lean-ctx 3.8.13

Context Runtime for AI Agents with CCP. 71 MCP tools, 10 read modes, 95+ compression patterns, cross-session memory (CCP), persistent AI knowledge with temporal facts + contradiction detection, multi-agent context sharing, LITM-aware positioning, AAAK compact format, adaptive compression with Thompson Sampling bandits. Supports 24+ AI tools. Reduces LLM token consumption by up to 99%.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
//! Shared tool lifecycle — ensures CLI and MCP paths have identical side effects.
//!
//! The MCP server dispatcher handles session, ledger, heatmap, intent detection,
//! and knowledge consolidation inline (via in-memory state). When the daemon is
//! unavailable, CLI commands call functions here to achieve the same coverage by
//! loading/saving state from disk.
//!
//! NOTE: When the daemon IS running, CLI routes through `daemon_client` which
//! calls the MCP server — these functions are NOT called in that path.

use crate::core::context_ir::{ContextIrSourceKindV1, ContextIrV1, RecordIrInput};
use crate::core::context_ledger::ContextLedger;
use crate::core::heatmap;
use crate::core::intent_engine::StructuredIntent;
use crate::core::session::SessionState;
use crate::core::stats;

/// How many recently-touched files form the "working set" a new read is
/// associated with for traversal (co-access) edges (#289). Small, so the signal
/// stays local to what the agent is actively juggling.
const TRAVERSAL_WINDOW: usize = 6;

/// Recent distinct file paths (excluding `current`), most-recent first, capped
/// to the traversal window — the working set a new read co-occurs with.
pub(crate) fn recent_working_set(session: &SessionState, current: &str) -> Vec<String> {
    let mut out: Vec<String> = Vec::new();
    for f in session.files_touched.iter().rev() {
        if f.path == current || out.contains(&f.path) {
            continue;
        }
        out.push(f.path.clone());
        if out.len() >= TRAVERSAL_WINDOW {
            break;
        }
    }
    out
}

/// Whether `root` is a usable project root for repo-relative normalization.
pub(crate) fn usable_root(root: Option<&str>) -> Option<&str> {
    root.filter(|r| !r.trim().is_empty() && *r != ".")
}

/// First 200 chars of `text` on a UTF-8 boundary — the exact excerpt bound the
/// MCP dispatcher applies before handing content to the IR store
/// (`server/call_tool.rs`), kept identical so CLI- and MCP-recorded IR items are
/// byte-compatible. `ContextIrV1::record` redacts and further caps it.
fn ir_excerpt(text: &str) -> &str {
    const MAX: usize = 200;
    if text.len() <= MAX {
        return text;
    }
    let mut end = MAX;
    while end > 0 && !text.is_char_boundary(end) {
        end -= 1;
    }
    &text[..end]
}

/// Record a file-read operation with full Context OS side effects.
///
/// `duration` and `output_excerpt` feed the Context IR lineage (#566); the MCP
/// dispatcher records both for every tool call but the shadow-mode `lean-ctx
/// read` subprocess used to drop them, so IR/`ctx_proof` exports were blind to
/// compressed shadow reads.
pub fn record_file_read(
    path: &str,
    mode: &str,
    original_tokens: usize,
    output_tokens: usize,
    is_cache_hit: bool,
    duration: std::time::Duration,
    output_excerpt: &str,
) {
    let saved = original_tokens.saturating_sub(output_tokens);
    let tool_key = format!("cli_{mode}");

    stats::record(&tool_key, original_tokens, output_tokens);
    heatmap::record_file_access(path, original_tokens, saved);
    // Verified ledger (#685): recorded explicitly now that the heatmap chokepoint
    // no longer bundles it. This direct-CLI path (daemon off) only has o200k
    // counts; the model-correct re-tokenization happens on the MCP read path,
    // which holds the source text. For the default O200kBase model these are
    // identical anyway.
    crate::core::savings_ledger::record_read_event(original_tokens, saved);

    // Project root the learning sinks below are scoped to. Defaults to "." (the
    // MCP path's `project_root_snapshot` fallback) so a rootless read still
    // trains a global model rather than being dropped.
    let mut learning_root = String::from(".");

    if let Some(mut session) = SessionState::load_latest() {
        session.touch_file(path, None, mode, original_tokens);
        if is_cache_hit {
            session.record_cache_hit();
        }

        if session.active_structured_intent.is_none() && session.files_touched.len() >= 2 {
            let touched: Vec<String> = session
                .files_touched
                .iter()
                .map(|ft| ft.path.clone())
                .collect();
            let inferred = StructuredIntent::from_file_patterns(&touched);
            if inferred.confidence >= 0.4 {
                session.active_structured_intent = Some(inferred);
            }
        }

        let project_root = session.project_root.clone();
        if let Some(root) = usable_root(project_root.as_deref()) {
            learning_root = root.to_string();
        }
        let calls = session.stats.total_tool_calls;

        // Traversal edges: associate this read with the recent working set so the
        // graph learns the files this task actually touches together (#289).
        let working_set = recent_working_set(&session, path);

        let _ = session.save();

        if let Some(root) = usable_root(project_root.as_deref()) {
            crate::core::cooccurrence::record_focus_access(root, path, &working_set);
        }
        maybe_consolidate(project_root.as_deref(), calls);
    }

    // Only real files belong in the context ledger (GL #512): directory
    // overviews and synthetic paths would show up as "files" in the pressure
    // table with eviction/pin semantics that make no sense for them.
    if std::path::Path::new(path).is_file() {
        let mut ledger = ContextLedger::load();
        ledger.record(path, mode, original_tokens, output_tokens);
        ledger.save();
    }

    // Learning sinks the MCP read path runs in a background thread but the CLI
    // path historically skipped — the mode predictor never trained, the
    // compression feedback loop stayed blind and dashboard anomaly signals were
    // missing for every shadow-mode (`view`/`grep` → `lean-ctx read`) hook read
    // (#550). Run inline: a single-shot CLI process must finish them before it
    // flushes and exits, so the off-hot-path thread the daemon uses is moot here.
    record_read_learning(
        path,
        mode,
        original_tokens,
        output_tokens,
        is_cache_hit,
        &learning_root,
    );

    // Context IR lineage (#566): the MCP dispatcher records provenance for every
    // tool call (`server/call_tool.rs`), but the shadow-mode hook's single-shot
    // `lean-ctx read` bypassed it. Disk-backed load→record→save persists the
    // entry before the process exits. `mode` rides the IR `pattern` slot to match
    // the MCP read path (which stores its `mode` arg there).
    let mut ir = ContextIrV1::load();
    ir.record(RecordIrInput {
        kind: ContextIrSourceKindV1::Read,
        tool: "ctx_read",
        client_name: None,
        agent_id: None,
        path: Some(path),
        command: None,
        pattern: Some(mode),
        input_tokens: original_tokens,
        output_tokens,
        duration,
        content_excerpt: ir_excerpt(output_excerpt),
    });
    ir.save();
}

/// Replicate the MCP read path's learning side effects (`registered/ctx_read.rs`
/// background thread) for the standalone CLI path (#550): mode-predictor
/// training, the compression feedback outcome and the per-call anomaly metric.
/// All three are disk-backed and therefore work from a single-shot process; the
/// in-memory-only detectors (loop/correction) and the bounce/adaptive signals
/// that require routing through `ctx_read::handle` are tracked separately.
fn record_read_learning(
    path: &str,
    resolved_mode: &str,
    original_tokens: usize,
    output_tokens: usize,
    is_cache_hit: bool,
    project_root: &str,
) {
    // Mode predictor: train auto-mode selection on the realized compression
    // density, exactly as the MCP background thread does.
    let sig = crate::core::mode_predictor::FileSignature::from_path(path, original_tokens);
    let density = if output_tokens > 0 {
        original_tokens as f64 / output_tokens as f64
    } else {
        1.0
    };
    let outcome = crate::core::mode_predictor::ModeOutcome {
        mode: resolved_mode.to_string(),
        tokens_in: original_tokens,
        tokens_out: output_tokens,
        density: density.min(1.0),
    };
    let mut predictor = crate::core::mode_predictor::ModePredictor::new();
    predictor.set_project_root(project_root);
    predictor.record(sig, outcome);
    predictor.save();

    // Compression feedback: the per-language outcome the adaptive thresholds and
    // bounce-aware tuning learn from. `total_turns`/`total_reads` are 1 — the
    // accurate count for this single-shot invocation, not a placeholder.
    let saved = original_tokens.saturating_sub(output_tokens);
    let ext = std::path::Path::new(path)
        .extension()
        .and_then(|e| e.to_str())
        .unwrap_or("")
        .to_string();
    let thresholds = crate::core::adaptive_thresholds::thresholds_for_path(path);
    let feedback_outcome = crate::core::feedback::CompressionOutcome {
        session_id: format!("{}", std::process::id()),
        language: ext,
        entropy_threshold: thresholds.bpe_entropy,
        jaccard_threshold: thresholds.jaccard,
        total_turns: 1,
        tokens_saved: saved as u64,
        tokens_original: original_tokens as u64,
        cache_hits: u32::from(is_cache_hit),
        total_reads: 1,
        // A compressed read only counts as task-completing when this extension
        // is not in a high-bounce state (#593); unknown stays optimistic so the
        // cold start matches the MCP path. 0.30 mirrors BOUNCE_RATE_THRESHOLD.
        task_completed: crate::core::bounce_tracker::global()
            .lock()
            .ok()
            .and_then(|bt| bt.bounce_rate_for_extension(path))
            .is_none_or(|rate| rate < 0.30),
        timestamp: chrono::Local::now().to_rfc3339(),
    };
    let mut store = crate::core::feedback::FeedbackStore::load();
    store.project_root = Some(project_root.to_string());
    store.record_outcome(feedback_outcome);

    // Anomaly detector: the same per-call metric the MCP post-dispatch records.
    // `save_debounced` writes on the first call of a fresh process (last-save
    // marker starts at 0), so the single shadow read persists before exit.
    crate::core::anomaly::record_metric("tokens_per_call", output_tokens as f64);
    crate::core::anomaly::save_debounced();
}

/// Record a search/grep operation with full Context OS side effects.
///
/// `modeled_baseline` (native-tool estimate, GL #479 D1) feeds the estimated
/// stats series; `observed_tokens` (raw measured match lines, no factor) feeds
/// the verified ledger (GL #479 D2). `pattern`/`path`/`duration`/`output_excerpt`
/// feed the Context IR lineage (#566).
pub fn record_search(
    modeled_baseline: usize,
    observed_tokens: usize,
    output_tokens: usize,
    pattern: &str,
    path: &str,
    duration: std::time::Duration,
    output_excerpt: &str,
) {
    stats::record("cli_grep", modeled_baseline, output_tokens);
    crate::core::savings_ledger::record_tool_event("cli_grep", observed_tokens, output_tokens);

    if let Some(mut session) = SessionState::load_latest() {
        session.record_command();
        let project_root = session.project_root.clone();
        let calls = session.stats.total_tool_calls;
        let _ = session.save();

        maybe_consolidate(project_root.as_deref(), calls);
    }

    // Per-call anomaly metric, mirroring the MCP post-dispatch (#550). Missing it
    // left dashboard signals blind to shadow-mode (`grep` → `lean-ctx grep`) hooks.
    crate::core::anomaly::record_metric("tokens_per_call", output_tokens as f64);
    crate::core::anomaly::save_debounced();

    // Context IR lineage for shadow-mode `grep` → `lean-ctx grep` (#566). The
    // raw matched-line estimate (`observed_tokens`) is the IR input so the stored
    // compression ratio reads matches-in / sent-out.
    let mut ir = ContextIrV1::load();
    ir.record(RecordIrInput {
        kind: ContextIrSourceKindV1::Search,
        tool: "ctx_search",
        client_name: None,
        agent_id: None,
        path: Some(path),
        command: None,
        pattern: Some(pattern),
        input_tokens: observed_tokens,
        output_tokens,
        duration,
        content_excerpt: ir_excerpt(output_excerpt),
    });
    ir.save();
}

/// Record a tree/ls operation with full Context OS side effects.
pub fn record_tree(original_tokens: usize, output_tokens: usize) {
    stats::record("cli_ls", original_tokens, output_tokens);

    if let Some(mut session) = SessionState::load_latest() {
        session.record_command();
        let _ = session.save();
    }
}

/// Record a shell command with full Context OS side effects.
/// Always records in stats (even for track-only 0-token calls) so the dashboard
/// command counter stays accurate. Adding 0 tokens does not inflate savings.
pub fn record_shell_command(original_tokens: usize, output_tokens: usize) {
    stats::record("cli_shell", original_tokens, output_tokens);
    // Shell compression is *measured* (raw output vs sent output), so it belongs
    // in the verified ledger too (GL #479 D2). Zero-saving calls are skipped.
    crate::core::savings_ledger::record_tool_event("cli_shell", original_tokens, output_tokens);

    if let Some(mut session) = SessionState::load_latest() {
        session.record_command();
        let project_root = session.project_root.clone();
        let calls = session.stats.total_tool_calls;
        let _ = session.save();

        if original_tokens > 0 {
            maybe_consolidate(project_root.as_deref(), calls);
        }
    }
}

/// Flush every buffered telemetry sink to disk.
///
/// The long-lived MCP daemon flushes these once at shutdown
/// (`cli/dispatch/server.rs`). Single-shot CLI commands — and the shadow-mode
/// hook subprocesses that spawn `lean-ctx read`/`grep` — exit immediately, so
/// without this the buffered heatmap, mode-predictor, feedback and threshold
/// writes are silently lost the moment the process ends: `lean-ctx heatmap`
/// stays empty and `lean-ctx gain` reports nothing for compressed reads (#550).
///
/// Centralized so the daemon shutdown, the parent watchdog and every CLI tool
/// command flush the *exact same* set — the historical per-arm copies had
/// drifted (the `read` arm flushed only `stats`, the `-c` arm four sinks, the
/// daemon nine), which is precisely how the gap went unnoticed.
pub fn flush_all() {
    stats::flush();
    heatmap::flush();
    crate::core::path_mode_memory::flush();
    crate::core::auto_mode_resolver::flush_sources();
    crate::core::edit_quality::flush();
    crate::core::mode_predictor::ModePredictor::flush();
    crate::core::feedback::FeedbackStore::flush();
    crate::core::threshold_learning::flush();
    crate::core::litm_calibration::flush();
}

// TODO(arch): crate::tools::autonomy is still referenced here. Move AutonomyState
// and should_auto_consolidate to core::autonomy_drivers for a clean layer boundary.
fn maybe_consolidate(project_root: Option<&str>, calls: u32) {
    let Some(root) = project_root else { return };
    let autonomy = crate::tools::autonomy::AutonomyState::new();
    if crate::tools::autonomy::should_auto_consolidate(&autonomy, calls) {
        let root = root.to_string();
        let _ = crate::core::consolidation_engine::consolidate_latest(
            &root,
            crate::core::consolidation_engine::ConsolidationBudgets::default(),
        );
    }
}

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

    // The record_* paths now drive process-global telemetry sinks (mode
    // predictor buffer, anomaly singleton) and read the data-dir env (#550), so
    // every test here takes the shared isolation lock to serialize that state
    // and keep its disk writes inside a throwaway dir.

    #[test]
    fn record_file_read_does_not_panic_without_session() {
        let _dir = crate::core::data_dir::isolated_data_dir();
        record_file_read(
            "/tmp/nonexistent.rs",
            "full",
            100,
            50,
            false,
            std::time::Duration::from_millis(1),
            "excerpt",
        );
    }

    #[test]
    fn record_search_does_not_panic_without_session() {
        let _dir = crate::core::data_dir::isolated_data_dir();
        record_search(
            500,
            200,
            150,
            "pattern",
            "/tmp",
            std::time::Duration::from_millis(1),
            "matches",
        );
    }

    #[test]
    fn record_tree_does_not_panic_without_session() {
        let _dir = crate::core::data_dir::isolated_data_dir();
        record_tree(100, 80);
    }

    #[test]
    fn record_shell_does_not_panic_without_session() {
        let _dir = crate::core::data_dir::isolated_data_dir();
        record_shell_command(500, 200);
    }

    #[test]
    fn flush_all_is_idempotent_and_safe_without_state() {
        let _dir = crate::core::data_dir::isolated_data_dir();
        // Empty buffers: flushing must be a harmless no-op, and calling it twice
        // (e.g. a CLI arm followed by an atexit path) must never panic.
        flush_all();
        flush_all();
    }

    #[test]
    fn cli_read_persists_learning_sinks_to_disk() {
        // #550 regression: a single-shot CLI read must leave the mode predictor,
        // compression feedback and heatmap on disk. The daemon used to be the
        // only path that flushed them, so shadow-mode hook reads (`view`/`grep` →
        // `lean-ctx read`) recorded nothing and `lean-ctx heatmap` stayed empty.
        let dir = crate::core::data_dir::isolated_data_dir();
        let file = dir.path().join("sample.rs");
        std::fs::write(&file, "fn main() {\n    println!(\"hi\");\n}\n").unwrap();
        let path = file.to_string_lossy();

        record_file_read(
            &path,
            "full",
            1000,
            200,
            false,
            std::time::Duration::from_millis(2),
            "sample.rs [3L]\nfn main() {}",
        );
        flush_all();

        let data = crate::core::data_dir::lean_ctx_data_dir().expect("data dir");
        let state = crate::core::paths::state_dir().expect("state dir");
        assert!(
            data.join("mode_stats.json").exists(),
            "mode predictor must persist after a CLI read + flush"
        );
        assert!(
            state.join("feedback.json").exists(),
            "compression feedback must persist after a CLI read + flush"
        );
        assert!(
            state.join("heatmap.json").exists(),
            "heatmap must persist after a CLI read + flush"
        );
    }

    #[test]
    fn cli_read_records_context_ir_lineage() {
        // #566: the MCP dispatcher records Context IR for every tool call, but the
        // shadow-mode `lean-ctx read` subprocess used to skip it, so IR/ctx_proof
        // exports were blind to compressed shadow reads. A single-shot CLI read
        // must now persist exactly one IR item (disk-backed load→record→save).
        let dir = crate::core::data_dir::isolated_data_dir();
        let file = dir.path().join("ir_sample.rs");
        std::fs::write(&file, "fn main() {}\n").unwrap();
        let path = file.to_string_lossy();

        record_file_read(
            &path,
            "full",
            1000,
            200,
            false,
            std::time::Duration::from_millis(3),
            "ir_sample.rs [1L]\nfn main() {}",
        );

        let ir = ContextIrV1::load();
        assert_eq!(ir.items.len(), 1, "exactly one IR item per CLI read");
        let item = &ir.items[0];
        assert_eq!(item.source.tool, "ctx_read");
        assert!(matches!(item.source.kind, ContextIrSourceKindV1::Read));
        assert!(
            item.source
                .path
                .as_deref()
                .unwrap_or("")
                .ends_with("ir_sample.rs"),
            "IR records the read path, got {:?}",
            item.source.path
        );
        assert_eq!(item.source.pattern.as_deref(), Some("full"));
        assert_eq!(item.input_tokens, 1000);
        assert_eq!(item.output_tokens, 200);
        assert!(item.duration_us > 0, "a real duration must be recorded");
        assert!(!item.content_excerpt.is_empty(), "excerpt must be captured");
    }

    #[test]
    fn cli_search_records_context_ir_lineage() {
        // #566: the shadow-mode `grep` → `lean-ctx grep` path records IR too.
        let _dir = crate::core::data_dir::isolated_data_dir();

        record_search(
            800,
            500,
            120,
            "fn handle",
            "src/",
            std::time::Duration::from_millis(4),
            "src/lib.rs:12: fn handle() {}",
        );

        let ir = ContextIrV1::load();
        assert_eq!(ir.items.len(), 1, "exactly one IR item per CLI search");
        let item = &ir.items[0];
        assert_eq!(item.source.tool, "ctx_search");
        assert!(matches!(item.source.kind, ContextIrSourceKindV1::Search));
        // Input is the raw matched-line estimate, not the modeled baseline.
        assert_eq!(item.input_tokens, 500);
        assert_eq!(item.output_tokens, 120);
        assert!(item.duration_us > 0, "a real duration must be recorded");
    }
}