switchroom 0.20.20 → 0.20.22

package/dist/agent-scheduler/index.js

@@ -11128,22 +11128,22 @@ var AgentMemorySchema = exports_external.object({
   observation_scopes: ObservationScopesSchema,
   observation_scope_strategy: ObservationScopeStrategySchema,
   recall: exports_external.object({
-    max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+    max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
     cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
     hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
     parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling — including " + "one set EQUAL to it — is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
     query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout — 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
     query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
     request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-    own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-    additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+    own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (≈ ⅓ of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+    additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (≈ ⅙ of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
     min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
     min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
     types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
     additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
     sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
     skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-    budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+    budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking — low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3× " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
     max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
     prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
     context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),

package/dist/auth-broker/index.js

@@ -11153,22 +11153,22 @@ var init_schema = __esm(() => {
     observation_scopes: ObservationScopesSchema,
     observation_scope_strategy: ObservationScopeStrategySchema,
     recall: exports_external.object({
-      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
       parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling — including " + "one set EQUAL to it — is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
       query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout — 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
       query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
       request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (≈ ⅓ of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (≈ ⅙ of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
       min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
       min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
       types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
       additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
       skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-      budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+      budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking — low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3× " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
       max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
       prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
       context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
@@ -19033,7 +19033,7 @@ function allocateAgentUid(name) {
 var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
 var HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT = "Relevant memories from past conversations (prioritize recent when " + "conflicting). Only use memories that are directly useful to continue " + "this conversation; ignore the rest:";
 var RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
-  budget: "low",
+  budget: "mid",
   maxTokens: 1024,
   preferObservations: true,
   contextTurns: 2,

package/dist/cli/notion-write-pretool.mjs

@@ -11889,22 +11889,22 @@ var AgentMemorySchema = exports_external.object({
   observation_scopes: ObservationScopesSchema,
   observation_scope_strategy: ObservationScopeStrategySchema,
   recall: exports_external.object({
-    max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+    max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
     cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
     hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
     parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling \u2014 including " + "one set EQUAL to it \u2014 is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
     query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout \u2014 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
     query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
     request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead \u2014 10 at the " + "shipped ceiling \u2014 because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-    own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "\u2014 the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-    additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` \u2014 same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+    own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "\u2014 the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (\u2248 \u2153 of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+    additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` \u2014 same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (\u2248 \u2159 of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
     min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' \u2014 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls \u2014 which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
     min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) \u2014 only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" \u2014 every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
     types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] \u2014 the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
     additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results \u2014 e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
     sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender \u2192 extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username \u2014 a leading @ is " + "optional \u2014 or numeric user_id), merged " + "into its own results \u2014 so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
     skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true \u2014 saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-    budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns \u2014 the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+    budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking \u2014 low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3\u00d7 " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
     max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` \u2014 on its own it buys nothing once the count cap " + "binds."),
     prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing \u2014 useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
     context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver \u2014 `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),

package/dist/cli/switchroom.js

@@ -2120,7 +2120,7 @@ var init_esm = __esm(() => {
 });
 
 // src/build-info.ts
-var VERSION = "0.20.20", COMMIT_SHA = "f444ea4a";
+var VERSION = "0.20.22", COMMIT_SHA = "ad3f50bd";
 
 // src/cli/resolve-version.ts
 import { existsSync, readFileSync } from "node:fs";
@@ -13754,22 +13754,22 @@ var init_schema = __esm(() => {
     observation_scopes: ObservationScopesSchema,
     observation_scope_strategy: ObservationScopeStrategySchema,
     recall: exports_external.object({
-      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
       parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling \u2014 including " + "one set EQUAL to it \u2014 is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
       query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout \u2014 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
       query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
       request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead \u2014 10 at the " + "shipped ceiling \u2014 because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "\u2014 the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` \u2014 same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "\u2014 the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (\u2248 \u2153 of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` \u2014 same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (\u2248 \u2159 of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
       min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' \u2014 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls \u2014 which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
       min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) \u2014 only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" \u2014 every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
       types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] \u2014 the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
       additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results \u2014 e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender \u2192 extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username \u2014 a leading @ is " + "optional \u2014 or numeric user_id), merged " + "into its own results \u2014 so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
       skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true \u2014 saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-      budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns \u2014 the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+      budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking \u2014 low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3\u00d7 " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
       max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` \u2014 on its own it buys nothing once the count cap " + "binds."),
       prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing \u2014 useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
       context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver \u2014 `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
@@ -21194,7 +21194,7 @@ var init_hindsight_recall_passthrough = __esm(() => {
   RECALL_BUDGETS = ["low", "mid", "high"];
   RECALL_TAGS_MATCH_MODES = ["any", "all", "any_strict", "all_strict"];
   RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
-    budget: "low",
+    budget: "mid",
     maxTokens: 1024,
     preferObservations: true,
     contextTurns: 2,
@@ -30161,6 +30161,16 @@ function applyHindsightHooksOverrides(pluginDestPath, tunables) {
     return;
   writeFileSyncIfChanged(hooksPath, next);
 }
+function recallSlotFloorFromCap(maxMemories, divisor) {
+  const cap = Number.isFinite(maxMemories) && maxMemories > 0 ? maxMemories : DEFAULT_RECALL_MAX_MEMORIES;
+  return Math.max(1, Math.round(cap / divisor));
+}
+function defaultRecallOwnBankMinSlots(maxMemories) {
+  return recallSlotFloorFromCap(maxMemories, RECALL_OWN_BANK_SLOT_DIVISOR);
+}
+function defaultRecallAdditionalBankMinSlots(maxMemories) {
+  return recallSlotFloorFromCap(maxMemories, RECALL_ADDITIONAL_BANK_SLOT_DIVISOR);
+}
 function applyHindsightSettingsOverrides(pluginDestPath, additionalBanks, retainConfig, recallTunables, recallCaps, observationScopeSettings) {
   const settingsPath = join12(pluginDestPath, "settings.json");
   if (!existsSync18(settingsPath))
@@ -30192,8 +30202,8 @@ function renderHindsightSettingsOverrides(raw, additionalBanks, retainConfig, re
   settings.recallParallelDeadlineSeconds = recallTunables.parallelDeadlineSeconds;
   settings.recallRequestTimeoutSeconds = recallTunables.requestTimeoutSeconds;
   settings.recallTypes = ["world", "experience", "observation"];
-  settings.recallOwnBankMinSlots = HINDSIGHT_OWN_BANK_MIN_SLOTS_DEFAULT;
-  settings.recallAdditionalBankMinSlots = HINDSIGHT_ADDITIONAL_BANK_MIN_SLOTS_DEFAULT;
+  settings.recallOwnBankMinSlots = defaultRecallOwnBankMinSlots(recallCaps.maxMemories);
+  settings.recallAdditionalBankMinSlots = defaultRecallAdditionalBankMinSlots(recallCaps.maxMemories);
   settings.recallSkipTrivial = true;
   settings.directiveCaptureNudge = true;
   settings.recallTagWeights = { ...HINDSIGHT_RECALL_TAG_WEIGHT_SEED };
@@ -30569,8 +30579,9 @@ function scaffoldAgent(name, agentConfigRaw, agentsDir, telegramConfig, switchro
   const hindsightRecallQueryMaxTokens = agentConfig.memory?.recall?.query_max_tokens ?? HINDSIGHT_RECALL_QUERY_MAX_TOKENS_DEFAULT;
   const rawRecallQueryStopTerms = agentConfig.memory?.recall?.query_stop_terms;
   const hindsightRecallQueryStopTermsJson = JSON.stringify(rawRecallQueryStopTerms ?? []);
-  const hindsightRecallOwnBankMinSlots = agentConfig.memory?.recall?.own_bank_min_slots ?? HINDSIGHT_OWN_BANK_MIN_SLOTS_DEFAULT;
-  const hindsightRecallAdditionalBankMinSlots = agentConfig.memory?.recall?.additional_bank_min_slots ?? HINDSIGHT_ADDITIONAL_BANK_MIN_SLOTS_DEFAULT;
+  const hindsightRecallEffectiveMaxMemories = resolveHindsightRecallCaps(agentConfig.memory?.recall).maxMemories;
+  const hindsightRecallOwnBankMinSlots = agentConfig.memory?.recall?.own_bank_min_slots ?? defaultRecallOwnBankMinSlots(hindsightRecallEffectiveMaxMemories);
+  const hindsightRecallAdditionalBankMinSlots = agentConfig.memory?.recall?.additional_bank_min_slots ?? defaultRecallAdditionalBankMinSlots(hindsightRecallEffectiveMaxMemories);
   const hindsightRecallMinScore = agentConfig.memory?.recall?.min_score ?? HINDSIGHT_RECALL_MIN_SCORE_DEFAULT;
   const hindsightRecallMinScoreScope = agentConfig.memory?.recall?.min_score_scope ?? HINDSIGHT_RECALL_MIN_SCORE_SCOPE_DEFAULT;
   const hindsightRecallPassthrough = resolveHindsightRecallPassthrough(agentConfig.memory?.recall);
@@ -31574,8 +31585,9 @@ function reconcileAgentInner(name, agentConfigRaw, agentsDir, telegramConfig, sw
   const hindsightRecallQueryMaxTokens = agentConfig.memory?.recall?.query_max_tokens ?? HINDSIGHT_RECALL_QUERY_MAX_TOKENS_DEFAULT;
   const rawRecallQueryStopTerms = agentConfig.memory?.recall?.query_stop_terms;
   const hindsightRecallQueryStopTermsJson = JSON.stringify(rawRecallQueryStopTerms ?? []);
-  const hindsightRecallOwnBankMinSlots = agentConfig.memory?.recall?.own_bank_min_slots ?? HINDSIGHT_OWN_BANK_MIN_SLOTS_DEFAULT;
-  const hindsightRecallAdditionalBankMinSlots = agentConfig.memory?.recall?.additional_bank_min_slots ?? HINDSIGHT_ADDITIONAL_BANK_MIN_SLOTS_DEFAULT;
+  const hindsightRecallEffectiveMaxMemories = resolveHindsightRecallCaps(agentConfig.memory?.recall).maxMemories;
+  const hindsightRecallOwnBankMinSlots = agentConfig.memory?.recall?.own_bank_min_slots ?? defaultRecallOwnBankMinSlots(hindsightRecallEffectiveMaxMemories);
+  const hindsightRecallAdditionalBankMinSlots = agentConfig.memory?.recall?.additional_bank_min_slots ?? defaultRecallAdditionalBankMinSlots(hindsightRecallEffectiveMaxMemories);
   const hindsightRecallMinScore = agentConfig.memory?.recall?.min_score ?? HINDSIGHT_RECALL_MIN_SCORE_DEFAULT;
   const hindsightRecallMinScoreScope = agentConfig.memory?.recall?.min_score_scope ?? HINDSIGHT_RECALL_MIN_SCORE_SCOPE_DEFAULT;
   const hindsightRecallPassthrough = resolveHindsightRecallPassthrough(agentConfig.memory?.recall);
@@ -32762,7 +32774,7 @@ plainer wins.
 Every turn that answers a user message ends with a user-visible \`reply\`
 \u2014 Telegram is all the user sees; your terminal output
 never reaches them.`, TELEGRAM_ENV_PLACEHOLDER = `# Set your bot token: TELEGRAM_BOT_TOKEN=your-token-here
-`, SWITCHROOM_TELEGRAM_MCP_TOOLS, HINDSIGHT_MCP_TOOLS, AGENT_CONFIG_MCP_TOOLS, HOSTD_MCP_TOOLS, WEBKITE_MCP_TOOLS, CONTEXT7_MCP_TOOLS, CONTEXT7_BLANKET_TOKENS, LEGACY_SWITCHROOM_MCP_TOKENS, LEGACY_HOSTD_BLANKET_TOKENS, DEFAULT_READ_ONLY_PREAPPROVED_TOOLS, NATIVE_WEBSEARCH_PREAPPROVED_TOOLS, WEBKITE_FLEET_DENY_TOOLS, INTERACTIVE_TUI_FLEET_DENY_TOOLS, WEBKITE_BINARY_CONTAINER_PATH = "/usr/local/bin/webkite", SWITCHROOM_DEFAULT_MAIN_MODEL = "claude-sonnet-5", SWITCHROOM_DEFAULT_THINKING_EFFORT = "low", CLAUDE_MD_YOURS_PLACEHOLDER, pendingBankOps, SWITCHROOM_OWNED_SETTINGS_KEYS, ALL_BUILTIN_TOOLS, HINDSIGHT_DEFAULT_RETAIN_EVERY_N_TURNS = 3, HINDSIGHT_DEFAULT_RETAIN_OVERLAP_TURNS = 1, HINDSIGHT_OWN_BANK_MIN_SLOTS_DEFAULT = 2, HINDSIGHT_ADDITIONAL_BANK_MIN_SLOTS_DEFAULT = 1, HINDSIGHT_RECALL_MIN_SCORE_DEFAULT = 0, HINDSIGHT_RECALL_MIN_SCORE_SCOPE_DEFAULT = "degraded", VALID_SETTINGS_DEFAULT_MODES, SWITCHROOM_DEFAULT_PERMISSION_MODE = "acceptEdits", DOCKER_TELEGRAM_PLUGIN_PATH = "/opt/switchroom/telegram-plugin", DOCKER_HOOKS_PATH, DOCKER_BUNDLED_HOOKS_PATH = "/opt/switchroom/hooks", DOCKER_BIN_PATH = "/opt/switchroom/bin", DOCKER_SECURITY_PLUGIN_PATH = "/opt/switchroom/security-plugin", DOCKER_CONFIG_PATH = "/state/config/switchroom.yaml", DOCKER_SWITCHROOM_CLI_PATH = "/usr/local/bin/switchroom", DOCKER_AUTH_BROKER_SOCKET = "/run/switchroom/auth-broker/sock", DOCKER_VAULT_BROKER_SOCKET = "/run/switchroom/broker/sock", DOCKER_AGENT_HOME = "/state/agent/home", CONTEXT7_MCP_URL = "https://mcp.context7.com/mcp", INTEGRATION_MCP_RESOLVERS;
+`, SWITCHROOM_TELEGRAM_MCP_TOOLS, HINDSIGHT_MCP_TOOLS, AGENT_CONFIG_MCP_TOOLS, HOSTD_MCP_TOOLS, WEBKITE_MCP_TOOLS, CONTEXT7_MCP_TOOLS, CONTEXT7_BLANKET_TOKENS, LEGACY_SWITCHROOM_MCP_TOKENS, LEGACY_HOSTD_BLANKET_TOKENS, DEFAULT_READ_ONLY_PREAPPROVED_TOOLS, NATIVE_WEBSEARCH_PREAPPROVED_TOOLS, WEBKITE_FLEET_DENY_TOOLS, INTERACTIVE_TUI_FLEET_DENY_TOOLS, WEBKITE_BINARY_CONTAINER_PATH = "/usr/local/bin/webkite", SWITCHROOM_DEFAULT_MAIN_MODEL = "claude-sonnet-5", SWITCHROOM_DEFAULT_THINKING_EFFORT = "low", CLAUDE_MD_YOURS_PLACEHOLDER, pendingBankOps, SWITCHROOM_OWNED_SETTINGS_KEYS, ALL_BUILTIN_TOOLS, HINDSIGHT_DEFAULT_RETAIN_EVERY_N_TURNS = 3, HINDSIGHT_DEFAULT_RETAIN_OVERLAP_TURNS = 1, RECALL_OWN_BANK_SLOT_DIVISOR = 3, RECALL_ADDITIONAL_BANK_SLOT_DIVISOR = 6, HINDSIGHT_RECALL_MIN_SCORE_DEFAULT = 0, HINDSIGHT_RECALL_MIN_SCORE_SCOPE_DEFAULT = "degraded", VALID_SETTINGS_DEFAULT_MODES, SWITCHROOM_DEFAULT_PERMISSION_MODE = "acceptEdits", DOCKER_TELEGRAM_PLUGIN_PATH = "/opt/switchroom/telegram-plugin", DOCKER_HOOKS_PATH, DOCKER_BUNDLED_HOOKS_PATH = "/opt/switchroom/hooks", DOCKER_BIN_PATH = "/opt/switchroom/bin", DOCKER_SECURITY_PLUGIN_PATH = "/opt/switchroom/security-plugin", DOCKER_CONFIG_PATH = "/state/config/switchroom.yaml", DOCKER_SWITCHROOM_CLI_PATH = "/usr/local/bin/switchroom", DOCKER_AUTH_BROKER_SOCKET = "/run/switchroom/auth-broker/sock", DOCKER_VAULT_BROKER_SOCKET = "/run/switchroom/broker/sock", DOCKER_AGENT_HOME = "/state/agent/home", CONTEXT7_MCP_URL = "https://mcp.context7.com/mcp", INTEGRATION_MCP_RESOLVERS;
 var init_scaffold = __esm(() => {
   init_source();
   init_cron_unit_name();

package/dist/host-control/main.js

@@ -11153,22 +11153,22 @@ var init_schema = __esm(() => {
     observation_scopes: ObservationScopesSchema,
     observation_scope_strategy: ObservationScopeStrategySchema,
     recall: exports_external.object({
-      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
       parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling — including " + "one set EQUAL to it — is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
       query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout — 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
       query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
       request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (≈ ⅓ of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (≈ ⅙ of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
       min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
       min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
       types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
       additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
       skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-      budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+      budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking — low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3× " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
       max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
       prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
       context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
@@ -21565,13 +21565,13 @@ function allocateAgentUid(name) {
 }
 
 // src/build-info.ts
-var VERSION = "0.20.20";
+var VERSION = "0.20.22";
 
 // src/setup/hindsight-recall-passthrough.ts
 var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
 var HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT = "Relevant memories from past conversations (prioritize recent when " + "conflicting). Only use memories that are directly useful to continue " + "this conversation; ignore the rest:";
 var RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
-  budget: "low",
+  budget: "mid",
   maxTokens: 1024,
   preferObservations: true,
   contextTurns: 2,

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

@@ -4593,22 +4593,22 @@ var init_schema = __esm(() => {
     observation_scopes: ObservationScopesSchema,
     observation_scope_strategy: ObservationScopeStrategySchema,
     recall: exports_external.object({
-      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
       parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling — including " + "one set EQUAL to it — is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
       query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout — 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
       query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
       request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (≈ ⅓ of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (≈ ⅙ of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
       min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
       min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
       types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
       additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
       skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-      budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+      budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking — low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3× " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
       max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
       prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
       context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
@@ -19466,7 +19466,7 @@ function allocateAgentUid(name) {
 var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
 var HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT = "Relevant memories from past conversations (prioritize recent when " + "conflicting). Only use memories that are directly useful to continue " + "this conversation; ignore the rest:";
 var RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
-  budget: "low",
+  budget: "mid",
   maxTokens: 1024,
   preferObservations: true,
   contextTurns: 2,

package/dist/vault/broker/server.js

@@ -4189,22 +4189,22 @@ var init_schema = __esm(() => {
     observation_scopes: ObservationScopesSchema,
     observation_scope_strategy: ObservationScopeStrategySchema,
     recall: exports_external.object({
-      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
+      max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. The vendor plugin ships " + "12; switchroom's resolver default is 8 when this is unset, and the " + "live fleet raises it via `defaults.memory.recall.max_memories`. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
       parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling — including " + "one set EQUAL to it — is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
       query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout — 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
       query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
       request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
-      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
-      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+      own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents DERIVE this " + "default from the count cap (≈ ⅓ of the effective `max_memories`, " + "floored at 1) so it tracks the cap instead of going stale when the " + "fleet raises it; an explicit value set here still wins over that " + "derived default."),
+      additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents DERIVE this default " + "from the count cap (≈ ⅙ of the effective `max_memories`, floored " + "at 1) so it tracks the cap instead of going stale; an explicit " + "value set here still wins. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
       min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
       min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
       types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
       additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
       skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
-      budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+      budget: exports_external.enum(["low", "mid", "high"]).optional().describe("How deep Hindsight's recall search goes. It tunes the CANDIDATE-" + "POOL depth retrieved before reranking — low (switchroom default) " + "= 100, mid = 300, high = 1000 candidates. The local cross-encoder " + "reranker (ms-marco-MiniLM) runs at EVERY tier; budget does not " + "gate it on or off, and there is no LLM rerank pass in the recall " + "path (the LLM serves reflect/retain/consolidation, not recall " + "ranking). Higher tiers cost more latency purely from the deeper " + "pool: mid measured ~5s of hook latency on real fleet turns (3× " + "the retrieval+rerank pool, not an added model call), the second-" + "largest contributor to perceived dead air after model TTFT. Raise " + "it for an agent whose recall quality matters more than its reply " + "latency (a research or audit role); leave it at low for chat."),
       max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
       prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
       context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
@@ -19707,7 +19707,7 @@ function allocateAgentUid(name) {
 var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
 var HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT = "Relevant memories from past conversations (prioritize recent when " + "conflicting). Only use memories that are directly useful to continue " + "this conversation; ignore the rest:";
 var RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
-  budget: "low",
+  budget: "mid",
   maxTokens: 1024,
   preferObservations: true,
   contextTurns: 2,

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": "0.20.20",
+  "version": "0.20.22",
   "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": {