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::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(String),
97 Approval {
98 id: String,
99 decision: crate::Decision,
100 },
101}
102
103/// A run in progress.
104///
105/// Yields events through [`Run::recv`] and settles into an [`Outcome`] through
106/// [`Run::finish`].
107///
108/// **Dropping a `Run` kills the agent.** That is the safe default for the hosts
109/// this crate targets: closing a window or cancelling a request should stop the
110/// work, not leave an agent running invisibly, spending quota and touching
111/// files with nobody watching. Call [`Run::detach`] when background execution is
112/// genuinely what you want.
113///
114/// On Unix, dropping **synchronously signals** the run's process group and then
115/// aborts the driver task. What it cannot do is *wait*: `Drop` cannot await, so
116/// it does not block until the child has exited or its readers have been
117/// joined. Use [`Run::cancel`] when you need to know the tree has actually gone
118/// before continuing, such as before touching the files it was working on. On
119/// Windows only the direct child is signalled.
120#[derive(Debug)]
121pub struct Run {
122 events: mpsc::Receiver<Event>,
123 /// Which agent this is, so `respond` can name it in an error.
124 agent: crate::Agent,
125 /// The typed command line, kept so both the plain and redacted views come
126 /// from the same source.
127 typed: Vec<crate::agent::Arg>,
128 /// The child's pid, so `Drop` can tear the group down itself rather than
129 /// depending on an aborted task being polled.
130 pid: Option<u32>,
131 /// Set by the driver once the child has been reaped, so `Drop` never
132 /// signals a pid the OS may since have handed to someone else.
133 reaped: std::sync::Arc<std::sync::atomic::AtomicBool>,
134 /// Lines on the way back to the agent: follow-up messages and approval
135 /// decisions share one channel because they share one stdin. `None` unless
136 /// the request opened it, which is what lets [`Run::send`] and
137 /// [`Run::respond`] refuse rather than silently do nothing.
138 to_agent: Option<mpsc::Sender<Control>>,
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
149impl Run {
150 /// The next event, or `None` once the agent has finished producing them.
151 pub async fn recv(&mut self) -> Option<Event> {
152 self.events.recv().await
153 }
154
155 /// Send another message while the agent is still working.
156 ///
157 /// The whole point of [`crate::Request::interactive`]: a user who types a
158 /// correction mid-turn should not have to wait for the turn to finish.
159 ///
160 /// The agent takes it at its **next step boundary**, not mid-token, so an
161 /// answer already being written finishes first and a long tool-using task
162 /// changes course at its next step. Verified against claude 2.1.212 and
163 /// codex-cli 0.145.0.
164 ///
165 /// # Ordering, and why there is no acknowledgement
166 ///
167 /// The caller already knows what it sent, so the intended pattern is to
168 /// append the message to the transcript immediately, below the user's
169 /// previous one, and carry on. This deliberately does not ask the agent to
170 /// echo the message back for sequencing: an echo would only tell a UI
171 /// something it already knew, and waiting for one would delay the very
172 /// thing this exists to make immediate.
173 ///
174 /// # Errors
175 /// [`Error::Unsupported`] on a run that did not open the channel with
176 /// [`crate::Request::interactive`]. [`Error::Cancelled`] once the channel
177 /// has closed, which happens when the turn settles or the run is torn down:
178 /// **a message sent after the turn ends is too late** and belongs in a new
179 /// run resuming the session, so this reports it rather than dropping it.
180 pub async fn send(&self, message: &str) -> Result<()> {
181 let Some(channel) = &self.to_agent else {
182 return Err(Error::Unsupported {
183 agent: self.agent,
184 what: "sending a follow-up on a run that is not interactive",
185 });
186 };
187 channel
188 .send(Control::Message(message.to_string()))
189 .await
190 .map_err(|_| Error::Cancelled {
191 bin: self.argv.first().cloned().unwrap_or_default(),
192 })
193 }
194
195 /// Answer an [`Event::ApprovalRequest`].
196 ///
197 /// The agent is blocked until this is called, so a consumer that receives an
198 /// approval request and never responds stalls the run until its timeout.
199 ///
200 /// The id must be the one from the request. The agent ignores an answer
201 /// carrying any other id and keeps waiting, so a mismatch presents as a
202 /// hang rather than an error; this passes the id straight through and does
203 /// not invent one.
204 ///
205 /// # Errors
206 /// [`Error::Unsupported`] on a run that did not ask for approvals, since
207 /// there is no channel to answer on. [`Error::Cancelled`] if the run has
208 /// already finished or been torn down, which is the same reason a decision
209 /// can no longer be delivered.
210 pub async fn respond(&self, id: &str, decision: &crate::Decision) -> Result<()> {
211 let Some(channel) = &self.to_agent else {
212 return Err(Error::Unsupported {
213 agent: self.agent,
214 what: "answering an approval on a run that did not request them",
215 });
216 };
217 channel
218 .send(Control::Approval {
219 id: id.to_string(),
220 decision: decision.clone(),
221 })
222 .await
223 .map_err(|_| Error::Cancelled {
224 bin: self.argv.first().cloned().unwrap_or_default(),
225 })
226 }
227
228 /// The exact command line that was spawned.
229 ///
230 /// **This contains the prompt and any session id.** Treat it as sensitive:
231 /// logging it verbatim puts user content into your logs. Use
232 /// [`Run::redacted_argv`] for diagnostics.
233 #[must_use]
234 pub fn argv(&self) -> &[String] {
235 &self.argv
236 }
237
238 /// The command line with every non-public value replaced by a placeholder.
239 ///
240 /// Prompts, system prompts, session ids and anything from
241 /// [`crate::Request::unchecked_args`] are removed; flag names are kept so
242 /// the command stays recognisable. Sensitivity is recorded where each
243 /// argument is built rather than inferred from the finished line, so a
244 /// bare positional prompt or an opaque raw argument is covered too.
245 #[must_use]
246 pub fn redacted_argv(&self) -> Vec<String> {
247 redact(&self.typed)
248 }
249
250 /// Wait for the run to finish.
251 ///
252 /// Drains any events still queued, so a caller that only wants the result
253 /// can call this without having consumed the stream.
254 ///
255 /// # Errors
256 /// Whatever the run failed with. See [`Error`].
257 pub async fn finish(mut self) -> Result<Outcome> {
258 // The driver owns teardown from here; `Drop` must not also fire.
259 self.pid = None;
260 while self.events.recv().await.is_some() {}
261 // Taking the handle disarms the `Drop` guard: this run is settling
262 // normally, not being abandoned.
263 let Some(task) = self.task.take() else {
264 unreachable!("the handle is only taken by a consuming method")
265 };
266 match task.await {
267 Ok(result) => result,
268 // The driver task panicked or was cancelled. The process itself
269 // started fine, so this is not a spawn failure and must not claim
270 // to be one.
271 Err(join) => Err(Error::Interrupted {
272 bin: self.argv.first().cloned().unwrap_or_default(),
273 detail: if join.is_panic() {
274 "the driver task panicked".into()
275 } else {
276 "the driver task was cancelled".into()
277 },
278 }),
279 }
280 }
281
282 /// Stop the run and wait until the agent is actually gone.
283 ///
284 /// Cooperative rather than an abort: the driver is asked to stop, signals
285 /// the process group, reaps the child and joins its readers, and only then
286 /// does this return. So when it returns the tree really has exited, which
287 /// matters if the next thing you do touches the files it was working on.
288 ///
289 /// Returns the partial [`Outcome`] if the run happened to finish first,
290 /// otherwise [`Error::Cancelled`].
291 ///
292 /// # Errors
293 /// [`Error::Cancelled`] in the normal case, or whatever the run failed with
294 /// if it failed before the request arrived.
295 pub async fn cancel(mut self) -> Result<Outcome> {
296 // The driver tears down cooperatively and this awaits it, so `Drop`
297 // must not race that with a kill of its own.
298 self.pid = None;
299 // Dropping the sender is itself the signal, so this cannot fail in a
300 // way that leaves the driver waiting.
301 drop(self.cancel.take());
302 let Some(task) = self.task.take() else {
303 unreachable!("the handle is only taken by a consuming method")
304 };
305 match task.await {
306 Ok(result) => result,
307 Err(join) => Err(Error::Interrupted {
308 bin: self.argv.first().cloned().unwrap_or_default(),
309 detail: if join.is_panic() {
310 "the driver task panicked".into()
311 } else {
312 "the driver task was cancelled".into()
313 },
314 }),
315 }
316 }
317
318 /// Let the run continue after this handle goes away.
319 ///
320 /// The opposite of the default. Nothing can observe or stop the agent
321 /// afterwards, so reach for this only when an unsupervised background run
322 /// is genuinely intended.
323 pub fn detach(mut self) {
324 // Disarm `Drop` before it runs, or detaching would immediately kill the
325 // run it exists to keep alive.
326 self.pid = None;
327 // Leak the cancel signal rather than dropping it: a dropped sender is
328 // read by the driver as "stop", which is the opposite of detaching.
329 if let Some(cancel) = self.cancel.take() {
330 std::mem::forget(cancel);
331 }
332 // Dropping the handle without aborting is what detaches a tokio task.
333 drop(self.task.take());
334 }
335}
336
337impl Drop for Run {
338 fn drop(&mut self) {
339 // Abandoned rather than finished, cancelled or detached.
340 //
341 // Kill the group here, directly. Signalling the driver and aborting it
342 // is not enough on its own: that leaves teardown waiting on the runtime
343 // to poll the aborted task so its guard runs, and a dropped `Run` was
344 // observed leaving grandchildren alive and sleeping on Linux while the
345 // same teardown worked from `cancel`. `Drop` cannot await, so it does
346 // the one thing it can do synchronously.
347 if let Some(pid) = self.pid
348 && !self.reaped.load(std::sync::atomic::Ordering::SeqCst)
349 {
350 kill_group_by_pid(pid);
351 }
352 drop(self.cancel.take());
353 if let Some(task) = self.task.take() {
354 task.abort();
355 }
356 }
357}
358
359/// Placeholder substituted for a sensitive argv value.
360const REDACTED: &str = "<redacted>";
361
362/// Render a typed command line for logging, keeping flag names and replacing
363/// every value that is not `Public`.
364///
365/// Derived from the sensitivity recorded where each argument was built, so it
366/// cannot miss a case the way matching on flag names and positions can.
367fn redact(argv: &[crate::agent::Arg]) -> Vec<String> {
368 use crate::agent::Sensitivity;
369
370 argv.iter()
371 .map(|arg| match arg.sensitivity {
372 Sensitivity::Public => arg.value.clone(),
373 _ => REDACTED.to_string(),
374 })
375 .collect()
376}
377
378/// Run `request` to completion, discarding the intermediate events.
379///
380/// # Errors
381/// See [`Error`]; notably [`Error::NotInstalled`], [`Error::Timeout`],
382/// [`Error::RateLimited`] and [`Error::Failed`].
383///
384/// [`Error::Unsupported`] for a request that asked for approvals: this entry
385/// point discards events, so an approval request would reach nobody and the run
386/// would sit blocked until its timeout. Use [`stream`] instead.
387pub async fn run(request: &Request) -> Result<Outcome> {
388 if request.plan().approvals {
389 return Err(Error::Unsupported {
390 agent: request.agent,
391 what: "approvals on a run whose events are discarded; use `stream`",
392 });
393 }
394 stream(request)?.finish().await
395}
396
397/// Start `request`, returning a handle that streams its events.
398///
399/// Returns as soon as the child is spawned; the work proceeds on a task.
400///
401/// # Errors
402/// [`Error::NotInstalled`] if the binary is missing, [`Error::Unsupported`] if
403/// the agent cannot honour the request, or [`Error::Spawn`] on an OS failure.
404#[allow(
405 clippy::too_many_lines,
406 reason = "one spawn boundary keeps command posture, pipes, process group, and driver selection together"
407)]
408pub fn stream(request: &Request) -> Result<Run> {
409 // `tokio::spawn` panics outside a runtime. A fallible signature must not
410 // hide that, so the context is checked and reported as an ordinary error.
411 let runtime = tokio::runtime::Handle::try_current().map_err(|_| Error::NoRuntime)?;
412
413 let initial_plan = request.plan();
414 let codex_app_server =
415 request.agent == crate::Agent::Codex && (initial_plan.duplex || initial_plan.approvals);
416
417 // Written before the argv is built, because the argv has to name it. The
418 // app-server protocol accepts the schema inline instead.
419 let schema_file = match (&request.schema, request.agent.caps().schema) {
420 (Some(_), _) if codex_app_server => None,
421 (Some(schema), crate::agent::SchemaSupport::File) => {
422 Some(SchemaFile::write(schema).map_err(|source| Error::Spawn {
423 bin: request.agent.bin().to_string(),
424 source,
425 })?)
426 }
427 _ => None,
428 };
429 let mut request = request.clone();
430 if let Some(file) = &schema_file {
431 request.schema_file = Some(file.0.display().to_string());
432 }
433 let request = &request;
434
435 let plan = request.plan();
436 let typed = request.typed_argv()?;
437 let argv: Vec<String> = typed.iter().map(|a| a.value.clone()).collect();
438
439 let mut command = Command::new(&argv[0]);
440 command
441 .args(&argv[1..])
442 .stdin(if plan.stdin_prompt || plan.duplex || plan.approvals {
443 // An interactive run needs stdin for the whole turn, not just to
444 // deliver a prompt: it is the channel follow-up messages and
445 // approval decisions travel back on.
446 Stdio::piped()
447 } else {
448 // Close stdin so an agent that would otherwise wait on it exits
449 // instead of hanging forever with nothing to read.
450 Stdio::null()
451 })
452 .stdout(Stdio::piped())
453 .stderr(Stdio::piped())
454 // Without this a killed run can leave the child alive holding the pipes.
455 .kill_on_drop(true);
456 if let Some(cwd) = &request.cwd {
457 command.current_dir(cwd);
458 }
459 // Narrow the environment first, then apply explicit variables, so an
460 // explicit `env()` always wins over the policy.
461 match &request.env_policy {
462 EnvPolicy::Inherit => {}
463 EnvPolicy::Minimal => {
464 command.env_clear();
465 inherit_named(&mut command, &request.agent.essential_env());
466 }
467 EnvPolicy::Only(names) => {
468 command.env_clear();
469 inherit_named(&mut command, names);
470 }
471 }
472 for (key, value) in &request.env {
473 command.env(key, value);
474 }
475
476 // Put the agent in its own process group so the whole tree can be signalled
477 // together. Killing only the CLI leaves the commands *it* spawned running:
478 // a build, a test run, a server, still holding files and credentials after
479 // the run is supposedly over.
480 // 0 means "make this child its own group leader". `tokio::process::Command`
481 // exposes this directly on unix.
482 #[cfg(unix)]
483 command.process_group(0);
484
485 // Reserve an assigned session id before the child exists. Doing it inside
486 // the driver leaves a window where a spawn that half-succeeds loses the
487 // binding, and this is the id the caller may already be showing in a UI.
488 if let Some(token) = preassigned_token(request) {
489 persist_session(request, &token)?;
490 }
491
492 let child = command.spawn().map_err(|source| {
493 // A missing binary is the common case and deserves an actionable error
494 // with an install hint. Reading it off the spawn avoids resolving PATH
495 // twice, and with it the window where the resolved path is replaced
496 // between the check and the exec.
497 if source.kind() == std::io::ErrorKind::NotFound {
498 Error::NotInstalled {
499 agent: request.agent,
500 bin: plan.bin.clone(),
501 hint: request.agent.install_hint(),
502 }
503 } else {
504 Error::Spawn {
505 bin: plan.bin.clone(),
506 source,
507 }
508 }
509 })?;
510
511 let request_agent = request.agent;
512 let pid = child.id();
513 let (tx, rx) = mpsc::channel(EVENT_BUFFER);
514 // Only created for an approvals run, so `respond` can tell "no channel" from
515 // "channel closed" and refuse the first rather than hanging on it.
516 let (decisions_tx, decisions_rx) = if plan.duplex || plan.approvals {
517 let (tx, rx) = mpsc::channel::<Control>(APPROVAL_BUFFER);
518 (Some(tx), Some(rx))
519 } else {
520 (None, None)
521 };
522 let (cancel_tx, cancel_rx) = tokio::sync::oneshot::channel();
523 let reaped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
524 let reaped_for_task = std::sync::Arc::clone(&reaped);
525 let request = request.clone();
526 let task = runtime.spawn(async move {
527 // Moved in so the file outlives the run and is removed with it.
528 let _schema_file = schema_file;
529 if codex_app_server {
530 drive_codex_app_server(child, request, tx, cancel_rx, reaped_for_task, decisions_rx)
531 .await
532 } else {
533 drive(child, request, tx, cancel_rx, reaped_for_task, decisions_rx).await
534 }
535 });
536 Ok(Run {
537 events: rx,
538 agent: request_agent,
539 typed,
540 pid,
541 reaped,
542 cancel: Some(cancel_tx),
543 to_agent: decisions_tx,
544 task: Some(task),
545 argv,
546 })
547}
548
549/// Copy the named variables from this process into `command`, skipping any that
550/// are unset so nothing is invented.
551fn inherit_named<S: AsRef<str>>(command: &mut Command, names: &[S]) {
552 for name in names {
553 if let Some(value) = std::env::var_os(name.as_ref()) {
554 command.env(name.as_ref(), value);
555 }
556 }
557}
558
559/// A schema file written for one run, removed when the run ends.
560///
561/// Codex reads its schema from disk, so the file has to outlive the spawn and
562/// not outlive the process. Tying it to a guard means every exit path removes
563/// it, including a cancel or a timeout, without each one remembering.
564struct SchemaFile(std::path::PathBuf);
565
566impl SchemaFile {
567 /// Write `schema` somewhere the agent can read it.
568 fn write(schema: &str) -> std::io::Result<SchemaFile> {
569 use std::io::Write as _;
570 use std::sync::atomic::{AtomicU64, Ordering};
571 static COUNTER: AtomicU64 = AtomicU64::new(0);
572
573 let path = std::env::temp_dir().join(format!(
574 "agent-abstraction-schema-{}-{}.json",
575 std::process::id(),
576 COUNTER.fetch_add(1, Ordering::Relaxed)
577 ));
578 let mut options = std::fs::OpenOptions::new();
579 options.write(true).create_new(true);
580 // A schema can encode what a caller is looking for, so it is no more
581 // public than the prompt.
582 #[cfg(unix)]
583 {
584 use std::os::unix::fs::OpenOptionsExt as _;
585 options.mode(0o600);
586 }
587 options.open(&path)?.write_all(schema.as_bytes())?;
588 Ok(SchemaFile(path))
589 }
590}
591
592impl Drop for SchemaFile {
593 fn drop(&mut self) {
594 let _ = std::fs::remove_file(&self.0);
595 }
596}
597
598/// Owns the child and tears down its whole process group when dropped.
599///
600/// `kill_on_drop` alone is not enough: it kills the CLI, leaving the commands
601/// *it* spawned running. Since aborting the driver task drops this guard, the
602/// same teardown covers cancellation, a dropped [`Run`] and a timeout, without
603/// each path having to remember to do it.
604struct ChildGuard {
605 child: Child,
606 /// Cleared once the child has been reaped, so a pid the OS may since have
607 /// recycled is never signalled.
608 armed: bool,
609}
610
611impl Drop for ChildGuard {
612 fn drop(&mut self) {
613 if self.armed {
614 kill_process_group(&self.child);
615 }
616 }
617}
618
619/// Feed the child, read both its pipes, and assemble the outcome.
620#[allow(
621 clippy::too_many_lines,
622 reason = "one linear lifecycle: feed, read, wait, classify. Splitting it \
623 would thread the child, parser, buffers and cancellation state \
624 through helpers and obscure the ordering that matters, such as \
625 killing the group before reaping."
626)]
627async fn drive(
628 child: Child,
629 request: Request,
630 events: mpsc::Sender<Event>,
631 cancel: tokio::sync::oneshot::Receiver<()>,
632 reaped: std::sync::Arc<std::sync::atomic::AtomicBool>,
633 decisions: Option<mpsc::Receiver<Control>>,
634) -> Result<Outcome> {
635 // From here on the child is owned by a guard, so every exit path from this
636 // task, including an abort, takes the process group with it.
637 let mut child = ChildGuard { child, armed: true };
638 let plan = request.plan();
639 let bin = plan.bin.clone();
640
641 // An approvals run owns stdin for the whole turn: the handshake and the
642 // prompt go out first, then it stays open carrying decisions until the run
643 // ends. Closing it after the prompt, as the plain piped path does, would
644 // take the answer channel with it.
645 let mut decision_task = None;
646 let mut close_stdin = None;
647 if plan.duplex || plan.approvals {
648 let Some(mut stdin) = child.child.stdin.take() else {
649 return Err(Error::Spawn {
650 bin: bin.clone(),
651 source: std::io::Error::other("stdin was not piped for an interactive run"),
652 });
653 };
654 let opening = format!(
655 "{}{}",
656 crate::approval::handshake(),
657 crate::approval::user_message(&request.agent.effective_prompt(&plan)),
658 );
659 stdin
660 .write_all(opening.as_bytes())
661 .await
662 .map_err(|source| Error::Spawn {
663 bin: bin.clone(),
664 source,
665 })?;
666 let _ = stdin.flush().await;
667 // Forwarding runs on its own task so a decision can be written while
668 // stdout is being read. It ends on whichever comes first: the channel
669 // closing, or the turn settling.
670 let (close_tx, mut close_rx) = tokio::sync::oneshot::channel::<()>();
671 close_stdin = Some(close_tx);
672 decision_task = decisions.map(|mut rx| {
673 tokio::spawn(async move {
674 loop {
675 tokio::select! {
676 reply = rx.recv() => {
677 let Some(control) = reply else { break };
678 let reply = match control {
679 Control::Message(message) => {
680 crate::approval::user_message(&message)
681 }
682 Control::Approval { id, decision } => decision.wire(&id),
683 };
684 if stdin.write_all(reply.as_bytes()).await.is_err() {
685 break;
686 }
687 let _ = stdin.flush().await;
688 }
689 // The turn is over. Dropping stdin is what lets claude
690 // exit rather than wait for another message.
691 _ = &mut close_rx => break,
692 }
693 }
694 drop(stdin);
695 })
696 });
697 }
698
699 // Deliver a piped prompt and close the pipe, or the agent waits on EOF.
700 if plan.stdin_prompt {
701 if let Some(mut stdin) = child.child.stdin.take() {
702 let prompt = request.agent.effective_prompt(&plan);
703 stdin
704 .write_all(prompt.as_bytes())
705 .await
706 .map_err(|source| Error::Spawn {
707 bin: bin.clone(),
708 source,
709 })?;
710 drop(stdin);
711 }
712 }
713
714 // Drain stderr on its own task: a full stderr pipe blocks the child even
715 // while stdout still has room.
716 // Aborted on every exit path from here, so a forwarder never survives the
717 // run it belongs to.
718 let _decision_guard = decision_task.map(AbortOnDrop);
719
720 let stderr = child.child.stderr.take();
721 let stderr_task = tokio::spawn(async move {
722 let mut buf = String::new();
723 if let Some(handle) = stderr {
724 let mut reader = BufReader::new(handle);
725 let mut line = String::new();
726 // Keep draining after the cap is hit: an undrained pipe blocks the
727 // child even though we no longer want the bytes.
728 while let Ok(Some(_)) = read_bounded_line(&mut reader, &mut line).await {
729 append_capped(&mut buf, &line);
730 }
731 }
732 buf
733 });
734
735 let stdout = child.child.stdout.take();
736 let mut parser = Parser::new(request.agent, plan.format);
737 // Raw stdout is retained only as a fallback answer for a run that exited
738 // cleanly without producing a structured one, and as evidence when
739 // classifying a failure. It is capped for the same reason as everything
740 // else here: an agent can stream for hours.
741 let mut raw = String::new();
742 // Tracks the first `Started`, so the binding is written once, and carries a
743 // store failure back out instead of discarding it.
744 let mut bound = false;
745 let mut persist_result: Result<()> = Ok(());
746
747 let read_stdout = async {
748 if let Some(handle) = stdout {
749 let mut reader = BufReader::new(handle);
750 let mut line = String::new();
751 while read_bounded_line(&mut reader, &mut line).await?.is_some() {
752 append_capped(&mut raw, &line);
753 let parsed = parser.push(&line);
754 // Close stdin as soon as the turn settles. Under stream-json
755 // input claude waits for another message otherwise, so the run
756 // would only end at its timeout even though the answer already
757 // arrived.
758 if parser.saw_terminal()
759 && let Some(close) = close_stdin.take()
760 {
761 let _ = close.send(());
762 }
763 for event in parsed {
764 // Bind a printed id the moment it appears rather than at the
765 // end. Codex announces its thread before answering, so a
766 // turn killed mid-answer stays resumable.
767 if let Event::Started { session, .. } = &event
768 && !bound
769 {
770 bound = true;
771 persist_result = persist_session(&request, session);
772 }
773 // A receiver that went away is not a failure: the run should
774 // still finish and produce its outcome.
775 if events.send(event).await.is_err() {
776 break;
777 }
778 }
779 }
780 }
781 Ok::<_, std::io::Error>(())
782 };
783
784 // Race three outcomes: the run finishing, the deadline, and a cancellation
785 // request. Reading and waiting are one future so a child that produces
786 // output forever is still bounded by the timeout.
787 let work = async {
788 read_stdout.await?;
789 child.child.wait().await
790 };
791 // A timeout is optional; `pending()` makes the un-timed case the same shape
792 // rather than duplicating the whole select.
793 let deadline = async {
794 match request.timeout {
795 Some(limit) => tokio::time::sleep(limit).await,
796 None => std::future::pending().await,
797 }
798 };
799
800 let status = tokio::select! {
801 // Biased so a finished run is reported as finished even if a deadline
802 // or cancellation lands in the same tick.
803 biased;
804 result = work => result,
805 () = deadline => {
806 // Order matters: signal the group *before* reaping. Reaping clears
807 // the child's pid, and the group kill needs that pid to target the
808 // group, so the other order silently leaves grandchildren running.
809 let partial = shut_down(&mut child, stderr_task).await;
810 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
811 return Err(Error::Timeout {
812 bin,
813 timeout: request.timeout.unwrap_or_default(),
814 partial: parser.finish().text,
815 })
816 .inspect_err(|_| drop(partial));
817 }
818 _ = cancel => {
819 // Cooperative teardown: the caller is waiting on this, so the tree
820 // is signalled, reaped and joined before returning.
821 shut_down(&mut child, stderr_task).await;
822 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
823 return Err(Error::Cancelled { bin });
824 }
825 }
826 .map_err(|source| Error::Spawn {
827 bin: bin.clone(),
828 source,
829 })?;
830
831 // The child has been reaped, so its pid must not be signalled again, by the
832 // guard here or by `Run::drop` racing this.
833 child.armed = false;
834 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
835
836 drop(events);
837 let stderr = stderr_task.await.unwrap_or_default();
838 let saw_structured = parser.saw_structured_record();
839 let saw_terminal = parser.saw_terminal_record();
840 let terminal = parser.finish();
841 let exit_code = status.code().unwrap_or(-1);
842
843 // Under a structured format, silently handing back raw stdout would turn a
844 // protocol failure into a plausible-looking answer. A run that recognized
845 // nothing, or never reached its terminal record, did not produce a result
846 // this crate can vouch for, so it is reported rather than papered over.
847 let structured = plan.format != crate::Format::Text;
848 if structured && exit_code == 0 {
849 if !saw_structured {
850 return Err(Error::Parse {
851 agent: request.agent,
852 detail: format!(
853 "no recognizable {} records in {} lines of output; the CLI's output shape has probably changed",
854 request.agent,
855 raw.lines().count()
856 ),
857 });
858 }
859 if !saw_terminal {
860 return Err(Error::Parse {
861 agent: request.agent,
862 detail: "the stream ended without its terminal record, so the turn did not complete"
863 .into(),
864 });
865 }
866 }
867
868 // Plain text has no structure to validate: the stream is the answer.
869 let mut terminal = terminal;
870 if terminal.text.is_empty() && !structured {
871 terminal.text = raw.trim().to_string();
872 }
873
874 // A provider refusal is not always an exit code. Claude can report a
875 // blocking `rate_limit_event` and still exit 0, and the crate promises that
876 // quota refusals surface as `Error::RateLimited`, so the terminal state is
877 // checked regardless of how the process exited.
878 let quota_blocked = terminal
879 .rate_limit
880 .as_ref()
881 .is_some_and(crate::outcome::RateLimit::is_blocking);
882 // An unauthenticated Claude run exits 0 and reports the problem in its
883 // result text, so checking only the exit code would hand back a successful
884 // Outcome whose answer is "Please run /login".
885 //
886 // Read from stderr and the agent's own prose rather than the raw stream, for
887 // the reason `classify` does the same with quota wording: a phrase hunted
888 // through structured output matches ids and field names, not statements.
889 //
890 // The answer is read at its opening only, by the same rule and the same
891 // helper `classify_run` uses. This gate used to read `terminal.text` whole
892 // while the classifier it guards read three lines, so the two could
893 // disagree: a healthy answer that merely discussed logging in opened the
894 // error path, no classifier would name it, and the turn fell out the far
895 // end as `Error::Failed` with exit code 0 and nothing to report. One rule,
896 // one helper, so that disagreement cannot exist.
897 let unauthenticated =
898 answer_reports_no_credentials(&terminal.text) || looks_unauthenticated(&stderr);
899 // The agent saying its turn failed is as much a failure as a non-zero exit,
900 // and Claude reports an unknown model exactly this way: exit 0, `is_error`
901 // true, and the explanation where the answer would be.
902 let turn_failed = terminal.stop == Stop::Error;
903 if exit_code != 0 || quota_blocked || unauthenticated || turn_failed {
904 let error = classify_run(request.agent, &bin, exit_code, &stderr, &raw, &terminal);
905 /*
906 * The backstop, and the reason a false positive can no longer cost an
907 * answer.
908 *
909 * Everything above is a heuristic reading of text the agent wrote, and
910 * a heuristic will be wrong eventually: these phrases are ordinary
911 * English, and an agent asked about rate limits or logging in answers
912 * in exactly the vocabulary that describes being rate limited or
913 * logged out. What must never follow from being wrong is discarding a
914 * finished answer.
915 *
916 * So a run whose process exited cleanly, whose terminal record says
917 * the turn completed, and which carries no parsed quota block is only
918 * ever failed by a classifier that can *name* the failure. A generic
919 * `Failed` on that run is the classifiers disagreeing with the gate,
920 * not evidence, and the answer stands.
921 */
922 let exited_clean = exit_code == 0 && !quota_blocked && !turn_failed;
923 if !exited_clean || names_a_failure(&error) {
924 return Err(error);
925 }
926 }
927
928 // A fork lands on a *new* id the agent only reveals at the end, so the name
929 // has to be repointed once the run settles. Everything else was bound above.
930 persist_result?;
931 // Resolved before the terminal is consumed by the Outcome below.
932 let structured = terminal.structured.clone().or_else(|| {
933 request
934 .schema
935 .as_ref()
936 .and_then(|_| serde_json::from_str(&terminal.text).ok())
937 });
938 if let Some(token) = &terminal.session
939 && !bound
940 {
941 persist_session(&request, token)?;
942 }
943 Ok(Outcome {
944 agent: request.agent,
945 session: terminal.session,
946 text: terminal.text,
947 usage: terminal.usage,
948 stop: terminal.stop,
949 rate_limit: terminal.rate_limit,
950 exit_code,
951 stderr,
952 unparsed: terminal.unparsed,
953 first_unparsed: terminal.first_unparsed,
954 // Claude reports the conforming value separately; Codex returns it as
955 // the answer text, so that is parsed only when a schema was asked for.
956 // Prose is never reinterpreted as data.
957 structured,
958 })
959}
960
961/// Drive one interactive Codex turn over app-server's JSON-RPC transport.
962///
963/// Unlike `codex exec`, app-server remains alive after a turn completes. This
964/// driver therefore treats `turn/completed` as the terminal record, closes the
965/// protocol pipe, and reaps the service process itself.
966#[allow(
967 clippy::too_many_lines,
968 reason = "one select loop owns the protocol, control channel, deadline, and child lifecycle"
969)]
970async fn drive_codex_app_server(
971 child: Child,
972 request: Request,
973 events: mpsc::Sender<Event>,
974 cancel: tokio::sync::oneshot::Receiver<()>,
975 reaped: std::sync::Arc<std::sync::atomic::AtomicBool>,
976 controls: Option<mpsc::Receiver<Control>>,
977) -> Result<Outcome> {
978 let mut child = ChildGuard { child, armed: true };
979 let plan = request.plan();
980 let bin = plan.bin.clone();
981 let Some(mut stdin) = child.child.stdin.take() else {
982 return Err(Error::Spawn {
983 bin,
984 source: std::io::Error::other("stdin was not piped for Codex app-server"),
985 });
986 };
987 let Some(stdout) = child.child.stdout.take() else {
988 return Err(Error::Spawn {
989 bin,
990 source: std::io::Error::other("stdout was not piped for Codex app-server"),
991 });
992 };
993 let Some(mut controls) = controls else {
994 return Err(Error::Spawn {
995 bin,
996 source: std::io::Error::other("Codex app-server has no host control channel"),
997 });
998 };
999
1000 let stderr = child.child.stderr.take();
1001 let stderr_task = tokio::spawn(async move {
1002 let mut buf = String::new();
1003 if let Some(handle) = stderr {
1004 let mut reader = BufReader::new(handle);
1005 let mut line = String::new();
1006 while let Ok(Some(_)) = read_bounded_line(&mut reader, &mut line).await {
1007 append_capped(&mut buf, &line);
1008 }
1009 }
1010 buf
1011 });
1012
1013 let mut protocol = crate::codex_app_server::Protocol::new(request.clone());
1014 for opening in protocol.opening() {
1015 stdin
1016 .write_all(opening.as_bytes())
1017 .await
1018 .map_err(|source| Error::Spawn {
1019 bin: bin.clone(),
1020 source,
1021 })?;
1022 }
1023 stdin.flush().await.map_err(|source| Error::Spawn {
1024 bin: bin.clone(),
1025 source,
1026 })?;
1027
1028 let mut reader = BufReader::new(stdout);
1029 let mut line = String::new();
1030 let mut raw = String::new();
1031 let mut pending = VecDeque::new();
1032 let mut bound = false;
1033 let mut persist_result: Result<()> = Ok(());
1034 let deadline = async {
1035 match request.timeout {
1036 Some(limit) => tokio::time::sleep(limit).await,
1037 None => std::future::pending().await,
1038 }
1039 };
1040 tokio::pin!(deadline);
1041 tokio::pin!(cancel);
1042
1043 while !protocol.finished {
1044 tokio::select! {
1045 biased;
1046 // User steering and approval answers outrank the agent's output.
1047 // app-server can keep stdout continuously ready with reasoning and
1048 // text deltas; reading it first in a biased select could starve a
1049 // correction precisely while Codex was busiest.
1050 control = controls.recv() => {
1051 let Some(control) = control else {
1052 continue;
1053 };
1054 pending.push_back(control);
1055 flush_codex_controls(&mut protocol, &mut pending, &mut stdin, &bin).await?;
1056 stdin.flush().await.map_err(|source| Error::Spawn {
1057 bin: bin.clone(), source
1058 })?;
1059 }
1060 record = read_bounded_line(&mut reader, &mut line) => {
1061 if record.map_err(|source| Error::Spawn { bin: bin.clone(), source })?.is_some() {
1062 append_capped(&mut raw, &line);
1063 if let Ok(value) = serde_json::from_str::<serde_json::Value>(&line) {
1064 let step = protocol.push(&value);
1065 for event in step.events {
1066 if let Event::Started { session, .. } = &event
1067 && !bound
1068 {
1069 bound = true;
1070 persist_result = persist_session(&request, session);
1071 }
1072 let _ = events.send(event).await;
1073 }
1074 for write in step.writes {
1075 stdin.write_all(write.as_bytes()).await.map_err(|source| {
1076 Error::Spawn { bin: bin.clone(), source }
1077 })?;
1078 }
1079 flush_codex_controls(&mut protocol, &mut pending, &mut stdin, &bin)
1080 .await?;
1081 stdin.flush().await.map_err(|source| Error::Spawn {
1082 bin: bin.clone(), source
1083 })?;
1084 } else {
1085 protocol.terminal.unparsed += 1;
1086 if protocol.terminal.first_unparsed.is_none() {
1087 protocol.terminal.first_unparsed = Some(line.clone());
1088 }
1089 }
1090 } else {
1091 protocol.failure.get_or_insert_with(|| {
1092 "app-server closed stdout before turn/completed".to_string()
1093 });
1094 protocol.finished = true;
1095 }
1096 }
1097 () = &mut deadline => {
1098 let partial = protocol.terminal.text.clone();
1099 shut_down(&mut child, stderr_task).await;
1100 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
1101 return Err(Error::Timeout {
1102 bin,
1103 timeout: request.timeout.unwrap_or_default(),
1104 partial,
1105 });
1106 }
1107 _ = &mut cancel => {
1108 shut_down(&mut child, stderr_task).await;
1109 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
1110 return Err(Error::Cancelled { bin });
1111 }
1112 }
1113 }
1114
1115 // app-server is a service rather than a one-shot process. EOF asks it to
1116 // stop cleanly; the short fallback prevents a completed turn from hanging
1117 // because a future CLI release keeps serving after its input closes.
1118 drop(stdin);
1119 if tokio::time::timeout(std::time::Duration::from_secs(2), child.child.wait())
1120 .await
1121 .is_err()
1122 {
1123 kill_process_group(&child.child);
1124 let _ = child.child.kill().await;
1125 }
1126 child.armed = false;
1127 reaped.store(true, std::sync::atomic::Ordering::SeqCst);
1128 drop(events);
1129 let stderr = stderr_task.await.unwrap_or_default();
1130
1131 persist_result?;
1132 if let Some(detail) = protocol.failure {
1133 return Err(Error::Parse {
1134 agent: request.agent,
1135 detail,
1136 });
1137 }
1138
1139 let terminal = protocol.terminal;
1140 if terminal.stop == Stop::Error {
1141 return Err(classify_run(
1142 request.agent,
1143 &bin,
1144 0,
1145 &stderr,
1146 &raw,
1147 &terminal,
1148 ));
1149 }
1150 let structured = terminal.structured.clone().or_else(|| {
1151 request
1152 .schema
1153 .as_ref()
1154 .and_then(|_| serde_json::from_str(&terminal.text).ok())
1155 });
1156 Ok(Outcome {
1157 agent: request.agent,
1158 session: terminal.session,
1159 text: terminal.text,
1160 usage: terminal.usage,
1161 stop: terminal.stop,
1162 rate_limit: terminal.rate_limit,
1163 exit_code: 0,
1164 stderr,
1165 unparsed: terminal.unparsed,
1166 first_unparsed: terminal.first_unparsed,
1167 structured,
1168 })
1169}
1170
1171/// Write every control whose protocol ids are available, preserving earlier
1172/// messages until thread and turn startup have both completed.
1173async fn flush_codex_controls(
1174 protocol: &mut crate::codex_app_server::Protocol,
1175 pending: &mut VecDeque<Control>,
1176 stdin: &mut tokio::process::ChildStdin,
1177 bin: &str,
1178) -> Result<()> {
1179 let mut waiting = VecDeque::new();
1180 while let Some(control) = pending.pop_front() {
1181 let encoded = match &control {
1182 Control::Message(message) => protocol.steer(message),
1183 Control::Approval { id, decision } => protocol.respond(id, decision),
1184 };
1185 if let Some(encoded) = encoded {
1186 stdin
1187 .write_all(encoded.as_bytes())
1188 .await
1189 .map_err(|source| Error::Spawn {
1190 bin: bin.to_string(),
1191 source,
1192 })?;
1193 } else {
1194 waiting.push_back(control);
1195 }
1196 }
1197 pending.append(&mut waiting);
1198 Ok(())
1199}
1200
1201/// Kill the process group, reap the child, and join the stderr reader.
1202///
1203/// The orderly teardown both cancellation and timeout share. Returns whatever
1204/// stderr had been captured, so a caller can still report why a run was stopped.
1205async fn shut_down(child: &mut ChildGuard, stderr_task: tokio::task::JoinHandle<String>) -> String {
1206 kill_process_group(&child.child);
1207 // Reap, so the caller is not left with a zombie once this returns.
1208 let _ = child.child.kill().await;
1209 child.armed = false;
1210 // The pipes are closed now that the child is gone, so this finishes
1211 // promptly rather than hanging the cancellation.
1212 stderr_task.await.unwrap_or_default()
1213}
1214
1215/// Turn a failure into the most specific error available, agent included so an
1216/// auth failure can carry the right login command.
1217fn classify_run(
1218 agent: crate::Agent,
1219 bin: &str,
1220 code: i32,
1221 stderr: &str,
1222 stdout: &str,
1223 terminal: &Terminal,
1224) -> Error {
1225 // Checked before quota and before a plain failure: a login problem is the
1226 // most specific reading of the output, and the only one a user can act on
1227 // directly.
1228 let named = |source: &str| Error::NotAuthenticated {
1229 agent,
1230 bin: bin.to_string(),
1231 message: first_meaningful_line(source).unwrap_or_default(),
1232 hint: agent.login_hint(),
1233 };
1234
1235 // The CLI's own channel, read whole: Copilot's notice runs to five lines.
1236 if looks_unauthenticated(stderr) {
1237 return named(stderr);
1238 }
1239
1240 /*
1241 * The agent's own answer, read only at the top.
1242 *
1243 * An agent with no credentials has nothing to say but the notice, so the
1244 * phrase is in its opening lines and the whole answer is those lines.
1245 * An agent that *writes about* logging in buries the same words in
1246 * paragraphs, and reading the whole answer counted that as a login
1247 * failure: a reply explaining why a publish had been refused mentioned
1248 * not being authenticated, so the run was reported as an auth error, the
1249 * hint told the user to run `/login`, and the answer itself was replaced
1250 * by the report. An agent's prose is not a diagnosis of the agent.
1251 */
1252 for source in [terminal.text.as_str(), stdout] {
1253 if answer_reports_no_credentials(source) {
1254 return named(&opening_lines(source, OPENING_LINES));
1255 }
1256 }
1257 classify(agent, bin, code, stderr, stdout, terminal)
1258}
1259
1260/// The longest an agent's answer may be and still be read as a notice.
1261///
1262/// A CLI that has been stopped says so briefly: Claude's is one sentence and a
1263/// reset time. An answer that *discusses* limits runs to paragraphs and uses
1264/// exactly the same words, so length is the only thing separating them.
1265const NOTICE_MAX: usize = 240;
1266
1267/// How far into an agent's own output a diagnosis may be read from.
1268///
1269/// Three rather than one, because a CLI is entitled to a banner line before it
1270/// says what is wrong, and three rather than more, because past that an agent
1271/// is answering the question it was asked.
1272const OPENING_LINES: usize = 3;
1273
1274/// The first `count` non-blank lines, trimmed and rejoined.
1275fn opening_lines(text: &str, count: usize) -> String {
1276 text.lines()
1277 .map(str::trim)
1278 .filter(|line| !line.is_empty())
1279 .take(count)
1280 .collect::<Vec<_>>()
1281 .join("\n")
1282}
1283
1284/// Whether an agent's own answer is a credentials notice rather than an answer
1285/// that happens to discuss credentials.
1286///
1287/// The single rule for reading an answer as a diagnosis of the run, shared by
1288/// the gate in `run` and by `classify_run`. They read the same text for the
1289/// same phrases and used to apply different rules to it: whole text at the
1290/// gate, opening lines in the classifier. A healthy answer about logging in
1291/// satisfied one and not the other, which opened the error path for a run no
1292/// classifier would then name.
1293///
1294/// Short *and* at the top, which is the same rule the quota branch applies,
1295/// and it takes both halves. Lines alone were not enough: asked to explain the
1296/// difference between a rate limit and an auth failure, a live Claude answered
1297/// in one 900-character paragraph, so "the first three lines" was the entire
1298/// essay and the phrase inside it convicted the run. Prose wraps at the
1299/// window, not at a newline, so length is what distinguishes a notice from an
1300/// answer. A real notice is a sentence: `Not logged in, please run /login`.
1301fn answer_reports_no_credentials(text: &str) -> bool {
1302 let opening = opening_lines(text, OPENING_LINES);
1303 opening.len() <= NOTICE_MAX && looks_unauthenticated(&opening)
1304}
1305
1306/// Whether a classifier named the failure rather than falling through to the
1307/// generic one.
1308///
1309/// `Error::Failed` is what `classify` returns when nothing more specific fits.
1310/// On a run that exited cleanly that is not a diagnosis, it is the absence of
1311/// one, and an answer must not be discarded for it.
1312fn names_a_failure(error: &Error) -> bool {
1313 !matches!(error, Error::Failed { .. })
1314}
1315
1316/// Whether text is an agent saying it has no usable credentials.
1317///
1318/// Narrow on purpose. Mislabelling an ordinary failure as an auth problem sends
1319/// someone to re-login over something unrelated, so these are phrases the CLIs
1320/// actually emit rather than every string containing "auth".
1321fn looks_unauthenticated(text: &str) -> bool {
1322 const PHRASES: &[&str] = &[
1323 // Claude, verified: an unauthenticated run answers exactly this.
1324 "not logged in",
1325 "please run /login",
1326 // Copilot, verified: it exits 1 with plain text, and none of the other
1327 // phrases here appear in it. Its wording shares no vocabulary with the
1328 // other two, which is why this had to be observed rather than guessed.
1329 "no authentication information",
1330 "invalid api key",
1331 "authentication_error",
1332 "unauthorized",
1333 "not authenticated",
1334 "no credentials",
1335 "credentials not found",
1336 "please log in",
1337 ];
1338 let lower = text.to_ascii_lowercase();
1339 PHRASES.iter().any(|needle| lower.contains(needle)) || mentions_status(&lower, "401")
1340}
1341
1342/// Whether `code` appears as a standalone token rather than inside a longer run
1343/// of characters.
1344///
1345/// `401` was previously matched as a bare substring, which made any Copilot
1346/// failure an auth failure whenever one of the UUIDs it prints happened to
1347/// contain those three digits: `"id":"1b0b1401-cb86-..."` was enough. That is
1348/// not rare, since a run emits several ids, so the misdiagnosis was
1349/// intermittent and told someone to re-login over an unrelated failure.
1350///
1351/// A status code is a word. Requiring non-alphanumeric neighbours keeps
1352/// `HTTP 401` and `(status 401)` while rejecting every hex blob, and a UUID
1353/// cannot produce a standalone `401` at all because its groups are four, eight
1354/// or twelve characters long.
1355fn mentions_status(haystack: &str, code: &str) -> bool {
1356 haystack.match_indices(code).any(|(at, _)| {
1357 let before = haystack[..at].chars().next_back();
1358 let after = haystack[at + code.len()..].chars().next();
1359 let free = |c: Option<char>| c.is_none_or(|c| !c.is_alphanumeric());
1360 free(before) && free(after)
1361 })
1362}
1363
1364/// Turn a non-zero exit into the most specific error available.
1365fn classify(
1366 agent: crate::Agent,
1367 bin: &str,
1368 code: i32,
1369 stderr: &str,
1370 stdout: &str,
1371 terminal: &Terminal,
1372) -> Error {
1373 let quota_signalled = terminal
1374 .rate_limit
1375 .as_ref()
1376 .is_some_and(crate::outcome::RateLimit::is_blocking);
1377 // Scanning the *raw* stream for quota wording is a false-positive machine:
1378 // under `stream-json` Claude prints a `rate_limit_event` record on every
1379 // run, including one whose status is `allowed`, so the substring
1380 // `rate_limit` is present in perfectly healthy output. Where the stream
1381 // parsed, the parsed signal and the agent's own prose decide; the raw scan
1382 // is only the fallback for output that produced neither.
1383 /*
1384 * The agent's own answer is evidence about the *topic*, not about the run.
1385 *
1386 * `terminal.text` is what the agent said. A turn that discusses quotas at
1387 * any length contains the vocabulary this function searches for, so a
1388 * finished, successful answer on that subject classified its own run as
1389 * blocked and replaced itself with a banner quoting one of its own
1390 * sentences. The same shape as the auth misclassification fixed in 0.4.2,
1391 * one branch further down the same function.
1392 *
1393 * So the run's own channels stay authoritative. `error_message` is the
1394 * CLI's own field rather than the model's words, and is read whole. The
1395 * answer is read only when it is short enough to *be* a notice: a run that
1396 * was really stopped has the notice and nothing else to say, in a couple
1397 * of lines, while an answer that discusses the subject runs to paragraphs.
1398 * Length is the one thing that separates them, because the vocabulary is
1399 * identical by definition.
1400 */
1401 let reported = terminal.error_message.clone().unwrap_or_default();
1402 let answered = if terminal.text.len() <= NOTICE_MAX {
1403 opening_lines(&terminal.text, OPENING_LINES)
1404 } else {
1405 String::new()
1406 };
1407 let prose = if terminal.text.is_empty() {
1408 format!("{reported}\n{}", opening_lines(stdout, OPENING_LINES))
1409 } else {
1410 format!("{reported}\n{answered}")
1411 };
1412 if quota_signalled || looks_rate_limited(stderr) || looks_rate_limited(&prose) {
1413 return Error::RateLimited {
1414 bin: bin.to_string(),
1415 message: first_meaningful_line(stderr)
1416 .or_else(|| first_meaningful_line(&prose))
1417 .unwrap_or_else(|| "usage limit reached".to_string()),
1418 };
1419 }
1420 // A rejected flag is not a failed request, it is this crate and the CLI
1421 // disagreeing about what the CLI accepts. Naming that is the difference
1422 // between "the run failed" and "your codex is a different version".
1423 if let Some(detail) = rejected_flag(stderr).or_else(|| rejected_flag(stdout)) {
1424 return Error::FlagRejected {
1425 bin: bin.to_string(),
1426 detail,
1427 };
1428 }
1429 // Checked before the generic failure but after quota and a rejected flag,
1430 // which are more specific readings of the same output.
1431 if terminal.stop == Stop::Error {
1432 return Error::AgentError {
1433 agent,
1434 bin: bin.to_string(),
1435 status: terminal.error_status,
1436 // Codex reports the reason apart from the answer; Claude puts it
1437 // where the answer would be.
1438 message: terminal
1439 .error_message
1440 .clone()
1441 .or_else(|| first_meaningful_line(&terminal.text))
1442 .or_else(|| first_meaningful_line(stderr))
1443 .unwrap_or_else(|| "the agent reported a failure without explaining it".into()),
1444 };
1445 }
1446
1447 Error::Failed {
1448 bin: bin.to_string(),
1449 code,
1450 // Fall back to stdout when stderr explains nothing. Codex reports a
1451 // rejected schema as an `{"type":"error"}` event on *stdout* while
1452 // stderr carries only "Reading additional input from stdin...", so
1453 // reporting stderr alone describes the failure as a status message.
1454 stderr: first_meaningful_line(stderr)
1455 .filter(|line| looks_explanatory(line))
1456 .or_else(|| first_meaningful_line(stdout))
1457 .or_else(|| first_meaningful_line(stderr))
1458 .unwrap_or_default(),
1459 }
1460}
1461
1462/// Whether a line plausibly explains a failure rather than narrating progress.
1463fn looks_explanatory(line: &str) -> bool {
1464 const NOISE: &[&str] = &[
1465 "reading additional input",
1466 "reading prompt",
1467 "waiting",
1468 "connecting",
1469 "loading",
1470 ];
1471 let lower = line.to_ascii_lowercase();
1472 !NOISE.iter().any(|noise| lower.contains(noise))
1473}
1474
1475/// The CLI's complaint, if it refused an argument.
1476///
1477/// The phrasings are clap's and commander's, which is what all three CLIs are
1478/// built on. Matched narrowly: a false positive would relabel a genuine failure
1479/// as a version problem and send someone chasing the wrong thing.
1480fn rejected_flag(text: &str) -> Option<String> {
1481 const REJECTIONS: &[&str] = &[
1482 "unexpected argument",
1483 "unknown option",
1484 "unrecognized option",
1485 "unknown flag",
1486 "invalid option",
1487 "unexpected option",
1488 ];
1489 let lower = text.to_ascii_lowercase();
1490 REJECTIONS
1491 .iter()
1492 .any(|needle| lower.contains(needle))
1493 .then(|| first_meaningful_line(text).unwrap_or_default())
1494}
1495
1496/// Whether text carries a provider quota refusal.
1497///
1498/// Deliberately a small set of unambiguous phrases: a false positive here would
1499/// relabel an ordinary failure as a quota problem and send a caller into a
1500/// pointless backoff.
1501fn looks_rate_limited(text: &str) -> bool {
1502 let lower = text.to_ascii_lowercase();
1503 [
1504 "rate limit",
1505 "rate_limit",
1506 "usage limit",
1507 "quota exceeded",
1508 "too many requests",
1509 ]
1510 .iter()
1511 .any(|needle| lower.contains(needle))
1512 // A status code is a word, and `429` as a bare substring is in every
1513 // line number, byte count, sha fragment and identifier that happens to
1514 // contain those digits. `401` was already given this treatment after it
1515 // matched inside a UUID and sent someone to re-login; this is the same
1516 // rule, applied to the code that had been left as a substring.
1517 || mentions_status(&lower, "429")
1518}
1519
1520/// The most useful line of a CLI's output for an error message.
1521///
1522/// Not simply the first non-blank one. CLIs open with progress and status
1523/// chatter, so the first line is often "Reading additional input from stdin..."
1524/// while the actual cause is further down. That turns a report into a
1525/// misdirection: it looks like an explanation and is not one.
1526///
1527/// So a line that looks like an error wins, and the first non-blank line is the
1528/// fallback when nothing does.
1529fn first_meaningful_line(text: &str) -> Option<String> {
1530 const ERROR_MARKERS: &[&str] = &[
1531 "error",
1532 "failed",
1533 "fatal",
1534 "panic",
1535 "denied",
1536 "invalid",
1537 "unexpected",
1538 "cannot",
1539 "unable",
1540 ];
1541 let lines: Vec<&str> = text
1542 .lines()
1543 .map(str::trim)
1544 .filter(|line| !line.is_empty())
1545 .collect();
1546
1547 lines
1548 .iter()
1549 .find(|line| {
1550 let lower = line.to_ascii_lowercase();
1551 ERROR_MARKERS.iter().any(|marker| lower.contains(marker))
1552 })
1553 .or_else(|| lines.first())
1554 .map(|line| (*line).to_string())
1555}
1556
1557/// Write the session binding back, reporting any store failure.
1558///
1559/// Called as soon as an id is known rather than only on a clean exit. Waiting
1560/// for success would lose the binding for exactly the runs where continuity
1561/// matters most: a timeout, a crash, or a cancelled turn.
1562fn persist_session(request: &Request, token: &str) -> Result<()> {
1563 let Some(binding) = &request.binding else {
1564 return Ok(());
1565 };
1566 binding
1567 .store
1568 .bind(request.agent, &binding.project, &binding.name, token)
1569 .map(|_| ())
1570}
1571
1572/// The id this run is already known by before it starts, if any.
1573///
1574/// Only a caller-assigned id qualifies: a printed id does not exist yet. This
1575/// is what makes an assigned session survive a run that never finishes.
1576fn preassigned_token(request: &Request) -> Option<String> {
1577 match &request.plan().cont {
1578 Continue::NewWith(id) => Some(id.clone()),
1579 _ => None,
1580 }
1581}
1582
1583/// Reported by an agent that exited cleanly but said nothing useful.
1584impl Outcome {
1585 /// Whether the agent produced any answer at all.
1586 #[must_use]
1587 pub fn is_empty(&self) -> bool {
1588 self.text.trim().is_empty() && self.stop == Stop::Completed
1589 }
1590}
1591
1592#[cfg(test)]
1593mod tests {
1594 use super::*;
1595 use crate::agent::Agent;
1596
1597 #[test]
1598 fn quota_phrases_are_recognized_and_ordinary_errors_are_not() {
1599 assert!(looks_rate_limited("Error: rate limit exceeded"));
1600 assert!(looks_rate_limited("HTTP 429 Too Many Requests"));
1601 assert!(looks_rate_limited("You have hit your usage limit"));
1602 // A plain failure must not be mistaken for a quota problem.
1603 assert!(!looks_rate_limited("error: no such file or directory"));
1604 assert!(!looks_rate_limited("model not found"));
1605 }
1606
1607 #[test]
1608 fn a_blocking_rate_limit_event_classifies_as_rate_limited() {
1609 let terminal = Terminal {
1610 rate_limit: Some(crate::outcome::RateLimit {
1611 status: "rejected".into(),
1612 window: Some("five_hour".into()),
1613 resets_at: None,
1614 overage_status: None,
1615 is_using_overage: None,
1616 }),
1617 ..Terminal::default()
1618 };
1619 assert!(matches!(
1620 classify(Agent::Claude, "claude", 1, "", "", &terminal),
1621 Error::RateLimited { .. }
1622 ));
1623 }
1624
1625 #[test]
1626 fn an_allowed_rate_limit_event_is_not_a_failure_cause() {
1627 let terminal = Terminal {
1628 rate_limit: Some(crate::outcome::RateLimit {
1629 status: "allowed".into(),
1630 window: None,
1631 resets_at: None,
1632 overage_status: None,
1633 is_using_overage: None,
1634 }),
1635 ..Terminal::default()
1636 };
1637 assert!(matches!(
1638 classify(Agent::Claude, "claude", 1, "boom", "", &terminal),
1639 Error::Failed { .. }
1640 ));
1641 }
1642
1643 /// The exact shape that made a Copilot run look unauthenticated: a UUID
1644 /// carrying the digits 401. Copilot prints several ids per run, so this
1645 /// misfired intermittently and told the user to re-login over a failure
1646 /// that had nothing to do with credentials.
1647 #[test]
1648 fn an_id_containing_401_is_not_an_auth_failure() {
1649 let line = r#"{"type":"session.mcp_server_status_changed","id":"1b0b1401-cb86-4276-9874-e84b94c96499"}"#;
1650 assert!(
1651 !looks_unauthenticated(line),
1652 "a hex blob is not a status code"
1653 );
1654 }
1655
1656 /// The needle still has to work where it was meant to. A status code is a
1657 /// word, and these are the forms an agent actually prints.
1658 #[test]
1659 fn a_real_401_is_still_recognized() {
1660 for text in [
1661 "HTTP 401",
1662 "request failed (status 401)",
1663 "401: unauthorized",
1664 "got a 401 from the API",
1665 ] {
1666 assert!(looks_unauthenticated(text), "should match: {text}");
1667 }
1668 }
1669
1670 /// Neighbouring digits mean it is part of some longer number, not a status.
1671 #[test]
1672 fn digits_around_401_keep_it_from_matching() {
1673 for text in ["error 4010", "code 1401", "seq 24019"] {
1674 assert!(!looks_unauthenticated(text), "should not match: {text}");
1675 }
1676 }
1677
1678 /// Verbatim from a healthy claude 2.1.205 run. Every `stream-json` run
1679 /// carries this record, and its status is `allowed`: nothing is refused.
1680 /// Scanning the raw stream for `rate_limit` matched it anyway, so any
1681 /// Claude failure was reported as a quota refusal, sending a caller to back
1682 /// off when the real cause was something they could fix.
1683 #[test]
1684 fn a_healthy_rate_limit_heartbeat_is_not_a_refusal() {
1685 let stdout = r#"{"type":"rate_limit_event","rate_limit_info":{"status":"allowed","resetsAt":1785331800,"rateLimitType":"five_hour","overageStatus":"rejected","isUsingOverage":false}}"#;
1686 let terminal = Terminal {
1687 stop: Stop::Error,
1688 error_status: Some(404),
1689 text: "There's an issue with the selected model (bogus-model-xyz).".into(),
1690 rate_limit: Some(crate::outcome::RateLimit {
1691 status: "allowed".into(),
1692 window: Some("five_hour".into()),
1693 resets_at: Some(1_785_331_800),
1694 overage_status: None,
1695 is_using_overage: None,
1696 }),
1697 ..Terminal::default()
1698 };
1699 let err = classify_run(Agent::Claude, "claude", 0, "", stdout, &terminal);
1700 assert!(
1701 matches!(err, Error::AgentError { .. }),
1702 "the heartbeat must not mask the real cause: {err:?}"
1703 );
1704 }
1705
1706 /// The counterpart: a refusal the parser did read must still be one, even
1707 /// though it arrives with the same zero exit code.
1708 #[test]
1709 fn a_rejected_quota_signal_is_still_a_refusal() {
1710 let terminal = Terminal {
1711 rate_limit: Some(crate::outcome::RateLimit {
1712 status: "rejected".into(),
1713 window: Some("five_hour".into()),
1714 resets_at: None,
1715 overage_status: None,
1716 is_using_overage: None,
1717 }),
1718 ..Terminal::default()
1719 };
1720 assert!(matches!(
1721 classify_run(Agent::Claude, "claude", 0, "", "", &terminal),
1722 Error::RateLimited { .. }
1723 ));
1724 }
1725
1726 /// Verbatim from a real run with an unknown model. Claude exits **0** with
1727 /// `subtype: "success"` while `is_error` is true and the explanation sits
1728 /// where the answer would be, so a caller checking only `Result::is_ok`
1729 /// renders "There's an issue with the selected model" as the answer.
1730 #[test]
1731 fn a_failed_turn_is_an_error_even_though_the_process_exited_cleanly() {
1732 let terminal = Terminal {
1733 stop: Stop::Error,
1734 error_status: Some(404),
1735 text: "There's an issue with the selected model (bogus-model-xyz). \
1736 It may not exist or you may not have access to it."
1737 .into(),
1738 ..Terminal::default()
1739 };
1740 let err = classify_run(Agent::Claude, "claude", 0, "", "", &terminal);
1741 let Error::AgentError {
1742 agent,
1743 status,
1744 message,
1745 ..
1746 } = &err
1747 else {
1748 panic!("expected AgentError, got {err:?}")
1749 };
1750 assert_eq!(*agent, Agent::Claude);
1751 assert_eq!(*status, Some(404), "the provider status must survive");
1752 assert!(message.contains("selected model"), "{message}");
1753 }
1754
1755 /// A quota refusal and a missing login are more specific readings of the
1756 /// same shape, so they must not be swallowed by the general case.
1757 #[test]
1758 fn a_failed_turn_does_not_mask_a_more_specific_cause() {
1759 let auth = Terminal {
1760 stop: Stop::Error,
1761 text: "Not logged in · Please run /login".into(),
1762 ..Terminal::default()
1763 };
1764 assert!(
1765 classify_run(Agent::Claude, "claude", 0, "", "", &auth).is_auth_failure(),
1766 "an unauthenticated failed turn must stay an auth failure"
1767 );
1768
1769 let quota = Terminal {
1770 stop: Stop::Error,
1771 rate_limit: Some(crate::outcome::RateLimit {
1772 status: "rejected".into(),
1773 window: None,
1774 resets_at: None,
1775 overage_status: None,
1776 is_using_overage: None,
1777 }),
1778 ..Terminal::default()
1779 };
1780 assert!(
1781 matches!(
1782 classify_run(Agent::Claude, "claude", 0, "", "", "a),
1783 Error::RateLimited { .. }
1784 ),
1785 "a quota-blocked failed turn must stay a rate limit"
1786 );
1787 }
1788
1789 /// Verified against the real CLI: with `USER` withheld, claude answers
1790 /// "Not logged in · Please run /login" and exits **0**. Checking only the
1791 /// exit code hands back a successful Outcome whose answer is a login
1792 /// prompt.
1793 #[test]
1794 fn an_unauthenticated_run_is_named_even_though_it_exits_zero() {
1795 let terminal = Terminal {
1796 text: "Not logged in · Please run /login".into(),
1797 ..Terminal::default()
1798 };
1799 let err = classify_run(Agent::Claude, "claude", 0, "", "", &terminal);
1800 let Error::NotAuthenticated { agent, hint, .. } = &err else {
1801 panic!("expected NotAuthenticated, got {err:?}")
1802 };
1803 assert_eq!(*agent, Agent::Claude);
1804 assert!(hint.contains("/login"), "{hint}");
1805 assert!(err.is_auth_failure());
1806 }
1807
1808 /// Verbatim from an unauthenticated Copilot run, captured by pointing it at
1809 /// an empty HOME. Its wording shares no phrase with Claude's or Codex's, so
1810 /// before this was observed the phrase list did not match it at all and a
1811 /// missing Copilot login was reported as a generic failure.
1812 #[test]
1813 fn copilots_own_unauthenticated_wording_is_recognized() {
1814 let stderr = "Error: No authentication information found.\n\n\
1815 Copilot can be authenticated with GitHub using an OAuth Token or a \
1816 Fine-Grained Personal Access Token.\n\n\
1817 To authenticate, you can use any of the following methods:\n\
1818 \u{2022} Start 'copilot' and run the '/login' command\n\
1819 \u{2022} Set the COPILOT_GITHUB_TOKEN, GH_TOKEN, or GITHUB_TOKEN \
1820 environment variable";
1821 let err = classify_run(
1822 Agent::Copilot,
1823 "copilot",
1824 1,
1825 stderr,
1826 "",
1827 &Terminal::default(),
1828 );
1829 let Error::NotAuthenticated { agent, hint, .. } = &err else {
1830 panic!("expected NotAuthenticated, got {err:?}")
1831 };
1832 assert_eq!(*agent, Agent::Copilot);
1833 assert!(hint.contains("copilot login"), "{hint}");
1834 }
1835
1836 /// Each agent's hint has to name its own login route, since they differ:
1837 /// Codex and Copilot have `login` subcommands, Claude does not.
1838 #[test]
1839 fn every_agent_offers_its_own_login_route() {
1840 for (agent, expected) in [
1841 (Agent::Claude, "setup-token"),
1842 (Agent::Codex, "codex login"),
1843 (Agent::Copilot, "copilot login"),
1844 ] {
1845 let err = classify_run(
1846 agent,
1847 agent.bin(),
1848 1,
1849 "error: unauthorized",
1850 "",
1851 &Terminal::default(),
1852 );
1853 let Error::NotAuthenticated { hint, .. } = &err else {
1854 panic!("{agent}: expected NotAuthenticated, got {err:?}")
1855 };
1856 assert!(hint.contains(expected), "{agent}: {hint}");
1857 }
1858 }
1859
1860 /// Reported from the field: a run was stopped, the user was told `claude`
1861 /// was not authenticated, and the answer was replaced by a login hint. The
1862 /// agent had been explaining why a `cargo publish` was refused, and its own
1863 /// prose contained the phrases this classifier looks for. An answer is not
1864 /// a diagnosis of the thing that produced it.
1865 #[test]
1866 fn an_agent_writing_about_authentication_is_not_an_auth_failure() {
1867 let answer = "The publish was refused before it ran.\n\n\
1868 What denied it was the auto mode classifier, not a missing \
1869 credential.\n\
1870 In auto mode there is no human to receive the prompt, so an \
1871 ask collapses into a refusal.\n\
1872 The message said the CLI was not authenticated, which is \
1873 unrelated: an unauthorized upload is exactly what the rule \
1874 is there to stop.";
1875 let terminal = Terminal {
1876 text: answer.into(),
1877 ..Terminal::default()
1878 };
1879 let err = classify_run(Agent::Claude, "claude", 1, "", answer, &terminal);
1880 assert!(
1881 !err.is_auth_failure(),
1882 "an answer that discusses auth was read as an auth failure: {err:?}"
1883 );
1884 }
1885
1886 /// The other half of the same rule: the notice itself still has to be
1887 /// caught, and it arrives as the agent's entire answer.
1888 #[test]
1889 fn the_notice_is_still_caught_when_it_is_the_whole_answer() {
1890 for text in [
1891 "Not logged in · Please run /login",
1892 // A banner first, which is why the opening is three lines deep.
1893 "claude 2.1.212\n\nNot logged in · Please run /login",
1894 ] {
1895 let terminal = Terminal {
1896 text: text.into(),
1897 ..Terminal::default()
1898 };
1899 let err = classify_run(Agent::Claude, "claude", 0, "", "", &terminal);
1900 assert!(
1901 err.is_auth_failure(),
1902 "{text:?} was not read as auth: {err:?}"
1903 );
1904 }
1905 }
1906
1907 /// The gate into the error path and the classifier behind it must read an
1908 /// answer by the same rule.
1909 ///
1910 /// Shaped after the report from the field: a healthy, finished turn about
1911 /// a stalled crate release, which mentions an auth error in a later
1912 /// paragraph because that was the subject. Read whole, as the gate used
1913 /// to, the phrase convicts the run; read at its opening, as everything
1914 /// now does, the answer is an answer. An agent with no credentials leads
1915 /// with the notice, which is what makes the opening the honest place to
1916 /// look.
1917 #[test]
1918 fn an_answer_mentioning_auth_late_does_not_open_the_error_path() {
1919 let answer = "So the duplicate pastes cost nothing.\n\
1920 The publish chain is done: 0.4.1 and 0.4.2 are both on the \
1921 registry and tagged.\n\
1922 Everything downstream already consumes them.\n\
1923 The only thing still open anywhere is the crate PR, \
1924 `pathscale/RustAgentAbstraction#18`, which is the auth error \
1925 that ate your reply: the run was reported as `not \
1926 authenticated` and the hint sent you to /login, while the \
1927 credentials were fine the whole time.";
1928 assert!(
1929 !answer_reports_no_credentials(answer),
1930 "an answer discussing auth opened the error path"
1931 );
1932 // And the notice itself, which is what the rule exists to catch.
1933 assert!(answer_reports_no_credentials(
1934 "Not logged in \u{b7} Please run /login"
1935 ));
1936 }
1937
1938 /// The backstop, which is what makes a false positive survivable at all.
1939 ///
1940 /// Every phrase check here is a heuristic over ordinary English and will
1941 /// be wrong eventually. When it is, the run reaches a classifier that
1942 /// cannot name any failure and returns the generic one. On a process that
1943 /// exited cleanly with a completed turn, that verdict is the absence of
1944 /// evidence rather than evidence, and the finished answer must stand.
1945 #[test]
1946 fn a_generic_failure_does_not_name_a_failure() {
1947 let unnamed = Error::Failed {
1948 bin: "claude".into(),
1949 code: 0,
1950 stderr: String::new(),
1951 };
1952 assert!(
1953 !names_a_failure(&unnamed),
1954 "a generic failure was treated as a diagnosis, which discards the answer"
1955 );
1956 // Everything a classifier can actually name still stands on its own.
1957 assert!(names_a_failure(&Error::RateLimited {
1958 bin: "claude".into(),
1959 message: "usage limit reached".into(),
1960 }));
1961 assert!(names_a_failure(&Error::NotAuthenticated {
1962 agent: Agent::Claude,
1963 bin: "claude".into(),
1964 message: "Not logged in".into(),
1965 hint: Agent::Claude.login_hint(),
1966 }));
1967 }
1968
1969 /// Reported from the field: a finished, successful turn that happened to be
1970 /// *about* usage limits classified its own run as blocked, and the banner
1971 /// quoted one of the answer's own sentences back as the provider's message.
1972 /// An answer is evidence about its topic, not about the run that produced it.
1973 #[test]
1974 fn an_agent_writing_about_limits_is_not_a_limit() {
1975 let answer = "The three retries cost nothing extra, so that is not where it \
1976 came from.\n\n\
1977 Providers do not rate limit on repetition: the limit is on \
1978 tokens per window, and a 429 is what you would see if one had \
1979 actually been reached. Each retry does re-send the whole \
1980 conversation, which is real spend, but spend is not the same \
1981 thing as a block and the run reported no blocking signal at \
1982 all.\n\
1983 Nothing here suggests the usage limit was reached, and the \
1984 banner quoted a sentence of this answer back as though a \
1985 provider had written it.";
1986 let terminal = Terminal {
1987 text: answer.into(),
1988 ..Terminal::default()
1989 };
1990 let err = classify(Agent::Claude, "claude", 1, "", answer, &terminal);
1991 assert!(
1992 !matches!(err, Error::RateLimited { .. }),
1993 "an answer discussing limits was read as one: {err:?}"
1994 );
1995 }
1996
1997 /// A status code is a word. These are the shapes that used to trip it.
1998 #[test]
1999 fn a_bare_429_in_prose_is_not_a_status_code() {
2000 assert!(!looks_rate_limited("see run.rs:4291 for the caller"));
2001 assert!(!looks_rate_limited("sha 8f429ac"));
2002 assert!(looks_rate_limited("HTTP 429"));
2003 assert!(looks_rate_limited("(status 429)"));
2004 assert!(looks_rate_limited("Error: too many requests"));
2005 }
2006
2007 /// Auth is the most specific reading, so it wins over a generic failure,
2008 /// but must not swallow unrelated errors.
2009 #[test]
2010 fn ordinary_failures_are_not_mistaken_for_auth_problems() {
2011 for stderr in [
2012 "error: no such file or directory",
2013 "model not found",
2014 "rate limit exceeded",
2015 "error: unexpected argument '--sandbox' found",
2016 ] {
2017 let err = classify_run(Agent::Codex, "codex", 1, stderr, "", &Terminal::default());
2018 assert!(
2019 !err.is_auth_failure(),
2020 "{stderr:?} was misread as an auth failure: {err:?}"
2021 );
2022 }
2023 }
2024
2025 /// The exact failure that cost a round of debugging: `codex exec resume`
2026 /// rejects `--sandbox`, which `Error::Failed` reported as a generic
2027 /// non-zero exit naming a flag rather than a version mismatch.
2028 #[test]
2029 fn a_rejected_flag_is_named_as_a_version_mismatch() {
2030 let err = classify(
2031 Agent::Codex,
2032 "codex",
2033 2,
2034 "error: unexpected argument '--sandbox' found",
2035 "",
2036 &Terminal::default(),
2037 );
2038 let Error::FlagRejected { bin, detail } = err else {
2039 panic!("expected FlagRejected, got {err:?}")
2040 };
2041 assert_eq!(bin, "codex");
2042 assert!(detail.contains("--sandbox"), "{detail}");
2043 }
2044
2045 #[test]
2046 fn ordinary_failures_are_not_mistaken_for_version_drift() {
2047 for stderr in [
2048 "error: no such file or directory",
2049 "model not found",
2050 "permission denied",
2051 ] {
2052 assert!(
2053 matches!(
2054 classify(Agent::Codex, "codex", 1, stderr, "", &Terminal::default()),
2055 Error::Failed { .. }
2056 ),
2057 "{stderr:?} should stay a plain failure"
2058 );
2059 }
2060 }
2061
2062 /// Real output from a failing codex run: the first line is status, the
2063 /// cause is below it. Reporting the first line looks like an explanation
2064 /// while pointing at the wrong thing.
2065 #[test]
2066 fn a_status_line_does_not_masquerade_as_the_cause() {
2067 let stderr = "Reading additional input from stdin...\n\
2068 error: invalid value 'nope' for '--sandbox <SANDBOX_MODE>'";
2069 let err = classify_run(Agent::Codex, "codex", 1, stderr, "", &Terminal::default());
2070 let Error::Failed {
2071 stderr: reported, ..
2072 } = err
2073 else {
2074 panic!("expected Failed, got {err:?}")
2075 };
2076 assert!(reported.contains("invalid value"), "reported {reported:?}");
2077 }
2078
2079 /// Codex reports a rejected schema as a JSON error event on **stdout**
2080 /// while stderr carries only a status line. Reporting stderr alone
2081 /// described the failure as "Reading additional input from stdin...",
2082 /// which is not what went wrong.
2083 #[test]
2084 fn a_cause_on_stdout_is_reported_when_stderr_only_narrates() {
2085 let stdout = r#"{"type":"error","message":"invalid_json_schema: 'additionalProperties' is required to be supplied and to be false."}"#;
2086 let err = classify_run(
2087 Agent::Codex,
2088 "codex",
2089 1,
2090 "Reading additional input from stdin...",
2091 stdout,
2092 &Terminal::default(),
2093 );
2094 let Error::Failed {
2095 stderr: reported, ..
2096 } = err
2097 else {
2098 panic!("expected Failed, got {err:?}")
2099 };
2100 assert!(
2101 reported.contains("additionalProperties"),
2102 "reported {reported:?}, which explains nothing"
2103 );
2104 }
2105
2106 #[test]
2107 fn failures_report_the_first_useful_line() {
2108 let err = classify(
2109 Agent::Claude,
2110 "claude",
2111 2,
2112 "\n\n real problem \nstack",
2113 "",
2114 &Terminal::default(),
2115 );
2116 let Error::Failed { code, stderr, .. } = err else {
2117 panic!("expected a plain failure")
2118 };
2119 assert_eq!(code, 2);
2120 assert_eq!(stderr, "real problem");
2121 }
2122
2123 /// Prompts and session ids ride the argv, and `Run::argv` invites logging
2124 /// it. The redacted form must keep the shape while dropping the content.
2125 #[test]
2126 fn redaction_removes_prompts_and_session_ids_but_keeps_flags() {
2127 let request = crate::Request::new(Agent::Claude, "my secret prompt")
2128 .system("secret system")
2129 .session_id("11111111-2222-3333-4444-555555555555");
2130 let safe = redact(&request.typed_argv().unwrap());
2131
2132 for secret in [
2133 "my secret prompt",
2134 "secret system",
2135 "11111111-2222-3333-4444-555555555555",
2136 ] {
2137 assert!(
2138 !safe.iter().any(|a| a.contains(secret)),
2139 "{secret:?} survived redaction: {safe:?}"
2140 );
2141 }
2142 // Still recognisable as the same command.
2143 assert_eq!(safe[0], "claude");
2144 assert!(safe.contains(&"--permission-mode".to_string()));
2145 assert!(safe.contains(&"--session-id".to_string()));
2146 }
2147
2148 #[test]
2149 fn codex_trailing_prompt_is_redacted_even_without_a_flag() {
2150 let request = crate::Request::new(Agent::Codex, "my secret prompt");
2151 let safe = redact(&request.typed_argv().unwrap());
2152 assert_eq!(safe.last().unwrap(), REDACTED);
2153 assert_eq!(safe[1], "exec", "the subcommand must survive");
2154 }
2155
2156 /// Redaction must cover the two shapes positional guesswork misses: Codex's
2157 /// bare trailing prompt, and raw arguments whose contents are unknowable.
2158 #[test]
2159 fn redaction_covers_positional_prompts_and_unchecked_arguments() {
2160 let request = crate::Request::new(Agent::Codex, "my secret prompt")
2161 .unchecked_args(["-c", "api_key=hunter2"]);
2162 let safe = redact(&request.typed_argv().unwrap());
2163 assert!(!safe.iter().any(|a| a.contains("my secret prompt")));
2164 assert!(
2165 !safe.iter().any(|a| a.contains("hunter2")),
2166 "unchecked arguments may hold secrets: {safe:?}"
2167 );
2168 assert_eq!(safe[1], "exec", "the subcommand must survive");
2169 }
2170
2171 /// A resume id is a capability: it continues someone's conversation.
2172 #[test]
2173 fn redaction_covers_the_codex_positional_resume_id() {
2174 let request = crate::Request::new(Agent::Codex, "hi").resume("thread-secret-9");
2175 let safe = redact(&request.typed_argv().unwrap());
2176 assert!(
2177 !safe.iter().any(|a| a.contains("thread-secret-9")),
2178 "{safe:?}"
2179 );
2180 assert!(safe.contains(&"resume".to_string()));
2181 }
2182
2183 /// `stream` is synchronous but spawns a task. Outside a runtime that would
2184 /// panic, which a `Result`-returning function must not do.
2185 #[test]
2186 fn stream_outside_a_runtime_errors_instead_of_panicking() {
2187 let err = stream(&crate::Request::new(Agent::Claude, "hi")).unwrap_err();
2188 assert!(matches!(err, Error::NoRuntime), "got {err:?}");
2189 }
2190
2191 #[tokio::test]
2192 async fn a_missing_binary_names_the_install_command() {
2193 let request = Request::new(Agent::Claude, "hi").bin("definitely-not-a-real-binary-xyz");
2194 let err = run(&request).await.unwrap_err();
2195 let Error::NotInstalled { hint, agent, .. } = err else {
2196 panic!("expected NotInstalled, got {err:?}")
2197 };
2198 assert_eq!(agent, Agent::Claude);
2199 assert!(hint.contains("claude-code"));
2200 }
2201
2202 #[test]
2203 fn transient_errors_are_distinguished_from_permanent_ones() {
2204 assert!(
2205 Error::RateLimited {
2206 bin: "claude".into(),
2207 message: String::new()
2208 }
2209 .is_transient()
2210 );
2211 assert!(
2212 !Error::NotInstalled {
2213 agent: Agent::Claude,
2214 bin: "claude".into(),
2215 hint: ""
2216 }
2217 .is_transient()
2218 );
2219 }
2220}