newt-core 0.7.1

Newt-Agent core types, errors, and the NeMoCode-style tier router
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
//! The `recall` tool seam — model-driven search over PAST conversations
//! (Step 17.5, #246).
//!
//! Mirrors the `save_note` architecture ([`super::note_sink`]): the loop
//! cannot name [`crate::store::ConversationStore`] directly without dragging
//! persistence wiring into `newt-core::agentic`, so the seam is a minimal
//! trait — the TUI injects [`StoreRecallSource`] (workspace-fenced by the
//! store itself, current conversation excluded) and passes it through
//! `ChatCtx` as `Option<&dyn RecallSource>`. `None` ⇒ the tool is not
//! advertised and the loop never searches — eval/headless callers are
//! unaffected.
//!
//! Design lineage (hermes-agent study, `session_search_tool.py`): coaching
//! schema text ("USE THIS PROACTIVELY when the user references prior
//! work"), snippets as the whole payload (no full content, no aux-LLM
//! recaps), and the 17.3 sanitizer pre-flight so a query of pure
//! operators comes back as coaching, never as a loop-aborting error.

use super::display::{print_tool_call, print_tool_output};
use crate::store::SearchHit;

/// Default hit count when the model omits `limit`.
const RECALL_DEFAULT_LIMIT: usize = 5;
/// Hard ceiling on hits per call — snippets ride back through the model's
/// context, so the cap is a token-budget guard (18.1), not a search knob.
const RECALL_MAX_LIMIT: usize = 10;

/// Read-only search over PAST conversations behind the `recall` tool.
///
/// Object-safe and shareable (the loop holds `&dyn RecallSource`; `Sync`
/// because the borrow crosses `.await` points). Implementations MUST be
/// scoped to the active workspace and MUST exclude the conversation the
/// model is currently in — what's said here is already in context, and a
/// recall hit on it would teach the model to search instead of read.
/// Hit order is the backend's bm25 rank (§6: never re-sorted by any
/// timestamp — wall-clock is a display claim, not an ordering key).
pub trait RecallSource: Send + Sync {
    /// Return up to `limit` best-first hits for `query` (plain keywords;
    /// the executor has already pre-flighted the 17.3 sanitizer).
    fn search(&self, query: &str, limit: usize) -> anyhow::Result<Vec<SearchHit>>;

    /// Return up to `limit` of THIS conversation's most recent durable turns —
    /// the deliberate OPPOSITE of [`Self::search`]'s exclusion filter (#714).
    ///
    /// The `recall` contract refuses the active conversation (what's said here
    /// is already in context); but after an interrupt + auto-resume the early
    /// turns compaction cut from the live window survive ONLY in the store and
    /// are reachable ONLY here. The `resume_context` tool uses this to let a
    /// resumed thread read its own pre-interrupt work. Hits are oldest-first
    /// (chronological), not bm25-ranked — this is a self-read, not a search.
    ///
    /// Default impl returns `Ok(vec![])` so existing mocks and headless callers
    /// (which have no conversation store) compile and behave unchanged — only
    /// [`StoreRecallSource`] overrides it.
    fn this_conversation_recent(&self, limit: usize) -> anyhow::Result<Vec<SearchHit>> {
        let _ = limit;
        Ok(Vec::new())
    }
}

/// The canonical [`RecallSource`]: [`crate::store::ConversationStore::search`]
/// (already fenced to the store's workspace key) minus the current
/// conversation. The TUI constructs one per turn next to its `NoteSink`.
pub struct StoreRecallSource<'a> {
    store: &'a crate::store::ConversationStore,
    current_conversation_id: &'a str,
}

/// Extra hits fetched beyond `limit` before the current conversation's own
/// turns are filtered out. If the current conversation holds more than this
/// many hits ranked above the cut, fewer than `limit` external hits return —
/// accepted: those would be the worst-ranked matches anyway.
const EXCLUSION_FETCH_HEADROOM: usize = 20;

impl<'a> StoreRecallSource<'a> {
    /// `current_conversation_id` is the conversation the model is in right
    /// now — its turns never appear in results (that's what context is for).
    pub fn new(
        store: &'a crate::store::ConversationStore,
        current_conversation_id: &'a str,
    ) -> Self {
        Self {
            store,
            current_conversation_id,
        }
    }
}

impl RecallSource for StoreRecallSource<'_> {
    fn search(&self, query: &str, limit: usize) -> anyhow::Result<Vec<SearchHit>> {
        // Over-fetch, drop the current conversation, truncate. bm25 order is
        // preserved end-to-end (§6 discipline: the store's rank IS the
        // ordering; nothing here re-sorts).
        let hits = self
            .store
            .search(query, limit.saturating_add(EXCLUSION_FETCH_HEADROOM))?;
        Ok(hits
            .into_iter()
            .filter(|h| h.conversation_id != self.current_conversation_id)
            .take(limit)
            .collect())
    }

    fn this_conversation_recent(&self, limit: usize) -> anyhow::Result<Vec<SearchHit>> {
        // The OPPOSITE of `search`'s filter (#714): load THIS conversation's own
        // record (the same path `/conversation restore` uses) and hand back its
        // last `limit` turns. On resume these are the pre-interrupt turns
        // compaction cut from the live window — durable in the store, refused by
        // `recall` by design, reachable only here.
        if limit == 0 {
            return Ok(Vec::new());
        }
        let record = self.store.load(self.current_conversation_id)?;
        let start = record.turns.len().saturating_sub(limit);
        Ok(record
            .turns
            .iter()
            .enumerate()
            .skip(start)
            .map(|(i, turn)| SearchHit {
                conversation_id: record.id.clone(),
                title: record.title.clone(),
                // Turns load without their §6 per-writer seq, so the 1-based
                // position is the ordering label shown to the model. This is a
                // self-read, not a recall rank — no bm25 score applies.
                seq: (i + 1) as i64,
                snippet: recent_turn_snippet(&turn.user, &turn.assistant),
                rank: 0.0,
            })
            .collect())
    }
}

/// Per-field char cap when compacting a turn into a `resume_context` snippet
/// (#714) — a long turn must not blow the model's send budget on resume.
const RESUME_TURN_FIELD_CAP: usize = 600;

/// Compact one turn (the user ask + the assistant reply) into a single readable
/// snippet for the `resume_context` self-read (#714): whitespace flattened (turn
/// text is multi-line) and each field length-capped.
fn recent_turn_snippet(user: &str, assistant: &str) -> String {
    format!(
        "you: {}\n    me: {}",
        flatten_and_cap(user, RESUME_TURN_FIELD_CAP),
        flatten_and_cap(assistant, RESUME_TURN_FIELD_CAP),
    )
}

/// Collapse internal whitespace to single spaces and truncate to `cap` chars
/// with a `[…]` marker — shared by [`recent_turn_snippet`].
fn flatten_and_cap(s: &str, cap: usize) -> String {
    let flat = s.split_whitespace().collect::<Vec<_>>().join(" ");
    if flat.chars().count() > cap {
        let head: String = flat.chars().take(cap).collect();
        format!("{head}[…]")
    } else {
        flat
    }
}

// ---------------------------------------------------------------------------
// Tool schema
// ---------------------------------------------------------------------------

/// The model-facing contract for `recall` — coaching text for a small local
/// LLM: what it searches (PAST conversations, this workspace), when to reach
/// for it (user references prior work), when NOT to (already in context),
/// and how to query (plain keywords, not boolean/FTS syntax). Kept tight:
/// schema tokens ride in every request (18.1).
const RECALL_DESCRIPTION: &str =
    "Search PAST conversations in this workspace. The current conversation is \
     never searched — what was said here is already in your context. Use this \
     when the user references earlier work ('like we did before', 'that bug \
     we fixed', 'where did we leave off') or resumes a topic you don't see in \
     context. Do NOT use it for information already in this conversation. \
     Query with plain keywords ('tokio panic retry'), not boolean operators \
     or quotes. Each result is a conversation id, title, and snippet with the \
     match marked «like this»; the user can reopen one with \
     /conversation restore <id>.";

/// The `recall` tool definition. NOT part of [`super::tool_definitions`]:
/// the loop advertises it only when a [`RecallSource`] is present, so
/// headless / eval callers (which pass `recall_source: None`) never see it.
pub fn recall_tool_definition() -> serde_json::Value {
    serde_json::json!({
        "type": "function",
        "function": {
            "name": "recall",
            "description": RECALL_DESCRIPTION,
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {
                        "type": "string",
                        "description": "Plain keywords describing what to find, e.g. \
                                        'fts5 sanitizer tests' — no boolean or quote syntax"
                    },
                    "limit": {
                        "type": "integer",
                        "description": "Max matches to return, 1-10 (default 5)"
                    }
                },
                "required": ["query"]
            }
        }
    })
}

// ---------------------------------------------------------------------------
// Dispatch
// ---------------------------------------------------------------------------

/// Execute one `recall` call against the source and return the result text
/// fed back to the model.
///
/// Outcome contract (every branch is a *tool result*, never a loop abort):
/// - hits → one block per hit: short id, title, `seq N`, snippet with the
///   FTS5 `>>>`/`<<<` markers rewritten to `«`/`»` (the 17.4 convention);
/// - a query the 17.3 sanitizer rejects → "no searchable terms" coaching;
/// - zero hits → "no matches…" (never an empty string);
/// - a real backend failure → `error: …`, verbatim, like every other tool.
pub(crate) fn execute_recall(
    args: &serde_json::Value,
    source: &dyn RecallSource,
    color: bool,
    tool_output_lines: usize,
) -> String {
    let query = args["query"].as_str().unwrap_or("").trim();
    // Absent / non-integer limit → default; present → clamped to [1, max].
    let limit = args["limit"]
        .as_u64()
        .map(|l| {
            usize::try_from(l)
                .unwrap_or(RECALL_MAX_LIMIT)
                .clamp(1, RECALL_MAX_LIMIT)
        })
        .unwrap_or(RECALL_DEFAULT_LIMIT);

    print_tool_call("recall", query, color);

    if query.is_empty() {
        return "error: recall requires `query` — plain keywords describing what to find"
            .to_string();
    }
    // Pre-flight the 17.3 sanitizer so an all-syntax query renders as
    // coaching the model can act on, not an error that ends the turn.
    if crate::store::sanitize_fts5_query(query).is_err() {
        let out = format!(
            "no searchable terms in {query:?} — every term was search syntax or \
             punctuation; try plain keywords (e.g. 'tokio panic retry')"
        );
        print_tool_output(&out, tool_output_lines, color);
        return out;
    }

    let hits = match source.search(query, limit) {
        Ok(hits) => hits,
        Err(e) => return format!("error: {e}"),
    };
    if hits.is_empty() {
        // #714: recall structurally excludes the CURRENT conversation, so on a
        // freshly-resumed thin context this dead-ends on the one conversation
        // that matters. Append the redirect off the dead-end. (Must keep the
        // "no matches in past conversations" prefix — `classify_phantom_reach`
        // keys the #717 telemetry on it.)
        let out = format!(
            "no matches in past conversations for {query:?} — try different keywords. \
             To recover THIS conversation's own earlier work, call resume_context."
        );
        print_tool_output(&out, tool_output_lines, color);
        return out;
    }

    let mut out = format!(
        "{} match(es) in past conversations (best first):",
        hits.len()
    );
    for hit in &hits {
        let title = hit.title.trim();
        let title = if title.is_empty() {
            "(untitled)"
        } else {
            title
        };
        out.push_str(&format!(
            "\n{}  {}  ·  seq {}\n    {}",
            short_id(&hit.conversation_id),
            title,
            hit.seq,
            readable_snippet(&hit.snippet),
        ));
    }
    print_tool_output(&out, tool_output_lines, color);
    out
}

/// First 12 characters of a conversation id — the 17.4 convention: enough
/// `{unix_nanos}` digits for 10ms granularity, and `resolve_id` accepts any
/// unique prefix, so the short id pastes into `/conversation restore`.
fn short_id(id: &str) -> &str {
    id.get(..12).unwrap_or(id)
}

/// Make a raw FTS5 snippet read naturally to a model: collapse internal
/// whitespace (turn text is multi-line) and rewrite the store's `>>>`/`<<<`
/// match markers to `«`/`»` — the same convention 17.4's `/recall` renders
/// for humans, so the model and the user see identical highlights.
fn readable_snippet(snippet: &str) -> String {
    snippet
        .split_whitespace()
        .collect::<Vec<_>>()
        .join(" ")
        .replace(">>>", "«")
        .replace("<<<", "»")
}

// ---------------------------------------------------------------------------
// Tests
// ---------------------------------------------------------------------------

#[cfg(test)]
pub(crate) mod tests {
    use super::*;
    use std::sync::Mutex;

    /// Scriptable mock: records every `(query, limit)` call, serves canned
    /// hits (truncated to `limit`), or a canned error. Shared with the
    /// dispatch tests in `agentic::tools`.
    #[derive(Default)]
    pub(crate) struct MockSource {
        pub calls: Mutex<Vec<(String, usize)>>,
        pub hits: Vec<SearchHit>,
        pub fail_with: Option<String>,
    }

    impl RecallSource for MockSource {
        fn search(&self, query: &str, limit: usize) -> anyhow::Result<Vec<SearchHit>> {
            self.calls.lock().unwrap().push((query.to_string(), limit));
            match &self.fail_with {
                Some(e) => Err(anyhow::anyhow!("{e}")),
                None => Ok(self.hits.iter().take(limit).cloned().collect()),
            }
        }
    }

    pub(crate) fn hit(id: &str, title: &str, seq: i64, snippet: &str) -> SearchHit {
        SearchHit {
            conversation_id: id.to_string(),
            title: title.to_string(),
            seq,
            snippet: snippet.to_string(),
            rank: -1.0,
        }
    }

    // -- schema text: the model-facing coaching ------------------------------

    #[test]
    fn schema_says_past_conversations_and_excludes_current() {
        let def = recall_tool_definition();
        let desc = def["function"]["description"].as_str().unwrap();
        assert!(desc.contains("Search PAST conversations in this workspace"));
        assert!(
            desc.contains("The current conversation is never searched"),
            "got: {desc}"
        );
        assert!(desc.contains("already in your context"), "got: {desc}");
    }

    #[test]
    fn schema_coaches_when_to_use_and_when_not() {
        let def = recall_tool_definition();
        let desc = def["function"]["description"].as_str().unwrap();
        assert!(desc.contains("'like we did before'"), "got: {desc}");
        assert!(desc.contains("'where did we leave off'"), "got: {desc}");
        assert!(
            desc.contains("Do NOT use it for information already in this conversation"),
            "got: {desc}"
        );
    }

    #[test]
    fn schema_coaches_plain_keywords_over_fts_syntax() {
        let def = recall_tool_definition();
        let desc = def["function"]["description"].as_str().unwrap();
        assert!(desc.contains("plain keywords"), "got: {desc}");
        assert!(
            desc.contains("not boolean operators or quotes"),
            "got: {desc}"
        );
    }

    #[test]
    fn schema_shape_query_required_limit_optional() {
        let def = recall_tool_definition();
        assert_eq!(def["function"]["name"], "recall");
        let required: Vec<&str> = def["function"]["parameters"]["required"]
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|v| v.as_str())
            .collect();
        assert_eq!(required, vec!["query"]);
        let props = &def["function"]["parameters"]["properties"];
        assert!(props["query"].is_object());
        assert_eq!(props["limit"]["type"], "integer");
    }

    // -- dispatch ------------------------------------------------------------

    #[test]
    fn happy_path_formats_short_id_title_seq_and_markers() {
        let source = MockSource {
            hits: vec![
                hit(
                    "1748563200123-aaaa-bbbb",
                    "fixing the sanitizer",
                    4,
                    "…the >>>sanitizer<<< drops dangling\noperators…",
                ),
                hit("1748563200456-cccc-dddd", "  ", 2, ">>>sanitizer<<< port"),
            ],
            ..Default::default()
        };
        let out = execute_recall(
            &serde_json::json!({"query": "sanitizer"}),
            &source,
            false,
            20,
        );
        assert!(
            out.starts_with("2 match(es) in past conversations (best first):"),
            "got: {out}"
        );
        // Short id (12 chars), title, seq — one header per hit.
        assert!(
            out.contains("174856320012  fixing the sanitizer  ·  seq 4"),
            "got: {out}"
        );
        // Markers converted per the 17.4 convention; whitespace flattened.
        assert!(
            out.contains("«sanitizer» drops dangling operators"),
            "got: {out}"
        );
        assert!(!out.contains(">>>"), "raw FTS5 markers leaked: {out}");
        // Empty title falls back, hit still rendered.
        assert!(out.contains("(untitled)  ·  seq 2"), "got: {out}");
        // Default limit reached the source.
        assert_eq!(
            *source.calls.lock().unwrap(),
            vec![("sanitizer".to_string(), 5)]
        );
    }

    #[test]
    fn limit_is_defaulted_and_clamped() {
        let source = MockSource::default();
        // Absent → default 5 (asserted above too); oversized → max 10; zero → 1.
        execute_recall(
            &serde_json::json!({"query": "x", "limit": 25}),
            &source,
            false,
            20,
        );
        execute_recall(
            &serde_json::json!({"query": "x", "limit": 0}),
            &source,
            false,
            20,
        );
        execute_recall(&serde_json::json!({"query": "x"}), &source, false, 20);
        let limits: Vec<usize> = source.calls.lock().unwrap().iter().map(|c| c.1).collect();
        assert_eq!(limits, vec![10, 1, 5]);
    }

    #[test]
    fn sanitizer_rejected_query_is_coaching_not_error() {
        let source = MockSource::default();
        let out = execute_recall(&serde_json::json!({"query": "AND (*)"}), &source, false, 20);
        assert!(out.contains("no searchable terms"), "got: {out}");
        assert!(out.contains("plain keywords"), "got: {out}");
        assert!(!out.starts_with("error:"), "must not abort-shape: {out}");
        assert!(
            source.calls.lock().unwrap().is_empty(),
            "rejected query must never reach the backend"
        );
    }

    #[test]
    fn zero_hits_says_no_matches_never_empty() {
        let source = MockSource::default();
        let out = execute_recall(
            &serde_json::json!({"query": "quetzalcoatl"}),
            &source,
            false,
            20,
        );
        assert!(out.contains("no matches"), "got: {out}");
        assert!(!out.is_empty());
        // The prefix `classify_phantom_reach` keys on is preserved (#717).
        assert!(
            out.starts_with("no matches in past conversations"),
            "got: {out}"
        );
        // #714: the dead-end now redirects to the self-recovery tool.
        assert!(out.contains("resume_context"), "got: {out}");
    }

    #[test]
    fn this_conversation_recent_default_impl_is_empty() {
        // #714: the trait's default impl returns empty so existing mocks /
        // headless callers (no conversation store) compile + behave unchanged.
        let source = MockSource {
            hits: vec![hit("conv-a", "t", 1, "snip")],
            ..Default::default()
        };
        assert!(
            source.this_conversation_recent(5).unwrap().is_empty(),
            "default impl ignores `hits` and returns empty"
        );
    }

    #[test]
    fn recent_turn_snippet_flattens_and_caps() {
        // #714: multi-line turn text collapses to one line per field.
        let s = recent_turn_snippet("fix\nthe   bug", "done\nshipping it");
        assert_eq!(s, "you: fix the bug\n    me: done shipping it");
        // Over-long fields truncate with the […] marker.
        let long = recent_turn_snippet(&"x".repeat(RESUME_TURN_FIELD_CAP + 50), "ok");
        assert!(long.contains("[…]"), "got: {long}");
        assert!(long.contains("me: ok"), "got: {long}");
    }

    #[test]
    fn missing_query_is_a_clear_error() {
        let source = MockSource::default();
        let out = execute_recall(&serde_json::json!({}), &source, false, 20);
        assert!(out.contains("requires `query`"), "got: {out}");
        assert!(source.calls.lock().unwrap().is_empty());
    }

    #[test]
    fn backend_failure_surfaces_as_tool_error_text() {
        let source = MockSource {
            fail_with: Some("database is on fire".to_string()),
            ..Default::default()
        };
        let out = execute_recall(
            &serde_json::json!({"query": "anything"}),
            &source,
            false,
            20,
        );
        assert_eq!(out, "error: database is on fire");
    }
}