_millis(10))
2469 .await
2470 .expect_err("the missing receipt must not wait forever");
2471 assert!(
2472 matches!(err, Error::ControlTimeout { .. }),
2473 "expected ControlTimeout, got {err:?}"
2474 );
2475 receiver.abort();
2476 }
2477
2478 #[test]
2479 fn transient_errors_are_distinguished_from_permanent_ones() {
2480 assert!(
2481 Error::RateLimited {
2482 bin: "claude".into(),
2483 message: String::new()
2484 }
2485 .is_transient()
2486 );
2487 assert!(
2488 Error::ControlTimeout {
2489 bin: "codex".into(),
2490 timeout: CONTROL_RECEIPT_TIMEOUT,
2491 }
2492 .is_transient()
2493 );
2494 assert!(
2495 !Error::NotInstalled {
2496 agent: Agent::Claude,
2497 bin: "claude".into(),
2498 hint: ""
2499 }
2500 .is_transient()
2501 );
2502 }
2503}