zeph-core 0.17.0

Core agent loop, configuration, context builder, metrics, and vault for Zeph
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
// SPDX-FileCopyrightText: 2026 Andrei G <bug-ops>
// SPDX-License-Identifier: MIT OR Apache-2.0

use std::collections::VecDeque;
use std::time::Duration;

use zeph_tools::{OverflowConfig, ResultCacheConfig, TafcConfig, ToolResultCache};

use super::DOOM_LOOP_WINDOW;

pub(crate) struct ToolOrchestrator {
    pub(super) doom_loop_history: Vec<u64>,
    pub(super) max_iterations: usize,
    pub(super) summarize_tool_output_enabled: bool,
    pub(super) overflow_config: OverflowConfig,
    /// Sliding window of recent (`tool_name`, `args_hash`) pairs for repeat-detection.
    /// Only LLM-initiated calls are recorded here — retry re-executions are excluded.
    /// Window capacity = 2 * `repeat_threshold`.
    pub(super) recent_tool_calls: VecDeque<(String, u64)>,
    /// Number of identical (`tool_name`, `args_hash`) appearances in the window required
    /// to trigger repeat-detection abort. 0 = disabled.
    pub(super) repeat_threshold: usize,
    /// Max retries for transient errors per tool call. 0 = disabled.
    pub(super) max_tool_retries: usize,
    /// Maximum wall-clock time (seconds) to spend on retries for a single tool call.
    /// 0 = no wall-clock budget (only `max_tool_retries` applies).
    pub(super) max_retry_duration_secs: u64,
    /// Pre-execution verifiers run before every native tool call (`TrustBench` pattern,
    /// issue #1630). Stored here rather than on `SecurityState` because they are tool-layer
    /// concerns: they inspect tool arguments at dispatch time, consistent with
    /// repeat-detection, rate-limiting, and overflow controls which also live here.
    pub(super) pre_execution_verifiers: Vec<Box<dyn zeph_tools::PreExecutionVerifier>>,
    /// Think-Augmented Function Calling configuration.
    pub(crate) tafc: TafcConfig,
    /// Session-scoped cache for tool results. Persists across tool rounds within a session;
    /// reset only on `/clear`. Unlike repeat-detection and doom-loop state (which reset per
    /// round), the cache is intentionally long-lived — its value comes from reuse across turns.
    pub(super) result_cache: ToolResultCache,
}

/// Truncate a tool name to at most 256 bytes, respecting UTF-8 char boundaries.
///
/// Used by both `push_tool_call` and `is_repeat` to ensure stored and queried
/// names always match when the original name exceeds the limit.
fn truncate_tool_name(name: &str) -> &str {
    const MAX_TOOL_NAME_BYTES: usize = 256;
    if name.len() <= MAX_TOOL_NAME_BYTES {
        return name;
    }
    let mut idx = MAX_TOOL_NAME_BYTES;
    while !name.is_char_boundary(idx) {
        idx -= 1;
    }
    &name[..idx]
}

impl ToolOrchestrator {
    #[must_use]
    pub(crate) fn new() -> Self {
        let repeat_threshold = 2_usize;
        Self {
            doom_loop_history: Vec::new(),
            max_iterations: 10,
            summarize_tool_output_enabled: false,
            overflow_config: OverflowConfig::default(),
            recent_tool_calls: VecDeque::with_capacity(2 * repeat_threshold),
            repeat_threshold,
            max_tool_retries: 2,
            max_retry_duration_secs: 30,
            pre_execution_verifiers: Vec::new(),
            tafc: TafcConfig::default(),
            result_cache: ToolResultCache::new(true, Some(Duration::from_secs(300))),
        }
    }

    /// Initialize the result cache from config.
    pub(crate) fn set_cache_config(&mut self, config: &ResultCacheConfig) {
        let ttl = if config.ttl_secs == 0 {
            None // ttl_secs = 0 → never expire
        } else {
            Some(Duration::from_secs(config.ttl_secs))
        };
        self.result_cache = ToolResultCache::new(config.enabled, ttl);
    }

    /// Clear the result cache. Called on `/clear`.
    pub(crate) fn clear_cache(&mut self) {
        self.result_cache.clear();
    }

    /// Returns a formatted cache stats string for `/cache-stats` command.
    pub(crate) fn cache_stats(&self) -> String {
        let cache = &self.result_cache;
        let status = if cache.is_enabled() {
            "enabled"
        } else {
            "disabled"
        };
        let hits = cache.hits();
        let misses = cache.misses();
        let total = hits + misses;
        #[allow(clippy::cast_precision_loss)]
        let hit_rate = if total > 0 {
            format!("{:.1}%", (hits as f64 / total as f64) * 100.0)
        } else {
            "n/a".to_owned()
        };
        let ttl_display = if cache.ttl_secs() == 0 {
            "never".to_owned()
        } else {
            format!("{}s", cache.ttl_secs())
        };
        format!(
            "Tool result cache: {status}\nEntries: {}, Hits: {hits}, Misses: {misses}, Hit rate: {hit_rate}\nTTL: {ttl_display}",
            cache.len(),
        )
    }

    pub(super) fn push_doom_hash(&mut self, hash: u64) {
        self.doom_loop_history.push(hash);
    }

    pub(super) fn clear_doom_history(&mut self) {
        self.doom_loop_history.clear();
    }

    /// Reset the repeat-detection sliding window between user turns.
    pub(super) fn clear_recent_tool_calls(&mut self) {
        self.recent_tool_calls.clear();
    }

    /// Returns `true` if the last `DOOM_LOOP_WINDOW` hashes are identical.
    pub(super) fn is_doom_loop(&self) -> bool {
        if self.doom_loop_history.len() < DOOM_LOOP_WINDOW {
            return false;
        }
        let recent = &self.doom_loop_history[self.doom_loop_history.len() - DOOM_LOOP_WINDOW..];
        recent.windows(2).all(|w| w[0] == w[1])
    }

    /// Record a tool call (LLM-initiated only — not retry re-executions).
    ///
    /// Maintains a sliding window of size `2 * repeat_threshold`.
    /// Tool names are truncated to 256 bytes to prevent unbounded memory growth
    /// from adversarially long names.
    pub(super) fn push_tool_call(&mut self, name: &str, args_hash: u64) {
        if self.repeat_threshold == 0 {
            return;
        }
        let window = 2 * self.repeat_threshold;
        if self.recent_tool_calls.len() >= window {
            self.recent_tool_calls.pop_front();
        }
        self.recent_tool_calls
            .push_back((truncate_tool_name(name).to_owned(), args_hash)); // lgtm[rust/cleartext-logging]
    }

    /// Returns `true` if the same `(name, args_hash)` pair appears `>= repeat_threshold`
    /// times in the current window.
    ///
    /// Applies the same 256-byte name truncation as `push_tool_call` so that long names
    /// are correctly matched against stored (truncated) entries.
    pub(super) fn is_repeat(&self, name: &str, args_hash: u64) -> bool {
        if self.repeat_threshold == 0 {
            return false;
        }
        let name = truncate_tool_name(name);
        let count = self
            .recent_tool_calls
            .iter()
            .filter(|(n, h)| n == name && *h == args_hash)
            .count();
        count >= self.repeat_threshold
    }
}

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

    #[test]
    fn new_defaults() {
        let o = ToolOrchestrator::new();
        assert!(o.doom_loop_history.is_empty());
        assert_eq!(o.max_iterations, 10);
        assert!(!o.summarize_tool_output_enabled);
        assert!(o.recent_tool_calls.is_empty());
        assert_eq!(o.repeat_threshold, 2);
        assert_eq!(o.max_tool_retries, 2);
        assert_eq!(o.max_retry_duration_secs, 30);
    }

    #[test]
    fn is_doom_loop_insufficient_history() {
        let mut o = ToolOrchestrator::new();
        o.push_doom_hash(42);
        o.push_doom_hash(42);
        assert!(!o.is_doom_loop());
    }

    #[test]
    fn is_doom_loop_identical_hashes() {
        let mut o = ToolOrchestrator::new();
        o.push_doom_hash(7);
        o.push_doom_hash(7);
        o.push_doom_hash(7);
        assert!(o.is_doom_loop());
    }

    #[test]
    fn is_doom_loop_mixed_hashes() {
        let mut o = ToolOrchestrator::new();
        o.push_doom_hash(1);
        o.push_doom_hash(2);
        o.push_doom_hash(3);
        assert!(!o.is_doom_loop());
    }

    #[test]
    fn is_doom_loop_only_recent_window_matters() {
        let mut o = ToolOrchestrator::new();
        o.push_doom_hash(1);
        o.push_doom_hash(2);
        o.push_doom_hash(9);
        o.push_doom_hash(9);
        o.push_doom_hash(9);
        assert!(o.is_doom_loop());
    }

    #[test]
    fn clear_doom_history_resets() {
        let mut o = ToolOrchestrator::new();
        o.push_doom_hash(5);
        o.push_doom_hash(5);
        o.push_doom_hash(5);
        assert!(o.is_doom_loop());
        o.clear_doom_history();
        assert!(!o.is_doom_loop());
        assert!(o.doom_loop_history.is_empty());
    }

    // Repeat-detection tests

    #[test]
    fn repeat_detection_no_repeat_before_threshold() {
        let mut o = ToolOrchestrator::new();
        o.push_tool_call("bash", 42);
        // Only 1 occurrence, threshold is 2 — not a repeat yet
        assert!(!o.is_repeat("bash", 42));
    }

    #[test]
    fn repeat_detection_triggers_at_threshold() {
        let mut o = ToolOrchestrator::new();
        o.push_tool_call("bash", 42);
        o.push_tool_call("bash", 42);
        // 2 occurrences >= threshold 2 → repeat
        assert!(o.is_repeat("bash", 42));
    }

    #[test]
    fn repeat_detection_different_args_no_repeat() {
        let mut o = ToolOrchestrator::new();
        o.push_tool_call("bash", 1);
        o.push_tool_call("bash", 2);
        assert!(!o.is_repeat("bash", 1));
        assert!(!o.is_repeat("bash", 2));
    }

    #[test]
    fn repeat_detection_different_tool_no_repeat() {
        let mut o = ToolOrchestrator::new();
        o.push_tool_call("bash", 42);
        o.push_tool_call("read", 42);
        assert!(!o.is_repeat("bash", 42));
        assert!(!o.is_repeat("read", 42));
    }

    #[test]
    fn repeat_detection_window_evicts_old_entries() {
        let mut o = ToolOrchestrator::new();
        // Window size = 2 * threshold = 4
        o.push_tool_call("bash", 42);
        o.push_tool_call("read", 1);
        o.push_tool_call("read", 2);
        o.push_tool_call("read", 3);
        // Now push another entry — "bash:42" should be evicted from front
        o.push_tool_call("read", 4);
        // "bash:42" was in the window once, now evicted → not a repeat
        assert!(!o.is_repeat("bash", 42));
    }

    #[test]
    fn repeat_detection_disabled_when_threshold_zero() {
        let mut o = ToolOrchestrator::new();
        o.repeat_threshold = 0;
        o.push_tool_call("bash", 42);
        o.push_tool_call("bash", 42);
        o.push_tool_call("bash", 42);
        // Threshold 0 means disabled
        assert!(!o.is_repeat("bash", 42));
    }

    // ── SEC-003: tool name truncation ─────────────────────────────────────────

    #[test]
    fn push_tool_call_long_name_truncated_to_256_bytes() {
        let mut o = ToolOrchestrator::new();
        // Name well above 256 bytes
        let long_name = "a".repeat(512);
        o.push_tool_call(&long_name, 99);
        let stored = &o.recent_tool_calls[0].0;
        assert_eq!(stored.len(), 256, "stored name must be exactly 256 bytes");
        assert!(
            stored.is_char_boundary(stored.len()),
            "truncation must land on char boundary"
        );
    }

    #[test]
    fn push_tool_call_unicode_name_truncated_at_char_boundary() {
        let mut o = ToolOrchestrator::new();
        // Each '日' is 3 bytes. 256 / 3 = 85 full chars = 255 bytes.
        // Appending one more gives 258 bytes total — must truncate to 255.
        let base: String = "".repeat(85); // 255 bytes
        let long_name = format!("{base}"); // 258 bytes — crosses 256-byte boundary
        o.push_tool_call(&long_name, 1);
        let stored = &o.recent_tool_calls[0].0;
        assert!(stored.len() <= 256, "stored name must not exceed 256 bytes");
        assert!(
            stored.is_char_boundary(stored.len()),
            "must be valid UTF-8 boundary"
        );
    }

    #[test]
    fn push_tool_call_short_name_not_truncated() {
        let mut o = ToolOrchestrator::new();
        let short_name = "shell";
        o.push_tool_call(short_name, 7);
        assert_eq!(o.recent_tool_calls[0].0, short_name);
    }

    #[test]
    fn push_tool_call_300_byte_name_truncated_to_at_most_256() {
        // SEC-003: specifically test with 300-byte name as the boundary case
        let mut o = ToolOrchestrator::new();
        let name_300 = "x".repeat(300);
        o.push_tool_call(&name_300, 42);
        let stored = &o.recent_tool_calls[0].0;
        assert!(
            stored.len() <= 256,
            "300-byte name must be stored as ≤256 bytes, got {}",
            stored.len()
        );
        assert!(stored.is_char_boundary(stored.len()));
    }

    // ── SEC-004: retry budget field and logic ─────────────────────────────────

    #[test]
    fn max_retry_duration_secs_default_is_30() {
        // SEC-004: verify the budget field is set at construction time
        let o = ToolOrchestrator::new();
        assert_eq!(o.max_retry_duration_secs, 30);
    }

    #[test]
    fn retry_budget_condition_zero_disables_check() {
        // SEC-004: when max_retry_duration_secs == 0, the budget check must be skipped.
        // This mirrors the `if max_retry_duration_secs > 0` guard in native.rs.
        let budget_secs: u64 = 0;
        // Simulate that 60 seconds have elapsed — budget check is disabled.
        let elapsed_secs: u64 = 60;
        let budget_exceeded = budget_secs > 0 && elapsed_secs >= budget_secs;
        assert!(
            !budget_exceeded,
            "budget=0 must disable the wall-clock check"
        );
    }

    #[test]
    fn retry_budget_condition_exceeded_triggers_break() {
        // SEC-004: simulate elapsed > budget with a real Instant to confirm the
        // condition that native.rs evaluates before breaking the retry loop.
        let budget_secs: u64 = 1;
        // Subtract 2 seconds to ensure elapsed >= budget without sleeping.
        let retry_start = std::time::Instant::now()
            .checked_sub(std::time::Duration::from_secs(2))
            .unwrap();
        let elapsed_secs = retry_start.elapsed().as_secs();
        let budget_exceeded = budget_secs > 0 && elapsed_secs >= budget_secs;
        assert!(
            budget_exceeded,
            "elapsed {elapsed_secs}s should exceed budget {budget_secs}s"
        );
    }

    // ── cache_stats() display ─────────────────────────────────────────────────

    #[test]
    fn cache_stats_disabled_shows_disabled_status() {
        let mut o = ToolOrchestrator::new();
        o.set_cache_config(&ResultCacheConfig {
            enabled: false,
            ttl_secs: 300,
        });
        let stats = o.cache_stats();
        assert!(
            stats.contains("disabled"),
            "expected 'disabled' in: {stats}"
        );
    }

    #[test]
    fn cache_stats_no_calls_shows_na_hit_rate() {
        let o = ToolOrchestrator::new();
        let stats = o.cache_stats();
        assert!(
            stats.contains("n/a"),
            "expected 'n/a' hit rate when total=0, got: {stats}"
        );
    }

    #[test]
    fn cache_stats_ttl_zero_shows_never() {
        let mut o = ToolOrchestrator::new();
        o.set_cache_config(&ResultCacheConfig {
            enabled: true,
            ttl_secs: 0,
        });
        let stats = o.cache_stats();
        assert!(
            stats.contains("never"),
            "expected 'never' TTL display for ttl_secs=0, got: {stats}"
        );
    }

    #[test]
    fn cache_stats_hit_rate_percentage() {
        use zeph_tools::CacheKey;
        let mut o = ToolOrchestrator::new();
        // Directly manipulate the cache to simulate 1 hit and 1 miss.
        // put() one entry, get() it (hit), then get() a missing key (miss).
        let output = zeph_tools::ToolOutput {
            tool_name: "read".to_owned(),
            summary: "contents".to_owned(),
            blocks_executed: 1,
            filter_stats: None,
            diff: None,
            streamed: false,
            terminal_id: None,
            locations: None,
            raw_response: None,
        };
        o.result_cache.put(CacheKey::new("read", 1), output);
        o.result_cache.get(&CacheKey::new("read", 1)); // hit
        o.result_cache.get(&CacheKey::new("read", 99)); // miss
        let stats = o.cache_stats();
        assert!(
            stats.contains("50.0%"),
            "expected 50.0% hit rate (1 hit / 2 total), got: {stats}"
        );
        assert!(
            stats.contains("Hits: 1"),
            "expected 'Hits: 1', got: {stats}"
        );
        assert!(
            stats.contains("Misses: 1"),
            "expected 'Misses: 1', got: {stats}"
        );
    }

    #[test]
    fn set_cache_config_ttl_mapping() {
        let mut o = ToolOrchestrator::new();
        // ttl_secs = 0 → None (never expire) → ttl_secs() returns 0
        o.set_cache_config(&ResultCacheConfig {
            enabled: true,
            ttl_secs: 0,
        });
        assert_eq!(o.result_cache.ttl_secs(), 0);

        // ttl_secs = 60 → Some(60s) → ttl_secs() returns 60
        o.set_cache_config(&ResultCacheConfig {
            enabled: true,
            ttl_secs: 60,
        });
        assert_eq!(o.result_cache.ttl_secs(), 60);
    }
}