trusty_memory/service/core_kg.rs
1//! `MemoryService` knowledge-graph / dream / activity methods.
2//!
3//! Why: the `MemoryService` impl exceeded the 500-SLOC production cap once the
4//! former monolithic `service.rs` was split (issue #607); its KG, dream-cycle,
5//! and activity-listing methods form a cohesive second half hosted here.
6//! What: a continuation `impl MemoryService` block whose methods were moved
7//! verbatim from the original single impl.
8//! Test: covered by the corresponding `web::tests` / `service::tests`.
9
10use crate::kg_write::{CachePolicy, KgWriteError};
11use crate::{ActivityFilter, ActivitySource, DaemonEvent};
12use std::sync::Arc;
13use trusty_common::memory_core::dream::{DreamConfig, Dreamer, PersistedDreamStats};
14use trusty_common::memory_core::palace::PalaceId;
15// #4670: ExpandDirection drives the progressive `kg/graph/neighbors` route.
16use trusty_common::memory_core::store::kg::{ExpandDirection, Triple};
17use trusty_common::memory_core::{PalaceHandle, PalaceRegistry};
18
19use super::core::KG_GRAPH_MAX_TRIPLES;
20use super::helpers::{list_palaces_blocking, refresh_gaps_cache};
21use super::types::{
22 DreamStatusPayload, KgAssertBody, KgGraphPayload, KgNeighborsPayload, KgNodeView,
23 KgSeedPayload, ServiceError, ServiceResult,
24};
25use super::MemoryService;
26
27/// The three adjacency-derived graph counts, computed on the blocking pool.
28///
29/// Why (#7106): `community_count` runs a full Louvain partition the first time
30/// it is asked at a given KG write generation. `kg_graph` and `kg_graph_seed`
31/// asked for it inline on an async worker, so one call against a large palace
32/// parked that worker for the whole partition — and before the memo, every
33/// call did it again. Hopping once for all three keeps the executor free
34/// without spending three tasks on work that shares a lock.
35/// What: `(node_count, edge_count, community_count)` as `u64`, read under
36/// `tokio::task::spawn_blocking`. A join failure degrades to zeros, the same
37/// rule the counts already use for a poisoned adjacency, rather than failing a
38/// graph render.
39/// Test: `kg_graph_returns_active_triples`, `kg_graph_seed_ranks_by_degree`.
40async fn adjacency_counts(handle: &Arc<PalaceHandle>) -> (u64, u64, u64) {
41 let handle = Arc::clone(handle);
42 tokio::task::spawn_blocking(move || {
43 (
44 handle.kg.node_count() as u64,
45 handle.kg.edge_count() as u64,
46 handle.kg.community_count() as u64,
47 )
48 })
49 .await
50 .unwrap_or_else(|e| {
51 tracing::warn!("kg adjacency counts task failed: {e}");
52 (0, 0, 0)
53 })
54}
55
56// ---------------------------------------------------------------------------
57// KG list bounds
58// ---------------------------------------------------------------------------
59//
60// #4776: these live on the service layer, not on either consumer, because both
61// readers of the subject list must agree on them — the HTTP explorer routes in
62// `web::kg_routes` (compiled only under `axum-server`) and the MCP
63// `kg_list_subjects` tool in `tools::kg_ops` (always compiled). Defining them
64// in `web` would put them on the far side of that feature gate from the tool.
65
66/// Default page size for KG subject listings when the caller omits `limit`.
67///
68/// Why: 50 is large enough to feel responsive in the KG Explorer and to answer
69/// "what is in this graph?" in one call, without dumping a full graph.
70pub(crate) const DEFAULT_KG_LIST_LIMIT: usize = 50;
71
72/// Hard ceiling on `limit` for KG subject listings.
73///
74/// Why: prevent a misconfigured client from asking the daemon to materialize
75/// thousands of rows in one go; matches the spec's max=200.
76pub(crate) const MAX_KG_LIST_LIMIT: usize = 200;
77
78impl MemoryService {
79 // -----------------------------------------------------------------
80 // Knowledge graph
81 // -----------------------------------------------------------------
82
83 /// Query the KG for all active triples whose subject matches.
84 pub async fn kg_query(&self, id: &str, subject: &str) -> ServiceResult<Vec<Triple>> {
85 let handle = self.open_handle(id)?;
86 handle
87 .kg
88 .query_active(subject)
89 .await
90 .map_err(|e| ServiceError::internal(format!("kg query: {e:#}")))
91 }
92
93 /// Assert a triple in the KG.
94 ///
95 /// Assert a triple through `POST /api/v1/palaces/{id}/kg`.
96 ///
97 /// Why: #4888 — this accepts an arbitrary predicate, so it can write a hot
98 /// one and must carry the same Tier S gate as the MCP tool. Not a
99 /// hypothetical path: `trusty-mpm`'s provisioner seeds its identity fact
100 /// through exactly this endpoint. #5524 — it also owed the prompt-cache
101 /// rebuild and never ran it, so a hot fact written here was stored and then
102 /// invisible to every later turn until some unrelated write rebuilt the
103 /// cache.
104 /// What: delegates the whole admission → assert → refresh sequence to
105 /// [`crate::kg_write::assert_triple`]. The variant split is what lets this
106 /// keep answering 400 for a refused write and 500 for a failed one.
107 /// Test: `http_kg_assert_endpoint_refreshes_prompt_cache`,
108 /// `http_kg_assert_endpoint_rejects_over_long_tier_s_object` in
109 /// `web::tests::prompt_tests`.
110 pub async fn kg_assert(&self, id: &str, body: KgAssertBody) -> ServiceResult<()> {
111 let handle = self.open_handle(id)?;
112 let triple = Triple {
113 subject: body.subject,
114 predicate: body.predicate,
115 object: body.object,
116 valid_from: chrono::Utc::now(),
117 valid_to: None,
118 confidence: body.confidence.unwrap_or(1.0),
119 provenance: body.provenance,
120 };
121 // #5524: route through the shared entry point so the prompt-cache
122 // rebuild cannot be forgotten here again.
123 crate::kg_write::assert_triple(&self.state, &handle, triple, CachePolicy::Inline)
124 .await
125 .map(|_| ())
126 .map_err(|e| match e {
127 KgWriteError::Admission(inner) => ServiceError::bad_request(format!("{inner:#}")),
128 other => ServiceError::internal(format!("{other}")),
129 })
130 }
131
132 /// Close the one active triple `(subject, predicate, object)`, leaving
133 /// every sibling object at that pair active. Returns the rows closed.
134 ///
135 /// Why: Issue #278 — the `DELETE /kg/triples/<id>` HTTP endpoint needs a
136 /// service-layer method so the HTTP handler stays a thin adapter. It took
137 /// no object and called the pair-level `KnowledgeGraph::retract`, whose
138 /// meaning is "close every active row at this pair", so a caller deleting
139 /// one triple lost the siblings it never named. Retraction is a soft close
140 /// — `close_active_row` copies the row to a `hist:` key first — so the
141 /// damage was recoverable, not silent data loss.
142 /// What: Opens the palace handle and calls
143 /// [`trusty_common::memory_core::store::kg::KnowledgeGraph::retract_triple`],
144 /// which keys on all three fields. Returns the closed count so the caller
145 /// can tell a retraction (`1`) from a miss (`0`); a miss is a genuine
146 /// no-op, which makes the call idempotent. Rebuilds the prompt cache when
147 /// a hot-predicate row was actually closed — otherwise a retracted Tier S
148 /// fact keeps being injected until the next write. This mirrors the
149 /// `kg_retract_triple` MCP tool so both surfaces agree.
150 /// Test: `kg_delete_triple_closes_one_object_and_keeps_siblings`,
151 /// `kg_delete_triple_returns_404_for_missing`,
152 /// `kg_delete_triple_rebuilds_prompt_cache_for_hot_predicate` in
153 /// `web::tests`. Only the third reaches the cache rebuild — the other two
154 /// retract under the predicate `is`, which is not hot.
155 pub async fn kg_retract_triple(
156 &self,
157 id: &str,
158 subject: &str,
159 predicate: &str,
160 object: &str,
161 ) -> ServiceResult<usize> {
162 let handle = self.open_handle(id)?;
163 let closed = handle
164 .kg
165 .retract_triple(subject, predicate, object)
166 .await
167 .map_err(|e| ServiceError::internal(format!("kg retract_triple: {e:#}")))?;
168 if closed > 0 && crate::prompt_facts::is_hot_predicate(predicate) {
169 // The write landed either way and the cache is only a
170 // denormalisation, so a rebuild failure is logged, not fatal.
171 // No test drives this arm: `rebuild_prompt_cache` skips a palace
172 // it cannot read and has no other fallible step, so it cannot
173 // currently return `Err`.
174 if let Err(e) = crate::prompt_facts::rebuild_prompt_cache(&self.state).await {
175 tracing::warn!("rebuild_prompt_cache after kg_retract_triple failed: {e:#}");
176 }
177 }
178 Ok(closed)
179 }
180
181 /// List distinct subjects in the KG.
182 pub async fn kg_list_subjects(&self, id: &str, limit: usize) -> ServiceResult<Vec<String>> {
183 let handle = self.open_handle(id)?;
184 handle
185 .kg
186 .list_subjects(limit)
187 .map_err(|e| ServiceError::internal(format!("kg list_subjects: {e:#}")))
188 }
189
190 /// List distinct subjects in the KG paired with their active-triple count.
191 pub async fn kg_list_subjects_with_counts(
192 &self,
193 id: &str,
194 limit: usize,
195 ) -> ServiceResult<Vec<(String, u64)>> {
196 let handle = self.open_handle(id)?;
197 handle
198 .kg
199 .list_subjects_with_counts(limit)
200 .map_err(|e| ServiceError::internal(format!("kg list_subjects_with_counts: {e:#}")))
201 }
202
203 /// Page through every active triple.
204 pub async fn kg_list_all(
205 &self,
206 id: &str,
207 limit: usize,
208 offset: usize,
209 ) -> ServiceResult<Vec<Triple>> {
210 let handle = self.open_handle(id)?;
211 handle
212 .kg
213 .list_active(limit, offset)
214 .await
215 .map_err(|e| ServiceError::internal(format!("kg list_active: {e:#}")))
216 }
217
218 /// Return the count of currently-active triples.
219 ///
220 /// #5384: a failed count read is a 500, not `{"active": 0}` — the badge
221 /// this feeds cannot tell those apart.
222 pub async fn kg_count(&self, id: &str) -> ServiceResult<usize> {
223 let handle = self.open_handle(id)?;
224 handle.kg.count_active_triples().map_err(|e| {
225 ServiceError::internal(format!("kg count_active_triples for palace {id}: {e:#}"))
226 })
227 }
228
229 /// Build the per-palace visual graph payload.
230 ///
231 /// Why (issue #4670): the `node_count` / `edge_count` / `community_count`
232 /// here are computed over the FULL adjacency while `triples` is capped at
233 /// [`KG_GRAPH_MAX_TRIPLES`]. That mismatch used to be invisible — the UI
234 /// rendered 5,000 triples under a "9,311 nodes" badge — and because
235 /// `list_active` orders by `valid_from` DESC the dropped triples were
236 /// silently the oldest. The payload now reports what it actually returned
237 /// alongside what exists, so truncation is machine-detectable.
238 /// What: unchanged query; adds `returned_triple_count`,
239 /// `active_triple_count`, and the derived `truncated` flag.
240 /// Test: `kg_graph_signals_truncation`, `kg_graph_returns_active_triples`.
241 pub async fn kg_graph(&self, id: &str) -> ServiceResult<KgGraphPayload> {
242 self.kg_graph_with_cap(id, KG_GRAPH_MAX_TRIPLES).await
243 }
244
245 /// [`Self::kg_graph`] with an explicit triple cap.
246 ///
247 /// Why (issue #4670): the truncation-signalling branch is only reachable
248 /// above `KG_GRAPH_MAX_TRIPLES` (5,000), and seeding 5,001 triples costs
249 /// ~90 s of test time — expensive enough that the branch would in practice
250 /// go untested. Taking the cap as a parameter makes it provable with five
251 /// triples and a cap of three, at no cost to the production call path.
252 /// What: the real implementation; `kg_graph` is a thin wrapper that passes
253 /// the production constant.
254 /// Test: `kg_graph_signals_truncation`.
255 pub async fn kg_graph_with_cap(
256 &self,
257 id: &str,
258 max_triples: usize,
259 ) -> ServiceResult<KgGraphPayload> {
260 let handle = self.open_handle(id)?;
261 let triples = handle
262 .kg
263 .list_active(max_triples, 0)
264 .await
265 .map_err(|e| ServiceError::internal(format!("kg list_active: {e:#}")))?;
266 // #4670: compare against the true active count, not the cap, so a
267 // palace sitting exactly on the cap is not falsely flagged.
268 // #5384: a failed read would come back as 0 and make `truncated` false
269 // for every payload, which is the flag's exact failure mode.
270 let active_triple_count = handle.kg.count_active_triples().map_err(|e| {
271 ServiceError::internal(format!("kg count_active_triples for palace {id}: {e:#}"))
272 })? as u64;
273 let returned_triple_count = triples.len() as u64;
274 // #7106: `community_count` runs a Louvain partition on the first read
275 // after a write, so it never runs on an async worker. The three
276 // adjacency-derived counts share one hop.
277 let (node_count, edge_count, community_count) = adjacency_counts(&handle).await;
278 Ok(KgGraphPayload {
279 triples,
280 node_count,
281 edge_count,
282 community_count,
283 returned_triple_count,
284 active_triple_count,
285 truncated: returned_triple_count < active_triple_count,
286 })
287 }
288
289 /// Top-`limit` nodes by degree plus the edges among them (issue #4670).
290 ///
291 /// Why: first paint must show the graph's skeleton, not 9,311 nodes in an
292 /// O(n²) layout. Measured on the live 8,266-triple palace, 90.2% of nodes
293 /// are degree-1 leaves and only 7.2% have degree >= 5, so a
294 /// top-degree slice carries essentially all of the visible structure and
295 /// everything else stays one click away.
296 /// What: runs `KnowledgeGraph::top_degree_subgraph` over the resident
297 /// adjacency (O(V log V + E), no disk I/O) and pairs the result with the
298 /// palace-wide totals the header needs to report honestly.
299 /// Test: `kg_graph_seed_ranks_by_degree`, `kg_graph_seed_clamps_limit`.
300 pub async fn kg_graph_seed(&self, id: &str, limit: usize) -> ServiceResult<KgSeedPayload> {
301 let handle = self.open_handle(id)?;
302 let (nodes, triples) = handle
303 .kg
304 .top_degree_subgraph(limit)
305 .map_err(|e| ServiceError::internal(format!("kg top_degree_subgraph: {e:#}")))?;
306 // #7106: same reason as `kg_graph_with_cap` — off the async worker.
307 let (node_count, edge_count, community_count) = adjacency_counts(&handle).await;
308 let returned_node_count = nodes.len() as u64;
309 Ok(KgSeedPayload {
310 nodes: nodes.into_iter().map(KgNodeView::from).collect(),
311 returned_triple_count: triples.len() as u64,
312 triples,
313 node_count,
314 edge_count,
315 community_count,
316 returned_node_count,
317 limit: limit as u64,
318 truncated: returned_node_count < node_count,
319 })
320 }
321
322 /// Direction-aware, hop-bounded expansion around one node (issue #4670).
323 ///
324 /// Why: click-to-expand needs "what points AT this node", which no HTTP
325 /// endpoint could answer — `kg_query` is a subject prefix scan. Bounding
326 /// the hops keeps one click on a hub from pulling the whole palace.
327 /// What: delegates to `KnowledgeGraph::expand_neighbors`. `direction`
328 /// and `max_hops` are already validated/clamped by the HTTP layer; they
329 /// are echoed back so the client can see what actually ran.
330 /// Test: `kg_neighbors_returns_incoming_edges`, `kg_neighbors_clamps_max_hops`.
331 pub async fn kg_neighbors(
332 &self,
333 id: &str,
334 node: &str,
335 direction: ExpandDirection,
336 max_hops: usize,
337 ) -> ServiceResult<KgNeighborsPayload> {
338 let handle = self.open_handle(id)?;
339 let (nodes, triples) = handle
340 .kg
341 .expand_neighbors(node, direction, max_hops)
342 .map_err(|e| ServiceError::internal(format!("kg expand_neighbors: {e:#}")))?;
343 Ok(KgNeighborsPayload {
344 origin: node.to_string(),
345 returned_node_count: nodes.len() as u64,
346 returned_triple_count: triples.len() as u64,
347 nodes: nodes.into_iter().map(KgNodeView::from).collect(),
348 triples,
349 direction: match direction {
350 ExpandDirection::In => "in",
351 ExpandDirection::Out => "out",
352 ExpandDirection::Both => "both",
353 }
354 .to_string(),
355 max_hops: max_hops as u64,
356 })
357 }
358
359 // -----------------------------------------------------------------
360 // Dream cycle
361 // -----------------------------------------------------------------
362
363 /// Aggregate dream stats across every persisted palace.
364 pub async fn dream_status_aggregate(&self) -> DreamStatusPayload {
365 let palaces = PalaceRegistry::list_palaces(&self.state.data_root).unwrap_or_default();
366 let mut out = DreamStatusPayload::default();
367 let mut latest: Option<chrono::DateTime<chrono::Utc>> = None;
368 for p in palaces {
369 let data_dir = self.state.data_root.join(p.id.as_str());
370 let snap = match PersistedDreamStats::load(&data_dir) {
371 Ok(Some(s)) => s,
372 _ => continue,
373 };
374 out.merged = out.merged.saturating_add(snap.stats.merged);
375 out.pruned = out.pruned.saturating_add(snap.stats.pruned);
376 out.compacted = out.compacted.saturating_add(snap.stats.compacted);
377 out.closets_updated = out
378 .closets_updated
379 .saturating_add(snap.stats.closets_updated);
380 out.duration_ms = out.duration_ms.saturating_add(snap.stats.duration_ms);
381 latest = match latest {
382 Some(t) if t >= snap.last_run_at => Some(t),
383 _ => Some(snap.last_run_at),
384 };
385 }
386 out.last_run_at = latest;
387 out
388 }
389
390 /// Per-palace dream stats snapshot.
391 pub async fn dream_status_for_palace(&self, id: &str) -> ServiceResult<DreamStatusPayload> {
392 let data_dir = self.state.data_root.join(id);
393 if !data_dir.exists() {
394 return Err(ServiceError::not_found(format!("palace not found: {id}")));
395 }
396 match PersistedDreamStats::load(&data_dir) {
397 Ok(Some(s)) => Ok(s.into()),
398 Ok(None) => Ok(DreamStatusPayload::default()),
399 Err(e) => Err(ServiceError::internal(format!("read dream stats: {e:#}"))),
400 }
401 }
402
403 /// Run a dream cycle across every palace.
404 ///
405 /// Why (issue #4637): like `recall_all`, this route is deliberately NOT
406 /// converted to `PalaceRegistry::peek`. Dreaming is a maintenance pass —
407 /// consolidating only the 64 palaces that happen to be cache-resident
408 /// would silently stop maintaining the other ~5,730, which is a worse
409 /// failure than a slow run because nothing reports it. Opening every
410 /// palace stays correct; what changes is that the blocking open no longer
411 /// runs inline on a tokio worker thread. A long dream run is expected —
412 /// this is an explicitly-triggered `POST`, not a page load.
413 /// What: lists palaces on the blocking pool, then per palace hops to
414 /// `spawn_blocking` for the open before awaiting the async dream cycle.
415 /// Test: `dream_run_aggregates_stats`.
416 pub async fn dream_run(&self) -> ServiceResult<DreamStatusPayload> {
417 let palaces = list_palaces_blocking(&self.state)
418 .await
419 .map_err(|e| ServiceError::internal(format!("{e:#}")))?;
420 let dreamer = Dreamer::new(DreamConfig::default());
421 let mut out = DreamStatusPayload::default();
422 for p in palaces {
423 // #4637: open_palace (not peek) is deliberate — a dream cycle must
424 // maintain every palace; the spawn_blocking hop keeps the cold open
425 // off the async executor.
426 let registry = std::sync::Arc::clone(&self.state.registry);
427 let root = self.state.data_root.clone();
428 let pid = p.id.clone();
429 let opened =
430 tokio::task::spawn_blocking(move || registry.open_palace(&root, &pid)).await;
431 let handle = match opened {
432 Ok(Ok(h)) => h,
433 Ok(Err(e)) => {
434 tracing::warn!(palace = %p.id, "dream_run: open failed: {e:#}");
435 continue;
436 }
437 Err(e) => {
438 tracing::warn!(palace = %p.id, "dream_run: join open failed: {e}");
439 continue;
440 }
441 };
442 match dreamer.dream_cycle(&handle).await {
443 Ok(stats) => {
444 out.merged = out.merged.saturating_add(stats.merged);
445 out.pruned = out.pruned.saturating_add(stats.pruned);
446 out.compacted = out.compacted.saturating_add(stats.compacted);
447 out.closets_updated = out.closets_updated.saturating_add(stats.closets_updated);
448 out.duration_ms = out.duration_ms.saturating_add(stats.duration_ms);
449 }
450 Err(e) => tracing::warn!(palace = %p.id, "dream_run: cycle failed: {e:#}"),
451 }
452 refresh_gaps_cache(&self.state, &handle).await;
453 }
454 out.last_run_at = Some(chrono::Utc::now());
455 self.state.emit(DaemonEvent::DreamCompleted {
456 palace_id: None,
457 merged: out.merged,
458 pruned: out.pruned,
459 compacted: out.compacted,
460 closets_updated: out.closets_updated,
461 duration_ms: out.duration_ms,
462 source: ActivitySource::Http,
463 });
464 self.state.emit(self.aggregate_status_event());
465 Ok(out)
466 }
467
468 // -----------------------------------------------------------------
469 // Activity log
470 // -----------------------------------------------------------------
471
472 /// Paginated activity-log read.
473 pub async fn list_activity(
474 &self,
475 filter: ActivityFilter,
476 limit: usize,
477 offset: usize,
478 ) -> ServiceResult<(Vec<crate::ActivityEntry>, u64)> {
479 let entries = self
480 .state
481 .activity_log
482 .list(&filter, limit, offset)
483 .map_err(|e| ServiceError::internal(format!("activity list: {e:#}")))?;
484 let total = self
485 .state
486 .activity_log
487 .count()
488 .map_err(|e| ServiceError::internal(format!("activity count: {e:#}")))?;
489 Ok((entries, total))
490 }
491
492 // -----------------------------------------------------------------
493 // Internal helper — open a palace handle, 404 only on a genuine absence.
494 // -----------------------------------------------------------------
495
496 /// Open the named palace.
497 ///
498 /// Why (#5549, ADR-0045): this mapped every `open_palace` failure to
499 /// `NotFound`, which the HTTP layer renders as 404. A denied or transient
500 /// read of `palace.json`, undecodable metadata, an open-queue timeout, or a
501 /// redb write-lock conflict then all reported that the palace does not
502 /// exist — erasing at the caller the distinction `load_palace` draws, and
503 /// across a much wider surface than the two rename paths: every
504 /// `/api/v1/palaces/{id}/kg*` endpoint, the drawer CRUD routes, and
505 /// per-palace recall reach this one helper.
506 /// What: returns `ServiceError::NotFound` only when
507 /// `PalaceRegistry::open_error_is_absent` confirms the palace is genuinely
508 /// not there, and `ServiceError::Internal` (500) otherwise.
509 /// Test: `unreadable_palace_is_500_not_404_at_the_service_open_handle`,
510 /// `unstattable_palace_is_500_not_404_at_the_service_open_handle`,
511 /// `absent_palace_is_still_404_at_both_open_handles`.
512 pub fn open_handle(&self, id: &str) -> ServiceResult<Arc<PalaceHandle>> {
513 self.state
514 .registry
515 .open_palace(&self.state.data_root, &PalaceId::new(id))
516 .map_err(|e| {
517 // #5549: every open failure mapped to 404, so a palace that
518 // could not be read was reported as one that is not there.
519 if PalaceRegistry::open_error_is_absent(&e) {
520 ServiceError::not_found(format!("palace not found: {id} ({e:#})"))
521 } else {
522 ServiceError::internal(format!("palace could not be loaded: {id} ({e:#})"))
523 }
524 })
525 }
526}