klieo-core 3.3.0

Core traits + runtime for the klieo agent framework.
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
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
//! Memory traits — short-term, long-term, episodic.

use crate::error::MemoryError;
use crate::ids::{FactId, RunId, ThreadId};
use crate::llm::Message;
use async_trait::async_trait;
use chrono::{DateTime, Utc};
use serde::{Deserialize, Serialize};

/// Outcome of a tool invocation as recorded in the episodic event stream.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(tag = "outcome", rename_all = "snake_case")]
pub enum ToolResult {
    /// Tool returned a successful JSON result.
    Ok {
        /// Result payload.
        value: serde_json::Value,
    },
    /// Tool returned an error message.
    Err {
        /// Error string.
        message: String,
    },
}

impl ToolResult {
    /// Build a successful result. Convenience for the
    /// `ToolResult::Ok { value }` struct-variant ceremony.
    pub fn ok(value: serde_json::Value) -> Self {
        Self::Ok { value }
    }

    /// Build an error result. Convenience for the
    /// `ToolResult::Err { message }` struct-variant ceremony.
    pub fn err(message: impl Into<String>) -> Self {
        Self::Err {
            message: message.into(),
        }
    }
}

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

    #[test]
    fn ok_factory_builds_ok_variant() {
        let v = serde_json::json!({"hit": true});
        let r = ToolResult::ok(v.clone());
        match r {
            ToolResult::Ok { value } => assert_eq!(value, v),
            _ => panic!("expected ToolResult::Ok variant"),
        }
    }

    #[test]
    fn err_factory_builds_err_variant() {
        let r = ToolResult::err("boom");
        match r {
            ToolResult::Err { message } => assert_eq!(message, "boom"),
            _ => panic!("expected ToolResult::Err variant"),
        }
    }
}

/// Conversation buffer scoped to a single thread.
///
/// ```
/// # tokio_test::block_on(async {
/// use klieo_core::test_utils::InMemoryShortTerm;
/// use klieo_core::{ShortTermMemory, Message, Role, ThreadId};
///
/// let m = InMemoryShortTerm::default();
/// let thread = ThreadId::new("t1");
/// m.append(thread.clone(), Message {
///     role: Role::User, content: "hi".into(),
///     tool_calls: vec![], tool_call_id: None,
/// }).await.unwrap();
/// let loaded = m.load(thread, 1024).await.unwrap();
/// assert_eq!(loaded.len(), 1);
/// # });
/// ```
#[async_trait]
pub trait ShortTermMemory: Send + Sync {
    /// Append a message to the thread's history.
    async fn append(&self, thread: ThreadId, msg: Message) -> Result<(), MemoryError>;

    /// Append a batch of messages, oldest first. The default appends each in
    /// turn; backends with bulk-insert support should override to collapse the
    /// N writes into one round-trip (the resume path replays a full history).
    async fn append_batch(
        &self,
        thread: ThreadId,
        messages: Vec<Message>,
    ) -> Result<(), MemoryError> {
        for msg in messages {
            self.append(thread.clone(), msg).await?;
        }
        Ok(())
    }

    /// Load up to `max_tokens` of the most-recent messages, oldest first.
    /// Implementations approximate token counts (provider-specific).
    async fn load(&self, thread: ThreadId, max_tokens: usize) -> Result<Vec<Message>, MemoryError>;

    /// Drop all messages for `thread`.
    async fn clear(&self, thread: ThreadId) -> Result<(), MemoryError>;
}

/// Namespacing for long-term memory facts.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub enum Scope {
    /// Workspace-scoped (multi-agent shared).
    Workspace(String),
    /// Per-agent scoped.
    Agent(String),
    /// Process-global (use sparingly).
    Global,
}

/// One stored fact in long-term memory.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[non_exhaustive]
pub struct Fact {
    /// Plain-text body, embedded for retrieval.
    pub text: String,
    /// Caller-supplied metadata, opaque to the store.
    #[serde(default)]
    pub metadata: serde_json::Value,
}

impl Fact {
    /// Prefer this over struct literals outside `klieo-core` (`#[non_exhaustive]`);
    /// metadata defaults to JSON null — attach it with [`Fact::with_metadata`].
    pub fn new(text: impl Into<String>) -> Self {
        Self {
            text: text.into(),
            metadata: serde_json::Value::Null,
        }
    }

    /// Override the default JSON-null metadata with an opaque caller value.
    pub fn with_metadata(mut self, metadata: serde_json::Value) -> Self {
        self.metadata = metadata;
        self
    }
}

/// Long-term semantic memory.
///
/// ```
/// # tokio_test::block_on(async {
/// use klieo_core::test_utils::InMemoryLongTerm;
/// use klieo_core::{Fact, LongTermMemory, Scope};
///
/// let m = InMemoryLongTerm::default();
/// let scope = Scope::Workspace("ws".into());
/// m.remember(scope.clone(), Fact::new("the sky is blue")).await.unwrap();
/// let hits = m.recall(scope, "sky", 1).await.unwrap();
/// assert_eq!(hits.len(), 1);
/// # });
/// ```
#[async_trait]
pub trait LongTermMemory: Send + Sync {
    /// Store a fact under `scope`. Returns a stable id.
    async fn remember(&self, scope: Scope, fact: Fact) -> Result<FactId, MemoryError>;

    /// Top-`k` semantic recall under `scope` for the supplied query.
    async fn recall(&self, scope: Scope, query: &str, k: usize) -> Result<Vec<Fact>, MemoryError>;

    /// Remove a stored fact.
    async fn forget(&self, id: FactId) -> Result<(), MemoryError>;
}

/// One event in the episodic event stream of a single agent run.
///
/// Marked `#[non_exhaustive]` so additive variants (e.g.
/// [`Self::SummaryCheckpoint`]) can be introduced without forcing a
/// SemVer-major bump. Match arms in downstream crates must include a
/// fallback `_ => …`.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[non_exhaustive]
pub enum Episode {
    /// Run started.
    Started {
        /// Agent name.
        agent: String,
    },
    /// LLM call completed.
    ///
    /// `provider`/`model`/`prompt_tokens`/`completion_tokens` are
    /// `Option` so legacy emit sites can leave them `None`; downstream
    /// projectors fall back to the `LlmIo` sidecar in `klieo-runlog`
    /// when the structured fields are absent.
    LlmCall {
        /// Total tokens reported by the provider (prompt + completion).
        tokens: u32,
        /// Wall-clock latency in milliseconds.
        latency_ms: u32,
        /// Provider identifier — e.g. `"ollama"`, `"openai"`,
        /// `"anthropic"`, `"gemini"`. `None` for older records or
        /// providers that don't expose a stable name.
        ///
        /// `#[serde(default)]` so legacy episodes serialised under
        /// klieo 0.6.x deserialise as `None`.
        #[serde(default)]
        provider: Option<String>,
        /// Model identifier — e.g. `"qwen2.5:14b"`, `"gpt-4o-mini"`.
        #[serde(default)]
        model: Option<String>,
        /// Prompt-side token count when the provider splits the
        /// breakdown; `None` falls back to `tokens` for total-only
        /// reports.
        #[serde(default)]
        prompt_tokens: Option<u32>,
        /// Completion-side token count when the provider splits the
        /// breakdown.
        #[serde(default)]
        completion_tokens: Option<u32>,
    },
    /// Tool call completed.
    ToolCall {
        /// Tool name.
        name: String,
        /// JSON arguments.
        args: serde_json::Value,
        /// Tool outcome.
        result: ToolResult,
    },
    /// Agent published a bus message.
    BusPublish {
        /// Subject.
        subject: String,
    },
    /// Agent received a bus message.
    BusReceive {
        /// Subject.
        subject: String,
    },
    /// Causal link: this run received a bus message caused by another run's
    /// publish. Recorded by `AgentContext::record_received` when the publisher
    /// threaded its run id via the `klieo-causation-run` bus header. Additive —
    /// absent on records from publishers that did not thread the causation header.
    BusCausalLink {
        /// Subject the causal handoff occurred on.
        subject: String,
        /// Run id of the publisher that caused this receive.
        caused_by_run: String,
    },
    /// Run completed successfully.
    Completed,
    /// Run failed.
    Failed {
        /// Error message.
        error: String,
    },
    /// Summarizer checkpoint completed.
    ///
    /// Emitted by [`crate::summarize::summarize_history`] in lieu of
    /// [`Self::LlmCall`] so the audit trail can distinguish summarizer
    /// overhead from substantive agent reasoning. Downstream
    /// observability (e.g. `klieo-runlog`) typically projects this as
    /// a separate step kind so cost / latency attribution stays
    /// faithful.
    SummaryCheckpoint {
        /// Number of older messages folded into the summary call.
        input_message_count: u32,
        /// Length of the resulting summary, in Unicode scalar values.
        summary_chars: u32,
        /// Wall-clock latency of the summarizer call.
        latency_ms: u32,
        /// Total tokens reported by the summarizer LLM (prompt +
        /// completion).
        tokens: u32,
    },
    /// Operational-layer event (klieo-ops). Body is an opaque
    /// `serde_json::Value` to keep klieo-core free of an ops dependency.
    /// klieo-ops provides typed serde conversion helpers via `OpsEvent`.
    Ops(serde_json::Value),
    /// Non-PII tenant attribution stamped at run entry when an external
    /// caller drives the run.
    ///
    /// `tenant_label` is a derived identifier (e.g. truncated SHA-256
    /// of the caller's `sub`) — never the raw principal, which lives
    /// only in server-side tracing/authorization. Emitted at most once
    /// per run, adjacent to [`Self::Started`], so the audit trail can
    /// attribute each run to its driving tenant without admitting PII
    /// into agent memory or LLM-visible context.
    RunAttributed {
        /// Derived non-PII attribution label for the driving caller.
        tenant_label: String,
    },
    /// Cross-hop provenance origin stamped at run entry when an
    /// authenticated external caller supplies a parent-chain anchor.
    ///
    /// `parent_anchor` is the caller's own provenance chain-entry id (or
    /// its run's episodic-root hash) — recorded **verbatim** so the value
    /// equals the caller's identifier and downstream tooling can stitch
    /// klieo→klieo lineage across deployments. It is a **caller-asserted,
    /// unverified** claim (klieo does not own or validate the caller's
    /// chain); it is co-recorded with [`Self::RunAttributed`] so the
    /// claim is attributable to the authenticated principal that made it.
    /// Emitted at most once per run, adjacent to [`Self::Started`]; never
    /// admitted into agent memory or LLM-visible context.
    RunOrigin {
        /// Verbatim caller-supplied cross-hop provenance anchor.
        parent_anchor: String,
    },
    /// A graphRAG recall performed during the run. Recorded by the
    /// recall-recording wrapper so the run view can surface — and
    /// deep-link — retrieval calls. `query` is redacted + length-bounded
    /// at the recording boundary before it reaches this episode.
    MemoryRecall {
        /// Redacted, length-bounded recall query text.
        query: String,
        /// Requested top-k.
        #[serde(default)]
        k: u32,
        /// Fact ids the recall returned.
        #[serde(default)]
        returned_fact_ids: Vec<FactId>,
    },
}

impl Episode {
    /// Construct an [`Episode::LlmCall`] with the legacy two-field shape
    /// (`tokens` + `latency_ms`), leaving the 0.7-added
    /// `provider`/`model`/`prompt_tokens`/`completion_tokens` fields as
    /// `None`.
    ///
    /// Use this when the emit site does not have the enriched fields to
    /// hand — e.g. test fixtures, providers that only report total
    /// tokens, or call sites being migrated incrementally. For full
    /// 0.7 emit semantics, construct the struct variant directly.
    ///
    /// ```
    /// use klieo_core::Episode;
    /// let ep = Episode::llm_call(42, 17);
    /// match ep {
    ///     Episode::LlmCall { tokens, latency_ms, provider, .. } => {
    ///         assert_eq!(tokens, 42);
    ///         assert_eq!(latency_ms, 17);
    ///         assert!(provider.is_none());
    ///     }
    ///     _ => unreachable!(),
    /// }
    /// ```
    pub fn llm_call(tokens: u32, latency_ms: u32) -> Self {
        Episode::LlmCall {
            tokens,
            latency_ms,
            provider: None,
            model: None,
            prompt_tokens: None,
            completion_tokens: None,
        }
    }
}

/// Filter passed to `EpisodicMemory::list_runs`.
#[derive(Debug, Clone, Default)]
pub struct RunFilter {
    /// Filter by agent name (substring match).
    pub agent: Option<String>,
    /// Inclusive lower bound on `started_at`.
    pub since: Option<DateTime<Utc>>,
    /// Inclusive upper bound on `started_at`.
    pub until: Option<DateTime<Utc>>,
    /// Maximum rows returned.
    pub limit: Option<usize>,
}

/// Index-row summary returned by `list_runs`.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct RunSummary {
    /// Run id.
    pub run_id: RunId,
    /// Agent name.
    pub agent: String,
    /// First-event timestamp.
    pub started_at: DateTime<Utc>,
    /// Last-event timestamp, if completed/failed.
    pub finished_at: Option<DateTime<Utc>>,
    /// Number of episodes.
    pub episode_count: u32,
}

/// Append-only event log of agent runs.
///
/// ```
/// # tokio_test::block_on(async {
/// use klieo_core::test_utils::InMemoryEpisodic;
/// use klieo_core::{Episode, EpisodicMemory, RunId};
///
/// let m = InMemoryEpisodic::default();
/// let run = RunId::new();
/// m.record(run, Episode::Started { agent: "a".into() }).await.unwrap();
/// let events = m.replay(run).await.unwrap();
/// assert_eq!(events.len(), 1);
/// # });
/// ```
#[async_trait]
pub trait EpisodicMemory: Send + Sync {
    /// Record an episode for `run`.
    async fn record(&self, run: RunId, event: Episode) -> Result<(), MemoryError>;

    /// Replay all episodes for `run` in order.
    async fn replay(&self, run: RunId) -> Result<Vec<Episode>, MemoryError>;

    /// Replay episodes for many runs.
    ///
    /// Returns one entry per requested run, in the requested order. A run with
    /// no recorded episodes yields an empty `Vec` so the caller's node set
    /// stays complete.
    ///
    /// # Performance
    ///
    /// The default implementation issues one `replay` call per run (N+1).
    /// Stores with a batched read **must** override this method — the SQLite
    /// impl uses a single `WHERE run_id IN (…)` query.
    async fn replay_many(&self, runs: &[RunId]) -> Result<Vec<(RunId, Vec<Episode>)>, MemoryError> {
        let mut out = Vec::with_capacity(runs.len());
        for &run in runs {
            out.push((run, self.replay(run).await?));
        }
        Ok(out)
    }

    /// List run summaries matching `filter`.
    async fn list_runs(&self, filter: RunFilter) -> Result<Vec<RunSummary>, MemoryError>;
}

/// Resolved trio of memory handles ready to drop into an
/// [`crate::agent::AgentContext`] or an `App`.
///
/// Impl crates (`klieo-memory-sqlite`, `klieo-memory-neo4j`,
/// `klieo-memory-qdrant`) provide `From` conversions where they cover
/// the full trio. Crates that only carry a subset (Neo4j covers
/// short + episodic; Qdrant covers long) compose via the `App` builder's
/// per-trait setters rather than a direct `From`.
#[derive(Clone)]
pub struct MemoryHandles {
    /// Short-term conversation memory.
    pub short_term: std::sync::Arc<dyn ShortTermMemory>,
    /// Long-term semantic memory.
    pub long_term: std::sync::Arc<dyn LongTermMemory>,
    /// Episodic event log.
    pub episodic: std::sync::Arc<dyn EpisodicMemory>,
}

impl MemoryHandles {
    /// Build directly from three already-`Arc`-wrapped handles.
    /// Most callers go through an impl crate's `From` instead.
    pub fn new(
        short_term: std::sync::Arc<dyn ShortTermMemory>,
        long_term: std::sync::Arc<dyn LongTermMemory>,
        episodic: std::sync::Arc<dyn EpisodicMemory>,
    ) -> Self {
        Self {
            short_term,
            long_term,
            episodic,
        }
    }
}

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

    #[test]
    fn fact_new_defaults_metadata_null() {
        let f = Fact::new("alice likes tea");
        assert_eq!(f.text, "alice likes tea");
        assert_eq!(f.metadata, serde_json::Value::Null);
        let f2 = Fact::new("x").with_metadata(serde_json::json!({"k":"v"}));
        assert_eq!(f2.metadata, serde_json::json!({"k":"v"}));
    }
}

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

    #[allow(dead_code)]
    fn _assert_dyn_short(_: &dyn ShortTermMemory) {}
    #[allow(dead_code)]
    fn _assert_dyn_long(_: &dyn LongTermMemory) {}
    #[allow(dead_code)]
    fn _assert_dyn_episodic(_: &dyn EpisodicMemory) {}

    /// Maps each variant to its published snake_case wire discriminant.
    ///
    /// The exhaustive match is the drift guard: a new `Episode` variant
    /// fails to compile here until its discriminant is added, which is the
    /// signal to also publish a payload schema under `docs/schemas/runlog/`
    /// and a fixture in `tests/schema_drift.rs`. `#[non_exhaustive]` does not
    /// force a wildcard inside the defining crate, so this stays exhaustive.
    fn kind_discriminant(episode: &Episode) -> &'static str {
        match episode {
            Episode::Started { .. } => "started",
            Episode::LlmCall { .. } => "llm_call",
            Episode::ToolCall { .. } => "tool_call",
            Episode::BusPublish { .. } => "bus_publish",
            Episode::BusReceive { .. } => "bus_receive",
            Episode::BusCausalLink { .. } => "bus_causal_link",
            Episode::Completed => "completed",
            Episode::Failed { .. } => "failed",
            Episode::SummaryCheckpoint { .. } => "summary_checkpoint",
            Episode::Ops(_) => "ops",
            Episode::RunAttributed { .. } => "run_attributed",
            Episode::RunOrigin { .. } => "run_origin",
            Episode::MemoryRecall { .. } => "memory_recall",
        }
    }

    fn one_sample_per_variant() -> Vec<Episode> {
        vec![
            Episode::Started {
                agent: String::new(),
            },
            Episode::LlmCall {
                tokens: 0,
                latency_ms: 0,
                provider: None,
                model: None,
                prompt_tokens: None,
                completion_tokens: None,
            },
            Episode::ToolCall {
                name: String::new(),
                args: serde_json::Value::Null,
                result: ToolResult::Ok {
                    value: serde_json::Value::Null,
                },
            },
            Episode::BusPublish {
                subject: String::new(),
            },
            Episode::BusReceive {
                subject: String::new(),
            },
            Episode::BusCausalLink {
                subject: String::new(),
                caused_by_run: String::new(),
            },
            Episode::Completed,
            Episode::Failed {
                error: String::new(),
            },
            Episode::SummaryCheckpoint {
                input_message_count: 0,
                summary_chars: 0,
                latency_ms: 0,
                tokens: 0,
            },
            Episode::Ops(serde_json::Value::Null),
            Episode::RunAttributed {
                tenant_label: String::new(),
            },
            Episode::RunOrigin {
                parent_anchor: String::new(),
            },
            Episode::MemoryRecall {
                query: String::new(),
                k: 0,
                returned_fact_ids: Vec::new(),
            },
        ]
    }

    /// The set of Rust `Episode` discriminants must equal the `kind` enum
    /// published in the envelope schema. Compile-time exhaustiveness of
    /// [`kind_discriminant`] plus this runtime equality close the drift loop
    /// from the type side; `tests/schema_drift.rs` validates payload shapes.
    #[test]
    fn episode_discriminants_match_published_envelope_schema() {
        let mut discriminants: Vec<&str> = one_sample_per_variant()
            .iter()
            .map(kind_discriminant)
            .collect();
        discriminants.sort_unstable();

        let schema_path = std::path::PathBuf::from(env!("CARGO_MANIFEST_DIR"))
            .join("../../docs/schemas/runlog/episode.schema.json");
        let text = std::fs::read_to_string(&schema_path)
            .unwrap_or_else(|err| panic!("read {}: {err}", schema_path.display()));
        let schema: serde_json::Value =
            serde_json::from_str(&text).expect("envelope schema parses");
        let mut published: Vec<&str> = schema["properties"]["kind"]["enum"]
            .as_array()
            .expect("envelope schema declares a kind enum")
            .iter()
            .map(|value| value.as_str().expect("kind enum is strings"))
            .collect();
        published.sort_unstable();

        assert_eq!(
            discriminants, published,
            "Episode variants have drifted from the published envelope kind enum",
        );
    }

    /// Episodes serialised under klieo 0.6.x carry only `tokens` and
    /// `latency_ms`. The four 0.7 fields must decode as `None` via
    /// `#[serde(default)]`. Without that attribute, every persisted
    /// 0.6 row breaks replay on upgrade.
    #[test]
    fn legacy_llm_call_json_deserialises_with_none_for_new_fields() {
        let legacy = serde_json::json!({
            "LlmCall": {
                "tokens": 42,
                "latency_ms": 17
            }
        });
        let ep: Episode = serde_json::from_value(legacy).expect("legacy LlmCall decodes");
        match ep {
            Episode::LlmCall {
                tokens,
                latency_ms,
                provider,
                model,
                prompt_tokens,
                completion_tokens,
            } => {
                assert_eq!(tokens, 42);
                assert_eq!(latency_ms, 17);
                assert!(provider.is_none());
                assert!(model.is_none());
                assert!(prompt_tokens.is_none());
                assert!(completion_tokens.is_none());
            }
            other => panic!("expected LlmCall, got {other:?}"),
        }
    }

    /// [`Episode::llm_call`] returns the struct variant with all four
    /// 0.7-added enrichment fields as `None`, preserving the legacy
    /// emit shape for callers that don't have provider/model split.
    #[test]
    fn llm_call_ctor_leaves_enrichment_fields_none() {
        match Episode::llm_call(42, 17) {
            Episode::LlmCall {
                tokens,
                latency_ms,
                provider,
                model,
                prompt_tokens,
                completion_tokens,
            } => {
                assert_eq!(tokens, 42);
                assert_eq!(latency_ms, 17);
                assert!(provider.is_none());
                assert!(model.is_none());
                assert!(prompt_tokens.is_none());
                assert!(completion_tokens.is_none());
            }
            other => panic!("expected LlmCall, got {other:?}"),
        }
    }

    /// 0.7 emit sites with the full field set round-trip through serde
    /// unchanged.
    #[test]
    fn enriched_llm_call_round_trips() {
        let original = Episode::LlmCall {
            tokens: 60,
            latency_ms: 17,
            provider: Some("ollama".into()),
            model: Some("qwen2.5:14b".into()),
            prompt_tokens: Some(40),
            completion_tokens: Some(20),
        };
        let json = serde_json::to_value(&original).expect("serialises");
        let back: Episode = serde_json::from_value(json).expect("deserialises");
        match back {
            Episode::LlmCall {
                provider, model, ..
            } => {
                assert_eq!(provider.as_deref(), Some("ollama"));
                assert_eq!(model.as_deref(), Some("qwen2.5:14b"));
            }
            other => panic!("expected LlmCall, got {other:?}"),
        }
    }

    /// [`Episode::MemoryRecall`] round-trips through serde, and a legacy
    /// record missing `k` / `returned_fact_ids` (recorded before either
    /// field existed) decodes via `#[serde(default)]` rather than failing
    /// replay.
    #[test]
    fn memory_recall_round_trips_and_defaults_legacy() {
        let ep = Episode::MemoryRecall {
            query: "q".into(),
            k: 5,
            returned_fact_ids: vec![FactId::new("fact_1")],
        };
        let json = serde_json::to_value(&ep).expect("serialises");
        let back: Episode = serde_json::from_value(json).expect("deserialises");
        match back {
            Episode::MemoryRecall {
                query,
                k,
                returned_fact_ids,
            } => {
                assert_eq!(query, "q");
                assert_eq!(k, 5);
                assert_eq!(returned_fact_ids, vec![FactId::new("fact_1")]);
            }
            other => panic!("expected MemoryRecall, got {other:?}"),
        }

        let legacy = serde_json::json!({ "MemoryRecall": { "query": "q" } });
        let ep: Episode = serde_json::from_value(legacy).expect("legacy MemoryRecall decodes");
        match ep {
            Episode::MemoryRecall {
                k,
                returned_fact_ids,
                ..
            } => {
                assert_eq!(k, 0);
                assert!(returned_fact_ids.is_empty());
            }
            other => panic!("expected MemoryRecall, got {other:?}"),
        }
    }
}