oneharness-core 0.3.1

Reusable engine behind the oneharness CLI: harness registry, hook rendering/installation, and harness config sync.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
//! Spawning harness subprocesses with timeouts and bounded parallelism.
//!
//! This layer only spawns and captures — it does no parsing, so the spawn path
//! and the extraction path stay independently testable. stdout and stderr are
//! drained on their own threads so a chatty harness can never deadlock the wait.

use std::path::PathBuf;
use std::process::{Command, Stdio};
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::Mutex;
use std::time::{Duration, Instant};

use wait_timeout::ChildExt;

use crate::domain::report::{Capture, Status};

/// A fully-specified subprocess to run.
#[derive(Clone)]
pub struct Job {
    pub argv: Vec<String>,
    pub cwd: Option<PathBuf>,
    pub env: Vec<(String, String)>,
    pub timeout: Duration,
}

/// One job's final result plus how many times it was invoked. `attempts` is 1
/// for a plain run, and more when a retry policy re-ran it (structured output).
pub struct Outcome {
    pub capture: Capture,
    pub attempts: u32,
}

/// Run `jobs` concurrently, at most `max_parallel` at a time, preserving order.
pub fn run_jobs(jobs: &[Job], max_parallel: usize) -> Vec<Capture> {
    run_jobs_with(jobs, max_parallel, |_, _, _| None)
        .into_iter()
        .map(|o| o.capture)
        .collect()
}

/// Like [`run_jobs`], but after each run consults `retry` to decide whether to
/// re-run the same job with a new argv. `retry(job_index, attempt, &capture)`
/// returns `Some(next_argv)` to run again (e.g. structured-output validation
/// failed, so re-prompt) or `None` to stop. `attempt` is the number of runs
/// completed so far (1 after the first), and the policy is responsible for its
/// own bound — the loop runs until it returns `None`. Only the argv changes
/// across attempts; cwd/env/timeout are reused. Returns each job's final capture
/// and total attempt count, in input order.
///
/// `retry` is `Sync` because it is shared across worker threads; the domain
/// validation it performs is pure, keeping this layer free of parsing logic.
pub fn run_jobs_with<F>(jobs: &[Job], max_parallel: usize, retry: F) -> Vec<Outcome>
where
    F: Fn(usize, u32, &Capture) -> Option<Vec<String>> + Sync,
{
    let n = jobs.len();
    if n == 0 {
        return Vec::new();
    }
    let workers = max_parallel.clamp(1, n);
    let next = AtomicUsize::new(0);
    let slots: Vec<Mutex<Option<Outcome>>> = (0..n).map(|_| Mutex::new(None)).collect();
    let retry = &retry;

    std::thread::scope(|scope| {
        for _ in 0..workers {
            scope.spawn(|| loop {
                let i = next.fetch_add(1, Ordering::SeqCst);
                if i >= n {
                    break;
                }
                let outcome = run_job_with_retry(&jobs[i], i, retry);
                *slots[i].lock().expect("slot mutex poisoned") = Some(outcome);
            });
        }
    });

    slots
        .into_iter()
        .map(|m| {
            m.into_inner()
                .expect("slot mutex poisoned")
                .expect("slot unfilled")
        })
        .collect()
}

/// Run one job, then loop while `retry` asks for another attempt with a new argv.
fn run_job_with_retry<F>(job: &Job, index: usize, retry: &F) -> Outcome
where
    F: Fn(usize, u32, &Capture) -> Option<Vec<String>>,
{
    let mut capture = run_job(job);
    let mut attempts = 1u32;
    while let Some(next_argv) = retry(index, attempts, &capture) {
        let next = Job {
            argv: next_argv,
            cwd: job.cwd.clone(),
            env: job.env.clone(),
            timeout: job.timeout,
        };
        capture = run_job(&next);
        attempts += 1;
    }
    Outcome { capture, attempts }
}

/// Resolve a program name to the binary actually spawned.
///
/// On Windows this is load-bearing: `CreateProcess` only auto-appends `.exe`, so
/// a bare name like `claude` never finds the `claude.cmd` shim that npm (and many
/// other installers) drop on `PATH`. Resolving via `which` is PATHEXT-aware, so it
/// locates the `.cmd`/`.bat` shim — and `std::process::Command`, given a path that
/// ends in `.cmd`/`.bat`, runs it through the command interpreter with the safe
/// argument escaping added in Rust 1.77 (CVE-2024-24576). Detection already uses
/// `which`, so a harness can be reported `available` yet still fail to spawn from a
/// bare name; this closes that gap. If resolution fails, fall back to the original
/// name so the spawn error stays accurate.
///
/// On non-Windows the name is returned unchanged: PATH lookup already handles
/// extensionless binaries, and we must not alter the established spawn behavior.
fn resolve_program(program: &str) -> std::ffi::OsString {
    #[cfg(windows)]
    {
        if let Ok(path) = which::which(program) {
            return path.into_os_string();
        }
    }
    program.into()
}

/// Decide the program and argument list to actually spawn for `argv`.
///
/// Normally this is just the [`resolve_program`]'d binary plus `argv[1..]`. On
/// Windows there is one rewrite: since Rust 1.77, `Command` runs a `.cmd`/`.bat`
/// through cmd.exe but refuses (with `InvalidInput`) any argument it cannot
/// escape for cmd.exe — a newline is the trigger. npm installs every JS harness
/// as a `claude.cmd` shim, so a multi-line argument (a rendered `--system`)
/// fails to spawn. When that exact situation is detected — a resolved `.cmd`/
/// `.bat` *and* a multi-line argument — parse the npm shim and invoke its real
/// target directly: a node interpreter plus script, or — as for claude-code,
/// whose bin is `bin/claude.exe` — the wrapped executable itself. That target is
/// a real `.exe`, so std's ordinary argument encoding carries the newline
/// through (only a `.cmd`/`.bat` goes through cmd.exe). Anything else
/// (single-line args, an unparseable shim, a non-shim `.cmd`) falls through
/// unchanged, so the established spawn path — and its error reporting — is byte
/// for byte what it was. The function is platform-shaped on purpose: the rewrite
/// only exists on Windows, and the pure shim parsing it relies on is tested on
/// every platform in `domain::shim`.
fn spawn_target(argv: &[String]) -> (std::ffi::OsString, Vec<String>) {
    let resolved = resolve_program(&argv[0]);
    let rest = argv[1..].to_vec();
    #[cfg(windows)]
    {
        if let Some(plan) = windows_shim_plan(std::path::Path::new(&resolved), &rest) {
            return plan;
        }
    }
    (resolved, rest)
}

/// The Windows shim rewrite for [`spawn_target`]: `Some((interpreter, args))`
/// when `resolved` is a `.cmd`/`.bat`, some argument is multi-line, and the file
/// parses as an npm shim; `None` (fall through to the `.cmd`) otherwise.
#[cfg(windows)]
fn windows_shim_plan(
    resolved: &std::path::Path,
    args: &[String],
) -> Option<(std::ffi::OsString, Vec<String>)> {
    // Only act when std would actually refuse the spawn: a multi-line argument.
    if !args.iter().any(|a| a.contains('\n') || a.contains('\r')) {
        return None;
    }
    let ext = resolved.extension()?.to_str()?.to_ascii_lowercase();
    if ext != "cmd" && ext != "bat" {
        return None;
    }
    let contents = std::fs::read_to_string(resolved).ok()?;
    let dir = resolved.parent()?.to_str()?;
    let target = crate::domain::shim::parse_cmd_shim(&contents, dir)?;
    let mut full = target.prefix_args;
    full.extend_from_slice(args);
    Some((resolve_program(&target.interpreter), full))
}

/// Run a single job, returning its raw capture. Never panics on harness behavior.
pub fn run_job(job: &Job) -> Capture {
    let start = Instant::now();
    let (program, args) = spawn_target(&job.argv);
    let mut command = Command::new(program);
    command
        .args(&args)
        .stdin(Stdio::null())
        .stdout(Stdio::piped())
        .stderr(Stdio::piped());
    if let Some(cwd) = &job.cwd {
        command.current_dir(cwd);
        // Mirror a shell `cd`: keep $PWD consistent with the working directory.
        // `current_dir` only chdir()s the child; the inherited $PWD stays stale.
        // Some tools (e.g. Bun-based CLIs like OpenCode) trust $PWD over getcwd()
        // to locate the project, so a stale $PWD points them at the wrong dir.
        // Use the logical path (no symlink resolution), like `cd` does.
        let pwd = if cwd.is_absolute() {
            cwd.clone()
        } else {
            std::env::current_dir()
                .map(|base| base.join(cwd))
                .unwrap_or_else(|_| cwd.clone())
        };
        command.env("PWD", pwd);
    }
    // Explicit --env entries win over the derived PWD above.
    for (key, value) in &job.env {
        command.env(key, value);
    }

    let mut child = match command.spawn() {
        Ok(child) => child,
        Err(err) => {
            return Capture {
                status: Status::SpawnError,
                exit_code: None,
                duration_ms: Some(start.elapsed().as_millis()),
                stdout: String::new(),
                stderr: String::new(),
                error: Some(format!(
                    "failed to spawn `{}`: {err}. Suggestion: check the binary exists and is executable (try `oneharness detect`)",
                    job.argv[0]
                )),
            };
        }
    };

    // Drain both pipes on their own threads so wait never blocks on a full buffer.
    let mut out = child.stdout.take().expect("piped stdout");
    let mut err = child.stderr.take().expect("piped stderr");
    let out_reader = std::thread::spawn(move || read_all(&mut out));
    let err_reader = std::thread::spawn(move || read_all(&mut err));

    let (status, exit_code, timed_out) = match child.wait_timeout(job.timeout) {
        Ok(Some(exit)) => {
            let code = exit.code();
            let status = if code == Some(0) {
                Status::Ok
            } else {
                Status::Nonzero
            };
            (status, code, false)
        }
        Ok(None) => {
            let _ = child.kill();
            let _ = child.wait();
            (Status::Timeout, None, true)
        }
        Err(_) => {
            let _ = child.kill();
            let _ = child.wait();
            (Status::SpawnError, None, false)
        }
    };

    let stdout = out_reader.join().unwrap_or_default();
    let stderr = err_reader.join().unwrap_or_default();
    let duration_ms = Some(start.elapsed().as_millis());

    let error = if timed_out {
        Some(format!(
            "`{}` exceeded the {}s timeout and was killed. Suggestion: raise --timeout or simplify the prompt",
            job.argv[0],
            job.timeout.as_secs()
        ))
    } else if status == Status::SpawnError {
        Some(format!("`{}` could not be waited on", job.argv[0]))
    } else {
        None
    };

    Capture {
        status,
        exit_code,
        duration_ms,
        stdout,
        stderr,
        error,
    }
}

fn read_all<R: std::io::Read>(reader: &mut R) -> String {
    let mut buf = Vec::new();
    let _ = std::io::Read::read_to_end(reader, &mut buf);
    String::from_utf8_lossy(&buf).into_owned()
}

/// What a streaming line callback asks the run to do next.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum StreamStep {
    /// Keep reading the harness's output.
    Continue,
    /// Stop now and tear down the child — the consumer has gone away (e.g. a
    /// broken stdout pipe: it short-circuited on an observed action).
    Stop,
}

/// Run one job, invoking `on_line` for each complete line of stdout **as it
/// arrives**, and return the same [`Capture`] the batch path would (accumulated
/// stdout/stderr, status, timing) so the caller can still emit a final envelope.
///
/// This is the streaming counterpart to [`run_job`]: it exists so a consumer can
/// observe a harness's normalized events incrementally and short-circuit the
/// moment it sees a disallowed action — instead of paying for a whole turn before
/// judging it. The parsing stays out of this layer: `on_line` (a pure
/// domain-driven closure in the command layer) decides what to emit and returns
/// [`StreamStep::Stop`] to end early (the command layer returns `Stop` when its
/// write to the consumer fails, i.e. the consumer closed the stream). On `Stop`
/// or timeout the child is killed; on normal EOF its exit is awaited. Never
/// panics on harness behavior — same contract as [`run_job`].
pub fn run_job_streaming<F>(job: &Job, mut on_line: F) -> Capture
where
    F: FnMut(&str) -> StreamStep,
{
    use std::io::BufRead;
    use std::sync::mpsc;

    let start = Instant::now();
    let (program, args) = spawn_target(&job.argv);
    let mut command = Command::new(program);
    command
        .args(&args)
        .stdin(Stdio::null())
        .stdout(Stdio::piped())
        .stderr(Stdio::piped());
    if let Some(cwd) = &job.cwd {
        command.current_dir(cwd);
        let pwd = if cwd.is_absolute() {
            cwd.clone()
        } else {
            std::env::current_dir()
                .map(|base| base.join(cwd))
                .unwrap_or_else(|_| cwd.clone())
        };
        command.env("PWD", pwd);
    }
    for (key, value) in &job.env {
        command.env(key, value);
    }

    let mut child = match command.spawn() {
        Ok(child) => child,
        Err(err) => {
            return Capture {
                status: Status::SpawnError,
                exit_code: None,
                duration_ms: Some(start.elapsed().as_millis()),
                stdout: String::new(),
                stderr: String::new(),
                error: Some(format!(
                    "failed to spawn `{}`: {err}. Suggestion: check the binary exists and is executable (try `oneharness detect`)",
                    job.argv[0]
                )),
            };
        }
    };

    let out = child.stdout.take().expect("piped stdout");
    let mut err = child.stderr.take().expect("piped stderr");
    let err_reader = std::thread::spawn(move || read_all(&mut err));

    // A reader thread turns blocking line reads into channel messages, so the
    // main loop can honor the wall-clock timeout (recv_timeout) between lines
    // without non-blocking I/O. Each message is one line *with* its terminator
    // preserved, so the accumulated stdout is byte-faithful to the batch path.
    let (tx, rx) = mpsc::channel::<String>();
    let out_reader = std::thread::spawn(move || {
        let mut reader = std::io::BufReader::new(out);
        loop {
            let mut line = String::new();
            match reader.read_line(&mut line) {
                Ok(0) => break, // EOF
                Ok(_) => {
                    if tx.send(line).is_err() {
                        break; // main loop gone
                    }
                }
                Err(_) => break,
            }
        }
    });

    let deadline = start + job.timeout;
    let mut stdout = String::new();
    let mut stopped = false;
    let mut timed_out = false;
    loop {
        let now = Instant::now();
        if now >= deadline {
            timed_out = true;
            break;
        }
        match rx.recv_timeout(deadline - now) {
            Ok(line) => {
                stdout.push_str(&line);
                // Feed the callback the line without its trailing newline(s).
                if on_line(line.trim_end_matches(['\n', '\r'])) == StreamStep::Stop {
                    stopped = true;
                    break;
                }
            }
            Err(mpsc::RecvTimeoutError::Timeout) => {
                timed_out = true;
                break;
            }
            Err(mpsc::RecvTimeoutError::Disconnected) => break, // EOF
        }
    }

    let (status, exit_code) = if stopped || timed_out {
        let _ = child.kill();
        let _ = child.wait();
        if timed_out {
            (Status::Timeout, None)
        } else {
            // Consumer-driven stop: not a harness failure. Report Ok so the
            // (best-effort) final envelope isn't misread as an error; the
            // consumer already has what it needed.
            (Status::Ok, None)
        }
    } else {
        match child.wait() {
            Ok(exit) => {
                let code = exit.code();
                let status = if code == Some(0) {
                    Status::Ok
                } else {
                    Status::Nonzero
                };
                (status, code)
            }
            Err(_) => (Status::SpawnError, None),
        }
    };

    // Drain any remaining buffered lines the reader already had, so the final
    // envelope's stdout is complete even if we broke out on stop/timeout.
    let _ = out_reader.join();
    while let Ok(line) = rx.try_recv() {
        stdout.push_str(&line);
    }
    let stderr = err_reader.join().unwrap_or_default();

    let error = if timed_out {
        Some(format!(
            "`{}` exceeded the {}s timeout and was killed. Suggestion: raise --timeout or simplify the prompt",
            job.argv[0],
            job.timeout.as_secs()
        ))
    } else {
        None
    };

    Capture {
        status,
        exit_code,
        duration_ms: Some(start.elapsed().as_millis()),
        stdout,
        stderr,
        error,
    }
}

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

    fn job(argv: &[&str]) -> Job {
        Job {
            argv: argv.iter().map(|s| s.to_string()).collect(),
            cwd: None,
            env: Vec::new(),
            timeout: Duration::from_secs(5),
        }
    }

    #[test]
    fn empty_jobs_returns_empty_without_spawning_workers() {
        // The no-work fast path: no jobs means no captures and no threads.
        assert!(run_jobs(&[], 4).is_empty());
    }

    #[test]
    fn spawn_error_is_data_not_a_panic() {
        // A binary that cannot be spawned must surface as a `SpawnError` capture
        // with a helpful message — never a crash. Run it through `run_jobs` so the
        // worker-pool path that fills the result slot is exercised too.
        let jobs = [job(&["/no/such/oneharness-binary-xyz", "arg"])];
        let captures = run_jobs(&jobs, 1);
        assert_eq!(captures.len(), 1);
        let cap = &captures[0];
        assert_eq!(cap.status, Status::SpawnError);
        assert!(cap.exit_code.is_none());
        assert!(cap.stdout.is_empty());
        assert!(cap.duration_ms.is_some());
        let msg = cap.error.as_deref().unwrap_or_default();
        assert!(msg.contains("failed to spawn"), "{msg}");
        assert!(msg.contains("oneharness-binary-xyz"), "{msg}");
    }

    #[test]
    fn run_jobs_with_retries_until_the_policy_stops() {
        // The structured-output loop in disguise: re-run with a fresh argv until
        // the policy returns None. Uses unspawnable binaries so it stays portable
        // and process-free, asserting on the attempt count and that the *final*
        // capture came from the last retry's argv (its error names that binary).
        let jobs = [job(&["/no/such/first"])];
        let outcomes = run_jobs_with(&jobs, 1, |i, attempt, cap| {
            assert_eq!(i, 0);
            assert_eq!(cap.status, Status::SpawnError);
            (attempt < 3).then(|| vec![format!("/no/such/retry-{attempt}")])
        });
        assert_eq!(outcomes.len(), 1);
        assert_eq!(outcomes[0].attempts, 3);
        let err = outcomes[0].capture.error.as_deref().unwrap_or_default();
        assert!(
            err.contains("retry-2"),
            "final capture should be last retry: {err}"
        );
    }

    #[test]
    fn run_jobs_with_a_no_op_policy_runs_once() {
        let jobs = [job(&["/no/such/binary"])];
        let outcomes = run_jobs_with(&jobs, 1, |_, _, _| None);
        assert_eq!(outcomes[0].attempts, 1);
        // Empty input is still the no-work fast path.
        assert!(run_jobs_with(&[], 4, |_, _, _| None).is_empty());
    }

    /// A portable job that prints three lines then exits 0 (`sh -c printf` on
    /// Unix, `cmd /c echo` on Windows), for exercising the streaming reader.
    fn three_line_job() -> Job {
        #[cfg(not(windows))]
        let argv = vec![
            "sh".to_string(),
            "-c".to_string(),
            "printf 'a\\nb\\nc\\n'".to_string(),
        ];
        #[cfg(windows)]
        let argv = vec![
            "cmd".to_string(),
            "/c".to_string(),
            "echo a& echo b& echo c".to_string(),
        ];
        Job {
            argv,
            cwd: None,
            env: Vec::new(),
            timeout: Duration::from_secs(10),
        }
    }

    #[test]
    fn streaming_delivers_each_line_and_accumulates_stdout() {
        // The happy path: every line reaches the callback in order, and the final
        // capture's stdout is the byte-faithful accumulation, status Ok.
        let mut lines = Vec::new();
        let cap = run_job_streaming(&three_line_job(), |line| {
            lines.push(line.to_string());
            StreamStep::Continue
        });
        assert_eq!(cap.status, Status::Ok);
        assert_eq!(cap.exit_code, Some(0));
        assert_eq!(lines.len(), 3, "got {lines:?}");
        // Trim to tolerate any shell quirks; the content is a/b/c in order.
        let trimmed: Vec<_> = lines.iter().map(|l| l.trim()).collect();
        assert_eq!(trimmed, vec!["a", "b", "c"]);
        for token in ["a", "b", "c"] {
            assert!(cap.stdout.contains(token), "stdout: {:?}", cap.stdout);
        }
    }

    #[test]
    fn streaming_stops_and_tears_down_on_callback_stop() {
        // Returning Stop after the first line ends the run immediately (the child
        // is killed); the consumer-driven stop is reported as Ok, not a failure.
        let mut count = 0u32;
        let cap = run_job_streaming(&three_line_job(), |_| {
            count += 1;
            StreamStep::Stop
        });
        assert_eq!(count, 1, "should stop after the first line");
        assert_eq!(cap.status, Status::Ok);
    }

    #[test]
    fn streaming_spawn_error_is_data_not_a_panic() {
        // A missing binary surfaces as a SpawnError capture, same as run_job.
        let job = job(&["/no/such/oneharness-stream-binary"]);
        let cap = run_job_streaming(&job, |_| StreamStep::Continue);
        assert_eq!(cap.status, Status::SpawnError);
        assert!(cap.error.as_deref().unwrap_or_default().contains("spawn"));
    }

    #[test]
    fn resolve_program_falls_back_to_the_name_when_unresolvable() {
        // A name that PATH lookup cannot resolve must come back unchanged on every
        // platform, so the spawn attempt — and its error message — stay accurate.
        let name = "oneharness-no-such-binary-zzz";
        assert_eq!(resolve_program(name), std::ffi::OsString::from(name));
    }

    #[cfg(windows)]
    #[test]
    fn windows_shim_plan_rewrites_a_cmd_only_for_multiline_args() {
        use std::io::Write;

        // A minimal npm-style `.cmd` shim written to a real temp file, so the
        // plan reads it the way `run_job` would.
        let dir = std::env::temp_dir().join(format!("oh-shim-{}", std::process::id()));
        std::fs::create_dir_all(&dir).unwrap();
        let cmd_path = dir.join("claude.cmd");
        let mut f = std::fs::File::create(&cmd_path).unwrap();
        write!(
            f,
            "SET \"_prog=node\"\r\n\"%_prog%\" \"%dp0%\\cli.js\" %*\r\n"
        )
        .unwrap();
        drop(f);

        let dir_str = dir.to_str().unwrap();
        let script = format!("{dir_str}\\cli.js");

        // A multi-line argument triggers the rewrite: node + script + the args.
        let multiline = vec!["-p".to_string(), "a\nb\nc".to_string()];
        let (prog, args) = windows_shim_plan(&cmd_path, &multiline).expect("multiline → rewrite");
        // `node` resolves to a real path ending in node.exe.
        assert!(
            std::path::Path::new(&prog)
                .to_string_lossy()
                .to_ascii_lowercase()
                .ends_with("node.exe"),
            "interpreter should resolve to node.exe, got {prog:?}"
        );
        assert_eq!(args, vec![script, "-p".to_string(), "a\nb\nc".to_string()]);

        // A single-line argument must NOT be rewritten — the `.cmd` spawns fine.
        let single = vec!["-p".to_string(), "hello".to_string()];
        assert!(windows_shim_plan(&cmd_path, &single).is_none());

        std::fs::remove_dir_all(&dir).ok();
    }

    #[cfg(windows)]
    #[test]
    fn windows_shim_plan_ignores_non_batch_programs() {
        // A real `.exe` (or anything not `.cmd`/`.bat`) is never rewritten, even
        // with a multi-line argument — std spawns it directly without cmd.exe.
        let exe = resolve_program("where");
        let multiline = vec!["x\ny".to_string()];
        assert!(windows_shim_plan(std::path::Path::new(&exe), &multiline).is_none());
    }

    #[cfg(windows)]
    #[test]
    fn resolve_program_finds_a_cmd_shim_on_windows() {
        // `where` is a stock Windows console command living at a real path; the
        // bare name must resolve to an absolute, existing file (the gap that left
        // npm `.cmd` shims unspawnable from a bare name).
        let resolved = resolve_program("where");
        let path = std::path::Path::new(&resolved);
        assert!(
            path.is_absolute(),
            "expected an absolute path, got {resolved:?}"
        );
        assert!(path.exists(), "resolved path should exist: {resolved:?}");
    }
}