choreo-daemon 0.2.2

Agentic coding assistant — daemon, TUI, and bridges
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
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
use super::{PreparedImage, ToolExecError, context::ToolContext, human_size, resolve_path};
use base64::{Engine as _, engine::general_purpose::STANDARD as BASE64};
use choreo_keystore::ServiceCredential;
use headless_chrome::Tab;
use headless_chrome::protocol::cdp::Page;
use headless_chrome::protocol::cdp::Page::CaptureScreenshotFormatOption;
use headless_chrome::{Browser, LaunchOptions};
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};
use std::ffi::OsStr;
use std::path::{Path, PathBuf};
use std::time::Duration;
use tracing::{debug, info, warn};
use url::Url;

/// What `retrieve_webpage` should produce from the rendered page.
#[derive(Debug, Clone, Copy, Deserialize, JsonSchema)]
#[serde(rename_all = "snake_case")]
pub enum WebpageAction {
    /// Fully rendered HTML (outerHTML); restricted to `selector` when given.
    Content,
    /// Human-readable plain text of the page (innerText); defaults to `<body>`.
    Text,
    /// A PNG screenshot, saved to `output_path` if given and shown inline.
    Screenshot,
    /// A PDF of the page; requires `output_path` (binary output can't be a string).
    Pdf,
}

impl WebpageAction {
    fn as_str(self) -> &'static str {
        match self {
            WebpageAction::Content => "content",
            WebpageAction::Text => "text",
            WebpageAction::Screenshot => "screenshot",
            WebpageAction::Pdf => "pdf",
        }
    }
}

#[derive(Debug, Default, Deserialize, JsonSchema)]
pub struct RetrieveWebpageArgs {
    /// URL to render (http, https, or file scheme; `file://` renders a local
    /// file directly in the browser).
    url: String,
    /// What to retrieve. Defaults to "content".
    action: Option<WebpageAction>,
    /// Milliseconds to wait after load before capturing (lets JS settle). 0 = none.
    wait_ms: Option<u64>,
    /// Navigation / element-wait timeout in milliseconds. Default `30_000`.
    timeout_ms: Option<u64>,
    /// Viewport width. Default 1280.
    width: Option<u32>,
    /// Viewport height. Default 800.
    height: Option<u32>,
    /// Screenshot: capture the full scrollable page (surface). Default true.
    full_page: Option<bool>,
    /// Restrict content/text/screenshot capture to this CSS selector.
    selector: Option<String>,
    /// Where to write the result for `screenshot` / `pdf`. Resolved against the
    /// session working directory. PDFs require this.
    output_path: Option<String>,
    /// Enable WebGL-capable rendering (new headless + software `SwiftShader`) so
    /// canvas/3D pages render instead of coming back blank. Off by default (the
    /// historical launch flags would disable WebGL). The result text reports the
    /// WebGL version obtained, or that no context could be created.
    webgl: Option<bool>,
}

/// Default viewport / nav timeout used when the caller omits them.
const DEFAULT_TIMEOUT_MS: u64 = 30_000;
/// Hard cap on a single call's navigation timeout so a mistyped/malicious
/// argument can't pin a worker thread for an unbounded stretch.
const MAX_TIMEOUT_MS: u64 = 120_000;
/// Hard cap on the post-load settle delay for the same reason.
const MAX_WAIT_MS: u64 = 30_000;

/// Binary names/known paths to search for, in *preference* order: chromium
/// first, then the various chrome bundle names, so a Chromium install wins
/// over Chrome when both are present.
const CANDIDATE_NAMES: &[&str] = &[
    "chromium",
    "chromium-browser",
    "google-chrome",
    "google-chrome-stable",
    "google-chrome-beta",
    "google-chrome-unstable",
    "chrome",
    "microsoft-edge",
    "brave-browser",
];

/// Resolve a locally-installed Chromium/Chrome binary, preferring chromium.
///
/// Lookup order:
/// 1. `CHROMIUM_BIN`, then `CHROME_BIN` env overrides (must exist).
/// 2. Names on `PATH` (chromium first).
/// 3. Known absolute install paths for the current OS.
///
/// Returns `None` when nothing usable is found — the caller reports that as a
/// clear error telling the operator to install Chromium (this tool deliberately
/// does NOT auto-download a browser).
fn resolve_browser_binary() -> Option<std::path::PathBuf> {
    for var in ["CHROMIUM_BIN", "CHROME_BIN"] {
        if let Some(p) = std::env::var_os(var) {
            let candidate = std::path::PathBuf::from(p);
            if is_executable(&candidate) {
                return Some(candidate);
            }
        }
    }

    // Search each PATH entry for a candidate executable name.
    if let Some(path_var) = std::env::var_os("PATH") {
        for dir in std::env::split_paths(&path_var) {
            for name in CANDIDATE_NAMES {
                let candidate = dir.join(name);
                if is_executable(&candidate) {
                    return Some(candidate);
                }
            }
        }
    }

    // Known absolute locations (best-effort per OS), chromium first.
    #[cfg(target_os = "macos")]
    {
        let mac_paths = [
            "/Applications/Chromium.app/Contents/MacOS/Chromium",
            "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome",
            "/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge",
        ];
        for p in mac_paths {
            let candidate = std::path::PathBuf::from(p);
            if is_executable(&candidate) {
                return Some(candidate);
            }
        }
    }
    #[cfg(target_os = "windows")]
    {
        for base in [
            "C:/Program Files/Chromium/Application/chrome.exe",
            "C:/Program Files/Google/Chrome/Application/chrome.exe",
            "C:/Program Files (x86)/Google/Chrome/Application/chrome.exe",
            "C:/Program Files (x86)/Microsoft/Edge/Application/msedge.exe",
        ] {
            let candidate = std::path::PathBuf::from(base);
            if is_executable(&candidate) {
                return Some(candidate);
            }
        }
    }

    None
}

/// True when `path` exists as a file *and* (on Unix) is executable, so we don't
/// hand the launcher a non-executable or directory candidate that would then
/// fail at spawn time. Windows has no executable bit; existence is enough.
fn is_executable(path: &std::path::Path) -> bool {
    if !path.is_file() {
        return false;
    }
    #[cfg(unix)]
    {
        use std::os::unix::fs::PermissionsExt;
        path.metadata()
            .is_ok_and(|m| m.permissions().mode() & 0o111 != 0)
    }
    #[cfg(not(unix))]
    {
        true
    }
}

/// How Chromium is launched. `Headless` is the historical default — legacy
/// headless with the crate's GPU-disabling `--disable-gpu`. `WebGl` opts into
/// new-headless with ANGLE/`SwiftShader` software GL so canvas/3D pages render
/// instead of coming back blank. Modelling this as an enum (rather than a bare
/// `bool`) makes the call site self-describing and leaves room for a future
/// launch variant to be an additive arm rather than a third boolean.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
enum LaunchMode {
    #[default]
    Headless,
    WebGl,
}

/// Extra Chromium switches that enable software-rendered WebGL, in the order
/// they are appended to argv. Single-sourced so the production launch path and
/// the unit tests that pin the flag set can never drift apart (a typo here is
/// caught by the same list the tests iterate).
///
/// - `--headless=new`: the real-browser rendering path WebGL needs — the crate's
///   bare `--headless` is legacy headless, with poor/absent WebGL support.
/// - `--use-gl=angle --use-angle=swiftshader`: route GL through ANGLE's
///   `SwiftShader` backend — portable software WebGL on hosts with no usable GPU.
/// - `--enable-unsafe-swiftshader`: REQUIRED on modern Chrome — without it the
///   `SwiftShader` fallback backing WebGL is refused and the context comes back
///   null.
/// - `--ignore-gpu-blocklist`: don't let a conservative GPU blocklist disable GL.
const WEBGL_ARGS: &[&str] = &[
    "--headless=new",
    "--use-gl=angle",
    "--use-angle=swiftshader",
    "--enable-unsafe-swiftshader",
    "--ignore-gpu-blocklist",
];

/// Build the Chromium launch options in a pure, testable way (no process is
/// spawned), so the flag set can be asserted in unit tests.
///
/// [`LaunchMode::Headless`] reproduces the historical configuration exactly:
/// legacy headless with the crate's default `--disable-gpu`.
/// [`LaunchMode::WebGl`] switches to new headless with ANGLE/`SwiftShader` so
/// canvas/3D content renders — see [`WEBGL_ARGS`] and the inline comments for
/// why each flag is required.
fn build_launch_options(
    binary: PathBuf,
    width: u32,
    height: u32,
    timeout_ms: u64,
    mode: LaunchMode,
) -> Result<LaunchOptions<'static>, ToolExecError> {
    let mut builder = LaunchOptions::default_builder();
    builder.path(Some(binary));
    builder.window_size(Some((width, height)));
    // Keep the DevTools socket alive well past the resolved navigation timeout
    // so a slow page can't get torn down mid-navigation; the idle grace is
    // derived from (and always larger than) the caller's timeout.
    builder.idle_browser_timeout(Duration::from_millis(
        timeout_ms.saturating_mul(2).max(60_000),
    ));

    match mode {
        LaunchMode::WebGl => {
            // The crate's `headless` flag only controls the bare `--headless`
            // switch, which it appends to argv *after* our custom `args`.
            // Chromium resolves duplicate switches to the LAST one, so leaving it
            // true would emit `… --headless=new … --headless` and silently drop
            // us back to legacy headless (which has poor/absent WebGL support).
            // Turn it off and supply `--headless=new` ourselves via `WEBGL_ARGS`.
            builder.headless(false);
            // Drop the crate's `--disable-gpu`: WebGL — even SwiftShader — runs
            // through the GPU/ANGLE path, and a global `--disable-gpu` can leave
            // `getContext('webgl')` returning null even though the flags look right.
            builder.enable_gpu(true);
            builder.args(WEBGL_ARGS.iter().copied().map(OsStr::new).collect());
        }
        LaunchMode::Headless => {
            // Historical default: legacy headless, GPU left disabled by the crate.
            builder.headless(true);
        }
    }

    builder.build().map_err(|e| ToolExecError(e.to_string()))
}

/// True when `url` has an http, https, or file scheme — the schemes a headless
/// browser can be asked to render. `file://` lets the browser read arbitrary
/// local files from the daemon's host; that reach is intended and is not gated
/// behind any option (the browser process already runs under the same OS-level
/// sandbox as the daemon), so `file://` URLs are passed through untouched.
fn validate_url(url: &str) -> Result<(), ToolExecError> {
    let parsed = Url::parse(url).map_err(|e| ToolExecError(format!("invalid URL '{url}': {e}")))?;
    match parsed.scheme() {
        "http" | "https" | "file" => Ok(()),
        other => Err(ToolExecError(format!(
            "unsupported URL scheme '{other}'; only http/https/file are allowed"
        ))),
    }
}

/// Build the JS expression that extracts text (innerText) from a bound node.
/// `selector` is JSON-embedded so it can't break out of the string literal.
fn text_expression(selector: Option<&str>) -> String {
    match selector {
        Some(sel) => {
            let sel = serde_json::to_string(sel).unwrap_or_else(|_| "\"body\"".to_string());
            format!(
                "(() => {{ const e = document.querySelector({sel}); return e ? e.innerText : ''; }})()"
            )
        }
        None => "(() => { const e = document.body; return e ? e.innerText : ''; })()".to_string(),
    }
}

/// Build the JS expression that extracts HTML (outerHTML) from a bound node.
fn html_expression(selector: &str) -> String {
    let sel = serde_json::to_string(selector).unwrap_or_else(|_| "\"html\"".to_string());
    format!("(() => {{ const e = document.querySelector({sel}); return e ? e.outerHTML : ''; }})()")
}

/// JS that measures the page's full scrollable content size, returned as a
/// `"<width>x<height>"` string. It's a *string* (not an array/object) on
/// purpose: the crate's `evaluate` hard-codes `returnByValue: false`, which
/// always yields `.value` for primitives but not for objects, so a primitive
/// result is the only one we can read back without a handle.
const PAGE_SIZE_JS: &str = "(() => { const d = document.documentElement; const w = Math.max(d.scrollWidth, d.clientWidth); \
     const h = Math.max(d.scrollHeight, d.clientHeight); return w + 'x' + h; })()";

/// Measure the full scrollable content size of the rendered page.
fn page_content_size(tab: &Tab) -> Result<(f64, f64), ToolExecError> {
    let obj = tab
        .evaluate(PAGE_SIZE_JS, false)
        .map_err(|e| ToolExecError(format!("failed to measure page size: {e:#}")))?;
    let raw = remote_text(&obj);
    let (w, h) = raw
        .split_once('x')
        .and_then(|(w, h)| Some((w.parse::<f64>().ok()?, h.parse::<f64>().ok()?)))
        .unwrap_or((0.0, 0.0));
    debug!(raw_width = %raw, "measured page content size");
    Ok((w, h))
}

/// JS that attempts to create a WebGL context and returns its version string,
/// or an empty string when no context can be created. The result is a
/// *primitive* string on purpose: the crate's `evaluate` hard-codes
/// `returnByValue: false`, which yields `.value` only for primitives — see
/// `PAGE_SIZE_JS` for the same rationale.
const WEBGL_PROBE_JS: &str = "(() => { const c = document.createElement('canvas'); \
     const gl = c.getContext('webgl2') || c.getContext('webgl'); \
     return gl ? gl.getParameter(gl.VERSION) : ''; })()";

/// Prefix introducing the WebGL status line appended to the tool result, so it
/// is always delimited from captured page content — the `content`/`text`
/// actions otherwise return exactly the page's HTML/text, and a bare trailing
/// note could be misread as part of it.
const WEBGL_NOTE_PREFIX: &str = "\n\n[webgl] ";

/// Ask the page whether a WebGL context is actually usable, returning its
/// version string (e.g. `"WebGL 2.0 (…)"`) or `None` when none was created.
///
/// A flag being present is not proof WebGL works: driver/host/Chrome-version
/// variance means the only reliable signal is asking the page itself. This is
/// why the tool probes rather than trusting the launch configuration.
fn probe_webgl(tab: &Tab) -> Option<String> {
    match tab.evaluate(WEBGL_PROBE_JS, false) {
        Ok(obj) => {
            let version = remote_text(&obj);
            let version = version.trim();
            if version.is_empty() {
                None
            } else {
                Some(version.to_string())
            }
        }
        Err(e) => {
            warn!(error = %e, "retrieve_webpage: WebGL probe failed");
            None
        }
    }
}

/// One-line report appended to the tool result when WebGL was requested: the
/// context version on success, or an explicit "unavailable" note (plus a
/// warning) when no context could be created, so a blank canvas render is never
/// mistaken for success.
fn webgl_note(tab: &Tab) -> String {
    if let Some(version) = probe_webgl(tab) {
        format!("WebGL: {version}")
    } else {
        warn!("retrieve_webpage: WebGL requested but no context was created");
        "WebGL: unavailable (no context created)".to_string()
    }
}

/// JS that measures an element's bounding box in **document space** (bounding
/// client rect plus current scroll offsets), returned as a `"x,y,w,h"` string.
/// A string (not an array/object) on purpose: the crate's `evaluate`
/// hard-codes `returnByValue: false`, which yields `.value` only for
/// primitives — see `PAGE_SIZE_JS` for the same rationale.
const ELEMENT_BOX_JS_TEMPLATE: &str = "(() => { const e = document.querySelector({sel}); if (!e) return ''; \
     const r = e.getBoundingClientRect(); \
     return (r.left + window.scrollX) + ',' + (r.top + window.scrollY) + ',' + r.width + ',' + r.height; })()";

/// Measure `selector`'s bounding box in document-space coordinates
/// (`(x, y, width, height)`). Returns `None` when the selector matches nothing
/// (the caller turns that into a user-facing "selector not found" error).
fn element_document_box(
    tab: &Tab,
    selector: &str,
) -> Result<Option<(f64, f64, f64, f64)>, ToolExecError> {
    let sel = serde_json::to_string(selector)
        .map_err(|e| ToolExecError(format!("failed to encode selector: {e}")))?;
    let expr = ELEMENT_BOX_JS_TEMPLATE.replace("{sel}", &sel);
    let obj = tab
        .evaluate(&expr, false)
        .map_err(|e| ToolExecError(format!("failed to measure element box: {e:#}")))?;
    let raw = remote_text(&obj);
    let parsed: Option<(f64, f64, f64, f64)> = raw
        .split(',')
        .map(str::parse::<f64>)
        .collect::<Result<Vec<_>, _>>()
        .ok()
        .and_then(|v| {
            <[f64; 4]>::try_from(v)
                .ok()
                .map(|a| (a[0], a[1], a[2], a[3]))
        });
    debug!(raw = %raw, "measured element document box");
    Ok(parsed)
}

/// Capture a screenshot of the whole visible viewport (or a single element, via
/// `selector`). When `full_page`, this returns the entire scrollable page.
///
/// `headless_chrome`'s `Tab::capture_screenshot` maps its last boolean to the CDP
/// `fromSurface` flag — *not* "full page" — and hard-codes
/// `captureBeyondViewport: None`, so it can only ever capture the current
/// viewport. For a real full-page shot we issue `Page.captureScreenshot`
/// ourselves with `captureBeyondViewport: true` plus a clip covering the whole
/// content area (measured in-page first).
fn capture_screenshot(
    tab: &Tab,
    selector: Option<&str>,
    full_page: bool,
) -> Result<Vec<u8>, ToolExecError> {
    if let Some(sel) = selector {
        // Element screenshots must NOT go through headless_chrome's
        // `Element::capture_screenshot`: it clips against the *viewport* with
        // `captureBeyondViewport` unset, so anything below the fold captures
        // blank body background. Instead, take the element's box in document
        // space and clip against it with `captureBeyondViewport: true` — the
        // same trick the full-page path uses below, so off-screen elements
        // render fully regardless of scroll position or sticky headers.
        let Some((x, y, w, h)) = element_document_box(tab, sel)? else {
            return Err(ToolExecError(format!(
                "selector '{sel}' matched no element"
            )));
        };
        if w <= 0.0 || h <= 0.0 {
            return Err(ToolExecError(format!(
                "selector '{sel}' matched an element with a zero-size box"
            )));
        }
        let result = tab
            .call_method(Page::CaptureScreenshot {
                format: Some(CaptureScreenshotFormatOption::Png),
                quality: None,
                clip: Some(Page::Viewport {
                    x,
                    y,
                    width: w,
                    height: h,
                    scale: 1.0,
                }),
                from_surface: Some(true),
                capture_beyond_viewport: Some(true),
                optimize_for_speed: None,
            })
            .map_err(|e| ToolExecError(format!("failed to screenshot element: {e:#}")))?;
        return BASE64
            .decode(result.data)
            .map_err(|e| ToolExecError(format!("element screenshot decode failed: {e}")));
    }

    if full_page {
        let (w, h) = page_content_size(tab)?;
        // A clip of (0,0,contentW,contentH) plus captureBeyondViewport lets
        // headless Chrome paint the regions outside the current viewport too.
        if w > 0.0 && h > 0.0 {
            let result = tab
                .call_method(Page::CaptureScreenshot {
                    format: Some(CaptureScreenshotFormatOption::Png),
                    quality: None,
                    clip: Some(Page::Viewport {
                        x: 0.0,
                        y: 0.0,
                        width: w,
                        height: h,
                        scale: 1.0,
                    }),
                    from_surface: Some(true),
                    capture_beyond_viewport: Some(true),
                    optimize_for_speed: None,
                })
                .map_err(|e| {
                    ToolExecError(format!("failed to capture full-page screenshot: {e:#}"))
                })?;
            return BASE64
                .decode(result.data)
                .map_err(|e| ToolExecError(format!("full-page screenshot decode failed: {e}")));
        }
        debug!(width = %w, height = %h, "content size was unusable; falling back to viewport shot");
    }

    tab.capture_screenshot(CaptureScreenshotFormatOption::Png, None, None, true)
        .map_err(|e| ToolExecError(format!("failed to capture screenshot: {e:#}")))
}

/// Decode PNG dimensions by reading only the 8-byte IHDR detail block, rather
/// than decoding the entire (potentially multi-MB) image. Screenshots here are
/// always PNG, so this is both faster and lighter than `image::load_from_memory`.
fn png_dimensions(bytes: &[u8]) -> Option<(u32, u32)> {
    // Signature (8) + chunk length (4) + "IHDR" (4) + width (4) + height (4).
    const SIGNATURE: [u8; 8] = [0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A];
    // Compose the length guard with the accesses (clippy::indexing_slicing):
    // `get` returns None for short/malformed input, which maps to the same
    // `None` result the len() check produced.
    if bytes.len() >= 24 && bytes.get(..8) == Some(&SIGNATURE) && bytes.get(12..16) == Some(b"IHDR")
    {
        let width = u32::from_be_bytes(bytes.get(16..20)?.try_into().ok()?);
        let height = u32::from_be_bytes(bytes.get(20..24)?.try_into().ok()?);
        Some((width, height))
    } else {
        None
    }
}

/// The `retrieve_webpage` tool's return value: a human-readable text handle
/// plus an optional screenshot, so the framework's `extract_image` hook reads
/// the image straight off the per-invocation return value (no shared state).
/// Only the Screenshot action sets `image`; Content/Text/Pdf set `None`. `impl
/// Serialize` emits only `text`, so the JSON tool result is a plain string
/// exactly as before.
#[derive(Debug)]
pub struct RetrieveWebpageReturn {
    /// The text handle (captured content, screenshot message, or saved-PDF
    /// message) shown to the model.
    pub text: String,
    /// The prepared screenshot handed to the client via `extract_image`, when
    /// the action was Screenshot.
    pub image: Option<PreparedImage>,
}

impl Serialize for RetrieveWebpageReturn {
    /// Serialize to just the text handle, keeping the JSON wire format a plain
    /// string (identical to the previous `Return = String`).
    fn serialize<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
        serializer.serialize_str(&self.text)
    }
}

impl JsonSchema for RetrieveWebpageReturn {
    fn schema_name() -> std::borrow::Cow<'static, str> {
        std::borrow::Cow::Borrowed("RetrieveWebpageReturn")
    }

    fn json_schema(_gen: &mut schemars::SchemaGenerator) -> schemars::Schema {
        schemars::json_schema!({ "type": "string" })
    }
}

/// A tool that renders a URL in a local headless Chromium/Chrome and returns
/// page content (HTML), plain text, a screenshot, or a PDF.
pub struct RetrieveWebpage {}

impl RetrieveWebpage {
    pub fn new() -> Self {
        RetrieveWebpage {}
    }
}

impl Default for RetrieveWebpage {
    fn default() -> Self {
        Self::new()
    }
}

impl super::Tool for RetrieveWebpage {
    type Args = RetrieveWebpageArgs;
    type Return = RetrieveWebpageReturn;
    type Error = ToolExecError;

    fn name(&self) -> &'static str {
        "retrieve_webpage"
    }
    fn description(&self) -> &'static str {
        "Render a URL in a local headless Chromium/Chrome and return page content (HTML), plain text, a screenshot (PNG), or a PDF. Runs locally and offline; requires a chromium/chrome binary already installed (prefers chromium; override with CHROMIUM_BIN). Screenshots are returned inline or saved to output_path; PDFs require output_path. Set webgl true for WebGL-capable rendering (new headless + software SwiftShader); the result then reports the WebGL version or that no context could be created. One-shot per call — no persistent session."
    }
    fn describe_invocation(&self, args: &Self::Args) -> String {
        let action = args
            .action
            .as_ref()
            .map_or(WebpageAction::Content.as_str(), |a| a.as_str());
        let mut parts = vec![format!(
            "Retrieving web page ({action}). URL: {}.",
            args.url
        )];
        if let Some(sel) = args.selector.as_deref() {
            parts.push(format!(" Selector: {sel}."));
        }
        if args.webgl.unwrap_or(false) {
            parts.push(" WebGL enabled.".to_string());
        }
        if let Some(out) = args.output_path.as_deref() {
            parts.push(format!(" Output: {out}."));
        }
        parts.concat()
    }

    fn return_string(ret: &Self::Return) -> String {
        ret.text.clone()
    }

    fn execute(
        &self,
        args: Self::Args,
        _x_credentials: Option<&ServiceCredential>,
        working_dir: Option<&Path>,
        _ctx: Option<&ToolContext>,
    ) -> Result<Self::Return, Self::Error> {
        let action = args.action.unwrap_or(WebpageAction::Content);
        let url = args.url.trim();
        let action_str = action.as_str();
        validate_url(url)?;
        info!(url, action = action_str, "retrieve_webpage: rendering page");

        let binary = resolve_browser_binary().ok_or_else(|| {
            warn!("retrieve_webpage: no chromium/chrome binary found on PATH");
            ToolExecError(
                "no chromium or chrome binary found on PATH (or in standard locations). \
                 Install Chromium/Chrome, or set CHROMIUM_BIN / CHROME_BIN to its path"
                    .to_string(),
            )
        })?;
        debug!(path = %binary.display(), "resolved browser binary");

        // Navigation timeout is clamped so a single call can't pin a worker
        // thread indefinitely (a mistyped or hostile argument would otherwise).
        let timeout_ms = args
            .timeout_ms
            .unwrap_or(DEFAULT_TIMEOUT_MS)
            .min(MAX_TIMEOUT_MS);
        let wait_ms = args.wait_ms.map(|ms| ms.min(MAX_WAIT_MS));

        let mode = if args.webgl.unwrap_or(false) {
            LaunchMode::WebGl
        } else {
            LaunchMode::Headless
        };

        // Launch a private, headless, one-shot browser instance with an
        // explicit path (so it uses the resolved chromium, never auto-detect)
        // and our viewport. `mode` selects the WebGL-capable flag set instead
        // of the historical legacy-headless/GPU-disabled default.
        let options = build_launch_options(
            binary,
            args.width.unwrap_or(1280),
            args.height.unwrap_or(800),
            timeout_ms,
            mode,
        )?;

        let browser = Browser::new(options)
            .map_err(|e| ToolExecError(format!("failed to launch headless browser: {e:#}")))?;
        debug!(?mode, "launched headless browser");

        // Run the whole capture in a closure so the `Browser` (and its Chromium
        // child process) is released on every path, success or error, when it
        // drops. Explicit `close()` is best-effort on top of that.
        let outcome = (|| -> Result<RetrieveWebpageReturn, ToolExecError> {
            let tab = browser
                .new_tab()
                .map_err(|e| ToolExecError(format!("failed to open a tab: {e:#}")))?;

            tab.set_default_timeout(Duration::from_millis(timeout_ms));
            tab.navigate_to(url)
                .map_err(|e| ToolExecError(format!("navigation to {url} failed: {e:#}")))?;
            tab.wait_until_navigated()
                .map_err(|e| ToolExecError(format!("page never finished loading: {e:#}")))?;

            // Optional settle delay so client-side JS can run before capture.
            // Only reached at runtime (never in unit tests, per repo rules).
            if let Some(ms) = wait_ms
                && ms > 0
            {
                std::thread::sleep(Duration::from_millis(ms));
            }
            debug!(timeout_ms, wait_ms, "page navigated; capturing");

            let mut result = Self::capture(&tab, &args, action, url, working_dir)?;
            // When WebGL was requested, report whether a context was actually
            // created — flags in argv are not proof the host can render WebGL,
            // so surface the real outcome regardless of which action ran. The
            // note is delimited (`WEBGL_NOTE_PREFIX`) so it can't be mistaken
            // for captured page content.
            if mode == LaunchMode::WebGl {
                result.text.push_str(WEBGL_NOTE_PREFIX);
                result.text.push_str(&webgl_note(&tab));
            }
            Ok(result)
        })();

        match &outcome {
            Ok(..) => info!(url, action = action_str, "retrieve_webpage: ok"),
            Err(e) => warn!(url, action = action_str, error = %e, "retrieve_webpage: failed"),
        }

        // The `Browser` is dropped here (and its Chromium child process is
        // terminated by the transport's `Drop`), releasing the instance whether
        // the capture succeeded or failed — see the closure above.
        outcome
    }

    fn extract_image(&self, ret: &Self::Return) -> Option<PreparedImage> {
        ret.image.clone()
    }
}

impl RetrieveWebpage {
    /// Perform the per-action capture on an already-navigated tab. Broken out
    /// of `execute` so each branch stays small and testable; handles the file
    /// write + inline-image hand-off for the binary actions (screenshot/pdf).
    /// The screenshot is returned inside the [`RetrieveWebpageReturn`] so the
    /// framework's `extract_image` hook reads it off the per-invocation return
    /// value (no shared-state parking).
    fn capture(
        tab: &Tab,
        args: &RetrieveWebpageArgs,
        action: WebpageAction,
        url: &str,
        working_dir: Option<&Path>,
    ) -> Result<RetrieveWebpageReturn, ToolExecError> {
        match action {
            WebpageAction::Content => match args.selector.as_deref() {
                Some(sel) => {
                    let obj = tab
                        .evaluate(&html_expression(sel), false)
                        .map_err(|e| ToolExecError(format!("failed to extract HTML: {e:#}")))?;
                    Ok(RetrieveWebpageReturn {
                        text: remote_text(&obj),
                        image: None,
                    })
                }
                None => tab
                    .get_content()
                    .map(|text| RetrieveWebpageReturn { text, image: None })
                    .map_err(|e| ToolExecError(format!("failed to get page HTML: {e:#}"))),
            },

            WebpageAction::Text => {
                let expr = text_expression(args.selector.as_deref());
                let obj = tab
                    .evaluate(&expr, false)
                    .map_err(|e| ToolExecError(format!("failed to extract text: {e:#}")))?;
                Ok(RetrieveWebpageReturn {
                    text: remote_text(&obj),
                    image: None,
                })
            }

            WebpageAction::Screenshot => {
                let bytes = capture_screenshot(
                    tab,
                    args.selector.as_deref(),
                    args.full_page.unwrap_or(true),
                )?;

                // Dimensions come from a cheap PNG-header read, not a full decode.
                let (width, height) = png_dimensions(&bytes).unwrap_or((0, 0));
                let size = bytes.len();
                let alt = Some(format!("Screenshot of {url}"));

                // Optionally persist to output_path, then carry the buffer in the
                // return value for `extract_image`. The write happens *before* the
                // buffer moves into the return, so a potentially multi-MB
                // screenshot is never cloned.
                let message = match args.output_path.as_deref() {
                    Some(out) => {
                        let path = resolve_path(out, working_dir);
                        write_bytes_with_dirs(&path, &bytes)?;
                        format!(
                            "captured screenshot ({width}x{height}, PNG, {size}); saved to {path}",
                            size = human_size(size as u64),
                            path = path.display(),
                        )
                    }
                    None => {
                        format!(
                            "captured screenshot ({width}x{height}, PNG, {size})",
                            size = human_size(size as u64),
                        )
                    }
                };

                // Always offer the screenshot inline to the client.
                Ok(RetrieveWebpageReturn {
                    text: message,
                    image: Some(PreparedImage {
                        mime_type: "image/png".to_string(),
                        data: bytes,
                        width,
                        height,
                        alt,
                    }),
                })
            }

            WebpageAction::Pdf => {
                let out = args.output_path.as_deref().ok_or_else(|| {
                    ToolExecError(
                        "pdf action requires output_path so the binary can be saved".to_string(),
                    )
                })?;
                let bytes = tab
                    .print_to_pdf(None)
                    .map_err(|e| ToolExecError(format!("failed to render PDF: {e:#}")))?;
                let path = resolve_path(out, working_dir);
                write_bytes_with_dirs(&path, &bytes)?;
                Ok(RetrieveWebpageReturn {
                    text: format!(
                        "saved PDF ({size}) to {path}",
                        size = human_size(bytes.len() as u64),
                        path = path.display(),
                    ),
                    image: None,
                })
            }
        }
    }
}

/// Pull the returned string out of a CDP `RemoteObject`: the protocol encodes
/// a primitive string result as a JSON value (`value.as_str()`).
fn extract_text(value: &serde_json::Value) -> String {
    value.as_str().map(str::to_owned).unwrap_or_default()
}

/// Convenience: unwrap a `RemoteObject`'s optional `value` and extract text.
fn remote_text(object: &headless_chrome::protocol::cdp::Runtime::RemoteObject) -> String {
    object.value.as_ref().map(extract_text).unwrap_or_default()
}

/// Write `bytes` to `path`, creating parent directories as needed.
fn write_bytes_with_dirs(path: &Path, bytes: &[u8]) -> Result<(), ToolExecError> {
    if let Some(parent) = path.parent()
        && !parent.as_os_str().is_empty()
    {
        std::fs::create_dir_all(parent)
            .map_err(|e| ToolExecError(format!("failed to create output dir: {e}")))?;
    }
    std::fs::write(path, bytes)
        .map_err(|e| ToolExecError(format!("failed to write '{}': {e}", path.display())))
}

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

    #[test]
    fn validate_url_accepts_http_https_and_file() {
        for u in [
            "https://example.com",
            "http://example.com/path?q=1",
            "file:///etc/passwd",
            "file:///tmp/foo.html",
            "file://localhost/etc/passwd",
        ] {
            assert!(validate_url(u).is_ok(), "{u} should be accepted");
        }
    }

    #[test]
    fn validate_url_rejects_other_schemes() {
        for u in [
            "javascript:alert(1)",
            "ftp://x",
            "data:text/html,hi",
            "not a url",
        ] {
            let err = validate_url(u).unwrap_err();
            assert_ne!(err.to_string(), "");
        }
    }

    #[test]
    fn text_expression_embeds_selector_safely() {
        // A selector containing a quote must be JSON-escaped, not interpolated,
        // so it cannot break out of the string literal.
        let expr = text_expression(Some("div[data-x=\"y\"]"));
        assert!(expr.contains("\\\"y\\\""));
        assert!(expr.contains("innerText"));
    }

    #[test]
    fn text_expression_defaults_to_body() {
        let expr = text_expression(None);
        assert!(expr.contains("document.body"));
        assert!(expr.contains("innerText"));
    }

    #[test]
    fn html_expression_uses_outer_html() {
        let expr = html_expression("#main");
        assert!(expr.contains("#main"));
        assert!(expr.contains("outerHTML"));
    }

    #[test]
    fn describe_invocation_defaults_to_content() {
        let tool = RetrieveWebpage::new();
        let args = RetrieveWebpageArgs {
            url: "https://example.com".to_string(),
            ..RetrieveWebpageArgs::default()
        };
        let desc = tool.describe_invocation(&args);
        assert!(desc.contains("content"));
        assert!(desc.contains("https://example.com"));
    }

    #[test]
    fn describe_invocation_includes_selector_and_output() {
        let tool = RetrieveWebpage::new();
        let args = RetrieveWebpageArgs {
            url: "https://example.com".to_string(),
            action: Some(WebpageAction::Screenshot),
            selector: Some("#main".to_string()),
            output_path: Some("shot.png".to_string()),
            ..RetrieveWebpageArgs::default()
        };
        let desc = tool.describe_invocation(&args);
        assert!(desc.contains("screenshot"));
        assert!(desc.contains("#main"));
        assert!(desc.contains("shot.png"));
    }

    #[test]
    fn extract_text_handles_string_and_absent_value() {
        assert_eq!(extract_text(&serde_json::json!("hello")), "hello");
        // Non-string / null values must degrade to an empty string.
        assert_eq!(extract_text(&serde_json::Value::Null), "");
        assert_eq!(extract_text(&serde_json::json!(42)), "");
    }

    #[test]
    fn png_dimensions_reads_ihdr() {
        // A minimal-but-valid-enough PNG header carrying 12x3456 dimensions.
        let mut png = vec![0x89, b'P', b'N', b'G', 0x0D, 0x0A, 0x1A, 0x0A];
        png.extend_from_slice(&[0, 0, 0, 13]); // IHDR chunk length
        png.extend_from_slice(b"IHDR");
        png.extend_from_slice(&12u32.to_be_bytes());
        png.extend_from_slice(&3456u32.to_be_bytes());
        assert_eq!(png_dimensions(&png), Some((12, 3456)));
    }

    #[test]
    fn png_dimensions_rejects_garbage() {
        assert_eq!(png_dimensions(b"not a png"), None);
        assert_eq!(png_dimensions(&[0u8; 32]), None);
    }

    #[test]
    fn launch_mode_defaults_to_headless() {
        assert_eq!(LaunchMode::default(), LaunchMode::Headless);
    }

    #[test]
    fn launch_options_webgl_enables_swiftshader_flags() {
        let opts = build_launch_options(
            std::path::PathBuf::from("/usr/bin/chromium"),
            1280,
            800,
            30_000,
            LaunchMode::WebGl,
        )
        .expect("build launch options");

        // WebGL needs new headless, so the crate's legacy `--headless` (which it
        // appends AFTER our custom flags and would win as the last duplicate
        // switch) must be off, and the crate's `--disable-gpu` must be dropped.
        assert!(!opts.headless, "webgl must not use legacy --headless");
        assert!(opts.enable_gpu, "webgl must drop --disable-gpu");

        // Assert against the same single-sourced list production uses, so the
        // test can never pass with a flag set that has drifted from `WEBGL_ARGS`.
        let args: Vec<&str> = opts.args.iter().filter_map(|a| a.to_str()).collect();
        for expected in WEBGL_ARGS {
            assert!(args.contains(expected), "missing {expected} in {args:?}");
        }
    }

    #[test]
    fn launch_options_default_matches_historical_config() {
        let opts = build_launch_options(
            std::path::PathBuf::from("/usr/bin/chromium"),
            1280,
            800,
            30_000,
            LaunchMode::Headless,
        )
        .expect("build launch options");

        assert!(opts.headless, "default keeps legacy headless");
        assert!(!opts.enable_gpu, "default keeps the crate's --disable-gpu");
        assert!(opts.args.is_empty(), "default must add no custom flags");
    }

    #[test]
    fn webgl_probe_js_requests_a_context_and_reads_version() {
        assert!(WEBGL_PROBE_JS.contains("getContext('webgl2')"));
        assert!(WEBGL_PROBE_JS.contains("getContext('webgl')"));
        assert!(WEBGL_PROBE_JS.contains("VERSION"));
    }

    #[test]
    fn describe_invocation_mentions_webgl_only_when_enabled() {
        let tool = RetrieveWebpage::new();
        let on = RetrieveWebpageArgs {
            url: "https://example.com".to_string(),
            webgl: Some(true),
            ..RetrieveWebpageArgs::default()
        };
        assert!(tool.describe_invocation(&on).contains("WebGL enabled"));

        let off = RetrieveWebpageArgs {
            url: "https://example.com".to_string(),
            ..RetrieveWebpageArgs::default()
        };
        assert!(!tool.describe_invocation(&off).contains("WebGL enabled"));
    }
}