zift 0.1.6

Scan codebases for embedded authorization logic and generate Policy as Code (Rego/OPA today)
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
//! Subprocess hook transport (Tier 3 of the deep-mode design).
//!
//! Spawns a user-supplied shell command per request, writes a single
//! JSON envelope to its stdin, and reads the deep-mode response schema
//! from its stdout. The escape hatch for any agent that doesn't fit the
//! MCP server (Tier 1) or the OpenAI-compatible HTTP client (Tier 2):
//!
//! ```text
//! claude -p --output-format json
//! aider --no-auto-commits
//! ./my-wrapper.sh         # arbitrary user script
//! ```
//!
//! Wire format on stdin (one line, then EOF):
//!
//! ```json
//! {"system": "...", "user": "...", "schema": { ... }}
//! ```
//!
//! Wire format expected on stdout (parsed verbatim, optional markdown
//! fence stripped):
//!
//! ```json
//! {"findings": [{"line_start": 12, "line_end": 18, ...}]}
//! ```
//!
//! Both shapes are identical to the HTTP transport's contract — that's
//! deliberate: agent CLIs that wrap real LLMs can route system/user
//! straight through, and we never fork the schema between transports.
//!
//! ## Cost tracking
//!
//! N/A. Subprocess agents don't return token counts in any standard
//! way; [`crate::deep::analyzer::TokenUsage::default`] short-circuits
//! the cost tracker. Users wanting a ceiling enforce it externally
//! (timeouts, ulimits, wrapper scripts).
//!
//! ## Security
//!
//! The user supplies an arbitrary shell command. If `.zift.toml` is
//! checked in and Zift is run by another user (CI, shared dev box),
//! that's a footgun — same threat as `.editorconfig`-style attacks. We
//! document the risk; we don't sandbox.

use crate::deep::analyzer::{AnalyzeResponse, Analyzer, TokenUsage};
use crate::deep::client::{strip_markdown_fence, truncate_for_log};
use crate::deep::config::DeepRuntime;
use crate::deep::error::DeepError;
use crate::deep::finding::SemanticFinding;
use crate::deep::prompt::RenderedPrompt;
use serde::Deserialize;
use std::io::{Read, Write};
use std::process::{Command, Stdio};
use std::sync::mpsc;
use std::thread;
use std::time::{Duration, Instant};

/// `Debug` is implemented manually so `Result<Self, _>::unwrap_err`
/// works in tests (the std `unwrap_err` requires `Self: Debug`) without
/// printing the raw command string. Users sometimes inline API keys or
/// bearer tokens directly in `agent_cmd` (e.g.
/// `claude -p --api-key sk-...`); a derived `Debug` would echo those
/// secrets through any panic, `unwrap_err`, or `?`-bubbled error log.
pub struct SubprocessClient {
    /// Shell command line, as supplied by the user. Passed to the
    /// platform shell (`sh -c` on Unix, `cmd /C` on Windows). Treated
    /// as potentially-sensitive — never formatted into errors or logs.
    cmd: String,
    /// Wall-clock ceiling for one request. On expiry the child is
    /// killed and [`DeepError::Timeout`] is returned.
    timeout: Duration,
}

impl std::fmt::Debug for SubprocessClient {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("SubprocessClient")
            .field("cmd", &"<redacted>")
            .field("timeout", &self.timeout)
            .finish()
    }
}

impl SubprocessClient {
    pub fn new(runtime: &DeepRuntime) -> Result<Self, DeepError> {
        let cmd = runtime
            .agent_cmd
            .clone()
            .ok_or_else(|| {
                // Belt-and-suspenders: `deep::config::build` already
                // validates this. If we ever reach here, the runtime
                // was hand-constructed with a bug — fail loud.
                DeepError::Config(
                    "subprocess analyzer constructed without agent_cmd \
                     (runtime invariant violated)"
                        .into(),
                )
            })?
            .trim()
            .to_string();
        if cmd.is_empty() {
            return Err(DeepError::Config(
                "subprocess agent_cmd is empty after trim".into(),
            ));
        }
        Ok(Self {
            cmd,
            timeout: Duration::from_secs(runtime.agent_timeout_secs),
        })
    }

    /// Spawn the agent, write the JSON envelope, read stdout to EOF,
    /// parse, and return.
    fn run_once(&self, prompt: &RenderedPrompt) -> Result<AnalyzeResponse, DeepError> {
        let envelope = build_envelope(prompt);

        // Spawn through the platform shell so users can supply pipelines
        // (`claude -p | jq ...`). On Unix that's `sh -c <cmd>`; on
        // Windows it's `cmd /C <cmd>`.
        let mut child = shell_command(&self.cmd)
            .stdin(Stdio::piped())
            .stdout(Stdio::piped())
            .stderr(Stdio::piped())
            .spawn()
            .map_err(|e| {
                // ENOENT / permission errors at spawn time are
                // operator-actionable misconfiguration (typo in
                // `agent_cmd`, missing binary), not transient. Surface
                // as `Config` so the orchestrator hard-fails the whole
                // deep run rather than silently skipping every
                // candidate — every spawn would fail identically.
                //
                // We deliberately do NOT include `self.cmd` in the
                // message: users sometimes inline API keys/tokens in
                // the command string, and this error can be logged or
                // surfaced verbatim by callers. The OS error itself
                // ("No such file or directory", "Permission denied")
                // is enough to diagnose typo/missing-binary cases.
                DeepError::Config(format!("failed to spawn agent_cmd: {e}"))
            })?;

        let mut stdin = child
            .stdin
            .take()
            .expect("stdin pipe was requested via Stdio::piped");
        let stdout = child
            .stdout
            .take()
            .expect("stdout pipe was requested via Stdio::piped");
        let stderr = child
            .stderr
            .take()
            .expect("stderr pipe was requested via Stdio::piped");

        // Writer thread: writes the envelope and drops stdin (closes
        // the pipe so the child sees EOF). On its own thread because
        // pipes block at ~64KB on Linux when the child isn't reading;
        // a synchronous write_all could deadlock against a buggy
        // agent that exits without consuming stdin.
        let envelope_bytes = envelope.into_bytes();
        let writer = thread::spawn(move || -> std::io::Result<()> {
            stdin.write_all(&envelope_bytes)?;
            stdin.flush()?;
            // Drop stdin → pipe closes → child sees EOF.
            drop(stdin);
            Ok(())
        });

        // Reader threads for stdout/stderr. Both must run on
        // background threads for the same backpressure reason as the
        // writer: a chatty agent that fills the pipe before exiting
        // would otherwise deadlock with the writer.
        let (stdout_tx, stdout_rx) = mpsc::channel::<std::io::Result<String>>();
        let _stdout_thread = thread::spawn(move || {
            let mut buf = String::new();
            let mut handle = stdout;
            let res = handle.read_to_string(&mut buf).map(|_| buf);
            // Receiver may have hung up if we timed out — ignore.
            let _ = stdout_tx.send(res);
        });

        let (stderr_tx, stderr_rx) = mpsc::channel::<String>();
        let _stderr_thread = thread::spawn(move || {
            let mut buf = String::new();
            let mut handle = stderr;
            let _ = handle.read_to_string(&mut buf);
            let _ = stderr_tx.send(buf);
        });

        // Bound the wait by polling `try_wait`. Without `wait_timeout`
        // this is the simplest portable approach; 50ms granularity is
        // fine since real agent latencies are seconds-to-minutes.
        let start = Instant::now();
        let exit = loop {
            match child.try_wait() {
                Ok(Some(status)) => break status,
                Ok(None) => {
                    if start.elapsed() >= self.timeout {
                        // Kill the entire process tree (group on Unix),
                        // reap the immediate child, and let the reader
                        // threads drain as the pipes close. Killing the
                        // whole group matters because `sh -c 'cmd'` on
                        // Linux dash forks `cmd` rather than execing
                        // into it — leaving the immediate `sh` reaped
                        // but `cmd` orphaned with our pipes open.
                        #[cfg(unix)]
                        kill_process_tree(&child);
                        #[cfg(not(unix))]
                        kill_process_tree(&mut child);
                        let _ = child.wait();
                        let _ = writer.join();
                        // Bound the drain so a misbehaving descendant
                        // that somehow survived `SIGKILL` (unkillable
                        // kernel state, ptrace stop, etc.) cannot hang
                        // the analyzer. 500ms is well past the kernel's
                        // signal-delivery latency in practice.
                        let drain_timeout = Duration::from_millis(500);
                        let _ = stdout_rx.recv_timeout(drain_timeout);
                        let _ = stderr_rx.recv_timeout(drain_timeout);
                        return Err(DeepError::Timeout {
                            secs: self.timeout.as_secs(),
                        });
                    }
                    thread::sleep(Duration::from_millis(50));
                }
                Err(e) => {
                    return Err(DeepError::Io(e));
                }
            }
        };

        // Drain background threads. EPIPE on writer is OK if the child
        // exited before reading stdin (e.g. "help" mode that ignores
        // input); only log it.
        if let Ok(Err(e)) = writer.join() {
            tracing::debug!("subprocess: writer error (likely EPIPE on early exit): {e}");
        }

        // Bound stdout/stderr reads by the remaining wall-clock budget.
        // `try_wait` above only watches the immediate shell child, so a
        // wrapper like `sh -c 'sleep 30 & printf "{...}"'` makes the
        // shell exit promptly while a backgrounded grandchild keeps
        // our pipes open. An unbounded `recv()` would then hang past
        // `agent_timeout_secs`. Using `recv_timeout(remaining)` keeps
        // the wall-clock contract intact end-to-end.
        let remaining = self.timeout.saturating_sub(start.elapsed());
        let stdout_buf = match stdout_rx.recv_timeout(remaining) {
            Ok(res) => res.map_err(DeepError::Io)?,
            Err(mpsc::RecvTimeoutError::Timeout) => {
                // Same teardown as the in-loop timeout branch: kill the
                // process group so any backgrounded descendant releases
                // our pipes, then drain stderr briefly.
                #[cfg(unix)]
                kill_process_tree(&child);
                #[cfg(not(unix))]
                kill_process_tree(&mut child);
                let _ = stderr_rx.recv_timeout(Duration::from_millis(500));
                return Err(DeepError::Timeout {
                    secs: self.timeout.as_secs(),
                });
            }
            Err(mpsc::RecvTimeoutError::Disconnected) => {
                return Err(DeepError::BadResponse(
                    "subprocess stdout reader disconnected".into(),
                ));
            }
        };
        // Stderr is best-effort: cap at a short timeout regardless of
        // remaining budget so a stuck stderr pipe (rare, but possible
        // with weird LD_PRELOAD shims) can't extend the request.
        let stderr_buf = stderr_rx
            .recv_timeout(Duration::from_millis(500))
            .unwrap_or_default();

        if !exit.success() {
            // Surface as `BadResponse` so the orchestrator skips this
            // candidate but keeps going — same per-candidate-skip path
            // as malformed JSON. Avoid leaking stderr verbatim into
            // the error message (it can echo prompt text); cap and log
            // to debug instead.
            tracing::debug!(
                exit = ?exit,
                stderr = %truncate_for_log(&stderr_buf),
                stdout = %truncate_for_log(&stdout_buf),
                "subprocess: agent_cmd exited nonzero",
            );
            return Err(DeepError::BadResponse(format!(
                "agent_cmd exited with {} (no findings parsed)",
                exit_status_brief(&exit),
            )));
        }

        // Parse stdout as our findings envelope. Same fence-stripping
        // and same truncated-debug-log discipline as the HTTP client —
        // model-or-CLI output may mirror prompt text and should not
        // appear verbatim in error strings.
        let cleaned = strip_markdown_fence(&stdout_buf);
        let parsed: FindingsEnvelope = serde_json::from_str(cleaned).map_err(|e| {
            tracing::debug!(
                error = %e,
                preview = %truncate_for_log(&stdout_buf),
                stderr_preview = %truncate_for_log(&stderr_buf),
                "subprocess: stdout was not valid findings JSON",
            );
            DeepError::BadResponse("agent_cmd output was not valid findings JSON".into())
        })?;

        Ok(AnalyzeResponse {
            findings: parsed.findings,
            // Subprocess agents don't report tokens in a standard way.
            // The cost tracker short-circuits on default usage.
            usage: TokenUsage::default(),
        })
    }
}

impl Analyzer for SubprocessClient {
    fn analyze(&self, prompt: &RenderedPrompt) -> Result<AnalyzeResponse, DeepError> {
        self.run_once(prompt)
    }
}

/// JSON envelope written verbatim to the agent's stdin.
fn build_envelope(prompt: &RenderedPrompt) -> String {
    // Use serde_json directly so any future field additions ride the
    // same canonical-JSON path as the HTTP transport's request body.
    serde_json::json!({
        "system": prompt.system,
        "user":   prompt.user,
        "schema": prompt.schema,
    })
    .to_string()
}

/// Construct a [`Command`] that runs `cmd` through the platform shell.
/// Unix → `sh -c`; Windows → `cmd /C`. Allowing a shell parse keeps the
/// CLI surface friendly (pipes, redirects, env-var expansion) at the
/// cost of inheriting whatever quoting the user's shell does — same
/// trade-off as `npm scripts` or `Makefile` recipes.
///
/// On Unix the child is placed in its own session/process group via
/// `setsid` in a pre-exec hook so [`kill_process_tree`] can later send
/// `SIGKILL` to the entire tree. Without that, `sh -c 'sleep 30'` on
/// Linux dash forks `sleep` as a grandchild — killing the immediate
/// child reaps `sh` but leaves `sleep` running with our pipes still
/// open, and the reader threads block until `sleep` finishes naturally.
#[cfg(unix)]
fn shell_command(cmd: &str) -> Command {
    use std::os::unix::process::CommandExt;
    let mut c = Command::new("sh");
    c.arg("-c").arg(cmd);
    // SAFETY: `setsid` is async-signal-safe and only mutates this
    // process's session/pgid — exactly the call documented as
    // permissible inside `pre_exec`. We do not allocate, lock, or
    // touch shared state here.
    unsafe {
        c.pre_exec(|| {
            if libc::setsid() == -1 {
                return Err(std::io::Error::last_os_error());
            }
            Ok(())
        });
    }
    c
}

#[cfg(windows)]
fn shell_command(cmd: &str) -> Command {
    let mut c = Command::new("cmd");
    c.arg("/C").arg(cmd);
    c
}

/// Kill the child and any grandchildren it spawned.
///
/// On Unix we send `SIGKILL` to the negated PID, which addresses the
/// process group (the child became its own group leader via `setsid`
/// in [`shell_command`]). This reaches every descendant — closing
/// inherited pipes promptly so the reader threads can drain. On
/// Windows we fall back to [`std::process::Child::kill`], which the
/// platform implements as `TerminateProcess` on the immediate child
/// only; the trade-off is acceptable here because the same Linux dash
/// vs. macOS bash divergence does not arise on Windows shells.
#[cfg(unix)]
fn kill_process_tree(child: &std::process::Child) {
    // SAFETY: `kill(2)` is async-signal-safe and stateless from our
    // perspective; the negative PID addresses the process group, and
    // an invalid PID just returns ESRCH which we ignore.
    unsafe {
        let pid = child.id() as libc::pid_t;
        libc::kill(-pid, libc::SIGKILL);
    }
}

#[cfg(not(unix))]
fn kill_process_tree(child: &mut std::process::Child) {
    let _ = child.kill();
}

/// Brief, allocation-free string form of [`std::process::ExitStatus`]
/// for the user-visible error message. `Display` for `ExitStatus`
/// prints "exit status: 1" / "signal: 9" already; just delegate.
fn exit_status_brief(status: &std::process::ExitStatus) -> String {
    status.to_string()
}

#[derive(Deserialize)]
struct FindingsEnvelope {
    findings: Vec<SemanticFinding>,
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::deep::candidate::{Candidate, CandidateKind};
    use crate::deep::config::DeepMode;
    use crate::deep::prompt::{PromptInputs, render};
    use crate::types::Language;
    use std::path::PathBuf;

    fn synth_runtime(cmd: &str, timeout_secs: u64) -> DeepRuntime {
        DeepRuntime {
            mode: DeepMode::Subprocess,
            base_url: String::new(),
            model: String::new(),
            api_key: None,
            max_cost_usd: None,
            cost_per_1k_input: None,
            cost_per_1k_output: None,
            request_timeout_secs: 120,
            max_candidates: 50,
            max_concurrent: 1,
            temperature: 0.0,
            max_prompt_chars: 16_000,
            excludes: Vec::new(),
            language_filter: Vec::new(),
            agent_cmd: Some(cmd.into()),
            agent_timeout_secs: timeout_secs,
        }
    }

    fn synth_prompt() -> RenderedPrompt {
        let cand = Candidate {
            kind: CandidateKind::ColdRegion,
            file: PathBuf::from("a.ts"),
            language: Language::TypeScript,
            line_start: 1,
            line_end: 5,
            source_snippet: "function isAdmin() { return true; }".into(),
            imports: Vec::new(),
            original_finding_id: None,
            seed_category: None,
        };
        render(&PromptInputs {
            candidate: &cand,
            structural_finding: None,
        })
    }

    #[test]
    fn new_rejects_empty_agent_cmd() {
        let mut rt = synth_runtime("nonempty", 60);
        rt.agent_cmd = Some("".into());
        let err = SubprocessClient::new(&rt).unwrap_err();
        assert!(matches!(err, DeepError::Config(_)));
    }

    #[test]
    fn new_rejects_whitespace_agent_cmd() {
        let mut rt = synth_runtime("nonempty", 60);
        rt.agent_cmd = Some("   ".into());
        let err = SubprocessClient::new(&rt).unwrap_err();
        assert!(matches!(err, DeepError::Config(_)));
    }

    #[test]
    fn new_rejects_runtime_without_agent_cmd() {
        // Defense-in-depth — `build` validates this, but a hand-rolled
        // runtime should still surface a clear error.
        let mut rt = synth_runtime("ignored", 60);
        rt.agent_cmd = None;
        let err = SubprocessClient::new(&rt).unwrap_err();
        assert!(
            matches!(err, DeepError::Config(ref msg) if msg.contains("invariant")),
            "expected Config(<invariant>), got: {err:?}",
        );
    }

    #[test]
    fn build_envelope_contains_system_user_and_schema() {
        let prompt = synth_prompt();
        let env = build_envelope(&prompt);
        let v: serde_json::Value = serde_json::from_str(&env).unwrap();
        assert!(v.get("system").is_some(), "envelope missing 'system'");
        assert!(v.get("user").is_some(), "envelope missing 'user'");
        assert!(v.get("schema").is_some(), "envelope missing 'schema'");
        // Schema round-trips structurally — must equal what render() produced,
        // because subprocess wrappers may forward it verbatim to a real LLM.
        assert_eq!(v.get("schema").unwrap(), &prompt.schema);
    }

    // -- Live-process tests are gated to Unix to avoid maintaining a
    // -- Rust fixture binary for Windows. The integration test crate
    // -- under tests/deep_subprocess_integration.rs covers the
    // -- end-to-end flow with shell-script fixtures.

    #[cfg(unix)]
    #[test]
    fn happy_path_with_cat_returning_canned_json() {
        // sh -c 'cat <<EOF ... EOF' ignores stdin and emits canned
        // JSON — proves the parse path without needing a real agent.
        let canned = r#"{"findings": []}"#;
        let cmd = format!("printf '%s' '{}'", canned);
        let rt = synth_runtime(&cmd, 10);
        let client = SubprocessClient::new(&rt).unwrap();
        let prompt = synth_prompt();
        let resp = client.analyze(&prompt).unwrap();
        assert!(resp.findings.is_empty());
        assert_eq!(resp.usage, TokenUsage::default());
    }

    #[cfg(unix)]
    #[test]
    fn nonzero_exit_surfaces_as_bad_response_for_skip() {
        // Per-candidate skip semantics: nonzero exit = skip this
        // candidate, keep going on others. Must NOT be Config (which
        // hard-fails the whole deep pass).
        let rt = synth_runtime("exit 7", 5);
        let client = SubprocessClient::new(&rt).unwrap();
        let err = client.analyze(&synth_prompt()).unwrap_err();
        assert!(
            matches!(err, DeepError::BadResponse(_)),
            "expected BadResponse for nonzero exit, got: {err:?}",
        );
    }

    #[cfg(unix)]
    #[test]
    fn malformed_stdout_surfaces_as_bad_response() {
        let rt = synth_runtime("printf 'this is not json'", 5);
        let client = SubprocessClient::new(&rt).unwrap();
        let err = client.analyze(&synth_prompt()).unwrap_err();
        assert!(
            matches!(err, DeepError::BadResponse(_)),
            "expected BadResponse for malformed stdout, got: {err:?}",
        );
    }

    #[cfg(unix)]
    #[test]
    fn timeout_kills_long_running_subprocess() {
        // 1s timeout, command sleeps 30s — must time out promptly,
        // not hang the test runner.
        let rt = synth_runtime("sleep 30", 1);
        let client = SubprocessClient::new(&rt).unwrap();
        let start = Instant::now();
        let err = client.analyze(&synth_prompt()).unwrap_err();
        let elapsed = start.elapsed();
        assert!(
            matches!(err, DeepError::Timeout { .. }),
            "expected Timeout, got: {err:?}",
        );
        // Generous upper bound — the polling cadence is 50ms and
        // killing the child should be near-instant.
        assert!(
            elapsed < Duration::from_secs(5),
            "timeout took too long: {elapsed:?}",
        );
    }

    #[cfg(unix)]
    #[test]
    fn spawn_failure_surfaces_as_config_for_hard_fail() {
        // Nonexistent command goes through `sh -c`, which exits 127.
        // That surfaces as BadResponse (per-candidate skip), NOT as a
        // spawn-time Config error — sh itself spawned just fine.
        // This test pins that behavior.
        let rt = synth_runtime("definitely-not-a-command-12345", 5);
        let client = SubprocessClient::new(&rt).unwrap();
        let err = client.analyze(&synth_prompt()).unwrap_err();
        assert!(
            matches!(err, DeepError::BadResponse(_)),
            "expected BadResponse (sh exited 127), got: {err:?}",
        );
    }

    #[cfg(unix)]
    #[test]
    fn markdown_fence_around_stdout_is_stripped() {
        // Wrapper CLIs sometimes fence JSON; we must still parse.
        let cmd = r#"printf '%s' '```json
{"findings": []}
```'"#;
        let rt = synth_runtime(cmd, 5);
        let client = SubprocessClient::new(&rt).unwrap();
        let resp = client.analyze(&synth_prompt()).unwrap();
        assert!(resp.findings.is_empty());
    }
}