switchroom 0.20.17 → 0.20.19

package/dist/host-control/main.js

@@ -11445,7 +11445,13 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced; and " + "HINDSIGHT_API_TEMPORAL_LANGUAGES — the language set dateparser is " + "restricted to during temporal query analysis, made live by switchroom's " + "temporal-language image patch (which ended a 200+-locale auto-detection " + "pass that blocked the shared asyncio loop on every recall); " + "comma-separated, unset means the image's baked `en`, set e.g. `en,es` to " + "restore i18n parsing; and " + "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS — the max chars of query text " + "dateparser.search_dates() is handed during temporal analysis (0 = " + "unlimited), made live by switchroom's temporal-offload image patch " + "(which moved the synchronous extraction off the shared asyncio loop onto " + "a single-worker thread and bounds its input, ending multi-second loop " + "stalls on multi-KB consolidation queries); unset means the image's baked " + "2000, set 0 to restore an unbounded full scan); and " + "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE — the Postgres " + "text-search regconfig the native BM25 arms and the search_vector index " + "write run to_tsvector/to_tsquery against; unset means upstream's " + "always-safe `english`, set `hindsight_english` to use switchroom's " + "junk-stopping config (dict + stopword file provisioned durably by the " + "image entrypoint, recall degrading to semantic-only if it is ever " + "missing); and " + "HINDSIGHT_API_BM25_MAX_QUERY_TERMS — the max number of query tokens the " + "keyword (BM25) recall arm searches for; unset means upstream's 0 = " + "unbounded, which makes a multi-KB consolidation query a disjunction " + "over hundreds of terms; set e.g. 64 to cap it, at the cost of " + "silently dropping the tail of a long query; plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+    recall_pool: exports_external.object({
+      enabled: exports_external.boolean().optional().describe("Turn on the recall/background WORKER SPLIT: a second container, " + "`switchroom-hindsight-recall`, serves recall/reflect only with N " + "uvicorn workers, while the existing `switchroom-hindsight` " + "container keeps pg0, the single background WorkerPoller, the " + "healthcheck and all consolidation/retain processing but moves off " + "the public port onto public+1. The public port (memory.config.url) " + "does not change — the pool binds it instead. Absent/false (the " + "default) → the single-container topology, unchanged. Opt-in only. " + "Raises recall throughput ~2.2x on the fleet (2.6→5.6 rps)."),
+      workers: exports_external.number().int().positive().optional().describe("uvicorn `--workers` for the recall pool (HINDSIGHT_API_WORKERS). " + "Default 4. Each worker holds its own GPU cross-encoder reranker, " + "so past ~4 workers contend for VRAM rather than adding throughput " + "on a single GPU — raise only on a bigger GPU or a CPU reranker. " + "The per-worker DB pools are auto-sized down from the pg0 " + "connection budget so N*(write+read)+60 background stays under " + "max_connections=250; a worker count that would overflow that " + "budget is rejected at launch."),
+      db_pool_max_size: exports_external.number().int().positive().optional().describe("Override the per-worker WRITE pool cap " + "(HINDSIGHT_API_DB_POOL_MAX_SIZE). Absent → auto-sized from the " + "worker count. Set this only to override the auto-sizing; the " + "connection-budget check still applies and rejects an overflow."),
+      read_db_pool_max_size: exports_external.number().int().positive().optional().describe("Override the per-worker READ pool cap " + "(HINDSIGHT_API_READ_DB_POOL_MAX_SIZE). Absent → auto-sized from " + "the worker count. Set this only to override the auto-sizing; the " + "connection-budget check still applies and rejects an overflow.")
+    }).optional().describe("Optional recall/background container split for the hindsight " + "singleton. Off by default; see `enabled`. Provisioned by " + "`switchroom memory setup` (src/setup/hindsight-recall-pool.ts) so it " + "survives `switchroom apply` — the hand-applied topology it replaces " + "did not."),
+    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced; and " + "HINDSIGHT_API_QUERY_ANALYZER_LANGUAGES — the language set dateparser is " + "restricted to during temporal query analysis (upstream #3154, which " + "superseded switchroom's own temporal-language image patch at v0.9.0); " + "comma-separated; switchroom EMITS `en` on every host because upstream's " + "unset default auto-detects across 200+ locales inline on the shared " + "asyncio loop (~165x slower per call), set e.g. `en,es` for i18n parsing; " + "and " + "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS — the max chars of query text " + "dateparser.search_dates() is handed during temporal analysis (0 = " + "unlimited), made live by switchroom's temporal-offload image patch " + "(which moved the synchronous extraction off the shared asyncio loop onto " + "a single-worker thread and bounds its input, ending multi-second loop " + "stalls on multi-KB consolidation queries); unset means the image's baked " + "2000, set 0 to restore an unbounded full scan); and " + "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE — the Postgres " + "text-search regconfig the native BM25 arms and the search_vector index " + "write run to_tsvector/to_tsquery against; unset means upstream's " + "always-safe `english`, set `hindsight_english` to use switchroom's " + "junk-stopping config (dict + stopword file provisioned durably by the " + "image entrypoint, recall degrading to semantic-only if it is ever " + "missing); and " + "HINDSIGHT_API_BM25_MAX_QUERY_TERMS — the max number of query tokens the " + "keyword (BM25) recall arm searches for; unset means upstream's 0 = " + "unbounded, which makes a multi-KB consolidation query a disjunction " + "over hundreds of terms; set e.g. 64 to cap it, at the cost of " + "silently dropping the tail of a long query; and " + "HINDSIGHT_API_EMBEDDINGS_PROVIDER with the four " + "HINDSIGHT_API_EMBEDDINGS_ONNX_{MODEL_ID,POOLING,QUERY_PREFIX," + "PASSAGE_PREFIX} keys — the ONNX-embeddings cutover: unset keeps " + "upstream's slower `local` (PyTorch) backend, set `onnx` plus the model " + "id/pooling to run the SAME model on ONNX-CPU at a fraction of the " + "latency; the two _PREFIX keys are the query/passage affixes and are " + "set to the empty string for bge parity (vectors are byte-identical " + "between the two backends, so no re-embedding is needed) — unusually " + "for this block an EMPTY value here is honoured, not dropped, because an " + "affix-free prefix is the intended value; and the v0.9.0 knobs " + "switchroom allowlists for reach but emits no value for — " + "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_{LOW,MID,HIGH} (per-recall-budget " + "rerank caps; unset/0 falls back to the global " + "HINDSIGHT_API_RERANKER_MAX_CANDIDATES switchroom does pin), " + "HINDSIGHT_API_EMBEDDINGS_MAX_INPUT_TOKENS (embedding-input truncation " + "ceiling; unset means the model's own limit), " + "HINDSIGHT_API_ENTITY_TRGM_SIMILARITY_THRESHOLD (entity-resolution fuzzy " + "match floor; unset means upstream's 0.15) and " + "HINDSIGHT_API_ENABLE_{TEMPORAL_RETRIEVAL,GRAPH_RETRIEVAL,RERANKING} " + "(global per-stage kill switches, all default true — the emergency lever " + "if one retrieval stage is implicated in a latency incident, at the cost " + "of recall quality); plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -21559,7 +21565,7 @@ function allocateAgentUid(name) {
 }
 
 // src/build-info.ts
-var VERSION = "0.20.17";
+var VERSION = "0.20.19";
 
 // src/setup/hindsight-recall-passthrough.ts
 var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
@@ -21758,6 +21764,7 @@ var HINDSIGHT_DEFAULT_LLM_SUPPORTS_MAX_ITEMS = "false";
 var HINDSIGHT_DEFAULT_RECENCY_DECAY_FUNCTION = "exponential";
 var HINDSIGHT_DEFAULT_RECENCY_DECAY_HALFLIFE_DAYS = 30;
 var HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION = "pg_search";
+var HINDSIGHT_DEFAULT_QUERY_ANALYZER_LANGUAGES = "en";
 var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -21824,7 +21831,11 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
     "HINDSIGHT_API_GRAPH_SEED_MIN_SIMILARITY",
     String(HINDSIGHT_DEFAULT_GRAPH_SEED_MIN_SIMILARITY)
   ],
-  ["HINDSIGHT_API_TEXT_SEARCH_EXTENSION", HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION]
+  ["HINDSIGHT_API_TEXT_SEARCH_EXTENSION", HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION],
+  [
+    "HINDSIGHT_API_QUERY_ANALYZER_LANGUAGES",
+    HINDSIGHT_DEFAULT_QUERY_ANALYZER_LANGUAGES
+  ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_GPU = [
   ["HINDSIGHT_API_RERANKER_LOCAL_FP16", HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16],
@@ -21859,10 +21870,26 @@ var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_API_RETAIN_WALL_TIMEOUT",
   "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT",
   "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S",
-  "HINDSIGHT_API_TEMPORAL_LANGUAGES",
   "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS",
   "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE",
-  "HINDSIGHT_API_BM25_MAX_QUERY_TERMS"
+  "HINDSIGHT_API_BM25_MAX_QUERY_TERMS",
+  "HINDSIGHT_API_EMBEDDINGS_PROVIDER",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_MODEL_ID",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_POOLING",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_QUERY_PREFIX",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_PASSAGE_PREFIX",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_LOW",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_MID",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_HIGH",
+  "HINDSIGHT_API_EMBEDDINGS_MAX_INPUT_TOKENS",
+  "HINDSIGHT_API_ENTITY_TRGM_SIMILARITY_THRESHOLD",
+  "HINDSIGHT_API_ENABLE_TEMPORAL_RETRIEVAL",
+  "HINDSIGHT_API_ENABLE_GRAPH_RETRIEVAL",
+  "HINDSIGHT_API_ENABLE_RERANKING"
+]);
+var HINDSIGHT_PERF_ALLOW_EMPTY_KEYS = new Set([
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_QUERY_PREFIX",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_PASSAGE_PREFIX"
 ]);
 var HINDSIGHT_WORKER_SLOT_TYPES = [
   "consolidation",
@@ -21876,6 +21903,15 @@ var HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES = new Map(HINDSIGHT_WORKER_SLOT_TYPES
   `HINDSIGHT_API_WORKER_${type.toUpperCase()}_MAX_SLOTS`,
   `HINDSIGHT_API_WORKER_${type.toUpperCase()}_RESERVED_SLOTS`
 ]));
+var HINDSIGHT_RECALL_POOL_ENV_KEYS = [
+  "HINDSIGHT_API_WORKERS",
+  "HINDSIGHT_API_WORKER_ENABLED",
+  "HINDSIGHT_API_DATABASE_URL",
+  "HINDSIGHT_API_DB_URL",
+  "HINDSIGHT_API_DB_POOL_MAX_SIZE",
+  "HINDSIGHT_API_READ_DB_POOL_MAX_SIZE",
+  "HINDSIGHT_API_RUN_MIGRATIONS_ON_STARTUP"
+];
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...[
     ...HINDSIGHT_PERF_DEFAULTS_UNGATED,
@@ -21884,7 +21920,8 @@ var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ].map(([k]) => k),
   ...HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS,
   ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.keys(),
-  ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.values()
+  ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.values(),
+  ...HINDSIGHT_RECALL_POOL_ENV_KEYS
 ]);
 
 // src/setup/hindsight-context-budget.ts

package/dist/vault/approvals/kernel-server.js

@@ -4885,7 +4885,13 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced; and " + "HINDSIGHT_API_TEMPORAL_LANGUAGES — the language set dateparser is " + "restricted to during temporal query analysis, made live by switchroom's " + "temporal-language image patch (which ended a 200+-locale auto-detection " + "pass that blocked the shared asyncio loop on every recall); " + "comma-separated, unset means the image's baked `en`, set e.g. `en,es` to " + "restore i18n parsing; and " + "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS — the max chars of query text " + "dateparser.search_dates() is handed during temporal analysis (0 = " + "unlimited), made live by switchroom's temporal-offload image patch " + "(which moved the synchronous extraction off the shared asyncio loop onto " + "a single-worker thread and bounds its input, ending multi-second loop " + "stalls on multi-KB consolidation queries); unset means the image's baked " + "2000, set 0 to restore an unbounded full scan); and " + "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE — the Postgres " + "text-search regconfig the native BM25 arms and the search_vector index " + "write run to_tsvector/to_tsquery against; unset means upstream's " + "always-safe `english`, set `hindsight_english` to use switchroom's " + "junk-stopping config (dict + stopword file provisioned durably by the " + "image entrypoint, recall degrading to semantic-only if it is ever " + "missing); and " + "HINDSIGHT_API_BM25_MAX_QUERY_TERMS — the max number of query tokens the " + "keyword (BM25) recall arm searches for; unset means upstream's 0 = " + "unbounded, which makes a multi-KB consolidation query a disjunction " + "over hundreds of terms; set e.g. 64 to cap it, at the cost of " + "silently dropping the tail of a long query; plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+    recall_pool: exports_external.object({
+      enabled: exports_external.boolean().optional().describe("Turn on the recall/background WORKER SPLIT: a second container, " + "`switchroom-hindsight-recall`, serves recall/reflect only with N " + "uvicorn workers, while the existing `switchroom-hindsight` " + "container keeps pg0, the single background WorkerPoller, the " + "healthcheck and all consolidation/retain processing but moves off " + "the public port onto public+1. The public port (memory.config.url) " + "does not change — the pool binds it instead. Absent/false (the " + "default) → the single-container topology, unchanged. Opt-in only. " + "Raises recall throughput ~2.2x on the fleet (2.6→5.6 rps)."),
+      workers: exports_external.number().int().positive().optional().describe("uvicorn `--workers` for the recall pool (HINDSIGHT_API_WORKERS). " + "Default 4. Each worker holds its own GPU cross-encoder reranker, " + "so past ~4 workers contend for VRAM rather than adding throughput " + "on a single GPU — raise only on a bigger GPU or a CPU reranker. " + "The per-worker DB pools are auto-sized down from the pg0 " + "connection budget so N*(write+read)+60 background stays under " + "max_connections=250; a worker count that would overflow that " + "budget is rejected at launch."),
+      db_pool_max_size: exports_external.number().int().positive().optional().describe("Override the per-worker WRITE pool cap " + "(HINDSIGHT_API_DB_POOL_MAX_SIZE). Absent → auto-sized from the " + "worker count. Set this only to override the auto-sizing; the " + "connection-budget check still applies and rejects an overflow."),
+      read_db_pool_max_size: exports_external.number().int().positive().optional().describe("Override the per-worker READ pool cap " + "(HINDSIGHT_API_READ_DB_POOL_MAX_SIZE). Absent → auto-sized from " + "the worker count. Set this only to override the auto-sizing; the " + "connection-budget check still applies and rejects an overflow.")
+    }).optional().describe("Optional recall/background container split for the hindsight " + "singleton. Off by default; see `enabled`. Provisioned by " + "`switchroom memory setup` (src/setup/hindsight-recall-pool.ts) so it " + "survives `switchroom apply` — the hand-applied topology it replaces " + "did not."),
+    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced; and " + "HINDSIGHT_API_QUERY_ANALYZER_LANGUAGES — the language set dateparser is " + "restricted to during temporal query analysis (upstream #3154, which " + "superseded switchroom's own temporal-language image patch at v0.9.0); " + "comma-separated; switchroom EMITS `en` on every host because upstream's " + "unset default auto-detects across 200+ locales inline on the shared " + "asyncio loop (~165x slower per call), set e.g. `en,es` for i18n parsing; " + "and " + "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS — the max chars of query text " + "dateparser.search_dates() is handed during temporal analysis (0 = " + "unlimited), made live by switchroom's temporal-offload image patch " + "(which moved the synchronous extraction off the shared asyncio loop onto " + "a single-worker thread and bounds its input, ending multi-second loop " + "stalls on multi-KB consolidation queries); unset means the image's baked " + "2000, set 0 to restore an unbounded full scan); and " + "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE — the Postgres " + "text-search regconfig the native BM25 arms and the search_vector index " + "write run to_tsvector/to_tsquery against; unset means upstream's " + "always-safe `english`, set `hindsight_english` to use switchroom's " + "junk-stopping config (dict + stopword file provisioned durably by the " + "image entrypoint, recall degrading to semantic-only if it is ever " + "missing); and " + "HINDSIGHT_API_BM25_MAX_QUERY_TERMS — the max number of query tokens the " + "keyword (BM25) recall arm searches for; unset means upstream's 0 = " + "unbounded, which makes a multi-KB consolidation query a disjunction " + "over hundreds of terms; set e.g. 64 to cap it, at the cost of " + "silently dropping the tail of a long query; and " + "HINDSIGHT_API_EMBEDDINGS_PROVIDER with the four " + "HINDSIGHT_API_EMBEDDINGS_ONNX_{MODEL_ID,POOLING,QUERY_PREFIX," + "PASSAGE_PREFIX} keys — the ONNX-embeddings cutover: unset keeps " + "upstream's slower `local` (PyTorch) backend, set `onnx` plus the model " + "id/pooling to run the SAME model on ONNX-CPU at a fraction of the " + "latency; the two _PREFIX keys are the query/passage affixes and are " + "set to the empty string for bge parity (vectors are byte-identical " + "between the two backends, so no re-embedding is needed) — unusually " + "for this block an EMPTY value here is honoured, not dropped, because an " + "affix-free prefix is the intended value; and the v0.9.0 knobs " + "switchroom allowlists for reach but emits no value for — " + "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_{LOW,MID,HIGH} (per-recall-budget " + "rerank caps; unset/0 falls back to the global " + "HINDSIGHT_API_RERANKER_MAX_CANDIDATES switchroom does pin), " + "HINDSIGHT_API_EMBEDDINGS_MAX_INPUT_TOKENS (embedding-input truncation " + "ceiling; unset means the model's own limit), " + "HINDSIGHT_API_ENTITY_TRGM_SIMILARITY_THRESHOLD (entity-resolution fuzzy " + "match floor; unset means upstream's 0.15) and " + "HINDSIGHT_API_ENABLE_{TEMPORAL_RETRIEVAL,GRAPH_RETRIEVAL,RERANKING} " + "(global per-stage kill switches, all default true — the emergency lever " + "if one retrieval stage is implicated in a latency incident, at the cost " + "of recall quality); plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -19653,6 +19659,7 @@ var HINDSIGHT_DEFAULT_LLM_SUPPORTS_MAX_ITEMS = "false";
 var HINDSIGHT_DEFAULT_RECENCY_DECAY_FUNCTION = "exponential";
 var HINDSIGHT_DEFAULT_RECENCY_DECAY_HALFLIFE_DAYS = 30;
 var HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION = "pg_search";
+var HINDSIGHT_DEFAULT_QUERY_ANALYZER_LANGUAGES = "en";
 var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -19719,7 +19726,11 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
     "HINDSIGHT_API_GRAPH_SEED_MIN_SIMILARITY",
     String(HINDSIGHT_DEFAULT_GRAPH_SEED_MIN_SIMILARITY)
   ],
-  ["HINDSIGHT_API_TEXT_SEARCH_EXTENSION", HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION]
+  ["HINDSIGHT_API_TEXT_SEARCH_EXTENSION", HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION],
+  [
+    "HINDSIGHT_API_QUERY_ANALYZER_LANGUAGES",
+    HINDSIGHT_DEFAULT_QUERY_ANALYZER_LANGUAGES
+  ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_GPU = [
   ["HINDSIGHT_API_RERANKER_LOCAL_FP16", HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16],
@@ -19754,10 +19765,26 @@ var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_API_RETAIN_WALL_TIMEOUT",
   "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT",
   "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S",
-  "HINDSIGHT_API_TEMPORAL_LANGUAGES",
   "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS",
   "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE",
-  "HINDSIGHT_API_BM25_MAX_QUERY_TERMS"
+  "HINDSIGHT_API_BM25_MAX_QUERY_TERMS",
+  "HINDSIGHT_API_EMBEDDINGS_PROVIDER",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_MODEL_ID",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_POOLING",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_QUERY_PREFIX",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_PASSAGE_PREFIX",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_LOW",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_MID",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_HIGH",
+  "HINDSIGHT_API_EMBEDDINGS_MAX_INPUT_TOKENS",
+  "HINDSIGHT_API_ENTITY_TRGM_SIMILARITY_THRESHOLD",
+  "HINDSIGHT_API_ENABLE_TEMPORAL_RETRIEVAL",
+  "HINDSIGHT_API_ENABLE_GRAPH_RETRIEVAL",
+  "HINDSIGHT_API_ENABLE_RERANKING"
+]);
+var HINDSIGHT_PERF_ALLOW_EMPTY_KEYS = new Set([
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_QUERY_PREFIX",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_PASSAGE_PREFIX"
 ]);
 var HINDSIGHT_WORKER_SLOT_TYPES = [
   "consolidation",
@@ -19771,6 +19798,15 @@ var HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES = new Map(HINDSIGHT_WORKER_SLOT_TYPES
   `HINDSIGHT_API_WORKER_${type.toUpperCase()}_MAX_SLOTS`,
   `HINDSIGHT_API_WORKER_${type.toUpperCase()}_RESERVED_SLOTS`
 ]));
+var HINDSIGHT_RECALL_POOL_ENV_KEYS = [
+  "HINDSIGHT_API_WORKERS",
+  "HINDSIGHT_API_WORKER_ENABLED",
+  "HINDSIGHT_API_DATABASE_URL",
+  "HINDSIGHT_API_DB_URL",
+  "HINDSIGHT_API_DB_POOL_MAX_SIZE",
+  "HINDSIGHT_API_READ_DB_POOL_MAX_SIZE",
+  "HINDSIGHT_API_RUN_MIGRATIONS_ON_STARTUP"
+];
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...[
     ...HINDSIGHT_PERF_DEFAULTS_UNGATED,
@@ -19779,7 +19815,8 @@ var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ].map(([k]) => k),
   ...HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS,
   ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.keys(),
-  ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.values()
+  ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.values(),
+  ...HINDSIGHT_RECALL_POOL_ENV_KEYS
 ]);
 
 // src/setup/hindsight-context-budget.ts

package/dist/vault/broker/server.js

@@ -4481,7 +4481,13 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced; and " + "HINDSIGHT_API_TEMPORAL_LANGUAGES — the language set dateparser is " + "restricted to during temporal query analysis, made live by switchroom's " + "temporal-language image patch (which ended a 200+-locale auto-detection " + "pass that blocked the shared asyncio loop on every recall); " + "comma-separated, unset means the image's baked `en`, set e.g. `en,es` to " + "restore i18n parsing; and " + "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS — the max chars of query text " + "dateparser.search_dates() is handed during temporal analysis (0 = " + "unlimited), made live by switchroom's temporal-offload image patch " + "(which moved the synchronous extraction off the shared asyncio loop onto " + "a single-worker thread and bounds its input, ending multi-second loop " + "stalls on multi-KB consolidation queries); unset means the image's baked " + "2000, set 0 to restore an unbounded full scan); and " + "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE — the Postgres " + "text-search regconfig the native BM25 arms and the search_vector index " + "write run to_tsvector/to_tsquery against; unset means upstream's " + "always-safe `english`, set `hindsight_english` to use switchroom's " + "junk-stopping config (dict + stopword file provisioned durably by the " + "image entrypoint, recall degrading to semantic-only if it is ever " + "missing); and " + "HINDSIGHT_API_BM25_MAX_QUERY_TERMS — the max number of query tokens the " + "keyword (BM25) recall arm searches for; unset means upstream's 0 = " + "unbounded, which makes a multi-KB consolidation query a disjunction " + "over hundreds of terms; set e.g. 64 to cap it, at the cost of " + "silently dropping the tail of a long query; plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+    recall_pool: exports_external.object({
+      enabled: exports_external.boolean().optional().describe("Turn on the recall/background WORKER SPLIT: a second container, " + "`switchroom-hindsight-recall`, serves recall/reflect only with N " + "uvicorn workers, while the existing `switchroom-hindsight` " + "container keeps pg0, the single background WorkerPoller, the " + "healthcheck and all consolidation/retain processing but moves off " + "the public port onto public+1. The public port (memory.config.url) " + "does not change — the pool binds it instead. Absent/false (the " + "default) → the single-container topology, unchanged. Opt-in only. " + "Raises recall throughput ~2.2x on the fleet (2.6→5.6 rps)."),
+      workers: exports_external.number().int().positive().optional().describe("uvicorn `--workers` for the recall pool (HINDSIGHT_API_WORKERS). " + "Default 4. Each worker holds its own GPU cross-encoder reranker, " + "so past ~4 workers contend for VRAM rather than adding throughput " + "on a single GPU — raise only on a bigger GPU or a CPU reranker. " + "The per-worker DB pools are auto-sized down from the pg0 " + "connection budget so N*(write+read)+60 background stays under " + "max_connections=250; a worker count that would overflow that " + "budget is rejected at launch."),
+      db_pool_max_size: exports_external.number().int().positive().optional().describe("Override the per-worker WRITE pool cap " + "(HINDSIGHT_API_DB_POOL_MAX_SIZE). Absent → auto-sized from the " + "worker count. Set this only to override the auto-sizing; the " + "connection-budget check still applies and rejects an overflow."),
+      read_db_pool_max_size: exports_external.number().int().positive().optional().describe("Override the per-worker READ pool cap " + "(HINDSIGHT_API_READ_DB_POOL_MAX_SIZE). Absent → auto-sized from " + "the worker count. Set this only to override the auto-sizing; the " + "connection-budget check still applies and rejects an overflow.")
+    }).optional().describe("Optional recall/background container split for the hindsight " + "singleton. Off by default; see `enabled`. Provisioned by " + "`switchroom memory setup` (src/setup/hindsight-recall-pool.ts) so it " + "survives `switchroom apply` — the hand-applied topology it replaces " + "did not."),
+    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced; and " + "HINDSIGHT_API_QUERY_ANALYZER_LANGUAGES — the language set dateparser is " + "restricted to during temporal query analysis (upstream #3154, which " + "superseded switchroom's own temporal-language image patch at v0.9.0); " + "comma-separated; switchroom EMITS `en` on every host because upstream's " + "unset default auto-detects across 200+ locales inline on the shared " + "asyncio loop (~165x slower per call), set e.g. `en,es` for i18n parsing; " + "and " + "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS — the max chars of query text " + "dateparser.search_dates() is handed during temporal analysis (0 = " + "unlimited), made live by switchroom's temporal-offload image patch " + "(which moved the synchronous extraction off the shared asyncio loop onto " + "a single-worker thread and bounds its input, ending multi-second loop " + "stalls on multi-KB consolidation queries); unset means the image's baked " + "2000, set 0 to restore an unbounded full scan); and " + "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE — the Postgres " + "text-search regconfig the native BM25 arms and the search_vector index " + "write run to_tsvector/to_tsquery against; unset means upstream's " + "always-safe `english`, set `hindsight_english` to use switchroom's " + "junk-stopping config (dict + stopword file provisioned durably by the " + "image entrypoint, recall degrading to semantic-only if it is ever " + "missing); and " + "HINDSIGHT_API_BM25_MAX_QUERY_TERMS — the max number of query tokens the " + "keyword (BM25) recall arm searches for; unset means upstream's 0 = " + "unbounded, which makes a multi-KB consolidation query a disjunction " + "over hundreds of terms; set e.g. 64 to cap it, at the cost of " + "silently dropping the tail of a long query; and " + "HINDSIGHT_API_EMBEDDINGS_PROVIDER with the four " + "HINDSIGHT_API_EMBEDDINGS_ONNX_{MODEL_ID,POOLING,QUERY_PREFIX," + "PASSAGE_PREFIX} keys — the ONNX-embeddings cutover: unset keeps " + "upstream's slower `local` (PyTorch) backend, set `onnx` plus the model " + "id/pooling to run the SAME model on ONNX-CPU at a fraction of the " + "latency; the two _PREFIX keys are the query/passage affixes and are " + "set to the empty string for bge parity (vectors are byte-identical " + "between the two backends, so no re-embedding is needed) — unusually " + "for this block an EMPTY value here is honoured, not dropped, because an " + "affix-free prefix is the intended value; and the v0.9.0 knobs " + "switchroom allowlists for reach but emits no value for — " + "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_{LOW,MID,HIGH} (per-recall-budget " + "rerank caps; unset/0 falls back to the global " + "HINDSIGHT_API_RERANKER_MAX_CANDIDATES switchroom does pin), " + "HINDSIGHT_API_EMBEDDINGS_MAX_INPUT_TOKENS (embedding-input truncation " + "ceiling; unset means the model's own limit), " + "HINDSIGHT_API_ENTITY_TRGM_SIMILARITY_THRESHOLD (entity-resolution fuzzy " + "match floor; unset means upstream's 0.15) and " + "HINDSIGHT_API_ENABLE_{TEMPORAL_RETRIEVAL,GRAPH_RETRIEVAL,RERANKING} " + "(global per-stage kill switches, all default true — the emergency lever " + "if one retrieval stage is implicated in a latency incident, at the cost " + "of recall quality); plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -20066,6 +20072,7 @@ var HINDSIGHT_DEFAULT_LLM_SUPPORTS_MAX_ITEMS = "false";
 var HINDSIGHT_DEFAULT_RECENCY_DECAY_FUNCTION = "exponential";
 var HINDSIGHT_DEFAULT_RECENCY_DECAY_HALFLIFE_DAYS = 30;
 var HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION = "pg_search";
+var HINDSIGHT_DEFAULT_QUERY_ANALYZER_LANGUAGES = "en";
 var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -20132,7 +20139,11 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
     "HINDSIGHT_API_GRAPH_SEED_MIN_SIMILARITY",
     String(HINDSIGHT_DEFAULT_GRAPH_SEED_MIN_SIMILARITY)
   ],
-  ["HINDSIGHT_API_TEXT_SEARCH_EXTENSION", HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION]
+  ["HINDSIGHT_API_TEXT_SEARCH_EXTENSION", HINDSIGHT_DEFAULT_TEXT_SEARCH_EXTENSION],
+  [
+    "HINDSIGHT_API_QUERY_ANALYZER_LANGUAGES",
+    HINDSIGHT_DEFAULT_QUERY_ANALYZER_LANGUAGES
+  ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_GPU = [
   ["HINDSIGHT_API_RERANKER_LOCAL_FP16", HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16],
@@ -20167,10 +20178,26 @@ var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_API_RETAIN_WALL_TIMEOUT",
   "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT",
   "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S",
-  "HINDSIGHT_API_TEMPORAL_LANGUAGES",
   "HINDSIGHT_API_TEMPORAL_MAX_QUERY_CHARS",
   "HINDSIGHT_API_TEXT_SEARCH_EXTENSION_NATIVE_LANGUAGE",
-  "HINDSIGHT_API_BM25_MAX_QUERY_TERMS"
+  "HINDSIGHT_API_BM25_MAX_QUERY_TERMS",
+  "HINDSIGHT_API_EMBEDDINGS_PROVIDER",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_MODEL_ID",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_POOLING",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_QUERY_PREFIX",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_PASSAGE_PREFIX",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_LOW",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_MID",
+  "HINDSIGHT_API_RERANKER_MAX_CANDIDATES_HIGH",
+  "HINDSIGHT_API_EMBEDDINGS_MAX_INPUT_TOKENS",
+  "HINDSIGHT_API_ENTITY_TRGM_SIMILARITY_THRESHOLD",
+  "HINDSIGHT_API_ENABLE_TEMPORAL_RETRIEVAL",
+  "HINDSIGHT_API_ENABLE_GRAPH_RETRIEVAL",
+  "HINDSIGHT_API_ENABLE_RERANKING"
+]);
+var HINDSIGHT_PERF_ALLOW_EMPTY_KEYS = new Set([
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_QUERY_PREFIX",
+  "HINDSIGHT_API_EMBEDDINGS_ONNX_PASSAGE_PREFIX"
 ]);
 var HINDSIGHT_WORKER_SLOT_TYPES = [
   "consolidation",
@@ -20184,6 +20211,15 @@ var HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES = new Map(HINDSIGHT_WORKER_SLOT_TYPES
   `HINDSIGHT_API_WORKER_${type.toUpperCase()}_MAX_SLOTS`,
   `HINDSIGHT_API_WORKER_${type.toUpperCase()}_RESERVED_SLOTS`
 ]));
+var HINDSIGHT_RECALL_POOL_ENV_KEYS = [
+  "HINDSIGHT_API_WORKERS",
+  "HINDSIGHT_API_WORKER_ENABLED",
+  "HINDSIGHT_API_DATABASE_URL",
+  "HINDSIGHT_API_DB_URL",
+  "HINDSIGHT_API_DB_POOL_MAX_SIZE",
+  "HINDSIGHT_API_READ_DB_POOL_MAX_SIZE",
+  "HINDSIGHT_API_RUN_MIGRATIONS_ON_STARTUP"
+];
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...[
     ...HINDSIGHT_PERF_DEFAULTS_UNGATED,
@@ -20192,7 +20228,8 @@ var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ].map(([k]) => k),
   ...HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS,
   ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.keys(),
-  ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.values()
+  ...HINDSIGHT_WORKER_RESERVED_SLOT_ALIASES.values(),
+  ...HINDSIGHT_RECALL_POOL_ENV_KEYS
 ]);
 
 // src/setup/hindsight-context-budget.ts

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "switchroom",
   "//version": "NOT the release version — source of truth is the git tag, resolved by scripts/build.mjs:resolveVersion() (see CLAUDE.md > Standard release process). This field is stale by design and only the Layer-4 dev/non-tag fallback for build.mjs + src/cli/resolve-version.ts; do NOT bump it expecting a release to pick it up. npm-pack tarball naming needs a real version — do that as an UNCOMMITTED pack-time bump (see release step 6), never a committed one.",
-  "version": "0.20.17",
+  "version": "0.20.19",
   "description": "Run Claude Code 24/7 on your Claude Pro/Max subscription over Telegram. Open-source alternative to OpenClaw and NanoClaw — no API keys.",
   "type": "module",
   "bin": {

package/profiles/_shared/dev-protocol.md.hbs

@@ -7,7 +7,7 @@ This governs HOW delegated work is done, and is **not a license to do it inline*
 - **Orient before you build — validate, don't assume.** Root-cause in real source — the repo's files at HEAD, not `dist/`, caches, or generated output. If the evidence contradicts your theory, the plan, or the task description, report the contradiction rather than force-fitting it. Cite `file:line`, commits, PRs.
 - **Infer before you ask.** Most questions are answerable from the code and history. If genuinely unsure, ask ONE question at a time, during planning; act autonomously during execution — make the reasonable call and note the assumption.
 - **Design report before implementation, when the change is large.** Architecturally significant, cross-cutting, or ambiguous-approach work earns an evidence-grounded design report and a red-team pass before code, and ships as focused single-concern PRs.
-- **Run `git fetch origin` first, then branch off `origin/main`.** Never branch off the working copy as-is: you do not start in the repo you are coding in, and the checkout on disk may be many commits stale. Scoped tests + lint locally; CI is the full-suite authority.
+- **Run `git fetch origin` first, then branch off `origin/main`.** Never use the dead `mekenthompson/switchroom` fork as a remote, base, or push target. Never branch off the working copy as-is: you do not start in the repo you are coding in, and the checkout on disk may be many commits stale; remove worktrees when done, don't orphan them. Scoped tests + lint locally; CI is the full-suite authority.
 - **Adversarial review of the diff.** Read every change as an adversary would: what breaks, what's untested, what's inconsistent.
 - **Blockers and majors block the merge; prefer fixing lows before merge too** — ship reliable code with a slight delay over filing tech debt that never gets addressed. Fix a low now unless it's genuinely out of scope or disproportionate to the change; only then file it as a tracked follow-up (filing is mandatory, and a filed low is a commitment to fix it, not a way to forget it). Fix what blocks, then merge on CI green. Do not count review rounds and do not run a mandatory re-review pass: verifying your own fix is part of making it, not a separate step. When that fix-vs-defer call is a genuine judgment call, don't decide it silently — surface it to the user as choice / user-impact / recommendation: name the options (fix now vs defer), their user-outcome impact, and which you recommend.
 - **Durable fixes over hack patches. Deterministic mechanisms over model-dependent behaviour** — if code can enforce a guarantee, don't leave it to prompt discipline. **Tests assert outcomes, not just code paths:** one that wouldn't fail on the bug it guards is not a test.

package/profiles/coding/CLAUDE.md.hbs

@@ -40,8 +40,8 @@ You are a senior software engineering agent. You write, review, debug, and archi
 
 Claude Code's file-based auto-memory is disabled. Use **Hindsight** MCP tools:
 
-- `mcp__hindsight__recall` — search past memories. Auto-fires on every message.
-- `mcp__hindsight__retain` — store important facts. Auto-retain runs in chunked mode — it fires every 3rd turn by default and only processes the recent window (~4 turns), so it's prompt and cheap.
+- `mcp__hindsight__recall` — search past memories. Auto-fires on most inbound messages — but not on bare acks, very short prompts, or stateless trivia ("what time is it"), so no memory block doesn't mean no memory.
+- `mcp__hindsight__retain` — store important facts. Auto-retain runs in chunked mode — it fires every Nth turn (N defaults to 3 but is per-agent — `config_get` shows your `memory.retain.every_n_turns`) and only processes that window plus one overlap turn, so it's prompt and cheap.
 - `mcp__hindsight__reflect` — synthesize across memories for complex queries.
 - `mcp__switchroom-telegram__mental_model_propose` — propose a mental model: a standing semantic summary refreshed over the bank (e.g. "codebase architecture"). Mental-model writes are operator-approved: this posts an approval card and persists on approval. Don't call `mcp__hindsight__create_mental_model`/`update_mental_model` directly (they're denied and redirected here).
 

package/profiles/default/CLAUDE.md.hbs

@@ -35,11 +35,11 @@ You are operating in the **{{topicName}}** {{#if topicEmoji}}{{topicEmoji}} {{/i
 
 **Claude Code's built-in file-based auto-memory is disabled for this agent.** Don't try to write `.md` files under `.claude/projects/.../memory/` or maintain a `MEMORY.md` index — that whole system is off. There's exactly one memory backend: **Hindsight**.
 
-Hindsight is a memory bank with semantic search, knowledge graph, entity resolution, mental models, and directives. You talk to it through MCP tools (all pre-approved):
+Hindsight is a memory bank with semantic search, knowledge graph, entity resolution, mental models, and directives. You talk to it through MCP tools — read / retain / reflect / directive are pre-approved (allow-listed, not a wildcard); the rest isn't: mental-model writes redirect (below), and destructive `delete_*` / `clear_*` ops raise an approval card.
 
 ### Day-to-day tools
-- `mcp__hindsight__recall` — semantic-search the bank for relevant past memories. Auto-fires on every inbound user message via the plugin's UserPromptSubmit hook (you'll see "Relevant memories from past conversations" in your context). Call manually when you need a more specific query than the auto-fired one.
-- `mcp__hindsight__retain` — store a new memory. The plugin auto-retains via the Stop hook in chunked mode — every 3rd turn by default, processing only the recent window (~4 turns) — so it captures memory promptly and survives restarts without re-sending the whole transcript, and you usually don't need this. Call manually for significant decisions, corrections, or facts you want immediately searchable.
+- `mcp__hindsight__recall` — semantic-search past memories. Auto-fires on MOST inbound messages via the UserPromptSubmit hook, but skips short prompts, bare acks, and trivia — a skipped turn injects nothing, so an absent memory block means nothing. Call manually for a specific query or a skipped turn needing memory.
+- `mcp__hindsight__retain` — store a new memory. The plugin auto-retains via the Stop hook — every Nth turn (that window plus one overlap), so it usually handles capture and survives restarts. N is per-agent (default 3, some 8; `config_get` shows your `memory.retain.every_n_turns`, so don't assume 3). Call manually for significant decisions, corrections, or facts you want immediately searchable.
 - `mcp__hindsight__reflect` — Hindsight's LLM-powered "answer this query using the bank's content + directives". Use when the user asks a question that requires synthesis across multiple past memories.
 
 ### Mental Models
@@ -74,7 +74,7 @@ Auto-retain (see above) covers routine capture; use manual `retain` for high-sig
 
 Auto-recall and auto-retain feed the bank but never *synthesize* — that's on you, only if you act on these triggers. Each has a backstop:
 
-- **Reflect instead of hand-assembling.** About to fire 2+ manual `recall`s for one answer ("summarize where Y stands")? Call `mcp__hindsight__reflect` instead. (Backstop: auto-recall injects the top hits every turn — reflect is the escalation.)
+- **Reflect instead of hand-assembling.** About to fire 2+ manual `recall`s for one answer ("summarize where Y stands")? Call `mcp__hindsight__reflect` instead. (Backstop: auto-recall injects the top hits on every non-skipped turn — reflect is the escalation.)
 - **Propose a model when you keep re-deriving.** Rebuilt the *same standing answer* across sessions? Propose a mental model via `mcp__switchroom-telegram__mental_model_propose(name, source_query)` (or run the `mental-model-curator` skill). Not for a one-off fact (`retain`) or identity (profile banks own that).
 - **Merge or retire directives when they pile up.** Directives cap at `MAX_DIRECTIVES=30` active per bank — past that the lowest-priority ones drop from recall (silently — the recall hook's stderr warning is swallowed by Claude Code; the visible signals are `directives_omitted` on the recall_log row and `switchroom doctor`). When they overlap or read stale, run the `mental-model-curator` merge/retire pass (deletes stay operator-approved). (Backstop: `switchroom doctor` WARNs at >24, FAILs at >30.)
 
@@ -83,7 +83,7 @@ Auto-recall and auto-retain feed the bank but never *synthesize* — that's on y
 By default, every restart starts a **fresh `claude` session** — the in-flight transcript is NOT carried over (`session_continuity.resume_mode: handoff`, the default since switchroom #362). Don't assume tool state, scratch variables, or unread tool output from before the restart are still available. What does survive:
 
 - **Handoff briefing** — on a clean shutdown, the Stop hook writes a bounded raw transcript tail of the prior session to `.handoff.md`. On boot, start.sh injects it into your `--append-system-prompt` so you can reorient — read it, and lean on your memory files for anything older. If `.handoff.md` is missing or stale (fresh agent, or pre-Stop-hook crash), `start.sh` runs `handoff-briefing.sh` to assemble `.handoff-briefing.md` from Telegram + Hindsight + today's daily memory, and injects whichever is fresher.
-- **Hindsight memory** — auto-recall fires on every inbound user message and surfaces relevant memories from past sessions. Long-term facts, decisions, and mental models live here, not in the transcript.
+- **Hindsight memory** — auto-recall fires on inbound user messages (minus skips) and surfaces memories from past sessions. Long-term facts, decisions, and mental models live here, not in the transcript.
 - **Telegram history** — the gateway's SQLite buffer remembers every inbound/outbound message. Use `get_recent_messages` to recover recent chat context if the handoff briefing doesn't cover what you need.
 - **Boot-resume inbound** — if your previous session was killed mid-turn, the gateway wakes you on its own with a synthesized inbound (you'll see `<channel source="resume_interrupted">` or `<channel source="resume_watchdog_timeout">`). You don't poll for this — it arrives as your first turn. Two cases, and the inbound text spells out which:
   - **`resume_interrupted`** (operator restart / SIGTERM / crash): pick the work back up and carry it to completion. Briefly tell the user you're resuming and roughly how long ago it was interrupted — then just do it. Do NOT ask whether to resume.

package/profiles/executive-assistant/CLAUDE.md.hbs

@@ -39,8 +39,8 @@ You help the user stay organized, prepared, and focused on high-leverage work. Y
 
 Claude Code's file-based auto-memory is disabled. Use **Hindsight** MCP tools:
 
-- `mcp__hindsight__recall` — search past memories. Auto-fires every message.
-- `mcp__hindsight__retain` — store important facts. Auto-retain runs in chunked mode — it fires every 3rd turn by default and only processes the recent window (~4 turns), so it's prompt and cheap.
+- `mcp__hindsight__recall` — search past memories. Auto-fires on most inbound messages — but not on bare acks, very short prompts, or stateless trivia ("what time is it"), so no memory block doesn't mean no memory.
+- `mcp__hindsight__retain` — store important facts. Auto-retain runs in chunked mode — it fires every Nth turn (N defaults to 3 but is per-agent — `config_get` shows your `memory.retain.every_n_turns`) and only processes that window plus one overlap turn, so it's prompt and cheap.
 - `mcp__switchroom-telegram__mental_model_propose` — propose a mental model: a standing semantic summary refreshed over the bank (e.g. "contacts", "active projects", "user preferences"). Mental-model writes are operator-approved: this posts an approval card and persists on approval. Don't call `mcp__hindsight__create_mental_model`/`update_mental_model` directly (they're denied and redirected here).
 
 Save proactively: contacts and their roles, scheduling preferences, project status, decisions with rationale, communication templates. Only Hindsight memories survive compaction.

package/profiles/health-coach/CLAUDE.md.hbs

@@ -33,8 +33,8 @@ Recommend the user consult a professional for: persistent pain/injury, medical c
 
 Claude Code's file-based auto-memory is disabled. Use **Hindsight** MCP tools:
 
-- `mcp__hindsight__recall` — search past memories. Auto-fires on every message.
-- `mcp__hindsight__retain` — store important facts. Auto-retain runs in chunked mode — it fires every 3rd turn by default and only processes the recent window (~4 turns), so it's prompt and cheap.
+- `mcp__hindsight__recall` — search past memories. Auto-fires on most inbound messages — but not on bare acks, very short prompts, or stateless trivia ("what time is it"), so no memory block doesn't mean no memory.
+- `mcp__hindsight__retain` — store important facts. Auto-retain runs in chunked mode — it fires every Nth turn (N defaults to 3 but is per-agent — `config_get` shows your `memory.retain.every_n_turns`) and only processes that window plus one overlap turn, so it's prompt and cheap.
 - `mcp__switchroom-telegram__mental_model_propose` — propose a mental model (e.g. a "fitness profile" — a standing semantic summary refreshed over the bank). Mental-model writes are operator-approved: this posts an approval card and persists on approval. Don't call `mcp__hindsight__create_mental_model`/`update_mental_model` directly (they're denied and redirected here).
 
 Save proactively: workout logs, goals, PRs, preferences, patterns (e.g. "poor sleep on Sundays"), injuries/limitations. Only Hindsight memories survive session compaction.