1use std::sync::atomic::{AtomicBool, Ordering};
2use std::sync::{Arc, Mutex};
3
4use rmcp::ErrorData;
5use rmcp::model::{ContentBlock, Tool};
6use serde_json::{Map, Value, json};
7
8use crate::server::tool_trait::{
9 McpTool, ToolContext, ToolOutput, get_bool, get_f64, get_int, get_str, get_str_array,
10 require_resolved_path,
11};
12use crate::tool_defs::tool_def;
13
14fn per_file_lock(path: &str) -> Arc<Mutex<()>> {
25 crate::core::path_locks::per_file_lock(path)
26}
27
28pub struct CtxReadTool;
29
30impl McpTool for CtxReadTool {
31 fn name(&self) -> &'static str {
32 "ctx_read"
33 }
34
35 fn tool_def(&self) -> Tool {
36 tool_def(
37 "ctx_read",
38 "Read source files. mode recommended — choose by intent (see `mode` below); defaults to auto when omitted.\n\
39 To UNDERSTAND code run ctx_compose FIRST; ctx_read after it identified files.\n\
40 anchored → edit by reference via ctx_patch (no exact-recall).",
41 json!({
42 "type": "object",
43 "properties": {
44 "path": { "type": "string", "description": "Absolute path" },
45 "paths": { "type": "array", "items": { "type": "string" }, "description": "Batch read" },
46 "mode": {
47 "type": "string",
48 "description": "Recommended (defaults to auto). full=verbatim(edit-ready) anchored=full+N:hh|anchors(edit via ctx_patch) raw=exact-bytes signatures=API map=structure auto=smart diff=git-delta lines:N-M=window (comma multi-selects: lines:5,10-20) reference=quotes task=focus"
49 },
50 "raw": { "type": "boolean", "description": "Verbatim (= mode=raw + fresh)" },
51 "start_line": { "type": "integer", "description": "1-based" },
52 "offset": { "type": "integer", "description": "start_line alias" },
53 "limit": { "type": "integer", "description": "Max lines" },
54 "fresh": { "type": "boolean", "description": "Bypass cache" },
55 "aggressiveness": { "type": "number", "description": "0.0–1.0 density (entropy/task)" },
56 "protect": { "type": "array", "items": { "type": "string" }, "description": "Symbols kept verbatim" }
57 },
58 "required": []
59 }),
60 )
61 }
62
63 fn handle(
64 &self,
65 args: &Map<String, Value>,
66 ctx: &ToolContext,
67 ) -> Result<ToolOutput, ErrorData> {
68 if args
71 .get("paths")
72 .and_then(|v| v.as_array())
73 .is_some_and(|a| !a.is_empty())
74 {
75 return super::ctx_multi_read::batch_read(args, ctx);
76 }
77
78 let path = if let Some(repo) = get_str(args, "repo") {
79 let root = crate::core::multi_repo::resolve_repo_root(&repo).ok_or_else(|| {
80 let known = crate::core::multi_repo::known_aliases().join(", ");
81 let known = if known.is_empty() {
82 "none registered — use ctx_multi_repo add_root".to_string()
83 } else {
84 known
85 };
86 ErrorData::invalid_params(
87 format!("unknown repo alias: {repo} (known: {known})"),
88 None,
89 )
90 })?;
91 let rel = get_str(args, "path").unwrap_or_else(|| ".".to_string());
92 crate::core::path_resolve::resolve_tool_path(Some(&root), None, &rel)
93 .map_err(|e| ErrorData::invalid_params(e, None))?
94 } else {
95 require_resolved_path(ctx, args, "path")?
96 };
97
98 match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
99 self.handle_inner(args, ctx, &path)
100 })) {
101 Ok(result) => result,
102 Err(_) => Err(ErrorData::internal_error(
103 format!(
104 "ctx_read panicked while processing '{path}'. This is a bug — please report it."
105 ),
106 None,
107 )),
108 }
109 }
110}
111
112impl CtxReadTool {
113 #[allow(clippy::unused_self)]
114 fn handle_inner(
115 &self,
116 args: &Map<String, Value>,
117 ctx: &ToolContext,
118 path: &str,
119 ) -> Result<ToolOutput, ErrorData> {
120 let session_lock = ctx
121 .session
122 .as_ref()
123 .ok_or_else(|| ErrorData::internal_error("session not available", None))?;
124 let cache_lock = ctx
125 .cache
126 .as_ref()
127 .ok_or_else(|| ErrorData::internal_error("cache not available", None))?;
128
129 let current_task = {
130 let deadline = std::time::Instant::now() + std::time::Duration::from_secs(8);
131 let mut attempt = 0u32;
132 loop {
133 if let Ok(guard) = session_lock.clone().try_read_owned() {
136 break guard.task.as_ref().map(|t| t.description.clone());
137 }
138 attempt += 1;
139 if std::time::Instant::now() >= deadline {
140 tracing::warn!(
141 "session read-lock timeout after {attempt} attempts in ctx_read for {path}"
142 );
143 break None;
144 }
145 std::thread::sleep(std::time::Duration::from_millis(25));
146 }
147 };
148 let task_ref = current_task.as_deref();
149 let arg_raw = get_bool(args, "raw").unwrap_or(false);
154 let explicit_mode_arg = resolve_raw_alias(arg_raw, get_str(args, "mode"));
155 let explicit_mode = explicit_mode_arg.is_some();
156 if let Some(ref requested) = explicit_mode_arg
164 && (requested.starts_with("lines:") || requested.starts_with("anchored:"))
165 && let Err(e) = requested.parse::<crate::tools::ctx_read::ReadMode>()
166 {
167 return Err(ErrorData::invalid_params(e.user_message(), None));
168 }
169 let configured_mode = (!explicit_mode)
170 .then(crate::core::auto_mode_resolver::configured_default_mode)
171 .flatten();
172 let learned_mode = if !explicit_mode && configured_mode.is_none() {
173 if let Ok(cache) = cache_lock.try_read() {
174 Some(crate::tools::ctx_smart_read::select_mode_with_task(
175 &cache, path, task_ref,
176 ))
177 } else {
178 tracing::debug!(
179 "cache lock contested during auto-mode selection for {path}; \
180 falling back to full"
181 );
182 None
183 }
184 } else {
185 None
186 };
187 let mut mode = crate::core::auto_mode_resolver::resolve_mode_precedence(
188 explicit_mode_arg,
189 configured_mode,
190 learned_mode,
191 "full",
192 );
193 let mut fresh = get_bool(args, "fresh").unwrap_or(false);
194 if arg_raw {
197 fresh = true;
198 }
199 let cache_policy = crate::server::compaction_sync::effective_cache_policy();
200 if cache_policy == "off" {
201 fresh = true;
202 }
203 let aggressiveness =
204 crate::core::aggressiveness::effective(get_f64(args, "aggressiveness"));
205 let protect = get_str_array(args, "protect").unwrap_or_default();
206 if !explicit_mode && let Some(a) = aggressiveness {
210 mode = crate::tools::ctx_read::ReadMode::Density(
213 crate::core::aggressiveness::AggressivenessProfile::from_level(a).density_target,
214 )
215 .to_string();
216 }
217 apply_line_window(
222 &mut mode,
223 &mut fresh,
224 explicit_mode,
225 get_int(args, "start_line"),
226 get_int(args, "offset"),
227 get_int(args, "limit"),
228 );
229
230 let pressure_action = ctx.pressure_snapshot.as_ref().map(|p| &p.recommendation);
231 let resolved_agent_id = ctx.agent_id.as_ref().and_then(|a| match a.try_read() {
232 Ok(guard) => guard.clone(),
233 Err(_) => None,
234 });
235 let gate_result = crate::server::context_gate::pre_dispatch_read_for_agent(
236 path,
237 &mode,
238 task_ref,
239 Some(&ctx.project_root),
240 pressure_action,
241 resolved_agent_id.as_deref(),
242 );
243 if gate_result.budget_blocked {
244 let msg = gate_result
245 .budget_warning
246 .unwrap_or_else(|| "Agent token budget exceeded".to_string());
247 return Err(ErrorData::invalid_params(msg, None));
248 }
249 let budget_warning = gate_result.budget_warning.clone();
250 let mut mode_override_note: Option<String> = None;
253 if mode != "raw"
254 && let Some(overridden) = gate_result.overridden_mode
255 {
256 if explicit_mode {
257 let reason = gate_result.reason.unwrap_or("context-gate");
258 mode_override_note = Some(format!(
259 "[mode overridden: {mode} -> {overridden}, reason={reason}]"
260 ));
261 }
262 mode = overridden;
263 }
264
265 let (instruction_mode, instruction_mode_note) = resolve_instruction_file_mode(path, &mode);
266 let (mut mode, degrade_warning) =
267 if instruction_mode_note.is_some() || instruction_mode != mode {
268 (instruction_mode, None)
269 } else if mode == "raw" || mode.starts_with("anchored") || mode.starts_with("lines:") {
270 (mode, None)
274 } else {
275 auto_degrade_read_mode(&mode)
276 };
277
278 let mut delta_explicit_note: Option<String> = None;
289 if !fresh
290 && explicit_mode
291 && (mode == "full" || mode == "full-compact" || mode.starts_with("lines:"))
292 && crate::core::config::Config::load().delta_explicit_effective()
293 && let Ok(cache) = cache_lock.try_read()
294 {
295 let decision = crate::tools::ctx_read::resolve_explicit_delta_mode(
296 &cache,
297 path,
298 &mode,
299 explicit_mode,
300 fresh,
301 true,
302 );
303 mode = decision.mode;
304 delta_explicit_note = decision.note;
305 }
306
307 if mode.starts_with("lines:") {
308 fresh = true;
309 }
310
311 if crate::core::binary_detect::is_llm_viewable_image(path) {
312 return read_image_file(path);
313 }
314 if crate::core::binary_detect::is_binary_file(path) {
315 let msg = crate::core::binary_detect::binary_file_message(path);
316 return Err(ErrorData::invalid_params(msg, None));
317 }
318 {
319 let cap = crate::core::limits::max_read_bytes() as u64;
320 if let Ok(meta) = std::fs::metadata(path)
321 && meta.len() > cap
322 {
323 let msg = format!(
324 "File too large ({} bytes, limit {} bytes via LCTX_MAX_READ_BYTES). \
325 Use mode=\"lines:1-100\" or start_line+limit for partial reads, \
326 mode=\"anchored\" with start_line+limit for edit-ready windows, \
327 or increase the limit.",
328 meta.len(),
329 cap
330 );
331 return Err(ErrorData::invalid_params(msg, None));
332 }
333 }
334
335 if !fresh
338 && let Ok(data_dir) = crate::core::data_dir::lean_ctx_data_dir()
339 && let Ok(mut cache) = cache_lock.try_write()
340 {
341 crate::server::compaction_sync::sync_if_compacted(&mut cache, &data_dir);
342 }
343
344 let read_timeout = std::time::Duration::from_secs(30);
348 let cancelled = Arc::new(AtomicBool::new(false));
349 let delivery_metadata = crate::core::config::Config::load()
351 .ocla
352 .delivery_enabled()
353 .then(|| crate::tools::ctx_read::file_blake3_prefix(path))
354 .flatten();
355 let (output, resolved_mode, original, is_cache_hit, file_ref, cache_stats) = {
356 let crp_mode = ctx.crp_mode;
357 let fast_result = 'fast: {
358 let file_lock = per_file_lock(path);
359 let Some(_file_guard) = file_lock.try_lock().ok() else {
360 break 'fast None;
361 };
362
363 if !fresh
374 && (mode == "full" || mode == "full-compact" || mode == "auto")
375 && let Ok(cache) = cache_lock.try_read()
376 && let Some(read_output) =
377 crate::tools::ctx_read::try_stub_hit_readonly(&cache, path)
378 {
379 let hit = read_output.is_cache_hit;
380 let content = read_output.content;
381 let rmode = read_output.resolved_mode;
382 let orig = cache.get(path).map_or(0, |e| e.original_tokens);
383 let fref = cache.file_ref_map().get(path).cloned();
384 let stats = cache.get_stats();
385 let stats_snapshot = (stats.total_reads(), stats.cache_hits());
386 break 'fast Some((content, rmode, orig, hit, fref, stats_snapshot));
387 }
388
389 None
393 };
394
395 if let Some(result) = fast_result {
396 result
397 } else {
398 let cache_lock = cache_lock.clone();
399 let mode = mode.clone();
400 let task_owned = current_task.clone();
401 let protect_owned = protect.clone();
402 let path_owned = path.to_string();
403 let cancel_flag = cancelled.clone();
404 let (tx, rx) = std::sync::mpsc::sync_channel(1);
405 std::thread::spawn(move || {
406 let file_lock = per_file_lock(&path_owned);
407
408 let _file_guard = {
409 let deadline =
410 std::time::Instant::now() + std::time::Duration::from_secs(25);
411 loop {
412 if cancel_flag.load(Ordering::Relaxed) {
413 return;
414 }
415 if let Ok(guard) = file_lock.try_lock() {
416 break guard;
417 }
418 if std::time::Instant::now() >= deadline {
419 tracing::error!(
420 "ctx_read: per-file lock timeout after 25s for {path_owned}"
421 );
422 let _ = tx.send((
423 format!("per-file lock contention for {path_owned} — retry in a moment"),
424 "error".to_string(), 0, false, None, (0, 0),
425 ));
426 return;
427 }
428 std::thread::sleep(std::time::Duration::from_millis(50));
429 }
430 };
431
432 if cancel_flag.load(Ordering::Relaxed) {
433 return;
434 }
435
436 if !fresh
443 && (mode == "full" || mode == "full-compact" || mode == "auto")
444 && let Ok(cache) = cache_lock.try_read()
445 && let Some(read_output) =
446 crate::tools::ctx_read::try_stub_hit_readonly(&cache, &path_owned)
447 {
448 let content = read_output.content;
449 let rmode = read_output.resolved_mode;
450 let orig = cache.get(&path_owned).map_or(0, |e| e.original_tokens);
451 let hit = true;
452 let fref = cache.file_ref_map().get(path_owned.as_str()).cloned();
453 let stats = cache.get_stats();
454 let stats_snapshot = (stats.total_reads(), stats.cache_hits());
455 let _ = tx.send((content, rmode, orig, hit, fref, stats_snapshot));
456 return;
457 }
458
459 if !crate::tools::ctx_read::effective_fresh_for_delivery(fresh)
463 && let Some((hash, mtime)) = delivery_metadata
464 && let Some(read_output) = crate::tools::ctx_read::try_cross_agent_stub(
465 &path_owned,
466 &mode,
467 hash,
468 mtime,
469 )
470 {
471 let _ = tx.send((
472 read_output.content,
473 read_output.resolved_mode,
474 0,
475 read_output.is_cache_hit,
476 None,
477 (0, 0),
478 ));
479 return;
480 }
481
482 let preread = match crate::tools::ctx_read::read_file_lossy(&path_owned) {
484 Ok(c) => Some(c),
485 Err(e) => {
486 tracing::warn!("ctx_read: cannot read {path_owned}: {e}");
487 None
488 }
489 };
490
491 if cancel_flag.load(Ordering::Relaxed) {
492 return;
493 }
494
495 let task_ref = task_owned.as_deref();
505 let tuning =
506 crate::tools::ctx_read::ReadTuning::resolve(aggressiveness, &protect_owned);
507
508 macro_rules! acquire_write {
510 ($deadline_secs:expr, $label:expr) => {{
511 let deadline = std::time::Instant::now()
512 + std::time::Duration::from_secs($deadline_secs);
513 loop {
514 if cancel_flag.load(Ordering::Relaxed) {
515 return;
516 }
517 if let Ok(guard) = cache_lock.try_write() {
518 break guard;
519 }
520 if std::time::Instant::now() >= deadline {
521 tracing::error!(
522 "ctx_read: cache write-lock timeout ({}) for {path_owned}",
523 $label,
524 );
525 let _ = tx.send((
526 format!(
527 "cache lock contention for {path_owned} — retry in a moment"
528 ),
529 "error".into(),
530 0,
531 false,
532 None,
533 (0, 0),
534 ));
535 return;
536 }
537 std::thread::sleep(std::time::Duration::from_millis(50));
538 }
539 }};
540 }
541
542 #[allow(clippy::large_enum_variant)]
546 enum PrepareOutcome {
547 Hit(String, String, usize, bool, Option<String>, (u64, u64)),
548 Compute {
549 file_ref: String,
550 resolved_mode: String,
551 content: String,
552 original_tokens: usize,
553 },
554 }
555
556 let outcome = {
557 let mut cache = acquire_write!(10, "prepare 10s");
558
559 if crate::core::plugins::PluginManager::has_listener("pre_read") {
560 crate::core::plugins::PluginManager::fire_hook_background(
561 crate::core::plugins::executor::HookPoint::PreRead {
562 path: path_owned.clone(),
563 },
564 );
565 }
566 if let Ok(mut bt) = crate::core::bounce_tracker::global().lock() {
567 bt.next_seq();
568 }
569
570 let file_ref = cache.get_file_ref(&path_owned);
571
572 let effective_fresh = fresh
573 || crate::tools::ctx_read::force_fresh_env()
574 || (crate::tools::ctx_read::is_subagent_context()
575 && !crate::core::conversation::scope_enabled());
576
577 let mode_eff = if mode != "raw"
578 && !mode.starts_with("lines:")
579 && crate::core::config::Config::load()
580 .proxy
581 .is_path_compress_protected(&path_owned)
582 {
583 "full".to_string()
584 } else {
585 mode.clone()
586 };
587
588 if effective_fresh {
589 cache.invalidate(&path_owned);
590 }
591
592 if !effective_fresh {
593 let stale = cache.get(&path_owned).is_some_and(|e| {
594 crate::core::cache::is_cache_entry_stale_verified(
595 &path_owned,
596 e.stored_mtime,
597 &e.hash,
598 )
599 });
600 if stale {
601 cache.invalidate(&path_owned);
602 }
603 }
604
605 let snap = cache
606 .get(&path_owned)
607 .map(|e| (e.original_tokens, e.content()));
608
609 if let Some((orig_tok, content_opt)) = snap {
610 let resolved = if mode_eff == "auto" {
611 tuning.auto_density_mode().unwrap_or_else(|| {
612 crate::tools::ctx_read::resolve_auto_mode(
613 Some(&cache),
614 &path_owned,
615 orig_tok,
616 None,
617 task_ref,
618 )
619 })
620 } else {
621 mode_eff
622 };
623
624 if (resolved == "full" || resolved == "full-compact")
625 && let Some(out) = crate::tools::ctx_read::try_stub_hit_readonly(
626 &cache,
627 &path_owned,
628 )
629 {
630 let orig = cache.get(&path_owned).map_or(0, |e| e.original_tokens);
631 let fref = cache.file_ref_map().get(path_owned.as_str()).cloned();
632 let s = cache.get_stats();
633 PrepareOutcome::Hit(
634 out.content,
635 out.resolved_mode,
636 orig,
637 true,
638 fref,
639 (s.total_reads(), s.cache_hits()),
640 )
641 } else if crate::tools::ctx_read::is_cacheable_mode(&resolved) {
642 let ck = crate::tools::ctx_read::compressed_cache_key(
643 &resolved,
644 crp_mode,
645 task_ref,
646 tuning.aggressiveness,
647 tuning.protect,
648 );
649 if let Some(hit) = cache.get_compressed(&path_owned, &ck).cloned() {
650 crate::core::auto_mode_resolver::count_source(
651 "compressed_cache_hit",
652 );
653 let hit = crate::core::redaction::redact_text_if_enabled(&hit);
654 let orig =
655 cache.get(&path_owned).map_or(0, |e| e.original_tokens);
656 let fref =
657 cache.file_ref_map().get(path_owned.as_str()).cloned();
658 let s = cache.get_stats();
659 PrepareOutcome::Hit(
660 hit,
661 resolved,
662 orig,
663 true,
664 fref,
665 (s.total_reads(), s.cache_hits()),
666 )
667 } else {
668 let c = content_opt
669 .or_else(|| preread.as_deref().map(String::from));
670 PrepareOutcome::Compute {
671 file_ref,
672 resolved_mode: resolved,
673 content: c.unwrap_or_default(),
674 original_tokens: orig_tok,
675 }
676 }
677 } else {
678 let c =
679 content_opt.or_else(|| preread.as_deref().map(String::from));
680 PrepareOutcome::Compute {
681 file_ref,
682 resolved_mode: resolved,
683 content: c.unwrap_or_default(),
684 original_tokens: orig_tok,
685 }
686 }
687 } else {
688 let raw = match preread {
689 Some(c) if !c.is_empty() => c,
690 _ => match crate::tools::ctx_read::read_file_lossy(&path_owned) {
691 Ok(c) if !c.is_empty() => c,
692 Ok(_) => {
693 tracing::debug!(
694 "ctx_read: skipping cache for empty content: {path_owned}"
695 );
696 let _ = tx.send((
697 format!("File is empty: {path_owned}"),
698 "error".into(),
699 0,
700 false,
701 None,
702 (0, 0),
703 ));
704 return;
705 }
706 Err(e) => {
707 tracing::debug!(
708 "ctx_read: skipping cache for empty content: {path_owned}"
709 );
710 let _ = tx.send((
711 format!("Cannot read file: {path_owned}: {e}"),
712 "error".into(),
713 0,
714 false,
715 None,
716 (0, 0),
717 ));
718 return;
719 }
720 },
721 };
722 let sr = cache.store(&path_owned, &raw);
723 let resolved = if mode_eff == "auto" {
724 tuning.auto_density_mode().unwrap_or_else(|| {
725 crate::tools::ctx_read::resolve_auto_mode(
726 None,
727 &path_owned,
728 sr.original_tokens,
729 Some(sr.line_count),
730 task_ref,
731 )
732 })
733 } else {
734 mode_eff
735 };
736 PrepareOutcome::Compute {
737 file_ref,
738 resolved_mode: resolved,
739 content: raw,
740 original_tokens: sr.original_tokens,
741 }
742 }
743 }; if let PrepareOutcome::Hit(c, rm, orig, hit, fref, ss) = outcome {
746 let mut cache = acquire_write!(10, "hit last_mode 10s");
749 if let Some(entry) = cache.get_mut(&path_owned) {
750 entry.last_mode.clone_from(&rm);
751 }
752 let _ = tx.send((c, rm, orig, hit, fref, ss));
753 return;
754 }
755 let PrepareOutcome::Compute {
756 file_ref,
757 resolved_mode,
758 content: compute_content,
759 original_tokens,
760 } = outcome
761 else {
762 unreachable!()
763 };
764
765 if cancel_flag.load(Ordering::Relaxed) {
766 return;
767 }
768
769 let short = crate::core::protocol::shorten_path(&path_owned);
774 let ext_s = std::path::Path::new(&*path_owned)
775 .extension()
776 .and_then(|e| e.to_str())
777 .unwrap_or("");
778
779 let (mut computed, rmode) = if resolved_mode == "full"
780 || resolved_mode == "full-compact"
781 {
782 if resolved_mode == "full-compact" {
783 let (out, _) = crate::tools::ctx_read::format_full_compact_output(
784 &compute_content,
785 );
786 (out, "full-compact".to_string())
787 } else {
788 let lc = compute_content.lines().count();
789 let (out, _) = crate::tools::ctx_read::format_full_output(
790 &file_ref,
791 &short,
792 ext_s,
793 &compute_content,
794 original_tokens,
795 lc,
796 task_ref,
797 );
798 let ft = crate::core::tokens::count_tokens(&out);
799 let out = crate::tools::ctx_read::cap_to_raw(
800 out,
801 ft,
802 &compute_content,
803 original_tokens,
804 );
805 (out, "full".to_string())
806 }
807 } else {
808 let (out, _) = crate::tools::ctx_read::process_mode_tuned(
809 &compute_content,
810 &resolved_mode,
811 &file_ref,
812 &short,
813 ext_s,
814 original_tokens,
815 crp_mode,
816 &path_owned,
817 task_ref,
818 tuning,
819 );
820 let out = if crate::tools::ctx_read::mode_allows_raw_cap(&resolved_mode) {
821 let ft = crate::core::tokens::count_tokens(&out);
822 crate::tools::ctx_read::cap_to_raw(
823 out,
824 ft,
825 &compute_content,
826 original_tokens,
827 )
828 } else {
829 out
830 };
831 (out, resolved_mode)
832 };
833
834 computed = crate::core::redaction::redact_text_if_enabled(&computed);
835
836 if cancel_flag.load(Ordering::Relaxed) {
837 return;
838 }
839
840 {
845 let deadline =
846 std::time::Instant::now() + std::time::Duration::from_secs(5);
847 let cache_guard = loop {
848 if cancel_flag.load(Ordering::Relaxed) {
849 return;
850 }
851 if let Ok(g) = cache_lock.try_write() {
852 break Some(g);
853 }
854 if std::time::Instant::now() >= deadline {
855 tracing::warn!(
856 "ctx_read: store-lock timeout (5s) for {path_owned}, returning without caching"
857 );
858 break None;
859 }
860 std::thread::sleep(std::time::Duration::from_millis(50));
861 };
862
863 if let Some(mut cache) = cache_guard {
864 if crate::tools::ctx_read::is_cacheable_mode(&rmode) {
865 let ck = crate::tools::ctx_read::compressed_cache_key(
866 &rmode,
867 crp_mode,
868 task_ref,
869 tuning.aggressiveness,
870 tuning.protect,
871 );
872 cache.set_compressed(&path_owned, &ck, computed.clone());
873 }
874 if rmode == "full" || rmode == "full-compact" {
875 cache.mark_full_delivered(&path_owned);
876 }
877 if let Some(entry) = cache.get_mut(&path_owned) {
878 entry.last_mode.clone_from(&rmode);
879 }
880 if let Ok(mut bt) = crate::core::bounce_tracker::global().lock() {
881 bt.record_read(
882 &path_owned,
883 &rmode,
884 crate::core::tokens::count_tokens(&computed),
885 original_tokens,
886 );
887 }
888 let orig = cache.get(&path_owned).map_or(0, |e| e.original_tokens);
889 let fref = cache.file_ref_map().get(path_owned.as_str()).cloned();
890 let s = cache.get_stats();
891 let _ = tx.send((
892 computed,
893 rmode,
894 orig,
895 false,
896 fref,
897 (s.total_reads(), s.cache_hits()),
898 ));
899 } else {
900 let _ =
901 tx.send((computed, rmode, original_tokens, false, None, (0, 0)));
902 }
903 }
904 });
905 if let Ok(result) = rx.recv_timeout(read_timeout) {
906 result
907 } else {
908 cancelled.store(true, Ordering::Relaxed);
909 tracing::error!("ctx_read timed out after {read_timeout:?} for {path}");
910 let msg = format!(
911 "ERROR: ctx_read timed out after {}s reading {path}. \
912 The file may be very large or a blocking I/O issue occurred. \
913 Try mode=\"lines:1-100\" for a partial read.",
914 read_timeout.as_secs()
915 );
916 return Err(ErrorData::internal_error(msg, None));
917 }
918 } };
920
921 if resolved_mode == "error" {
922 return Err(ErrorData::invalid_params(output, None));
923 }
924
925 let output_tokens = crate::core::tokens::count_tokens(&output);
926 let saved = original.saturating_sub(output_tokens);
927
928 if !is_cache_hit {
929 if let Some((hash, mtime)) = delivery_metadata {
930 crate::tools::ctx_read::record_cross_agent_delivery(
931 path,
932 hash,
933 mtime,
934 0,
935 output_tokens,
936 None,
937 None,
938 );
939 }
940 }
941
942 let mut ensured_root: Option<String> = None;
944 let mut traversal_working_set: Vec<String> = Vec::new();
945 let mut prefetch_paths: Vec<String> = Vec::new();
946 let project_root_snapshot;
947 {
948 let deadline = std::time::Instant::now() + std::time::Duration::from_secs(10);
949 let session_guard = loop {
950 if let Ok(g) = session_lock.clone().try_write_owned() {
951 break Some(g);
952 }
953 if std::time::Instant::now() >= deadline {
954 break None;
955 }
956 std::thread::sleep(std::time::Duration::from_millis(25));
957 };
958 if let Some(mut session) = session_guard {
959 session.touch_file(path, file_ref.as_deref(), &resolved_mode, original);
960 prefetch_paths = session.prefetch_predictions(3);
961 traversal_working_set =
964 crate::core::tool_lifecycle::recent_working_set(&session, path);
965 let file_summary = extract_file_summary(&output, path);
966 if !file_summary.is_empty() {
967 session.set_file_summary(path, &file_summary);
968 }
969 if is_cache_hit {
970 session.record_cache_hit();
971 }
972 if session.active_structured_intent.is_none() && session.files_touched.len() >= 2 {
973 let touched: Vec<String> = session
974 .files_touched
975 .iter()
976 .map(|f| f.path.clone())
977 .collect();
978 let inferred =
979 crate::core::intent_engine::StructuredIntent::from_file_patterns(&touched);
980 if inferred.confidence >= 0.4 {
981 session.active_structured_intent = Some(inferred);
982 }
983 }
984 if session.task.is_none() && session.stats.files_read % 5 == 0 {
985 session.auto_infer_task();
986 }
987 let root_missing = session
988 .project_root
989 .as_deref()
990 .is_none_or(|r| r.trim().is_empty());
991 if root_missing && let Some(root) = crate::core::protocol::detect_project_root(path)
992 {
993 session.project_root = Some(root.clone());
994 ensured_root = Some(root);
995 }
996 project_root_snapshot = session
997 .project_root
998 .clone()
999 .unwrap_or_else(|| ".".to_string());
1000 } else {
1001 tracing::warn!(
1002 "session write-lock timeout (5s) in ctx_read post-update for {path}"
1003 );
1004 project_root_snapshot = ctx.project_root.clone();
1005 }
1006 }
1007 if let Some(root) = ensured_root.as_deref() {
1008 crate::core::index_orchestrator::ensure_all_background(root);
1009 }
1010
1011 if !prefetch_paths.is_empty() {
1012 crate::core::context_prefetch::warm_predictions(&prefetch_paths, Some(&cache_lock));
1013 }
1014
1015 {
1021 let path_bg = path.to_string();
1022 let resolved_mode_bg = resolved_mode.clone();
1023 let project_root_bg = project_root_snapshot.clone();
1024 let (turns, hits) = cache_stats;
1025 let ledger_cache = (crate::core::savings_ledger::ledger_family()
1031 != crate::core::tokens::TokenizerFamily::O200kBase)
1032 .then(|| cache_lock.clone());
1033 let ledger_output = ledger_cache.as_ref().map(|_| output.clone());
1034 std::thread::spawn(move || {
1035 let _ = std::panic::catch_unwind(std::panic::AssertUnwindSafe(move || {
1038 crate::core::heatmap::record_file_access(&path_bg, original, saved);
1039
1040 {
1045 use crate::core::savings_ledger as ledger;
1046 let (lbase, lsaved) = match (&ledger_cache, &ledger_output) {
1047 (Some(cl), Some(out)) => match cl.try_read().ok().and_then(|c| {
1048 c.get(&path_bg)
1049 .and_then(crate::core::cache::CacheEntry::content)
1050 }) {
1051 Some(raw) => {
1052 let lo = ledger::count_for_ledger(&raw);
1053 (lo, lo.saturating_sub(ledger::count_for_ledger(out)))
1054 }
1055 None => (original, saved),
1056 },
1057 _ => (original, saved),
1058 };
1059 ledger::record_read_event(lbase, lsaved, None, None);
1060 }
1061
1062 if let Some(root) =
1065 crate::core::tool_lifecycle::usable_root(Some(project_root_bg.as_str()))
1066 {
1067 crate::core::cooccurrence::record_focus_access(
1068 root,
1069 &path_bg,
1070 &traversal_working_set,
1071 );
1072 }
1073 let sig =
1074 crate::core::mode_predictor::FileSignature::from_path(&path_bg, original);
1075 let density = if output_tokens > 0 {
1076 original as f64 / output_tokens as f64
1077 } else {
1078 1.0
1079 };
1080 let outcome = crate::core::mode_predictor::ModeOutcome {
1081 mode: resolved_mode_bg,
1082 tokens_in: original,
1083 tokens_out: output_tokens,
1084 density: density.min(1.0),
1085 };
1086 let mut predictor = crate::core::mode_predictor::ModePredictor::new();
1087 predictor.set_project_root(&project_root_bg);
1088 predictor.record(sig, outcome);
1089 predictor.save();
1090
1091 let ext = std::path::Path::new(&path_bg)
1092 .extension()
1093 .and_then(|e| e.to_str())
1094 .unwrap_or("")
1095 .to_string();
1096 let thresholds =
1097 crate::core::adaptive_thresholds::thresholds_for_path(&path_bg);
1098 let feedback_outcome = crate::core::feedback::CompressionOutcome {
1099 session_id: format!("{}", std::process::id()),
1100 language: ext,
1101 entropy_threshold: thresholds.bpe_entropy,
1102 jaccard_threshold: thresholds.jaccard,
1103 total_turns: turns as u32,
1104 tokens_saved: saved as u64,
1105 tokens_original: original as u64,
1106 cache_hits: hits as u32,
1107 total_reads: turns as u32,
1108 task_completed: crate::core::bounce_tracker::global()
1116 .lock()
1117 .ok()
1118 .and_then(|bt| bt.bounce_rate_for_extension(&path_bg))
1119 .is_none_or(|rate| rate < 0.30),
1120 timestamp: chrono::Local::now().to_rfc3339(),
1121 };
1122 let mut store = crate::core::feedback::FeedbackStore::load();
1123 store.project_root = Some(project_root_bg);
1124 store.record_outcome(feedback_outcome);
1125 }));
1126 });
1127 }
1128 if let Some(aid) = resolved_agent_id.as_deref() {
1129 crate::core::agent_budget::record_consumption(aid, output_tokens);
1130 }
1131
1132 let graph_hint = if !is_cache_hit
1136 && !resolved_mode.starts_with("lines:")
1137 && crate::core::profiles::active_profile()
1138 .output_hints
1139 .related_hint()
1140 {
1141 crate::tools::ctx_read::graph_related_hint(path)
1142 } else {
1143 None
1144 };
1145
1146 let hints_suffix = if crate::core::profiles::active_profile()
1150 .output_hints
1151 .cross_source_hint()
1152 {
1153 let graph_db =
1154 crate::core::property_graph::graph_dir(&ctx.project_root).join("graph.db");
1155 let graph = graph_db
1156 .exists()
1157 .then(|| crate::core::property_graph::CodeGraph::open(&ctx.project_root))
1158 .transpose()
1159 .ok()
1160 .flatten();
1161 graph.map_or_else(String::new, |graph| {
1162 let edges = graph.all_cross_source_edges();
1163 if edges.is_empty() {
1164 String::new()
1165 } else {
1166 let ranges = scoped_read_ranges(&resolved_mode);
1167 let relative_path =
1168 crate::core::graph_index::graph_relative_key(path, &ctx.project_root);
1169 let hints = crate::core::cross_source_hints::hints_for_file_matching(
1170 path,
1171 &edges,
1172 &ctx.project_root,
1173 |hint| {
1174 ranges.as_ref().is_none_or(|ranges| {
1175 hint_intersects_ranges(hint, ranges, &graph, &relative_path)
1176 })
1177 },
1178 );
1179 crate::core::cross_source_hints::format_hints(&hints)
1180 }
1181 })
1182 } else {
1183 String::new()
1184 };
1185
1186 let rules_suffix = {
1190 let client_id = ctx
1191 .client_name
1192 .as_ref()
1193 .map(|c| c.blocking_read().clone())
1194 .unwrap_or_default();
1195 crate::core::rule_discovery::rules_suffix_for_read(path, &ctx.project_root, &client_id)
1196 };
1197 let mut warnings = Vec::new();
1198 if let Some(ref w) = budget_warning {
1199 warnings.push(w.as_str());
1200 }
1201 if let Some(ref w) = degrade_warning {
1202 warnings.push(w.as_str());
1203 }
1204 if let Some(ref w) = delta_explicit_note {
1205 warnings.push(w.as_str());
1206 }
1207 if let Some(ref w) = mode_override_note {
1208 warnings.push(w.as_str());
1209 }
1210 if let Some(ref w) = instruction_mode_note {
1211 warnings.push(w.as_str());
1212 }
1213 let graph_suffix = graph_hint.map(|h| format!("\n{h}")).unwrap_or_default();
1214 let final_output = if !warnings.is_empty() {
1217 format!(
1218 "{}\n\n{output}{hints_suffix}{graph_suffix}{rules_suffix}",
1219 warnings.join("\n")
1220 )
1221 } else if hints_suffix.is_empty() && graph_suffix.is_empty() && rules_suffix.is_empty() {
1222 output
1223 } else {
1224 format!("{output}{hints_suffix}{graph_suffix}{rules_suffix}")
1225 };
1226 let final_output = if crate::core::profiles::active_profile()
1230 .output_hints
1231 .proactive_context()
1232 {
1233 let proactive_query = format!(
1234 "ctx_read path={path} mode={resolved_mode} task={}",
1235 task_ref.unwrap_or_default()
1236 );
1237 if let Some(block) =
1238 crate::core::relevance_tracker::proactive_context_for_path(&proactive_query, path)
1239 {
1240 format!("{final_output}{block}")
1241 } else {
1242 final_output
1243 }
1244 } else {
1245 final_output
1246 };
1247
1248 let final_tokens = crate::core::tokens::count_tokens(&final_output);
1255 let verified_saved = original.saturating_sub(final_tokens);
1256
1257 Ok(ToolOutput {
1258 text: final_output,
1259 original_tokens: original,
1260 saved_tokens: verified_saved,
1261 mode: Some(resolved_mode),
1262 path: Some(path.to_string()),
1263 changed: false,
1264 shell_outcome: None,
1265 content_blocks: None,
1266 })
1267 }
1268}
1269
1270#[path = "ctx_read_window.rs"]
1271mod window;
1272#[allow(unused_imports)]
1273use window::{
1275 apply_line_window, hint_intersects_ranges, lines_mode, resolve_instruction_file_mode,
1276 resolve_line_window, resolve_raw_alias, scoped_read_ranges,
1277};
1278
1279fn apply_verdict(
1280 mode: &str,
1281 verdict: crate::core::degradation_policy::DegradationVerdictV1,
1282) -> (String, bool) {
1283 use crate::core::degradation_policy::DegradationVerdictV1;
1284 match verdict {
1285 DegradationVerdictV1::Ok => (mode.to_string(), false),
1286 DegradationVerdictV1::Warn => match mode {
1287 "full" => ("map".to_string(), true),
1288 other => (other.to_string(), false),
1289 },
1290 DegradationVerdictV1::Throttle => match mode {
1291 "full" | "map" => ("signatures".to_string(), true),
1292 other => (other.to_string(), false),
1293 },
1294 DegradationVerdictV1::Block => {
1295 if mode == "signatures" {
1296 ("signatures".to_string(), false)
1297 } else {
1298 ("signatures".to_string(), true)
1299 }
1300 }
1301 }
1302}
1303
1304fn auto_degrade_read_mode(mode: &str) -> (String, Option<String>) {
1305 if crate::core::config::Config::load().no_degrade_effective() {
1306 return (mode.to_string(), None);
1307 }
1308 let profile = crate::core::profiles::active_profile();
1309 if !profile.degradation.enforce_effective() {
1310 return (mode.to_string(), None);
1311 }
1312 let policy = crate::core::degradation_policy::evaluate_v1_for_tool("ctx_read", None);
1313 let (new_mode, degraded) = apply_verdict(mode, policy.decision.verdict);
1314 let warning = if degraded {
1315 Some(format!(
1316 "⚠ Context pressure: mode={mode} was downgraded to mode={new_mode} \
1317 (verdict: {:?}). Use start_line=1 to bypass, or run ctx_compress to free budget.",
1318 policy.decision.verdict
1319 ))
1320 } else {
1321 None
1322 };
1323 (new_mode, warning)
1324}
1325
1326fn extract_file_summary(output: &str, path: &str) -> String {
1327 let hint = crate::core::auto_findings::extract_content_hint(output);
1328 if !hint.is_empty() {
1329 return hint;
1330 }
1331 let ext = std::path::Path::new(path)
1332 .extension()
1333 .and_then(|e| e.to_str())
1334 .unwrap_or("");
1335 let line_count = output.lines().count();
1336 if line_count > 5 {
1337 format!("{ext} file, {line_count} lines")
1338 } else {
1339 String::new()
1340 }
1341}
1342
1343#[cfg(test)]
1345#[path = "ctx_read_inline_tests.rs"]
1346mod tests;
1347
1348fn read_image_file(path: &str) -> Result<ToolOutput, ErrorData> {
1353 use crate::core::binary_detect::{IMAGE_MAX_BYTES, image_mime_type};
1354 use base64::Engine;
1355
1356 let metadata = std::fs::metadata(path)
1357 .map_err(|e| ErrorData::invalid_params(format!("Cannot read image: {e}"), None))?;
1358
1359 if metadata.len() > IMAGE_MAX_BYTES {
1360 return Err(ErrorData::invalid_params(
1361 format!(
1362 "Image too large ({:.1} MB, limit {:.0} MB). Resize or use a smaller image.",
1363 metadata.len() as f64 / 1024.0 / 1024.0,
1364 IMAGE_MAX_BYTES as f64 / 1024.0 / 1024.0,
1365 ),
1366 None,
1367 ));
1368 }
1369
1370 let mime_type = image_mime_type(path)
1371 .ok_or_else(|| ErrorData::invalid_params("Unsupported image format".to_string(), None))?;
1372
1373 let bytes = std::fs::read(path)
1374 .map_err(|e| ErrorData::invalid_params(format!("Cannot read image: {e}"), None))?;
1375
1376 let base64_data = base64::prelude::BASE64_STANDARD.encode(&bytes);
1377 let short_name = std::path::Path::new(path)
1378 .file_name()
1379 .and_then(|n| n.to_str())
1380 .unwrap_or(path);
1381
1382 let text_block = ContentBlock::text(format!(
1383 "[Image: {} ({} KB, {})]",
1384 short_name,
1385 bytes.len() / 1024,
1386 mime_type
1387 ));
1388 let image_block = ContentBlock::image(base64_data, mime_type);
1389
1390 Ok(ToolOutput::image(
1391 vec![text_block, image_block],
1392 path.to_string(),
1393 ))
1394}
1395
1396#[cfg(test)]
1397#[path = "ctx_read_repo_param_tests.rs"]
1398mod repo_param_tests;