File size: 40,312 Bytes
9d0fd45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

"""read_file parsers Β· core fragment (concatenated first)."""
import hashlib
import json
import os as _os
import re as _re
import subprocess
import sys

_CACHE_DIR = "/workspace/.readdoc_cache"  # bind-mounted, persists across commands; missing/unwritable β†’ silently re-render

# Diagnostic trace: only when env READDOC_TRACE is set, record each step's routing and
# backend calls and dump them to stderr wrapped in sentinels when main() ends β€”
# stdout is always clean markdown (what the LLM sees); the trace is for debugging only.
_TRACE: list = []


def _trace(ev: dict) -> None:
    if _os.environ.get("READDOC_TRACE"):
        _TRACE.append(ev)


def _dump_trace() -> None:
    if _TRACE:
        sys.stderr.write("\n<<<READDOC_TRACE\n" + json.dumps(_TRACE, ensure_ascii=False)
                         + "\nREADDOC_TRACE>>>\n")


def _ensure(mod: str, pkg: str) -> None:
    try:
        __import__(mod)
    except ImportError:
        subprocess.run(
            [sys.executable, "-m", "pip", "install", "-q", pkg],
            stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
        )


# Vision reading (charts / figures / text inside an image): the reader calls the gateway
# or a self-hosted VLM (OpenAI-compatible chat/completions) itself and inlines the
# result into the readout β€” view_image is a low-level tool not exposed to the main LLM,
# so the reader has to read and return images on its own.
# Config (env, reachable inside the sandbox): READDOC_VISION_URL (include /v1) /
# READDOC_VISION_MODEL / READDOC_VISION_KEY.
_VISION_PROMPT = (
    "Reproduce ALL content of this image faithfully and completely; do not summarize or guess. "
    "First, one line: what it is (chart/diagram/table/form/photo/screenshot). Then: transcribe text "
    "verbatim (exact numbers/units/labels); for any table preserve rows/columns and which cell each "
    "value belongs to; for a chart/diagram give title, axes, legend, series and the values/relationships "
    "it conveys; for purely visual elements describe only what carries information. Keep reading order. "
    "Use [illegible] rather than guessing."
)
# Figures only (for mixed pages): body text and ordinary tables were already extracted
# by OCR, so this only adds charts/figures and avoids duplicating the text.
_VISION_FIGURE_PROMPT = (
    "This is a full page image that may contain charts/plots/diagrams/flowcharts/infographics, "
    "possibly alongside body text and plain tables. Extract ONLY the visual figures β€” for EACH figure: "
    "its title, axis labels and scales, legend, data series, and the concrete values or relationships it "
    "conveys (read approximate values off the axes when not labeled). Do NOT transcribe ordinary paragraph "
    "text, headings, or plain data tables β€” those are captured elsewhere. Process figures in reading order; "
    "use [illegible] rather than guessing. If the page has no real figure (only text/logos), reply exactly: NO_FIGURE."
)


def _vision_read(img_bytes: bytes, mime: str = "image/png", question: str | None = None,
                 timeout: int = 120):
    """One image β†’ VLM text. READDOC_VISION_URL unset or a failure β†’ None (the caller
    falls back).
    timeout: lower for batch reads (45s) so the whole batch converges inside the outer
    sandbox timeout."""
    base = _os.environ.get("READDOC_VISION_URL", "").rstrip("/")
    if not base:
        return None
    import base64 as _b64
    import json as _json
    import urllib.request as _ur
    key = _os.environ.get("READDOC_VISION_KEY", "EMPTY")
    model = _os.environ.get("READDOC_VISION_MODEL", "")
    import time as _t
    b = _b64.b64encode(img_bytes).decode("ascii")
    # Reasoning models (Qwen3 and friends) burn the budget on reasoning by default β†’
    # max_tokens exhausted, content empty, and very slow
    # (measured: 130s for 4096 tok, still finish_reason=length with content=None).
    # Turning thinking off answers directly β€” faster and more accurate.
    # Most OpenAI-compatible servers ignore unknown fields; if an endpoint rejects it,
    # set READDOC_VISION_THINK=1 to disable this switch.
    payload = {"model": model, "max_tokens": 4096, "temperature": 0.2, "messages": [
        {"role": "user", "content": [
            {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b}"}},
            {"type": "text", "text": question or _VISION_PROMPT}]}]}
    if _os.environ.get("READDOC_VISION_THINK", "0") != "1":
        payload["chat_template_kwargs"] = {"enable_thinking": False}
    t0 = _t.time()
    try:
        req = _ur.Request(base + "/chat/completions", data=_json.dumps(payload).encode("utf-8"),
                          headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"})
        r = _json.loads(_ur.urlopen(req, timeout=timeout).read())
        out = r["choices"][0]["message"]["content"]
        _trace({"stage": "vision", "model": model, "bytes": len(img_bytes),
                "ms": int((_t.time() - t0) * 1000), "ok": bool(out),
                "figure_only": question is not None})
        return out
    except Exception as exc:
        if "chat_template_kwargs" in payload:
            payload.pop("chat_template_kwargs", None)
            try:
                req = _ur.Request(base + "/chat/completions", data=_json.dumps(payload).encode("utf-8"),
                                  headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"})
                r = _json.loads(_ur.urlopen(req, timeout=timeout).read())
                out = r["choices"][0]["message"]["content"]
                _trace({"stage": "vision", "model": model, "bytes": len(img_bytes),
                        "ms": int((_t.time() - t0) * 1000), "ok": bool(out),
                        "figure_only": question is not None, "retry": True})
                return out
            except Exception:
                pass
        _trace({"stage": "vision", "model": model, "bytes": len(img_bytes),
                "ms": int((_t.time() - t0) * 1000), "ok": False, "error": str(exc)})
        return None


def _ext(path: str) -> str:
    return path.rsplit(".", 1)[-1].lower() if "." in path else ""


# ------------- Unstructured files: sniff and route (text / imageΒ·audioΒ·video / unsupported binary) -------------
# Decided by content, not by an extension whitelist, so unlisted formats are still
# handled. Rule: never pipe binary bytes into the text channel.

# Bytes allowed in the text verdict: common control chars (BEL/BS/TAB/LF/FF/CR/ESC)
# plus 0x20-0xFF (covers utf-8 / latin-1 high bytes)
_TEXT_BYTES = bytes([7, 8, 9, 10, 12, 13, 27]) + bytes(range(0x20, 0x100))

# These formats are returned verbatim with **no line numbers**: json/yaml have to stay
# parseable and md has to stay renderable, so line numbers are pure noise;
# they only help for code / logs / plain txt (file:line references, str_replace edits).
_NO_LINENO_EXTS = {"json", "yaml", "yml", "md", "markdown"}


def _is_text(head: bytes) -> bool:
    """Does the first block look like text? A text BOM (utf-8/16/32) settles it as text
    (utf-16's NUL does not count as binary);
    otherwise a NUL β†’ binary; failing that, by "share of non-text control bytes < 30%"
    (the same test file(1) and git use)."""
    if not head:
        return True
    if (head[:4] in (b"\xff\xfe\x00\x00", b"\x00\x00\xfe\xff")
            or head[:3] == b"\xef\xbb\xbf" or head[:2] in (b"\xff\xfe", b"\xfe\xff")):
        return True
    if b"\x00" in head:
        return False
    nontext = head.translate(None, _TEXT_BYTES)
    return len(nontext) / len(head) < 0.30


def _magic_mime(head: bytes):
    """Binary magic bytes β†’ mime (only to name image/audio/video); None when unknown."""
    if head[:8] == b"\x89PNG\r\n\x1a\n":
        return "image/png"
    if head[:3] == b"\xff\xd8\xff":
        return "image/jpeg"
    if head[:6] in (b"GIF87a", b"GIF89a"):
        return "image/gif"
    if head[:2] == b"BM":
        return "image/bmp"
    if head[:4] in (b"II*\x00", b"MM\x00*"):
        return "image/tiff"
    if head[:4] == b"RIFF":
        sub = head[8:12]
        if sub == b"WEBP":
            return "image/webp"
        if sub == b"WAVE":
            return "audio/wav"
        if sub == b"AVI ":
            return "video/x-msvideo"
    if head[:3] == b"ID3" or head[:2] in (b"\xff\xfb", b"\xff\xf3", b"\xff\xf2"):
        return "audio/mpeg"
    if head[:4] == b"fLaC":
        return "audio/flac"
    if head[:4] == b"OggS":
        return "audio/ogg"
    if head[4:8] == b"ftyp":
        return "audio/mp4" if head[8:11] == b"M4A" else "video/mp4"
    if head[:4] == b"\x1aE\xdf\xa3":
        return "video/x-matroska"
    if head[:3] == b"FLV":
        return "video/x-flv"
    return None


def _decode_bytes(raw: bytes) -> str:
    """Bytes β†’ text, CJK-friendly. Order: BOM β†’ strict utf-8 β†’ charset_normalizer
    detection
    + scoring of verified CJK/Cyrillic/Western candidates β†’ utf-8 replace as the floor.
    The old chain's BOM-less utf-16 attempt (any even-length byte string can "succeed"
    into garbage) and its latin-1 floor
    (SJIS mapped byte-by-byte into Latin mojibake) were the root cause of garbled
    non-major-language text, and are gone."""
    if not raw:
        return ""
    if raw[:3] == b"\xef\xbb\xbf":
        return raw.decode("utf-8-sig", errors="replace")
    if raw[:4] in (b"\xff\xfe\x00\x00", b"\x00\x00\xfe\xff"):
        return raw.decode("utf-32", errors="replace")
    if raw[:2] in (b"\xff\xfe", b"\xfe\xff"):
        return raw.decode("utf-16", errors="replace")
    try:
        return raw.decode("utf-8")
    except UnicodeDecodeError:
        pass

    # Not utf-8: score the candidates. charset_normalizer's verdict competes rather than
    # deciding β€” measured, it misreads short Shift-JIS as EUC-KR, and calls
    # KOI8-R / Thai shift_jis (chaos=0 but coherence=0, i.e. no linguistic evidence at
    # all). Score every strictly-decodable candidate by character range:
    # full-width kana = strong Japanese signal; an all-half-width-katakana document is
    # the hallmark of a single-byte encoding misread by cp932, so it scores negative;
    # the normalizer candidate gets a small bonus as a tie-breaker (when it is right, it
    # should win).
    def _score(t: str, bonus: float = 0.0) -> float:
        if not t:
            return -1.0
        s = 0.0
        upper = lower = 0
        non_ascii = 0
        ascii_alpha = 0
        scripts = {
            "kana": 0, "hangul": 0, "cjk": 0, "latin": 0,
            "greek": 0, "cyrillic": 0, "arabic": 0, "hebrew": 0, "thai": 0,
        }
        sample = t[:8000]                   # score the first 8K chars; keeps big files fast
        for ch in sample:
            o = ord(ch)
            if o < 0x80:
                if ch.isalpha():
                    ascii_alpha += 1
                continue
            non_ascii += 1
            if ch.isupper():
                upper += 1
            elif ch.islower():
                lower += 1
            if 0x3040 <= o <= 0x30FF:      # full-width hiragana/katakana: strong Japanese signal
                s += 1
                scripts["kana"] += 1
            elif 0xAC00 <= o <= 0xD7A3:
                s += 1
                scripts["hangul"] += 1
            elif 0x4E00 <= o <= 0x9FFF:
                s += 1
                scripts["cjk"] += 1
            elif 0x00C0 <= o <= 0x024F:
                s += 1
                scripts["latin"] += 1
            elif 0x0370 <= o <= 0x03FF:
                s += 1
                scripts["greek"] += 1
            elif 0x0400 <= o <= 0x052F:
                s += 1
                scripts["cyrillic"] += 1
            elif 0x0590 <= o <= 0x05FF:
                s += 1
                scripts["hebrew"] += 1
            elif 0x0600 <= o <= 0x06FF:
                s += 1
                scripts["arabic"] += 1
            elif 0x0E00 <= o <= 0x0E7F:
                s += 1
                scripts["thai"] += 1
            elif 0xFF61 <= o <= 0xFF9F:    # half-width katakana: mis-decode hallmark (real Japanese is mostly full-width)
                s -= 2
            elif o == 0xFFFD or 0xE000 <= o <= 0xF8FF or 0x80 <= o <= 0x9F:
                s -= 5                      # replacement char / private use / C1 control: traces of a mis-decode
        # Denominator = number of non-ASCII chars: a single-byte codec misreading a
        # double-byte stream produces 2x the characters and halves the ASCII-space share,
        # so using total length would inflate the mis-decode's density past the correct one.
        score = s / max(non_ascii, 1) + bonus
        # Inverted-case penalty: the hallmark of KOI8 ↔ cp125x cross-decoding (the two
        # families swap case ranges, producing "lowercase initial, rest uppercase").
        # Normal text has far more lowercase than uppercase; caseless scripts are unaffected.
        cased = upper + lower
        if cased >= 10 and upper / cased > 0.6:
            score -= 0.5
        # Mixed writing systems are usually a strong signal of a misread double-byte
        # encoding (e.g. GBK β†’ CJK + Hangul).
        # Real Japanese does mix kanji and kana, so treat the two as one system only when
        # the kana share is high enough β€” one stray
        # kana must not launder mojibake. Give the dominant script a small purity bonus and
        # penalise the share held by the rest.
        if scripts["kana"] >= 2 and scripts["kana"] * 5 >= scripts["cjk"]:
            scripts["kana"] += scripts["cjk"]
            scripts["cjk"] = 0
        script_total = sum(scripts.values())
        if script_total:
            purity = max(scripts.values()) / script_total
            score += 0.5 * purity - 1.5 * (1.0 - purity)
            # When a double-byte stream (Big5 and similar) is misread by a single-byte
            # codec, stray ASCII letters get wedged between non-Latin characters;
            # a correct decode usually keeps whole CJK characters. Penalise lightly by the
            # interleaving ratio, so genuinely
            # ASCII/Latin-dominant Western candidates are not affected.
            if scripts["latin"] == 0 and ascii_alpha:
                score -= ascii_alpha / (script_total + ascii_alpha)
        return score

    scored: list = []
    try:
        _ensure("charset_normalizer", "charset-normalizer")
        from charset_normalizer import from_bytes
        best = from_bytes(raw).best()
        if best is not None:
            t = str(best)
            # ``charset_normalizer`` often returns coherence=0 guesses for very short text.
            # Chinese and Western single-byte short strings are especially prone to
            # same-script ties
            # (GB18030 "ζ΅‹θ―•" β†’ Big5 "聆彸"; cp1252 β†’ cp1250). Let it into the tie-break only
            # when coherence is genuinely non-zero, or when Korean/Japanese is detected via
            # their distinctive glyphs; otherwise defer to the verified ordering below.
            coherence = float(getattr(best, "coherence", 0.0) or 0.0)
            language = str(getattr(best, "language", "") or "").lower()
            if coherence > 0.1 or language in {"japanese", "korean"}:
                scored.append((_score(t, bonus=0.1), t))
    except Exception:
        pass
    # Verified range: multi-byte CJK + Cyrillic + Western. Inside one single-byte family
    # (koi8 vs cp1251) no language model can separate them, so tie-break by
    # market share. Unverified Hebrew/Arabic/Greek/Thai candidates are deliberately absent
    # rather than letting tuple order masquerade as language identification.
    for enc in ("cp932", "gb18030", "big5", "cp949", "euc_jp",
                "cp1251", "koi8_r", "cp1252", "iso8859_2"):
        try:
            t = raw.decode(enc)
        except (UnicodeDecodeError, UnicodeError, LookupError):
            continue
        scored.append((_score(t), t))
    if scored:
        return max(scored, key=lambda x: x[0])[1]
    return raw.decode("utf-8", errors="replace")


def _text_to_md(path: str) -> str:
    """Text file β†’ full content with line numbers (cat -n style, added by the parser);
    encoding via _decode_bytes
    (BOM / utf-8 / charset detection, CJK-friendly)."""
    with open(path, "rb") as f:
        raw = f.read()
    text = _decode_bytes(raw)
    if _ext(path) in _NO_LINENO_EXTS:
        # json/yaml/md: verbatim (parseable / renderable), no line numbers, no header note
        return text if text.strip() else "(empty file)"
    head = "<!-- text readout: leading 'N\\t' line numbers are parser-added, not file content. -->"
    lines = text.splitlines()
    if not lines:
        return head + "\n\n(empty file)"
    w = len(str(len(lines)))
    body = "\n".join(f"{i:>{w}}\t{ln}" for i, ln in enumerate(lines, 1))
    return head + "\n" + body


def _classify_read(path: str) -> str:
    """One path for every unstructured file: text β†’ numbered full content; image β†’ VLM
    read (type only when unconfigured);
    audio/video β†’ type only; anything else binary β†’ unsupported (raises).

    A read failure (cannot open, or an unknown binary that cannot be read as text) always
    raises, and main() turns that into one
    fail-closed path (stderr + non-zero exit); reporting a media type is valid output and
    returns normally.
    """
    name = _os.path.basename(path)
    with open(path, "rb") as f:  # OSError propagates up β†’ main() catches it
        head = f.read(65536)
    if _is_text(head):
        return _text_to_md(path)
    mime = _magic_mime(head)
    if mime and mime.split("/")[0] == "image":
        # A standalone image goes straight to the VLM (scans, charts, screenshots, photos all get read); vision unconfigured or failing β†’ fall back to reporting the type.
        try:
            with open(path, "rb") as f:
                data = f.read()
            vt = _vision_read(data, mime)
        except Exception:
            vt = None
        if vt:
            return f"<!-- image {name} | read via vision -->\n\n{vt}"
        return (f"file: {name}\ntype: {mime}\n"
                "note: image β€” vision unavailable (set READDOC_VISION_URL/MODEL/KEY); not read.")
    if mime:
        kind = mime.split("/")[0]  # audio / video
        return (f"file: {name}\ntype: {mime}\n"
                f"note: binary {kind} β€” type detected only, not read as text.")
    # No text and no known media type: no document content can be read β€” raise instead of
    # returning a notice, or save_to would
    # store the notice as the document and report "saved N bytes".
    raise ValueError(f"unsupported binary {name} (application/octet-stream) "
                     "β€” cannot be read as text.")


# ---------------- Pagination (hard-boundary cutting + continuation notes) ----------------
# The markdown each format renders already carries hard boundary lines, so pagination is
# post-processing and the renderers stay untouched.
# One boundary marker for all: an HTML comment <!-- page/slide/sheet/... --> (not a
# heading, so it can never collide with content heading levels;
# visible in raw text, easy to match). docx injects no marker and uses the document's
# own headings as boundaries.
_BOUNDARY_RE = {
    "xlsx": r"(?m)^<!-- (sheet:|range |charts)",
    "pptx": r"(?m)^<!-- (slide |needs VLM)",
    "pdf": r"(?m)^<!-- page ",
    "docx": r"(?m)^#{1,6} ",
    "image_batch": r"(?m)^===== image ",   # batch image read: window on whole images, never cut a transcription in half
}


def _split_blocks(md: str, fmt: str):
    """md β†’ list of hard-boundary block start offsets (the first block starts at 0; legend / lead-in belong to it)."""
    pat = _BOUNDARY_RE.get(fmt)
    starts = [m.start() for m in _re.finditer(pat, md)] if pat else []
    if not starts or starts[0] != 0:
        starts = [0, *starts]
    return starts


def _weak_cut(md: str, start: int, hard_end: int) -> int:
    """Degraded cut point when one block exceeds the budget: line boundaries (an xlsx grid row is one line, never cut inside a cell row)."""
    nl = md.rfind("\n", start + 1, hard_end)
    return nl + 1 if nl > start else hard_end


def _resume_ctx(md: str, fmt: str, block_start: int, resume_at: int) -> str:
    """Rebuild context when resuming inside a block: the block's first line (sheet / slide / page title); for xlsx also the nearest column-letter header row."""
    head_line = md[block_start:md.find("\n", block_start) + 1].rstrip()
    # When the boundary marker is an HTML comment, put (continued) inside it so the
    # comment stays valid
    if head_line.endswith("-->"):
        cont = head_line[:-3].rstrip() + " (continued) -->"
    else:
        cont = f"{head_line} (continued)"
    lines = [cont]
    if fmt == "xlsx":
        seg = md[block_start:resume_at]
        for ln in reversed(seg.splitlines()):
            if ln.startswith("|   |"):  # grid column-letter header
                lines.append(ln)
                lines.append("| --- |" + " --- |" * (ln.count("|") - 2))
                break
    return "\n".join(lines) + "\n"


def _paginate(md: str, fmt: str, offset: int, max_chars: int) -> str:
    """Return one page of md as [offset, …]: whole blocks until the budget is spent; always
    cut on a hard boundary
    (sole exception: a single block larger than the budget β†’ line boundary, noted in the
    message).
    Continuation notes are added at both ends; when one page holds the whole document and
    offset=0, the output is byte-identical to the source (zero additions)."""
    total = len(md)
    if max_chars <= 0:  # unset = send everything (offset still applies)
        return md[offset:] if offset else md
    if offset >= total:
        return (f"[read_file] offset {offset} >= total {total} chars; nothing left. "
                f"The document was fully covered by earlier reads.")
    starts = _split_blocks(md, fmt)
    # Snap the start: an offset landing mid-block (the model passed an arbitrary value, or
    # the last cut was line-level) β†’ keep the exact position and re-add context
    import bisect
    bi = bisect.bisect_right(starts, offset) - 1
    start = max(offset, 0)
    mid_block_start = start > starts[bi]

    budget_end = start + max_chars
    end = start
    j = bi
    forced_weak = False
    if mid_block_start:  # finish the remainder of the current block first
        be = starts[j + 1] if j + 1 < len(starts) else total
        if be <= budget_end:
            end = be
            j += 1
        else:
            end = _weak_cut(md, start, budget_end)
            forced_weak = True
    if not forced_weak:
        while j < len(starts):
            be = starts[j + 1] if j + 1 < len(starts) else total
            if be - start > max_chars:
                break
            end = be
            j += 1
        if end == start:  # the very first whole block already exceeds the budget β†’ degrade to line boundaries
            end = _weak_cut(md, start, budget_end)
            forced_weak = True

    body = md[start:end]
    done = end >= total
    if done and offset == 0:
        return body  # the whole document fits one page: identical to the old behaviour, zero additions

    parts = []
    if offset > 0:
        note = f"[read_file] continued read: chars {start}-{end} of {total}."
        parts.append(note + "\n")
        if mid_block_start:
            parts.append(_resume_ctx(md, fmt, starts[bi], start))
        parts.append("\n")
    parts.append(body)
    if not done:
        cut_desc = ("inside a section (single section exceeds max_chars; cut at a line "
                    "boundary)" if forced_weak else "at a section boundary")
        # Map of what is left: the unread blocks' title lines + their sizes
        remain = []
        k = bisect.bisect_right(starts, end) - 1
        if starts[k] < end:
            k += 1
        for idx in range(k, min(k + 10, len(starts))):
            s = starts[idx]
            e = starts[idx + 1] if idx + 1 < len(starts) else total
            title = md[s:md.find("\n", s) + 1].strip() or "(untitled)"
            remain.append(f"  - {title} (~{e - max(s, end)} chars)")
        more = len(starts) - k - len(remain)
        if more > 0:
            remain.append(f"  - … and {more} more sections")
        parts.append(
            f"\n\n[read_file] PARTIAL READ: returned chars {start}-{end} of {total}, "
            f"cut {cut_desc}. NOT finished β€” call read_file again with offset={end} "
            f"to continue."
            + ("\nRemaining:\n" + "\n".join(remain) if remain else "")
        )
    return "".join(parts)


# ------------- Render cache (persists under /workspace; any read/write failure just re-renders) -------------

def _render_cached(path: str, render) -> str:
    """Cache the full render keyed by (path, size, mtime), so every continuation page comes
    from the same md and parsing / recalc happens once.
    Any cache read/write failure silently falls back to rendering directly."""
    cf = None
    try:
        st = _os.stat(path)
        key = hashlib.md5(f"{path}|{st.st_size}|{st.st_mtime_ns}".encode()).hexdigest()
        cf = _os.path.join(_CACHE_DIR, key + ".md")
        if _os.path.exists(cf):
            return open(cf, encoding="utf-8").read()
    except OSError:
        cf = None
    md = render()
    if cf:
        try:
            _os.makedirs(_CACHE_DIR, exist_ok=True)
            tmp = cf + ".tmp"
            with open(tmp, "w", encoding="utf-8") as f:
                f.write(md)
            _os.replace(tmp, cf)
        except OSError:
            pass
    return md


# Look parser functions up by name via globals() (the per-format fragments are
# concatenated after this one, so they exist at runtime).
# Legacy binary formats (doc/ppt/xls) are absent from this table: openpyxl / pandoc /
# python-pptx do not recognise OLE2, so
# main() first converts them to OOXML through _legacy_to_ooxml (the soffice bridge) and
# then enters the matching reader.
_DISPATCH_NAMES = {
    "xlsx": "_xlsx_to_md", "xlsm": "_xlsx_to_md",
    "csv": "_csv_to_md", "tsv": "_csv_to_md",
    "docx": "_docx_to_md",
    "pptx": "_pptx_to_md", "pdf": "_pdf_to_md",
}
_FMT_NORM = {"xlsm": "xlsx"}
_LEGACY_TO_OOXML = {"doc": "docx", "ppt": "pptx", "xls": "xlsx"}


def _legacy_to_ooxml(path, target_ext):
    """Legacy Office binary (doc/ppt/xls) β†’ an OOXML temp copy (LibreOffice); None on
    failure.
    The copy is used only for this parse; the markdown result is still cached under the
    original path via _render_cached."""
    import shutil as _sh
    import tempfile as _tf
    if not _sh.which("soffice"):
        return None
    td = _tf.mkdtemp(prefix="legacy_read_")
    env = dict(_os.environ)
    home = env.get("HOME", "")
    if not (home and _os.path.isdir(home) and _os.access(home, _os.W_OK)):
        env["HOME"] = "/tmp"   # /root may not exist inside the bwrap sandbox, and soffice needs a writable HOME
    try:
        r = subprocess.run(["soffice", "--headless", "--norestore", "--convert-to",
                            target_ext, "--outdir", td, _os.path.abspath(path)],
                           capture_output=True, timeout=90, env=env)  # < the outer 120s, so this
        # function's own failure message reaches the model instead of being swallowed by a generic outer timeout
        out = _os.path.join(td, _os.path.splitext(_os.path.basename(path))[0]
                            + "." + target_ext)
        if r.returncode == 0 and _os.path.exists(out):
            return out
    except Exception:
        pass
    # Conversion failed: clean up the temp dir we created (on the success path the caller cleans up after parsing)
    _sh.rmtree(td, ignore_errors=True)
    return None


_IMG_EXTS = {"png", "jpg", "jpeg", "gif", "bmp", "tif", "tiff", "webp"}
_IMG_MIME = {"png": "image/png", "jpg": "image/jpeg", "jpeg": "image/jpeg",
             "gif": "image/gif", "bmp": "image/bmp", "tif": "image/tiff",
             "tiff": "image/tiff", "webp": "image/webp"}


def _resolve_image_batch(path):
    """path is a glob (contains * ? [) or a comma-separated list β†’ (existing image files,
    skipped paths).
    Skipped = a named path that does not exist or is not an image (a glob expansion only
    contains existing entries, so skipped comes mostly from comma lists)."""
    import glob as _g
    cands = []
    if any(c in path for c in "*?["):
        cands = sorted(_g.glob(path))
    elif "," in path:
        cands = [q.strip() for q in path.split(",") if q.strip()]
    imgs = [q for q in cands if _os.path.isfile(q) and _ext(q) in _IMG_EXTS]
    skipped = [q for q in cands if q not in imgs]
    return imgs, skipped


# A batch read runs at most two waves: 4 concurrent x 2 waves x 45s per image = 90s worst
# case, leaving 30s of the outer 120s for
# process start, file reads, caching and result assembly. READDOC_BATCH_MAX lowers it
# further,
# but even with lower concurrency it never exceeds two waves.
_BATCH_MAX = max(1, int(_os.environ.get("READDOC_BATCH_MAX", "8") or "8"))
_BATCH_TIMEOUT = 45
_BATCH_MAX_WAVES = 2


def _batch_concurrency():
    return max(1, int(_os.environ.get("READDOC_VISION_CONCURRENCY", "4") or "4"))


def _batch_limit():
    return max(1, min(_BATCH_MAX, _batch_concurrency() * _BATCH_MAX_WAVES))


def _batch_image_read(paths, skipped=(), limit=None):
    """Batch images: one vision call **per image** (keeps fidelity β€” never packs several
    images into a single message) + parallel + reassembled in order.
    Concurrency READDOC_VISION_CONCURRENCY (default 4); per-image vision timeout 45s (so
    the whole batch converges inside the outer
    sandbox timeout). One read_file over N rendered pages collapses N agent turns into 1."""
    from concurrent.futures import ThreadPoolExecutor
    conc = _batch_concurrency()
    limit = max(1, _batch_limit() if limit is None else int(limit))
    truncated = paths[limit:]
    paths = paths[:limit]

    def _one(p):
        try:
            with open(p, "rb") as f:
                data = f.read()
            return _vision_read(
                data, _IMG_MIME.get(_ext(p), "image/png"), timeout=_BATCH_TIMEOUT
            )
        except Exception:
            return None

    with ThreadPoolExecutor(max_workers=min(conc, len(paths))) as ex:
        vts = list(ex.map(_one, paths))
    head = f"(batch image read β€” {len(paths)} images, one vision call per image, in parallel)"
    if skipped:
        head += ("\n<!-- skipped (not found / not an image): "
                 + ", ".join(_os.path.basename(s) for s in skipped) + " -->")
    if truncated:
        head += (f"\n<!-- NOTE: {len(truncated)} more images matched but only the first "
                 f"{limit} were read this call. Read the rest with another read_file "
                 f"call, e.g. a comma-separated list starting at "
                 f"{_os.path.basename(truncated[0])}. -->")
    out = [head]
    for i, (p, vt) in enumerate(zip(paths, vts, strict=False), 1):
        nm = _os.path.basename(p)
        out.append(f"\n===== image {i}/{len(paths)}: {nm} =====\n"
                   + (vt if vt else "(vision unavailable/failed for this image)"))
    return "\n".join(out), all(vt is not None for vt in vts)


def _batch_cache_key(pattern, imgs, limit):
    """Batch-read cache key: the full match set + file state + vision endpoint/model + this call's image cap."""
    parts = [
        pattern,
        f"vision_url={_os.environ.get('READDOC_VISION_URL', '').rstrip('/')}",
        f"vision_model={_os.environ.get('READDOC_VISION_MODEL', '')}",
        f"limit={limit}",
    ]
    for p in imgs:
        try:
            st = _os.stat(p)
            parts.append(f"{p}|{st.st_size}|{st.st_mtime_ns}")
        except OSError:
            parts.append(p)
    return _os.path.join(_CACHE_DIR,
                         hashlib.md5("\n".join(parts).encode()).hexdigest() + ".md")


def main() -> None:
    # argv: path [max_chars] [cell_range|-] [offset] [pdf_mode] [pages|-]
    #   cell_range β€” xlsx point lookup;  pdf_mode/pages β€” pdf mode and range;  offset β€” generic pagination
    if len(sys.argv) < 2:
        sys.stdout.write("[read_file error] usage: <path> [max_chars] [cell_range|-] "
                         "[offset] [pdf_mode] [pages|-]")
        return
    path = sys.argv[1]
    max_chars = int(sys.argv[2]) if len(sys.argv) > 2 else 0  # unset = send everything
    cell_range = sys.argv[3] if len(sys.argv) > 3 and sys.argv[3] != "-" else None
    offset = int(sys.argv[4]) if len(sys.argv) > 4 else 0
    pdf_mode = sys.argv[5] if len(sys.argv) > 5 and sys.argv[5] != "-" else "auto"
    pages = sys.argv[6] if len(sys.argv) > 6 and sys.argv[6] != "-" else None
    # Multi-image batch: path is a glob (build/pg-*.png) or a comma-separated list β†’
    # vision per image in parallel, reassembled. Results: (1) enter the render cache
    # (offset continuation re-calls no VLM); (2) pass through _paginate (windowed on image
    # boundaries, governed by
    # max_chars/offset β€” everything is returned at once when it fits, cut only when it
    # does not, never silently truncated).
    _imgs, _skipped = _resolve_image_batch(path)
    if any(c in path for c in "*?[") and not _imgs:
        non_image_count = len(_skipped)
        sys.stdout.write(
            "[read_file error] image glob matched no images "
            f"({non_image_count} non-image match"
            f"{'es' if non_image_count != 1 else ''}): {path}"
        )
        _dump_trace()
        return
    # 2+ images β†’ batch read; exactly 1 but with skipped paths also goes batch (the output must report skipped, not swallow it)
    if len(_imgs) >= 2 or (len(_imgs) == 1 and _skipped):
        _trace({"stage": "batch_image", "n": len(_imgs), "skipped": len(_skipped)})
        limit = _batch_limit()
        cf = _batch_cache_key(path, _imgs, limit)
        md = None
        try:
            if _os.path.exists(cf):
                md = open(cf, encoding="utf-8").read()
        except OSError:
            pass
        if md is None:
            md, cacheable = _batch_image_read(_imgs, _skipped, limit)
            if cacheable:
                try:
                    _os.makedirs(_CACHE_DIR, exist_ok=True)
                    _tmp = cf + ".tmp"
                    with open(_tmp, "w", encoding="utf-8") as f:
                        f.write(md)
                    _os.replace(_tmp, cf)
                except OSError:
                    pass
        # max_chars is honoured exactly as passed (0 = everything). The larger default
        # window for batch reads is chosen by the tool layer
        # (read_file.py) when the caller did not pass one explicitly; the reader never
        # overrides an explicit value.
        sys.stdout.write(_paginate(md, "image_batch", offset, max_chars))
        _dump_trace()
        return
    if len(_imgs) == 1:
        # A glob / comma list matching exactly one image with nothing skipped: fall back to a plain single-file read (old behaviour)
        path = _imgs[0]
    ext = _ext(path)
    trace_on = bool(_os.environ.get("READDOC_TRACE"))
    # Legacy Office binary: the soffice bridge converts to an OOXML copy and the original
    # reader runs on it; the readout header notes the conversion source.
    # Cache first: the full render is cached under the **original path** (the dispatch
    # section writes the same key) β€” a hit returns immediately, so
    # continuing through a large .doc no longer re-runs soffice (seconds each) per offset.
    legacy_note = ""
    read_path = path
    _legacy_tmp = None
    if ext in _LEGACY_TO_OOXML:
        tgt = _LEGACY_TO_OOXML[ext]
        try:
            st = _os.stat(path)
            _cf0 = _os.path.join(_CACHE_DIR, hashlib.md5(
                f"{path}|{st.st_size}|{st.st_mtime_ns}".encode()).hexdigest() + ".md")
            if _os.path.exists(_cf0):
                _trace({"stage": "legacy_convert", "ext": ext, "cache": "hit"})
                sys.stdout.write(_paginate(open(_cf0, encoding="utf-8").read(),
                                           _FMT_NORM.get(tgt, tgt), offset, max_chars))
                _dump_trace()
                return
        except OSError:
            pass
        _trace({"stage": "legacy_convert", "ext": ext, "to": tgt})
        conv = _legacy_to_ooxml(path, tgt)
        if conv is None:
            sys.stdout.write(f"[read_file error] legacy .{ext} requires LibreOffice "
                             "conversion, which is unavailable/failed in this sandbox.")
            _dump_trace()
            return
        legacy_note = f"`converted from .{ext} via LibreOffice; fidelity best-effort`\n\n"
        read_path, ext = conv, tgt
        _legacy_tmp = _os.path.dirname(conv)   # converted-copy dir, cleaned after parsing (keeps /tmp from growing)
    if ext not in _DISPATCH_NAMES:
        # Unstructured document: sniff and route β€” text β†’ numbered content (paginated); image/audio/video β†’ type only; anything else binary β†’ unsupported
        _trace({"stage": "dispatch", "ext": ext, "fmt": "text", "reader": "_classify_read"})
        try:
            md = _render_cached(path, lambda: _classify_read(path))
        except Exception as e:
            # Fail closed like the structured route below: any read failure
            # (bad open / unsupported binary) goes to stderr + exit non-zero,
            # else under read_file's save_to (`... > <file>`) the error text is
            # captured AS the saved document and reported as a success. Media
            # type notices are valid output β€” _classify_read returns those.
            sys.stderr.write(f"[read_file error] {type(e).__name__}: {e}")
            _dump_trace()
            sys.exit(1)
        sys.stdout.write(_paginate(md, "text", offset, max_chars))
        _dump_trace()
        return
    fmt = _FMT_NORM.get(ext, ext)
    fn = globals()[_DISPATCH_NAMES[ext]]
    _trace({"stage": "dispatch", "ext": ext, "fmt": fmt, "reader": _DISPATCH_NAMES[ext]})
    # Parameterised reads (output varies with the parameters) skip the full-text cache;
    # only a plain full read is cached.
    # Tracing also bypasses the cache: a cache hit skips the routing code, leaving the
    # trace empty.
    pdf_param = ext == "pdf" and (pdf_mode != "auto" or pages)
    try:
        # Parse read_path (the converted copy for legacy formats); the cache key stays the original path
        if cell_range and ext in ("xlsx", "xlsm"):
            md = fn(read_path, cell_range)
        elif pdf_param:
            md = fn(read_path, pdf_mode, pages)
        elif ext == "pdf":
            md = fn(read_path, "auto", None) if trace_on else \
                _render_cached(path, lambda: fn(read_path, "auto", None))
        else:
            if trace_on:
                md = fn(read_path)
            elif legacy_note:
                # Cache the exact string that pagination indexes.  The first
                # page and continuation reads share this cache key, so omitting
                # the conversion note here shifts every later offset by its
                # length and silently skips document content.
                md = _render_cached(path, lambda: legacy_note + fn(read_path))
                legacy_note = ""
            else:
                md = _render_cached(path, lambda: fn(read_path))
    except Exception as e:
        # Parse error must NOT go to stdout: under read_file's ``save_to`` the
        # command is ``... > <file>``, so an error written to stdout would be
        # captured AS the saved document and read_file would report "saved N
        # bytes" (a silent success on a failed parse). Emit to stderr and exit
        # non-zero so read_file's ``exit_code != 0`` branch surfaces the error.
        sys.stderr.write(f"[read_file error parsing .{ext}] {type(e).__name__}: {e}")
        _dump_trace()
        sys.exit(1)
    finally:
        if _legacy_tmp:   # the legacy converted copy is removed as soon as it is used, so the sandbox /tmp does not grow per read
            import shutil as _sh
            _sh.rmtree(_legacy_tmp, ignore_errors=True)
    sys.stdout.write(_paginate(legacy_note + md, fmt, offset, max_chars))
    _dump_trace()