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