agent-harness 0.4.0-alpha.3

Drive LLM coding agents — Claude Code, OpenAI Codex, and local Ollama / OpenAI-compatible models — from Rust behind one trait, with a normalized streaming event vocabulary. Bring your own agent too.
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
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
//! bob's stream-json parser — bob's wire format → the neutral
//! [`crate::RunEvent`] vocabulary.
//!
//! The normalized event types and the generic `normalize_process_event`
//! skeleton live in [`crate::events`]; this module is bob's
//! adapter-side decoder on top of them. bob emits one JSON object per
//! line with a snake_case `type` discriminator — see [`parse_bob_line`]
//! for the grounded schema. Reasoning is streamed inline as
//! `<thinking>…</thinking>` and routed by the stateful [`BobStreamParser`].

use serde_json::Value;

use crate::{
    normalize_process_event, ByteRange, ParsedLine, ProcessEvent, RunEvent, SessionInfo,
    SuggestedEdit, ToolCallEnd, ToolCallStart, ToolKind, UsageInfo,
};

/// Classify a bob (Roo-style) tool name into the neutral [`ToolKind`].
/// Public so a consumer that interprets bob's *raw* events directly — rather
/// than through `RunEvent` — can share this one table instead of re-encoding
/// bob's vocabulary downstream.
pub fn bob_tool_kind(name: &str) -> ToolKind {
    match name {
        "read_file" => ToolKind::Read,
        "write_file" | "write_to_file" => ToolKind::Write,
        "apply_diff" | "insert_content" | "search_and_replace" => ToolKind::Edit,
        "search_files" | "list_files" | "list_code_definition_names" => ToolKind::Search,
        "execute_command" => ToolKind::Execute,
        _ => ToolKind::Other,
    }
}

/// bob's adapter-side normalization: parse bob's `--output-format
/// stream-json` stdout via [`parse_bob_line`].
pub fn normalize_bob_event(event: ProcessEvent) -> Vec<RunEvent> {
    normalize_process_event(event, parse_bob_line)
}

/// Parse one line of bob's `--output-format stream-json` into the shared
/// [`ParsedLine`]. Grounded in bob's *empirical* event schema (the
/// `bob-agents` reference + "bob shell usage" findings), not guessed: bob
/// emits one JSON object per line with a snake_case `type` discriminator —
/// `init` / `message{role,content,delta}` / `tool_use{tool_id,tool_name,
/// parameters}` / `tool_result{tool_id,status,output}` / `result{stats}`.
///
/// Mapping: an assistant `message` → text (the echoed `user` prompt is
/// skipped — a real fix vs. the old role-blind heuristic); `tool_use` →
/// a structured [`ToolCallStart`] (bob's edit tools — write_file /
/// apply_diff / insert_content — surface as tool-cards too; reconstructing
/// previewable diffs from their `parameters` is a separate follow-up);
/// `tool_result` → [`ToolCallEnd`] (ok unless `status == "error"`).
/// `init` / `result` are lifecycle (process start/exit drives
/// Started/Exited). A non-JSON line passes through as raw text.
/// Unrecognized shapes fall back to the legacy suggested-edits heuristic
/// so a bob build that emits edit arrays still surfaces them.
pub fn parse_bob_line(line: &str) -> ParsedLine {
    let trimmed = line.trim();
    if trimmed.is_empty() {
        return ParsedLine::default();
    }

    let payload: Value = match serde_json::from_str(trimmed) {
        Ok(value) => value,
        // Not JSON — bob occasionally prints prose / stderr-ish lines.
        // Pass the raw (untrimmed) line through as text.
        Err(_) => {
            return ParsedLine {
                text: Some(line.to_owned()),
                ..ParsedLine::default()
            }
        }
    };

    let Some(record) = payload.as_object() else {
        return ParsedLine::default();
    };

    match record.get("type").and_then(Value::as_str) {
        // Assistant text (`delta: true` marks a streaming chunk; both
        // chunk and full message carry the text in `content`). The echoed
        // user prompt (role "user") is not surfaced.
        Some("message") => {
            if record.get("role").and_then(Value::as_str) == Some("assistant") {
                if let Some(content) = pick_string(record, "content") {
                    return ParsedLine {
                        text: Some(content),
                        ..ParsedLine::default()
                    };
                }
            }
            ParsedLine::default()
        }
        // Tool call start → structured ToolStart (tool_id + tool_name).
        Some("tool_use") => {
            let name = pick_string(record, "tool_name").unwrap_or_else(|| "tool".to_owned());
            // bob delivers its final answer via the `attempt_completion`
            // tool (grounded in a real run) — surface its `result` as the
            // answer text, not a bare tool-card.
            if name == "attempt_completion" {
                return match record
                    .get("parameters")
                    .and_then(Value::as_object)
                    .and_then(|p| p.get("result"))
                    .and_then(Value::as_str)
                    .filter(|s| !s.is_empty())
                {
                    Some(result) => ParsedLine {
                        text: Some(result.to_owned()),
                        ..ParsedLine::default()
                    },
                    None => ParsedLine::default(),
                };
            }
            let tool_call_id = pick_string(record, "tool_id").unwrap_or_default();
            // The call's arguments, lifted verbatim (parameters object →
            // compact JSON) so the UI can show what the tool was asked to do.
            let input = record.get("parameters").map(value_to_display_string);
            let tool_kind = bob_tool_kind(&name);
            ParsedLine {
                tool_start: Some(ToolCallStart { tool_call_id, name, input, tool_kind }),
                ..ParsedLine::default()
            }
        }
        // Tool call end → ToolEnd, matched by tool_id; ok unless the
        // status is explicitly "error". `output` carries the tool's result.
        Some("tool_result") => {
            let tool_call_id = pick_string(record, "tool_id").unwrap_or_default();
            let ok = record.get("status").and_then(Value::as_str) != Some("error");
            let output = record
                .get("output")
                .map(value_to_display_string)
                .filter(|s| !s.is_empty());
            ParsedLine {
                tool_end: Some(ToolCallEnd { tool_call_id, ok, output }),
                ..ParsedLine::default()
            }
        }
        // init → session identity (id + model), arriving a beat after the
        // engine's `Started`.
        Some("init") => ParsedLine {
            session: Some(SessionInfo {
                session_id: pick_string(record, "session_id"),
                model: pick_string(record, "model"),
            }),
            ..ParsedLine::default()
        },
        // result → token usage. bob reports a single `total_tokens` in
        // `stats` (no input/output split); coins (`session_costs`) are
        // bob-specific and intentionally NOT lifted into the neutral Usage.
        Some("result") => {
            let total_tokens = record
                .get("stats")
                .and_then(Value::as_object)
                .and_then(|s| s.get("total_tokens"))
                .and_then(Value::as_u64);
            ParsedLine {
                usage: total_tokens.map(|t| UsageInfo {
                    total_tokens: Some(t),
                    ..UsageInfo::default()
                }),
                ..ParsedLine::default()
            }
        }
        // Anything else: unknown. Fall back to the legacy suggested-edits
        // heuristic so a bob build that emits edit arrays still surfaces them.
        _ => {
            let edits = parse_suggested_edits(record);
            if edits.is_empty() {
                ParsedLine::default()
            } else {
                let n = edits.len();
                ParsedLine {
                    edits,
                    activity: Some(format!("{n} suggested edit{}", if n == 1 { "" } else { "s" })),
                    ..ParsedLine::default()
                }
            }
        }
    }
}

/// Stateful wrapper over [`parse_bob_line`] for a single bob run. bob
/// streams its reasoning inline as `<thinking>…</thinking>` within the
/// assistant `message` content (grounded in a real run — the tags arrive
/// as their own deltas), so routing that reasoning to the Thinking
/// stream requires tracking the open/closed state *across* lines. The
/// per-line dispatch (text / tool events / answer) stays in
/// `parse_bob_line`; this only re-routes assistant text through the
/// thinking-tag state machine. One instance per run (see `BobHarness`).
#[derive(Debug, Default)]
pub struct BobStreamParser {
    in_thinking: bool,
    /// True while inside a `[using tool …]` narration echo whose closing `]`
    /// has not yet arrived — bob prints that echo inline right before each
    /// structured `tool_use`, so it's redundant with the ToolStart card and is
    /// dropped. The echo can span deltas, so the state carries across lines.
    suppressing_echo: bool,
}

impl BobStreamParser {
    /// Parse one stdout line, routing any assistant text through the
    /// `<thinking>` state machine into [`ParsedLine::text`] /
    /// [`ParsedLine::thinking`].
    pub fn parse_line(&mut self, line: &str) -> ParsedLine {
        let mut parsed = parse_bob_line(line);
        if let Some(content) = parsed.text.take() {
            let (visible, mut thinking) = self.route_thinking(&content);
            // Drop bob's `[using tool …]` narration echo (redundant with the
            // structured ToolStart card; may span deltas).
            let visible = visible.and_then(|t| self.narration_after_echo(t));
            // bob's *answer* is always the `attempt_completion` result (verified
            // in code AND ask mode — the assistant `message` carries only
            // `<thinking>` + preamble prose, e.g. "I'll read … then enrich …").
            // So an assistant message's leftover prose is narration, not the
            // answer: fold it into the collapsed thinking trace, leaving only the
            // attempt_completion text as the visible message. A `tool_use` /
            // non-message line (the answer) keeps its text.
            if line_is_assistant_message(line) {
                if let Some(t) = visible {
                    thinking = Some(match thinking {
                        Some(a) => a + &t,
                        None => t,
                    });
                }
            } else {
                parsed.text = visible;
            }
            parsed.thinking = match (thinking, parsed.thinking.take()) {
                (Some(a), Some(b)) => Some(a + &b),
                (a, b) => a.or(b),
            };
        }
        parsed
    }

    /// Split an assistant content chunk into (visible text, thinking),
    /// honoring `<thinking>`/`</thinking>` markers and the carried-over
    /// `in_thinking` state. Handles tags split across chunks and multiple
    /// tags within one chunk.
    fn route_thinking(&mut self, content: &str) -> (Option<String>, Option<String>) {
        const OPEN: &str = "<thinking>";
        const CLOSE: &str = "</thinking>";
        let mut text = String::new();
        let mut thinking = String::new();
        let mut rest = content;
        loop {
            if self.in_thinking {
                match rest.find(CLOSE) {
                    Some(i) => {
                        thinking.push_str(&rest[..i]);
                        self.in_thinking = false;
                        rest = &rest[i + CLOSE.len()..];
                    }
                    None => {
                        thinking.push_str(rest);
                        break;
                    }
                }
            } else {
                match rest.find(OPEN) {
                    Some(i) => {
                        text.push_str(&rest[..i]);
                        self.in_thinking = true;
                        rest = &rest[i + OPEN.len()..];
                    }
                    None => {
                        text.push_str(rest);
                        break;
                    }
                }
            }
        }
        (
            (!text.is_empty()).then_some(text),
            (!thinking.is_empty()).then_some(thinking),
        )
    }

    /// Drop bob's `[using tool …]` narration echo. bob prints it inline right
    /// before the structured `tool_use` event, so it's redundant with the
    /// ToolStart card. The echo can span deltas (its `]` arriving in a later
    /// chunk), so `suppressing_echo` carries the drop across lines. Ported from
    /// the pre-migration `BobChatMapper`.
    fn narration_after_echo(&mut self, text: String) -> Option<String> {
        if self.suppressing_echo {
            // Still inside an unterminated echo — drop until its `]`.
            if text.contains(']') {
                self.suppressing_echo = false;
            }
            return None;
        }
        if text.trim_start().starts_with("[using tool") {
            // Start of the echo. If its `]` isn't in this delta, keep dropping
            // subsequent deltas until it arrives.
            if !text.contains(']') {
                self.suppressing_echo = true;
            }
            return None;
        }
        Some(text)
    }
}

/// Is this stdout line bob's assistant `message` (its narration / preamble
/// prose) — as opposed to a `tool_use` (incl. `attempt_completion`, the answer),
/// a `tool_result`, or a lifecycle line? Used to fold the message prose into the
/// collapsed thinking trace: bob's answer is always the `attempt_completion`
/// result, never the message text. A cheap re-parse of the line envelope (bob's
/// stdout is bounded, not a hot loop).
fn line_is_assistant_message(line: &str) -> bool {
    serde_json::from_str::<Value>(line.trim())
        .ok()
        .map(|v| {
            v.get("type").and_then(Value::as_str) == Some("message")
                && v.get("role").and_then(Value::as_str) == Some("assistant")
        })
        .unwrap_or(false)
}

/// Render a JSON value as a display string for a tool's `input`/`output`:
/// a JSON string is taken verbatim (no surrounding quotes); any other shape
/// (object / array / number) is serialized compactly. Lets the adapter lift
/// bob's `parameters` / `tool_result.output` into the neutral layer without
/// imposing a schema on them.
fn value_to_display_string(value: &Value) -> String {
    match value {
        Value::String(s) => s.clone(),
        other => other.to_string(),
    }
}

/// Non-empty string field, else `None` (mirrors TS `pickString`).
fn pick_string(record: &serde_json::Map<String, Value>, key: &str) -> Option<String> {
    match record.get(key) {
        Some(Value::String(s)) if !s.is_empty() => Some(s.clone()),
        _ => None,
    }
}

/// String field allowing empty (mirrors TS `pickStringValue` — used
/// for replacements, which may legitimately be the empty string for
/// a deletion).
fn pick_string_value(record: &serde_json::Map<String, Value>, key: &str) -> Option<String> {
    match record.get(key) {
        Some(Value::String(s)) => Some(s.clone()),
        _ => None,
    }
}

fn parse_suggested_edits(record: &serde_json::Map<String, Value>) -> Vec<SuggestedEdit> {
    let mut edits = Vec::new();
    if let Some(direct) = parse_suggested_edit(record) {
        edits.push(direct);
    }
    for key in ["edits", "suggestedEdits", "suggestions"] {
        let Some(Value::Array(items)) = record.get(key) else {
            continue;
        };
        for item in items {
            if let Some(obj) = item.as_object() {
                if let Some(parsed) = parse_suggested_edit(obj) {
                    edits.push(parsed);
                }
            }
        }
    }
    edits
}

fn parse_suggested_edit(record: &serde_json::Map<String, Value>) -> Option<SuggestedEdit> {
    let file_path = pick_string(record, "filePath")
        .or_else(|| pick_string(record, "path"))
        .or_else(|| pick_string(record, "file"))?;

    // Range may be nested under `range` or flat on the record.
    let range_record = match record.get("range").and_then(Value::as_object) {
        Some(nested) => nested,
        None => record,
    };
    let start = range_record.get("start").and_then(Value::as_u64)?;
    let end = range_record.get("end").and_then(Value::as_u64)?;

    let replacement = pick_string_value(record, "replacement")
        .or_else(|| pick_string_value(record, "replaceWith"))
        .or_else(|| pick_string_value(record, "insert"))
        .or_else(|| pick_string_value(record, "newText"))?;

    let title = pick_string(record, "title")
        .or_else(|| pick_string(record, "summary"))
        .or_else(|| pick_string(record, "description"));

    Some(SuggestedEdit {
        file_path,
        range: ByteRange { start, end },
        replacement,
        title,
    })
}

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

    #[test]
    fn blank_line_yields_nothing() {
        assert!(parse_bob_line("   ").is_empty());
    }

    #[test]
    fn non_json_passes_through_as_text() {
        let parsed = parse_bob_line("hello world");
        assert_eq!(parsed.text.as_deref(), Some("hello world"));
        assert!(parsed.edits.is_empty());
    }

    #[test]
    fn assistant_message_becomes_text() {
        let parsed =
            parse_bob_line(r#"{"type":"message","role":"assistant","content":"hi there"}"#);
        assert_eq!(parsed.text.as_deref(), Some("hi there"));
        assert!(parsed.activity.is_none());
    }

    #[test]
    fn user_message_is_skipped() {
        // The echoed user prompt must not surface as assistant text.
        let parsed = parse_bob_line(r#"{"type":"message","role":"user","content":"my prompt"}"#);
        assert!(parsed.is_empty());
    }

    #[test]
    fn assistant_delta_chunk_becomes_text() {
        // `delta: true` marks a streaming chunk; the text is still in
        // `content`.
        let parsed = parse_bob_line(
            r#"{"type":"message","role":"assistant","content":"chunk","delta":true}"#,
        );
        assert_eq!(parsed.text.as_deref(), Some("chunk"));
    }

    #[test]
    fn flat_suggested_edit_parses() {
        let line = r#"{"filePath":"notes/a.md","start":3,"end":7,"replacement":"X","title":"fix"}"#;
        let parsed = parse_bob_line(line);
        assert_eq!(parsed.edits.len(), 1);
        let edit = &parsed.edits[0];
        assert_eq!(edit.file_path, "notes/a.md");
        assert_eq!(edit.range, ByteRange { start: 3, end: 7 });
        assert_eq!(edit.replacement, "X");
        assert_eq!(edit.title.as_deref(), Some("fix"));
        // No text → activity reports the edit count.
        assert_eq!(parsed.activity.as_deref(), Some("1 suggested edit"));
    }

    #[test]
    fn nested_range_and_array_edits_parse() {
        let line = r#"{"edits":[{"path":"a.md","range":{"start":0,"end":1},"newText":""},
                                 {"file":"b.md","range":{"start":2,"end":4},"insert":"yo"}]}"#;
        let parsed = parse_bob_line(line);
        assert_eq!(parsed.edits.len(), 2);
        assert_eq!(parsed.edits[0].replacement, ""); // empty replacement = deletion, allowed
        assert_eq!(parsed.edits[1].replacement, "yo");
        assert_eq!(parsed.activity.as_deref(), Some("2 suggested edits"));
    }

    #[test]
    fn tool_use_becomes_tool_start() {
        let parsed = parse_bob_line(
            r#"{"type":"tool_use","tool_id":"tool-1","tool_name":"execute_command","parameters":{"command":"ls"}}"#,
        );
        let start = parsed.tool_start.expect("tool_start");
        assert_eq!(start.tool_call_id, "tool-1");
        assert_eq!(start.name, "execute_command");
        // The parameters object is lifted verbatim as the call's input.
        assert_eq!(start.input.as_deref(), Some(r#"{"command":"ls"}"#));
        assert!(parsed.activity.is_none());
    }

    #[test]
    fn edit_tools_surface_as_tool_start() {
        // bob's edit tools (apply_diff / insert_content / write_file) flow
        // through as tool-cards too.
        let start = parse_bob_line(
            r#"{"type":"tool_use","tool_id":"t9","tool_name":"apply_diff","parameters":{"path":"a.md"}}"#,
        )
        .tool_start
        .expect("tool_start");
        assert_eq!(start.name, "apply_diff");
    }

    #[test]
    fn tool_result_becomes_tool_end() {
        let ok = parse_bob_line(
            r#"{"type":"tool_result","tool_id":"tool-1","status":"success","output":"done"}"#,
        )
        .tool_end
        .expect("tool_end");
        assert_eq!(ok.tool_call_id, "tool-1");
        assert!(ok.ok);
        // The tool's result is lifted as the end event's output.
        assert_eq!(ok.output.as_deref(), Some("done"));

        let err = parse_bob_line(
            r#"{"type":"tool_result","tool_id":"tool-2","status":"error","output":"boom"}"#,
        )
        .tool_end
        .expect("tool_end");
        assert!(!err.ok);
    }

    #[test]
    fn init_yields_session_and_result_yields_usage() {
        // init → Session (id + model); no text/tool content.
        let init = parse_bob_line(r#"{"type":"init","session_id":"s1","model":"premium"}"#);
        let session = init.session.expect("session");
        assert_eq!(session.session_id.as_deref(), Some("s1"));
        assert_eq!(session.model.as_deref(), Some("premium"));
        assert!(init.text.is_none() && init.tool_start.is_none());

        // result → Usage (neutral tokens only; coins stay bob-specific).
        let result = parse_bob_line(
            r#"{"type":"result","status":"success","stats":{"total_tokens":1280,"session_costs":3,"tool_calls":2}}"#,
        );
        let usage = result.usage.expect("usage");
        assert_eq!(usage.total_tokens, Some(1280));
        assert_eq!(usage.input_tokens, None);
        assert_eq!(usage.output_tokens, None);

        // A result with no token count → no usage (nothing to report).
        assert!(parse_bob_line(r#"{"type":"result","status":"success","stats":{"tool_calls":2}}"#)
            .is_empty());
    }

    #[test]
    fn incomplete_edit_is_ignored() {
        // Missing `end` → not a valid edit.
        let parsed = parse_bob_line(r#"{"filePath":"a.md","start":3,"replacement":"X"}"#);
        assert!(parsed.edits.is_empty());
    }

    #[test]
    fn normalize_stdout_text_event() {
        let events = normalize_bob_event(ProcessEvent::Stdout {
            run_id: "r1".to_owned(),
            line: r#"{"type":"message","role":"assistant","content":"hi"}"#.to_owned(),
        });
        assert_eq!(events.len(), 1);
        assert!(matches!(
            &events[0],
            RunEvent::Text { run_id, delta } if run_id == "r1" && delta == "hi"
        ));
    }

    #[test]
    fn normalize_bob_tool_events() {
        let start = normalize_bob_event(ProcessEvent::Stdout {
            run_id: "r1".to_owned(),
            line: r#"{"type":"tool_use","tool_id":"t1","tool_name":"write_file"}"#.to_owned(),
        });
        assert!(matches!(
            start.as_slice(),
            [RunEvent::ToolStart { tool_call_id, title, .. }]
                if tool_call_id == "t1" && title == "write_file"
        ));
        let end = normalize_bob_event(ProcessEvent::Stdout {
            run_id: "r1".to_owned(),
            line: r#"{"type":"tool_result","tool_id":"t1","status":"success"}"#.to_owned(),
        });
        assert!(matches!(
            end.as_slice(),
            [RunEvent::ToolEnd { tool_call_id, ok, .. }] if tool_call_id == "t1" && *ok
        ));
    }

    #[test]
    fn attempt_completion_becomes_answer_text() {
        // bob's final answer is delivered via the attempt_completion tool,
        // not plain message content — surface it as text, not a card.
        let parsed = parse_bob_line(
            r#"{"type":"tool_use","tool_id":"tool-2","tool_name":"attempt_completion","parameters":{"result":"The answer is 42."}}"#,
        );
        assert_eq!(parsed.text.as_deref(), Some("The answer is 42."));
        assert!(parsed.tool_start.is_none());
    }

    #[test]
    fn using_tool_echo_is_dropped() {
        // bob narrates each tool call inline as `[using tool …]` right before
        // the structured tool_use; that echo is redundant with the ToolStart
        // card and must not appear in the message text.
        let mut parser = BobStreamParser::default();
        let msg =
            |c: &str| serde_json::json!({"type":"message","role":"assistant","content":c}).to_string();
        assert!(parser.parse_line(&msg("[using tool write_to_file: notes/x.md]")).text.is_none());
        // Surrounding narration still flows — to the thinking trace (only the
        // attempt_completion result is the visible answer).
        assert_eq!(parser.parse_line(&msg("All set.")).thinking.as_deref(), Some("All set."));
    }

    #[test]
    fn chunked_using_tool_echo_is_dropped_across_deltas() {
        // The echo's closing `]` can arrive in a later delta; suppression
        // carries across lines until it does.
        let mut parser = BobStreamParser::default();
        let msg =
            |c: &str| serde_json::json!({"type":"message","role":"assistant","content":c}).to_string();
        assert!(parser.parse_line(&msg("[using tool read_file: a/very")).text.is_none());
        assert!(parser.parse_line(&msg("/long/path.md")).text.is_none());
        assert!(parser.parse_line(&msg("]")).text.is_none());
        // Suppression ended; narration resumes — into the thinking trace.
        assert_eq!(parser.parse_line(&msg("ok")).thinking.as_deref(), Some("ok"));
    }

    #[test]
    fn message_prose_routes_to_thinking_only_attempt_completion_is_text() {
        // Neither bob's <thinking> reasoning NOR its plain preamble prose (both
        // ride inside assistant `message`s) is the answer — bob answers via the
        // attempt_completion tool. So the persistent parser folds ALL message
        // content into `thinking`, and only the attempt_completion text becomes
        // the visible message. (The <thinking> tags arrive as their own deltas;
        // state carries across them.)
        let mut parser = BobStreamParser::default();
        let msg = |content: &str| {
            serde_json::json!({ "type": "message", "role": "assistant", "content": content, "delta": true })
                .to_string()
        };
        // Opening tag (its own delta): the content after <thinking> is reasoning.
        let open = parser.parse_line(&msg("<thinking>\n"));
        assert_eq!(open.thinking.as_deref(), Some("\n"));
        assert!(open.text.is_none());
        // Mid-block chunk → thinking (state carried across deltas).
        let mid = parser.parse_line(&msg("the user wants X"));
        assert_eq!(mid.thinking.as_deref(), Some("the user wants X"));
        assert!(mid.text.is_none());
        // Close tag + trailing PROSE in one delta: the prose is narration, not
        // the answer → it also folds into thinking, never text.
        let close = parser.parse_line(&msg("</thinking>I'll update the file."));
        assert_eq!(close.thinking.as_deref(), Some("I'll update the file."));
        assert!(close.text.is_none());
        // Further plain message prose → thinking too.
        let more = parser.parse_line(&msg(" Working on it."));
        assert_eq!(more.thinking.as_deref(), Some(" Working on it."));
        assert!(more.text.is_none());
        // Only the attempt_completion result is the visible answer.
        let answer = parser.parse_line(
            r#"{"type":"tool_use","tool_id":"t","tool_name":"attempt_completion","parameters":{"result":"Done."}}"#,
        );
        assert_eq!(answer.text.as_deref(), Some("Done."));
        assert!(answer.thinking.is_none());
    }

    #[test]
    fn grounded_against_real_bob_capture() {
        // Verbatim shapes captured from `bob 1.0.4 -o stream-json`
        // (timestamps trimmed) — locks the parser to bob's real format.
        let mut parser = BobStreamParser::default();
        // init → Session (id + model), not empty.
        let session = parser
            .parse_line(r#"{"type":"init","session_id":"s","model":"premium"}"#)
            .session
            .expect("session");
        assert_eq!(session.session_id.as_deref(), Some("s"));
        assert_eq!(session.model.as_deref(), Some("premium"));
        // Echoed user prompt → skipped.
        assert!(parser
            .parse_line(r#"{"type":"message","role":"user","content":"list files"}"#)
            .is_empty());
        // Assistant reasoning wrapped in <thinking> → thinking, not text.
        assert_eq!(
            parser
                .parse_line(
                    r#"{"type":"message","role":"assistant","content":"<thinking>\n","delta":true}"#
                )
                .thinking
                .as_deref(),
            Some("\n")
        );
        let _ = parser.parse_line(
            r#"{"type":"message","role":"assistant","content":"</thinking>\n","delta":true}"#,
        );
        // Real tool_use (list_files) → ToolStart, parameters lifted as input.
        let start = parser
            .parse_line(r#"{"type":"tool_use","tool_name":"list_files","tool_id":"tool-1","parameters":{"dir_path":"/x/docs"}}"#)
            .tool_start
            .expect("tool_start");
        assert_eq!(start.tool_call_id, "tool-1");
        assert_eq!(start.name, "list_files");
        assert_eq!(start.input.as_deref(), Some(r#"{"dir_path":"/x/docs"}"#));
        // Real tool_result → ToolEnd(ok), output lifted.
        let end = parser
            .parse_line(r#"{"type":"tool_result","tool_id":"tool-1","status":"success","output":"Listed 11 item(s)."}"#)
            .tool_end
            .expect("tool_end");
        assert!(end.ok);
        assert_eq!(end.output.as_deref(), Some("Listed 11 item(s)."));
        // attempt_completion → the answer text.
        let answer = parser.parse_line(
            r#"{"type":"tool_use","tool_id":"tool-2","tool_name":"attempt_completion","parameters":{"result":"The docs directory contains 10 files."}}"#,
        );
        assert_eq!(answer.text.as_deref(), Some("The docs directory contains 10 files."));
        assert!(answer.tool_start.is_none());
        // result → ignored.
        assert!(parser
            .parse_line(r#"{"type":"result","status":"success","stats":{"tool_calls":2}}"#)
            .is_empty());
    }

    #[test]
    fn normalize_passes_through_lifecycle_events() {
        assert!(matches!(
            normalize_bob_event(ProcessEvent::Started { run_id: "r".into() }).as_slice(),
            [RunEvent::Started { .. }]
        ));
        assert!(matches!(
            normalize_bob_event(ProcessEvent::Exited {
                run_id: "r".into(),
                exit_code: Some(0),
                cancelled: false
            })
            .as_slice(),
            [RunEvent::Exited { exit_code: Some(0), cancelled: false, .. }]
        ));
    }
}