lean-ctx 3.9.5

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
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
//! Quality loop v1 (#494): compression-caused edit failures feed back into
//! mode selection.
//!
//! `BounceTracker`/`path_mode_memory` close the loop for *re-read* bounces,
//! but an edit that fails because the file was last read in a compressed mode
//! (`old_string` not found — the body simply wasn't in context) taught the
//! system nothing. This module records edit outcomes correlated with the last
//! read mode and feeds two signals back into `auto_mode_resolver::resolve`:
//!
//! 1. **Per-path escalation** — after a compression-correlated edit failure
//!    the *next* auto read of that file resolves to `full` (one-shot, 1 h TTL).
//! 2. **Per-(extension × mode) penalty** — modes whose edit-failure rate for a
//!    file type crosses the risky threshold resolve to `full` until the rate
//!    recovers (hysteresis, see below).
//!
//! Risk formula (documented in `docs/contracts/quality-loop-v1.md`):
//! a (ext, mode) pair becomes risky when `fails >= 2 && fails / (fails +
//! successes) >= 0.25`, and stops being risky only when the rate drops below
//! `0.15` — two thresholds so one lucky edit doesn't flap the decision.
//!
//! Storage: `~/.lean-ctx/edit_quality.json`, atomic write (tmp+rename),
//! loaded once per process, flushed periodically like `path_mode_memory`.

use std::collections::HashMap;
use std::path::PathBuf;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Mutex, OnceLock};

use serde::{Deserialize, Serialize};

const STORE_FILE: &str = "edit_quality.json";
/// (ext, mode) pairs without a failure for this long are dropped on load.
const DECAY_SECS: u64 = 30 * 24 * 3600;
/// Pending per-path escalations expire after this long.
const ESCALATION_TTL_SECS: u64 = 3600;
/// Hard caps; oldest entries are evicted first.
const MAX_PAIRS: usize = 200;
const MAX_PENDING: usize = 100;
const FLUSH_EVERY: usize = 10;

/// Risky when the failure share reaches this rate (with >= 2 fails)…
const RISKY_ENTER_RATE: f64 = 0.25;
/// …and recovers only once the rate drops below this (hysteresis).
const RISKY_EXIT_RATE: f64 = 0.15;
const RISKY_MIN_FAILS: u32 = 2;

static STORE: OnceLock<Mutex<EditQualityStore>> = OnceLock::new();
static RECORD_CALLS: AtomicUsize = AtomicUsize::new(0);

#[derive(Debug, Clone, Serialize, Deserialize, Default)]
pub struct PairStats {
    pub fails: u32,
    pub successes: u32,
    pub risky: bool,
    pub last_fail_unix: u64,
}

impl PairStats {
    fn fail_rate(&self) -> f64 {
        let total = self.fails + self.successes;
        if total == 0 {
            return 0.0;
        }
        f64::from(self.fails) / f64::from(total)
    }

    /// Applies the documented enter/exit thresholds after every outcome.
    fn update_risky(&mut self) {
        if self.risky {
            if self.fail_rate() < RISKY_EXIT_RATE {
                self.risky = false;
            }
        } else if self.fails >= RISKY_MIN_FAILS && self.fail_rate() >= RISKY_ENTER_RATE {
            self.risky = true;
        }
    }
}

#[derive(Debug, Clone, Serialize, Deserialize, Default)]
pub struct EditQualityStore {
    /// Key: `"{ext}|{mode}"` (e.g. `"rs|map"`).
    pub pairs: HashMap<String, PairStats>,
    /// Normalized path -> unix time of the compression-correlated edit fail.
    pub pending_escalations: HashMap<String, u64>,
    /// Normalized path -> unix time of an anchored-edit (`ctx_patch`) staleness
    /// miss. The next auto read of that path resolves to `anchored` (not `full`),
    /// so the model gets fresh line anchors to retry by reference (#1008).
    /// `#[serde(default)]` keeps stores written before anchored editing loadable.
    #[serde(default)]
    pub pending_anchored_escalations: HashMap<String, u64>,
    /// All-time counter of consumed escalations (observability).
    #[serde(default)]
    pub escalations_served: u64,
    #[serde(skip)]
    dirty: bool,
}

fn pair_key(ext: &str, mode: &str) -> String {
    format!("{ext}|{mode}")
}

/// Evict the oldest entries of a `path -> timestamp` pending map down to
/// [`MAX_PENDING`]; flips `dirty` when anything was dropped. Shared by the
/// `full` and `anchored` escalation maps.
fn evict_pending_to_cap(map: &mut HashMap<String, u64>, dirty: &mut bool) {
    if map.len() <= MAX_PENDING {
        return;
    }
    let mut items: Vec<(String, u64)> = map.iter().map(|(k, ts)| (k.clone(), *ts)).collect();
    items.sort_by_key(|(_, ts)| *ts);
    let drop_n = map.len() - MAX_PENDING;
    for (key, _) in items.into_iter().take(drop_n) {
        map.remove(&key);
    }
    *dirty = true;
}

impl EditQualityStore {
    fn load_from_disk() -> Self {
        let Ok(raw) = std::fs::read_to_string(store_path()) else {
            return Self::default();
        };
        let mut store: Self = serde_json::from_str(&raw).unwrap_or_default();
        store.decay(now_unix());
        store
    }

    fn decay(&mut self, now: u64) {
        let before = self.pairs.len()
            + self.pending_escalations.len()
            + self.pending_anchored_escalations.len();
        self.pairs
            .retain(|_, s| now.saturating_sub(s.last_fail_unix) <= DECAY_SECS);
        self.pending_escalations
            .retain(|_, ts| now.saturating_sub(*ts) <= ESCALATION_TTL_SECS);
        self.pending_anchored_escalations
            .retain(|_, ts| now.saturating_sub(*ts) <= ESCALATION_TTL_SECS);
        if self.pairs.len()
            + self.pending_escalations.len()
            + self.pending_anchored_escalations.len()
            != before
        {
            self.dirty = true;
        }
    }

    fn evict_to_caps(&mut self) {
        if self.pairs.len() > MAX_PAIRS {
            let mut items: Vec<(String, u64)> = self
                .pairs
                .iter()
                .map(|(k, s)| (k.clone(), s.last_fail_unix))
                .collect();
            items.sort_by_key(|(_, ts)| *ts);
            let drop_n = self.pairs.len() - MAX_PAIRS;
            for (key, _) in items.into_iter().take(drop_n) {
                self.pairs.remove(&key);
            }
            self.dirty = true;
        }
        evict_pending_to_cap(&mut self.pending_escalations, &mut self.dirty);
        evict_pending_to_cap(&mut self.pending_anchored_escalations, &mut self.dirty);
    }

    pub fn record_failure(&mut self, ext: &str, mode: &str, now: u64) {
        let entry = self.pairs.entry(pair_key(ext, mode)).or_default();
        entry.fails = entry.fails.saturating_add(1);
        entry.last_fail_unix = now;
        entry.update_risky();
        self.dirty = true;
        self.evict_to_caps();
    }

    pub fn record_success(&mut self, ext: &str, mode: &str) {
        let entry = self.pairs.entry(pair_key(ext, mode)).or_default();
        entry.successes = entry.successes.saturating_add(1);
        entry.update_risky();
        self.dirty = true;
    }

    pub fn set_pending_escalation(&mut self, norm_path: &str, now: u64) {
        self.pending_escalations.insert(norm_path.to_string(), now);
        self.dirty = true;
        self.evict_to_caps();
    }

    /// Consumes the escalation for this path if present and not expired.
    pub fn take_pending_escalation(&mut self, norm_path: &str, now: u64) -> bool {
        Self::take_from(
            &mut self.pending_escalations,
            norm_path,
            now,
            &mut self.escalations_served,
            &mut self.dirty,
        )
    }

    pub fn set_pending_anchored_escalation(&mut self, norm_path: &str, now: u64) {
        self.pending_anchored_escalations
            .insert(norm_path.to_string(), now);
        self.dirty = true;
        self.evict_to_caps();
    }

    /// Consumes the anchored escalation for this path if present and not expired.
    pub fn take_pending_anchored_escalation(&mut self, norm_path: &str, now: u64) -> bool {
        Self::take_from(
            &mut self.pending_anchored_escalations,
            norm_path,
            now,
            &mut self.escalations_served,
            &mut self.dirty,
        )
    }

    /// Shared one-shot consume: remove `norm_path`, count it served when still
    /// within [`ESCALATION_TTL_SECS`], else drop it silently.
    fn take_from(
        map: &mut HashMap<String, u64>,
        norm_path: &str,
        now: u64,
        served: &mut u64,
        dirty: &mut bool,
    ) -> bool {
        match map.remove(norm_path) {
            Some(ts) if now.saturating_sub(ts) <= ESCALATION_TTL_SECS => {
                *served += 1;
                *dirty = true;
                true
            }
            Some(_) => {
                *dirty = true;
                false
            }
            None => false,
        }
    }

    pub fn is_risky(&self, ext: &str, mode: &str) -> bool {
        self.pairs
            .get(&pair_key(ext, mode))
            .is_some_and(|s| s.risky)
    }

    pub fn save(&self) -> std::io::Result<()> {
        let path = store_path();
        if let Some(parent) = path.parent() {
            std::fs::create_dir_all(parent)?;
        }
        let json = serde_json::to_string(self)?;
        let tmp = path.with_extension("tmp");
        std::fs::write(&tmp, json)?;
        std::fs::rename(&tmp, &path)
    }
}

fn store_path() -> PathBuf {
    crate::core::data_dir::lean_ctx_data_dir()
        .unwrap_or_else(|_| PathBuf::from("."))
        .join(STORE_FILE)
}

fn now_unix() -> u64 {
    std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
        .map_or(0, |d| d.as_secs())
}

fn global() -> &'static Mutex<EditQualityStore> {
    STORE.get_or_init(|| Mutex::new(EditQualityStore::load_from_disk()))
}

fn ext_of(path: &str) -> String {
    std::path::Path::new(path)
        .extension()
        .and_then(|e| e.to_str())
        .unwrap_or("")
        .to_string()
}

/// Process-global: record the outcome of an edit, correlated with the mode of
/// the last read of that file. `last_mode` must be the recorded read mode
/// (empty = file was never read through lean-ctx → no signal, skipped).
/// Compression-correlated failures additionally arm the one-shot per-path
/// escalation so the next auto read of `path` resolves to `full`.
pub fn record_edit_outcome(path: &str, last_mode: &str, success: bool) {
    record_outcome_with(path, last_mode, success, Escalation::Full);
}

/// Like [`record_edit_outcome`], but a failure is a `ctx_patch` anchor-staleness
/// miss: the recovery is a *fresh anchored read* (the model edits by reference),
/// so the next auto read escalates to `anchored` instead of `full` (#1008).
pub fn record_anchored_edit_outcome(path: &str, last_mode: &str, success: bool) {
    record_outcome_with(path, last_mode, success, Escalation::Anchored);
}

/// Which read mode the *next* auto read escalates to after a correlated edit
/// failure. Both are high-signal "the context the model edited against was
/// wrong" events; they differ only in the recovery view handed back.
#[derive(Clone, Copy)]
enum Escalation {
    /// str_replace miss → give the real body (`full`).
    Full,
    /// anchored miss → give fresh line anchors (`anchored`).
    Anchored,
}

impl Escalation {
    /// The read mode that fully neutralizes this failure class, hence the value
    /// to *not* re-arm against (escalating `full→full` / `anchored→anchored` is a
    /// no-op).
    fn target_mode(self) -> &'static str {
        match self {
            Escalation::Full => "full",
            Escalation::Anchored => "anchored",
        }
    }
}

fn record_outcome_with(path: &str, last_mode: &str, success: bool, esc: Escalation) {
    if last_mode.is_empty() {
        return;
    }
    let ext = ext_of(path);
    let Ok(mut store) = global().lock() else {
        return;
    };
    if success {
        store.record_success(&ext, last_mode);
    } else {
        let now = now_unix();
        store.record_failure(&ext, last_mode, now);
        if last_mode != esc.target_mode() {
            let norm = crate::core::pathutil::normalize_tool_path(path);
            match esc {
                Escalation::Full => store.set_pending_escalation(&norm, now),
                Escalation::Anchored => store.set_pending_anchored_escalation(&norm, now),
            }
            // Quality signal (#538): edit failures after a stale read are the
            // strongest "the model's view was wrong" evidence we have — they also
            // penalize the bandit arm that produced the read (#593).
            crate::core::adaptive_thresholds::record_quality_signal(
                path,
                crate::core::threshold_learning::QualitySignal::EditFail,
            );
            // Stigmergy (#540): edit failures mark the path as Stuck ("context
            // drifted"), the explicit anchor-miss signal called for in #1008.
            let scent_path = norm.clone();
            std::thread::spawn(move || {
                crate::core::scent_field::deposit(
                    crate::core::scent_field::scent_agent_id(),
                    crate::core::scent_field::ScentKind::Stuck,
                    &scent_path,
                    1.0,
                );
            });
        }
    }
    maybe_flush(&mut store);
}

/// Process-global: one-shot check-and-consume of the per-path `full` escalation.
pub fn take_pending_escalation(path: &str) -> bool {
    consume_escalation(path, false)
}

/// Process-global: one-shot check-and-consume of the per-path `anchored`
/// escalation (armed by [`record_anchored_edit_outcome`]).
pub fn take_pending_anchored_escalation(path: &str) -> bool {
    consume_escalation(path, true)
}

fn consume_escalation(path: &str, anchored: bool) -> bool {
    let norm = crate::core::pathutil::normalize_tool_path(path);
    let Ok(mut store) = global().lock() else {
        return false;
    };
    let now = now_unix();
    let hit = if anchored {
        store.take_pending_anchored_escalation(&norm, now)
    } else {
        store.take_pending_escalation(&norm, now)
    };
    if hit {
        maybe_flush(&mut store);
    }
    hit
}

/// Process-global: is `mode` currently risky for files with this extension?
pub fn is_risky_mode(path: &str, mode: &str) -> bool {
    let ext = ext_of(path);
    global().lock().is_ok_and(|s| s.is_risky(&ext, mode))
}

/// Snapshot for `ctx_metrics`: (risky pairs, per-pair stats, escalations served).
pub fn metrics_snapshot() -> serde_json::Value {
    let Ok(store) = global().lock() else {
        return serde_json::json!({});
    };
    let mut pairs: Vec<serde_json::Value> = store
        .pairs
        .iter()
        .map(|(key, s)| {
            serde_json::json!({
                "pair": key,
                "fails": s.fails,
                "successes": s.successes,
                "fail_rate": (s.fail_rate() * 1000.0).round() / 1000.0,
                "risky": s.risky,
            })
        })
        .collect();
    pairs.sort_by(|a, b| {
        let fa = a["fail_rate"].as_f64().unwrap_or(0.0);
        let fb = b["fail_rate"].as_f64().unwrap_or(0.0);
        fb.partial_cmp(&fa).unwrap_or(std::cmp::Ordering::Equal)
    });
    serde_json::json!({
        "pairs": pairs,
        "pending_escalations": store.pending_escalations.len(),
        "pending_anchored_escalations": store.pending_anchored_escalations.len(),
        "escalations_served": store.escalations_served,
    })
}

pub fn flush() {
    if let Ok(store) = global().lock()
        && store.dirty
    {
        let _ = store.save();
    }
}

fn maybe_flush(store: &mut EditQualityStore) {
    let n = RECORD_CALLS.fetch_add(1, Ordering::Relaxed) + 1;
    if n.is_multiple_of(FLUSH_EVERY) && store.dirty && store.save().is_ok() {
        store.dirty = false;
    }
}

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

    #[test]
    fn risky_after_two_majority_fails_with_hysteresis() {
        let mut s = EditQualityStore::default();
        s.record_failure("rs", "map", 1000);
        assert!(!s.is_risky("rs", "map"), "one fail is not a pattern");
        s.record_failure("rs", "map", 1001);
        assert!(s.is_risky("rs", "map"), "2 fails, rate 1.0 >= 0.25");

        // Rate must drop below 0.15 to recover: 2 fails need > 11 successes.
        for _ in 0..11 {
            s.record_success("rs", "map");
        }
        assert!(s.is_risky("rs", "map"), "2/13 ≈ 0.154 still risky");
        s.record_success("rs", "map");
        assert!(!s.is_risky("rs", "map"), "2/14 ≈ 0.143 < 0.15 recovers");
    }

    #[test]
    fn entering_risky_needs_quarter_rate_not_just_two_fails() {
        let mut s = EditQualityStore::default();
        for _ in 0..7 {
            s.record_success("ts", "signatures");
        }
        s.record_failure("ts", "signatures", 1000);
        s.record_failure("ts", "signatures", 1001);
        // 2 fails / 9 total ≈ 0.22 < 0.25 — healthy mode stays usable.
        assert!(!s.is_risky("ts", "signatures"));
        s.record_failure("ts", "signatures", 1002);
        // 3/10 = 0.30 — now risky.
        assert!(s.is_risky("ts", "signatures"));
    }

    #[test]
    fn penalty_is_per_extension_not_global() {
        let mut s = EditQualityStore::default();
        s.record_failure("rs", "map", 1000);
        s.record_failure("rs", "map", 1001);
        assert!(s.is_risky("rs", "map"));
        assert!(!s.is_risky("py", "map"), "py|map untouched");
        assert!(!s.is_risky("rs", "signatures"), "rs|signatures untouched");
    }

    #[test]
    fn escalation_is_one_shot_and_expires() {
        let mut s = EditQualityStore::default();
        s.set_pending_escalation("src/a.rs", 1000);
        assert!(s.take_pending_escalation("src/a.rs", 1100));
        assert!(
            !s.take_pending_escalation("src/a.rs", 1101),
            "consumed — second read is normal again"
        );
        assert_eq!(s.escalations_served, 1);

        s.set_pending_escalation("src/b.rs", 1000);
        assert!(
            !s.take_pending_escalation("src/b.rs", 1000 + ESCALATION_TTL_SECS + 1),
            "expired escalations are dropped, not served"
        );
        assert_eq!(s.escalations_served, 1);
    }

    #[test]
    fn anchored_escalation_is_independent_and_one_shot() {
        // #1008: the anchored map is separate from the `full` map — arming one
        // must never consume the other, so str_replace and ctx_patch recoveries
        // don't cross-talk.
        let mut s = EditQualityStore::default();
        s.set_pending_anchored_escalation("src/a.rs", 1000);
        assert!(
            !s.take_pending_escalation("src/a.rs", 1100),
            "anchored arming must not satisfy a full escalation"
        );
        assert!(s.take_pending_anchored_escalation("src/a.rs", 1100));
        assert!(
            !s.take_pending_anchored_escalation("src/a.rs", 1101),
            "anchored escalation is one-shot"
        );
        assert_eq!(s.escalations_served, 1);
    }

    #[test]
    fn anchored_outcome_arms_anchored_not_full() {
        // A miss after an anchored read arms only the anchored escalation.
        let mut s = EditQualityStore::default();
        s.record_failure("rs", "anchored", 1000);
        s.set_pending_anchored_escalation("src/x.rs", 1000);
        assert!(s.pending_escalations.is_empty());
        assert_eq!(s.pending_anchored_escalations.len(), 1);
    }

    #[test]
    fn store_without_anchored_field_deserializes() {
        // Back-compat (#1008): a store written before anchored editing has no
        // `pending_anchored_escalations` key; `#[serde(default)]` must fill it.
        let legacy = r#"{"pairs":{},"pending_escalations":{"old.rs":42}}"#;
        let s: EditQualityStore = serde_json::from_str(legacy).unwrap();
        assert!(s.pending_anchored_escalations.is_empty());
        assert!(s.pending_escalations.contains_key("old.rs"));
    }

    #[test]
    fn decay_drops_stale_pairs_and_pendings() {
        let mut s = EditQualityStore::default();
        s.record_failure("rs", "map", 1000);
        s.record_failure("go", "map", 5000);
        s.set_pending_escalation("old.rs", 1000);
        s.set_pending_escalation("fresh.rs", 5000);
        s.decay(5000 + DECAY_SECS - 10);
        assert!(!s.pairs.contains_key("rs|map"));
        assert!(s.pairs.contains_key("go|map"));
        // Pendings use the much shorter escalation TTL.
        assert!(s.pending_escalations.is_empty());
    }

    #[test]
    fn eviction_keeps_newest() {
        let mut s = EditQualityStore::default();
        for i in 0..(MAX_PAIRS + 10) {
            s.record_failure(&format!("e{i}"), "map", 1000 + i as u64);
        }
        assert_eq!(s.pairs.len(), MAX_PAIRS);
        assert!(!s.pairs.contains_key("e0|map"));
        for i in 0..(MAX_PENDING + 5) {
            s.set_pending_escalation(&format!("f{i}.rs"), 1000 + i as u64);
        }
        assert_eq!(s.pending_escalations.len(), MAX_PENDING);
        assert!(!s.pending_escalations.contains_key("f0.rs"));
    }

    #[test]
    fn roundtrip_serialization() {
        let mut s = EditQualityStore::default();
        s.record_failure("rs", "map", 42);
        s.set_pending_escalation("x.rs", 42);
        let json = serde_json::to_string(&s).unwrap();
        let back: EditQualityStore = serde_json::from_str(&json).unwrap();
        assert_eq!(back.pairs.get("rs|map").unwrap().fails, 1);
        assert!(back.pending_escalations.contains_key("x.rs"));
    }
}