switchroom 0.19.24 → 0.19.26

package/dist/host-control/main.js

@@ -19221,9 +19221,15 @@ var AgentMemorySchema = exports_external.object({
   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)."),
     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."),
-    min_overlap: exports_external.number().min(0).max(1).optional().describe("Minimum containment token overlap [0.0–1.0] between the user " + "prompt and a memory's text for the memory to be injected. " + "A cheap FLOOR that removes candidates with (near-)zero lexical " + "relationship to the prompt — NOT a precision control: at the " + "recommended 0.10 the gate is near-passthrough, and the " + "effective precision control is the engine rerank plus the " + "max_memories head-slice. 0.0 disables (default — current " + "behaviour). Use 0.10; values at or above 0.20 measurably " + "starve recall — on production replay 0.20 leaves ~41.9% of " + "turns with NO memories at all, re-creating the bug #3541 " + "fixed. Observe the `overlap_dropped` field via " + "`switchroom memory recall-log`."),
+    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`."),
     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 an 8s timeout 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)."),
+    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."),
     topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
@@ -19446,7 +19452,7 @@ var HindsightConfigSchema = exports_external.object({
     reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
     consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
   }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-  env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+  env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_MAX_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
 });
 var MicrosoftWorkspaceConfigSchema = exports_external.object({
   microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -19582,7 +19588,13 @@ var profileFields = {
     recall: exports_external.object({
       max_memories: exports_external.number().int().min(0).optional(),
       cache_ttl_secs: exports_external.number().int().min(0).optional(),
-      min_overlap: exports_external.number().min(0).max(1).optional(),
+      hook_timeout_seconds: exports_external.number().int().min(1).optional(),
+      parallel_deadline_seconds: exports_external.number().int().min(1).optional(),
+      query_max_tokens: exports_external.number().int().min(0).optional(),
+      query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional(),
+      request_timeout_seconds: exports_external.number().int().min(1).optional(),
+      own_bank_min_slots: exports_external.number().int().min(0).optional(),
+      additional_bank_min_slots: exports_external.number().int().min(0).optional(),
       additional_banks: exports_external.array(exports_external.string()).optional(),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional()
     }).optional()
@@ -19616,7 +19628,8 @@ var profileFields = {
     memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "memory must be a Docker size string like '6g', '512m', '1.5g'").optional().describe("Hard memory cap (Docker `mem_limit` → cgroup memory.max). When the " + "container exceeds this, the kernel OOM-kills processes in the cgroup. " + "Format: '6g', '1.5g', '512m'. When unset at every cascade layer the " + "compose generator falls back to the hard-coded per-profile defaults " + "in src/agents/compose.ts (klanker 6g, coding 2g, conversational 1.5g, " + "lightweight 1g, default 1.5g)."),
     memory_reservation: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "memory_reservation must be a Docker size string like '4g', '256m'").optional().describe("Soft memory floor (Docker `mem_reservation` → cgroup memory.low). " + "Under host-wide memory pressure, the kernel protects at least this " + "much from being reclaimed from the cgroup. Must be ≤ memory. Use to " + "keep an agent RAM-resident when the host has other tenants that " + "might push the box (Coolify apps, build jobs). Default: unset."),
     pids_limit: exports_external.number().int().positive().optional().describe("Max processes the cgroup can spawn (cgroup pids.max). Prevents " + "fork bombs and runaway test runners. Counts every process in the " + "cgroup including bash subprocesses, claude itself, sidecars, and " + "any test/build worker. A typical agent at idle uses ~30 PIDs; " + "`npm test`-style workloads can spike to 200+. Set generously " + "(2000 is a comfortable cap for test-running agents). Default: " + "unset (no cgroup pid cap)."),
-    cpus: exports_external.number().positive().optional().describe("CPU quota (Docker `cpus`). Fractional values OK (e.g. 0.5, 2.0). " + "When unset at every cascade layer the compose generator falls " + "back to the per-profile default (klanker/coding 2.0, default 1.0, " + "lightweight 0.5).")
+    cpus: exports_external.number().positive().optional().describe("CPU quota (Docker `cpus`). Fractional values OK (e.g. 0.5, 2.0). " + "When unset at every cascade layer the compose generator falls " + "back to the per-profile default (klanker/coding 2.0, default 1.0, " + "lightweight 0.5)."),
+    tmp_size: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional().describe("Size of the agent container's RAM-backed `/tmp` (Docker " + "`tmpfs: - /tmp:size=<this>,mode=1777`). The container root FS is " + "read-only, so /tmp is the only scratch space the agent and its " + "sub-agents get for repo clones and toolchain caches (bunx/npx/pip) " + "— a fan-out of several sub-agents exhausts the 1g default. A tmpfs " + "only consumes host RAM for the pages actually written, so raising " + "the ceiling costs nothing until it is used; those pages are still " + "charged to the container's `memory` cap, so raise both together. " + "Format: '1g', '4g', '512m'. Default when unset at every cascade " + "layer: '1g'.")
   }).optional().describe("Per-agent resource limits. Cascades through defaults → profile → " + "per-agent with per-field merge (agent wins on each field independently). " + "Any field left unset at every layer falls back to the hard-coded " + "per-profile defaults in src/agents/compose.ts."),
   experimental: exports_external.object({
     legacy_pty: exports_external.boolean().optional().describe("Opt out of the default tmux supervisor (#725) and run the agent under " + "the legacy PTY supervisor instead. Default: false (tmux is the default)."),
@@ -19709,7 +19722,8 @@ var AgentSchema = exports_external.object({
     memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).optional(),
     memory_reservation: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).optional(),
     pids_limit: exports_external.number().int().positive().optional(),
-    cpus: exports_external.number().positive().optional()
+    cpus: exports_external.number().positive().optional(),
+    tmp_size: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional()
   }).optional()
 }).superRefine((agent, ctx) => {
   if (agent.dm_only !== true)
@@ -21123,8 +21137,11 @@ function allocateAgentUid(name) {
 }
 
 // src/build-info.ts
-var VERSION = "0.19.24";
+var VERSION = "0.19.26";
 
+// src/setup/hindsight-recall-tunables.ts
+var RECALL_DEADLINE_HEADROOM_SECONDS = 2;
+var MIN_RECALL_HOOK_TIMEOUT_SECONDS = RECALL_DEADLINE_HEADROOM_SECONDS + 1;
 // src/cli/agent-config.ts
 import { join as join2 } from "node:path";
 import { homedir as homedir2 } from "node:os";
@@ -21216,6 +21233,27 @@ function assertPositive(value, label) {
   }
 }
 
+// src/setup/hindsight-pg-defaults.ts
+var HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION = 8 * 1024;
+var HINDSIGHT_PG_APP_ANON_MIB = 2560;
+var HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB = 2048;
+var HINDSIGHT_PG_SHARED_BUFFERS_BUDGET_MIB = HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION - HINDSIGHT_PG_APP_ANON_MIB - HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB;
+var HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB = 1536;
+var HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB = 4096;
+function pgMib(mib) {
+  return `${mib}MB`;
+}
+var HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV = "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE";
+var HINDSIGHT_PG_SHARED_BUFFERS_ENV = "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS";
+var HINDSIGHT_PG_DEFAULTS = [
+  [
+    HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV,
+    pgMib(HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB)
+  ],
+  [HINDSIGHT_PG_SHARED_BUFFERS_ENV, pgMib(HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB)]
+];
+var HINDSIGHT_PG_ENV_KEYS = new Set(HINDSIGHT_PG_DEFAULTS.map(([k]) => k));
+
 // src/setup/hindsight-perf-defaults.ts
 var HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION = 150;
 var HINDSIGHT_DEFAULT_RECALL_MAX_CANDIDATES_PER_SOURCE = Math.ceil(HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION * 0.4);
@@ -21230,6 +21268,15 @@ var HINDSIGHT_DEFAULT_RERANKER_LOCAL_BATCH_SIZE = 128;
 var HINDSIGHT_DEFAULT_LLM_STRICT_SCHEMA = "true";
 var HINDSIGHT_DEFAULT_LLM_MAX_RETRIES = 2;
 var HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM = 2;
+var HINDSIGHT_DEFAULT_MAX_OBSERVATIONS_PER_SCOPE = 1000;
+var HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING = "true";
+var HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES = 150;
+var HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4;
+var HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT = 8;
+var HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S = 600;
+var HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND = 500;
+var HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT = 6;
+var HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS = 1;
 var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -21247,6 +21294,42 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_CONSOLIDATION_LLM_PARALLELISM",
     String(HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM)
+  ],
+  [
+    "HINDSIGHT_API_MAX_OBSERVATIONS_PER_SCOPE",
+    String(HINDSIGHT_DEFAULT_MAX_OBSERVATIONS_PER_SCOPE)
+  ],
+  [
+    "HINDSIGHT_API_RERANKER_LOCAL_BUCKET_BATCHING",
+    HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING
+  ],
+  [
+    "HINDSIGHT_API_RERANKER_MAX_CANDIDATES",
+    String(HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES)
+  ],
+  [
+    "HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT",
+    String(HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT)
+  ],
+  [
+    "HINDSIGHT_API_RECALL_MAX_CONCURRENT",
+    String(HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT)
+  ],
+  [
+    "HINDSIGHT_API_REFLECT_WALL_TIMEOUT",
+    String(HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S)
+  ],
+  [
+    "HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS",
+    String(HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS)
+  ],
+  [
+    "HINDSIGHT_API_WORKER_CONSOLIDATION_SLOT_LIMIT",
+    String(HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT)
+  ],
+  [
+    "HINDSIGHT_API_CONSOLIDATION_MAX_MEMORIES_PER_ROUND",
+    String(HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND)
   ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_GPU = [
@@ -21270,7 +21353,8 @@ var HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM = [
   ["HINDSIGHT_API_LLM_MAX_RETRIES", String(HINDSIGHT_DEFAULT_LLM_MAX_RETRIES)]
 ];
 var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
-  "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY"
+  "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY",
+  "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP"
 ]);
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...[
@@ -21281,27 +21365,6 @@ var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS
 ]);
 
-// src/setup/hindsight-pg-defaults.ts
-var HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION = 8 * 1024;
-var HINDSIGHT_PG_APP_ANON_MIB = 2560;
-var HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB = 2048;
-var HINDSIGHT_PG_SHARED_BUFFERS_BUDGET_MIB = HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION - HINDSIGHT_PG_APP_ANON_MIB - HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB;
-var HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB = 1536;
-var HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB = 4096;
-function pgMib(mib) {
-  return `${mib}MB`;
-}
-var HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV = "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE";
-var HINDSIGHT_PG_SHARED_BUFFERS_ENV = "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS";
-var HINDSIGHT_PG_DEFAULTS = [
-  [
-    HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV,
-    pgMib(HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB)
-  ],
-  [HINDSIGHT_PG_SHARED_BUFFERS_ENV, pgMib(HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB)]
-];
-var HINDSIGHT_PG_ENV_KEYS = new Set(HINDSIGHT_PG_DEFAULTS.map(([k]) => k));
-
 // src/setup/hindsight-context-budget.ts
 var HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE = 3000;
 var HINDSIGHT_RETAIN_MAX_COMPLETION_FLOOR = HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE + 72;

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

@@ -4164,9 +4164,15 @@ var init_schema = __esm(() => {
     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)."),
       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."),
-      min_overlap: exports_external.number().min(0).max(1).optional().describe("Minimum containment token overlap [0.0–1.0] between the user " + "prompt and a memory's text for the memory to be injected. " + "A cheap FLOOR that removes candidates with (near-)zero lexical " + "relationship to the prompt — NOT a precision control: at the " + "recommended 0.10 the gate is near-passthrough, and the " + "effective precision control is the engine rerank plus the " + "max_memories head-slice. 0.0 disables (default — current " + "behaviour). Use 0.10; values at or above 0.20 measurably " + "starve recall — on production replay 0.20 leaves ~41.9% of " + "turns with NO memories at all, re-creating the bug #3541 " + "fixed. Observe the `overlap_dropped` field via " + "`switchroom memory recall-log`."),
+      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`."),
       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 an 8s timeout 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)."),
+      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."),
       topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
@@ -4389,7 +4395,7 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_MAX_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -4525,7 +4531,13 @@ var init_schema = __esm(() => {
       recall: exports_external.object({
         max_memories: exports_external.number().int().min(0).optional(),
         cache_ttl_secs: exports_external.number().int().min(0).optional(),
-        min_overlap: exports_external.number().min(0).max(1).optional(),
+        hook_timeout_seconds: exports_external.number().int().min(1).optional(),
+        parallel_deadline_seconds: exports_external.number().int().min(1).optional(),
+        query_max_tokens: exports_external.number().int().min(0).optional(),
+        query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional(),
+        request_timeout_seconds: exports_external.number().int().min(1).optional(),
+        own_bank_min_slots: exports_external.number().int().min(0).optional(),
+        additional_bank_min_slots: exports_external.number().int().min(0).optional(),
         additional_banks: exports_external.array(exports_external.string()).optional(),
         sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional()
       }).optional()
@@ -4559,7 +4571,8 @@ var init_schema = __esm(() => {
       memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "memory must be a Docker size string like '6g', '512m', '1.5g'").optional().describe("Hard memory cap (Docker `mem_limit` → cgroup memory.max). When the " + "container exceeds this, the kernel OOM-kills processes in the cgroup. " + "Format: '6g', '1.5g', '512m'. When unset at every cascade layer the " + "compose generator falls back to the hard-coded per-profile defaults " + "in src/agents/compose.ts (klanker 6g, coding 2g, conversational 1.5g, " + "lightweight 1g, default 1.5g)."),
       memory_reservation: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "memory_reservation must be a Docker size string like '4g', '256m'").optional().describe("Soft memory floor (Docker `mem_reservation` → cgroup memory.low). " + "Under host-wide memory pressure, the kernel protects at least this " + "much from being reclaimed from the cgroup. Must be ≤ memory. Use to " + "keep an agent RAM-resident when the host has other tenants that " + "might push the box (Coolify apps, build jobs). Default: unset."),
       pids_limit: exports_external.number().int().positive().optional().describe("Max processes the cgroup can spawn (cgroup pids.max). Prevents " + "fork bombs and runaway test runners. Counts every process in the " + "cgroup including bash subprocesses, claude itself, sidecars, and " + "any test/build worker. A typical agent at idle uses ~30 PIDs; " + "`npm test`-style workloads can spike to 200+. Set generously " + "(2000 is a comfortable cap for test-running agents). Default: " + "unset (no cgroup pid cap)."),
-      cpus: exports_external.number().positive().optional().describe("CPU quota (Docker `cpus`). Fractional values OK (e.g. 0.5, 2.0). " + "When unset at every cascade layer the compose generator falls " + "back to the per-profile default (klanker/coding 2.0, default 1.0, " + "lightweight 0.5).")
+      cpus: exports_external.number().positive().optional().describe("CPU quota (Docker `cpus`). Fractional values OK (e.g. 0.5, 2.0). " + "When unset at every cascade layer the compose generator falls " + "back to the per-profile default (klanker/coding 2.0, default 1.0, " + "lightweight 0.5)."),
+      tmp_size: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional().describe("Size of the agent container's RAM-backed `/tmp` (Docker " + "`tmpfs: - /tmp:size=<this>,mode=1777`). The container root FS is " + "read-only, so /tmp is the only scratch space the agent and its " + "sub-agents get for repo clones and toolchain caches (bunx/npx/pip) " + "— a fan-out of several sub-agents exhausts the 1g default. A tmpfs " + "only consumes host RAM for the pages actually written, so raising " + "the ceiling costs nothing until it is used; those pages are still " + "charged to the container's `memory` cap, so raise both together. " + "Format: '1g', '4g', '512m'. Default when unset at every cascade " + "layer: '1g'.")
     }).optional().describe("Per-agent resource limits. Cascades through defaults → profile → " + "per-agent with per-field merge (agent wins on each field independently). " + "Any field left unset at every layer falls back to the hard-coded " + "per-profile defaults in src/agents/compose.ts."),
     experimental: exports_external.object({
       legacy_pty: exports_external.boolean().optional().describe("Opt out of the default tmux supervisor (#725) and run the agent under " + "the legacy PTY supervisor instead. Default: false (tmux is the default)."),
@@ -4652,7 +4665,8 @@ var init_schema = __esm(() => {
       memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).optional(),
       memory_reservation: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).optional(),
       pids_limit: exports_external.number().int().positive().optional(),
-      cpus: exports_external.number().positive().optional()
+      cpus: exports_external.number().positive().optional(),
+      tmp_size: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional()
     }).optional()
   }).superRefine((agent, ctx) => {
     if (agent.dm_only !== true)
@@ -19080,6 +19094,10 @@ function allocateAgentUid(name) {
   return AGENT_UID_MIN + u32 % range;
 }
 
+// src/setup/hindsight-recall-tunables.ts
+var RECALL_DEADLINE_HEADROOM_SECONDS = 2;
+var MIN_RECALL_HOOK_TIMEOUT_SECONDS = RECALL_DEADLINE_HEADROOM_SECONDS + 1;
+
 // src/agents/scaffold.ts
 init_schema();
 
@@ -19194,6 +19212,27 @@ function assertPositive(value, label) {
 // src/setup/host-capabilities.ts
 init_paths();
 
+// src/setup/hindsight-pg-defaults.ts
+var HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION = 8 * 1024;
+var HINDSIGHT_PG_APP_ANON_MIB = 2560;
+var HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB = 2048;
+var HINDSIGHT_PG_SHARED_BUFFERS_BUDGET_MIB = HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION - HINDSIGHT_PG_APP_ANON_MIB - HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB;
+var HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB = 1536;
+var HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB = 4096;
+function pgMib(mib) {
+  return `${mib}MB`;
+}
+var HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV = "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE";
+var HINDSIGHT_PG_SHARED_BUFFERS_ENV = "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS";
+var HINDSIGHT_PG_DEFAULTS = [
+  [
+    HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV,
+    pgMib(HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB)
+  ],
+  [HINDSIGHT_PG_SHARED_BUFFERS_ENV, pgMib(HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB)]
+];
+var HINDSIGHT_PG_ENV_KEYS = new Set(HINDSIGHT_PG_DEFAULTS.map(([k]) => k));
+
 // src/setup/hindsight-perf-defaults.ts
 var HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION = 150;
 var HINDSIGHT_DEFAULT_RECALL_MAX_CANDIDATES_PER_SOURCE = Math.ceil(HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION * 0.4);
@@ -19208,6 +19247,15 @@ var HINDSIGHT_DEFAULT_RERANKER_LOCAL_BATCH_SIZE = 128;
 var HINDSIGHT_DEFAULT_LLM_STRICT_SCHEMA = "true";
 var HINDSIGHT_DEFAULT_LLM_MAX_RETRIES = 2;
 var HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM = 2;
+var HINDSIGHT_DEFAULT_MAX_OBSERVATIONS_PER_SCOPE = 1000;
+var HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING = "true";
+var HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES = 150;
+var HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4;
+var HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT = 8;
+var HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S = 600;
+var HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND = 500;
+var HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT = 6;
+var HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS = 1;
 var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -19225,6 +19273,42 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_CONSOLIDATION_LLM_PARALLELISM",
     String(HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM)
+  ],
+  [
+    "HINDSIGHT_API_MAX_OBSERVATIONS_PER_SCOPE",
+    String(HINDSIGHT_DEFAULT_MAX_OBSERVATIONS_PER_SCOPE)
+  ],
+  [
+    "HINDSIGHT_API_RERANKER_LOCAL_BUCKET_BATCHING",
+    HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING
+  ],
+  [
+    "HINDSIGHT_API_RERANKER_MAX_CANDIDATES",
+    String(HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES)
+  ],
+  [
+    "HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT",
+    String(HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT)
+  ],
+  [
+    "HINDSIGHT_API_RECALL_MAX_CONCURRENT",
+    String(HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT)
+  ],
+  [
+    "HINDSIGHT_API_REFLECT_WALL_TIMEOUT",
+    String(HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S)
+  ],
+  [
+    "HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS",
+    String(HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS)
+  ],
+  [
+    "HINDSIGHT_API_WORKER_CONSOLIDATION_SLOT_LIMIT",
+    String(HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT)
+  ],
+  [
+    "HINDSIGHT_API_CONSOLIDATION_MAX_MEMORIES_PER_ROUND",
+    String(HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND)
   ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_GPU = [
@@ -19248,7 +19332,8 @@ var HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM = [
   ["HINDSIGHT_API_LLM_MAX_RETRIES", String(HINDSIGHT_DEFAULT_LLM_MAX_RETRIES)]
 ];
 var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
-  "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY"
+  "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY",
+  "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP"
 ]);
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...[
@@ -19259,27 +19344,6 @@ var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS
 ]);
 
-// src/setup/hindsight-pg-defaults.ts
-var HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION = 8 * 1024;
-var HINDSIGHT_PG_APP_ANON_MIB = 2560;
-var HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB = 2048;
-var HINDSIGHT_PG_SHARED_BUFFERS_BUDGET_MIB = HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION - HINDSIGHT_PG_APP_ANON_MIB - HINDSIGHT_PG_PAGE_CACHE_FLOOR_MIB;
-var HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB = 1536;
-var HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB = 4096;
-function pgMib(mib) {
-  return `${mib}MB`;
-}
-var HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV = "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE";
-var HINDSIGHT_PG_SHARED_BUFFERS_ENV = "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS";
-var HINDSIGHT_PG_DEFAULTS = [
-  [
-    HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE_ENV,
-    pgMib(HINDSIGHT_PG_DEFAULT_EFFECTIVE_CACHE_SIZE_MIB)
-  ],
-  [HINDSIGHT_PG_SHARED_BUFFERS_ENV, pgMib(HINDSIGHT_PG_DEFAULT_SHARED_BUFFERS_MIB)]
-];
-var HINDSIGHT_PG_ENV_KEYS = new Set(HINDSIGHT_PG_DEFAULTS.map(([k]) => k));
-
 // src/setup/hindsight-context-budget.ts
 var HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE = 3000;
 var HINDSIGHT_RETAIN_MAX_COMPLETION_FLOOR = HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE + 72;