switchroom 0.20.18 → 0.20.20

package/dist/agent-scheduler/index.js

@@ -11420,6 +11420,12 @@ var HindsightConfigSchema = exports_external.object({
     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."),
+  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).")
 });
 var MicrosoftWorkspaceConfigSchema = exports_external.object({

package/dist/auth-broker/index.js

@@ -11445,6 +11445,12 @@ 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."),
+    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({
@@ -19359,6 +19365,15 @@ 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,
@@ -19367,7 +19382,8 @@ 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/cli/notion-write-pretool.mjs

@@ -12181,6 +12181,12 @@ var HindsightConfigSchema = exports_external.object({
     reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent \u2192 uses global."),
     consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent \u2192 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."),
+  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 \u2014 the pool binds it instead. Absent/false (the " + "default) \u2192 the single-container topology, unchanged. Opt-in only. " + "Raises recall throughput ~2.2x on the fleet (2.6\u21925.6 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 \u2014 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 \u2192 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 \u2192 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` \u2014 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 \u2014 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 \u2014 a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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 \u2014 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) \u2014 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 \u2014 " + "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 \u2014 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 \u2014 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 \u2014 a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
 });
 var MicrosoftWorkspaceConfigSchema = exports_external.object({