switchroom 0.19.23 → 0.19.25

package/dist/agent-scheduler/index.js

@@ -11107,7 +11107,11 @@ 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`."),
+    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. Plugin default is 12, matching the UserPromptSubmit " + "hook ceiling; the shared `recallParallelDeadlineSeconds` (10) " + "is the tighter outer guard in the default configuration, so " + "this is a per-request safety net. Was a hardcoded 8 in the " + "plugin before #3757."),
+    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)."),
     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`."),
@@ -11320,17 +11324,19 @@ var HindsightPerOpLlmSchema = exports_external.object({
   model: exports_external.string().min(1).optional().describe("Per-op model (upstream `HINDSIGHT_API_<OP>_LLM_MODEL`). Absent → " + "inherit the global `hindsight.llm.model`."),
   provider: exports_external.string().min(1).optional().describe("Per-op provider (upstream `HINDSIGHT_API_<OP>_LLM_PROVIDER`). " + "Absent → inherit the global `hindsight.llm.provider`."),
   base_url: exports_external.string().min(1).optional().describe("Per-op base URL (upstream `HINDSIGHT_API_<OP>_LLM_BASE_URL`). " + "Optional passthrough; absent → inherit the global."),
-  api_key: exports_external.string().min(1).optional().describe("Per-op API key (upstream `HINDSIGHT_API_<OP>_LLM_API_KEY`). Literal " + "or `vault:` reference. Optional passthrough; absent → inherit global.")
+  api_key: exports_external.string().min(1).optional().describe("Per-op API key (upstream `HINDSIGHT_API_<OP>_LLM_API_KEY`). Literal " + "or `vault:` reference. Optional passthrough; absent → inherit global."),
+  context_window: exports_external.number().int().positive().optional().describe("Context window (tokens) of the backend serving THIS op. NOT an " + "upstream env var — switchroom derives the op's token budget " + "(consolidation batch size / max-completion caps / reflect " + "max-context cap) from it so a single call can never overflow the " + "window. Absent → inherit " + "`hindsight.llm.context_window`, else a per-provider default " + "(conservative for non-`claude-code` providers, which usually mean " + "a local llama.cpp/Ollama slot; a self-hosted `base_url` — loopback, " + "RFC1918, `.local`/`.internal` — forces the conservative default too, " + "regardless of the provider NAME, since the endpoint is where the " + "traffic actually terminates). All three lanes (`retain`, " + "`reflect`, `consolidation`) are budgeted independently.")
 }).describe("Per-operation LLM override. Every field optional; an unset field (or " + "an omitted op block) inherits the global `hindsight.llm.*`, which is " + "already the engine's fallback — switchroom emits only the vars set.");
 var HindsightConfigSchema = exports_external.object({
   llm: exports_external.object({
     provider: exports_external.string().min(1).optional().describe("Hindsight LLM provider (upstream `HINDSIGHT_API_LLM_PROVIDER`). " + "Defaults to `claude-code` (subscription-honest, broker-fed OAuth). " + "Any litellm-routable provider the upstream image supports is valid. " + "Serves as the GLOBAL default for every op absent a per-op override."),
     model: exports_external.string().min(1).optional().describe("Hindsight LLM model (upstream `HINDSIGHT_API_LLM_MODEL`). Defaults " + "to HINDSIGHT_DEFAULT_MODEL. Any model your LiteLLM proxy can route " + "is valid, e.g. `openrouter/z-ai/glm-5.2` when routing through the " + "fleet proxy. With provider=claude-code this value is ALSO exported " + "as `ANTHROPIC_MODEL` to the claude subprocess. Serves as the GLOBAL " + "default for every op absent a per-op override."),
+    context_window: exports_external.number().int().positive().optional().describe("GLOBAL context window (tokens) of the backend serving hindsight's " + "LLM ops — the declared size switchroom derives every token budget " + "from. Set this to the real window of whatever you point " + "`hindsight.llm` at (e.g. 32768 for a llama.cpp slot launched with " + "`-c 65536 -np 2`, 131072 for a large-window OpenRouter model). " + "Absent → a per-provider default: 200000 for `claude-code`, a " + "conservative 32768 for everything else. Overflowing a local " + "backend's window does NOT error — llama.cpp context-shift silently " + "drops the system prompt and the model answers conversationally " + "with HTTP 200 — so this value is what makes the failure " + "detectable at setup time instead of never."),
     retain: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `retain` LLM op (memory ingestion). Emits " + "`HINDSIGHT_API_RETAIN_LLM_*`. Absent → uses the global model/provider."),
     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, " + "LLM_MAX_CONCURRENT, RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), 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)."),
@@ -11466,7 +11472,11 @@ 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(),
+      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()
@@ -11500,7 +11510,8 @@ 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]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional().describe("Size of the agent container's RAM-backed `/tmp` (Docker " + "`tmpfs: - /tmp:size=<this>,mode=1777`). The container root FS is " + "read-only, so /tmp is the only scratch space the agent and its " + "sub-agents get for repo clones and toolchain caches (bunx/npx/pip) " + "— a fan-out of several sub-agents exhausts the 1g default. A tmpfs " + "only consumes host RAM for the pages actually written, so raising " + "the ceiling costs nothing until it is used; those pages are still " + "charged to the container's `memory` cap, so raise both together. " + "Format: '1g', '4g', '512m'. Default when unset at every cascade " + "layer: '1g'.")
   }).optional().describe("Per-agent resource limits. Cascades through defaults → profile → " + "per-agent with per-field merge (agent wins on each field independently). " + "Any field left unset at every layer falls back to the hard-coded " + "per-profile defaults in src/agents/compose.ts."),
   experimental: exports_external.object({
     legacy_pty: exports_external.boolean().optional().describe("Opt out of the default tmux supervisor (#725) and run the agent under " + "the legacy PTY supervisor instead. Default: false (tmux is the default)."),
@@ -11593,7 +11604,8 @@ var AgentSchema = exports_external.object({
     memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).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)
@@ -14136,7 +14148,6 @@ class AuthBrokerUnreachableError extends Error {
     this.name = "AuthBrokerUnreachableError";
   }
 }
-
 class AuthBrokerClient {
   socketPath;
   timeoutMs;

package/dist/auth-broker/index.js

@@ -16749,7 +16749,11 @@ 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`."),
+    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. Plugin default is 12, matching the UserPromptSubmit " + "hook ceiling; the shared `recallParallelDeadlineSeconds` (10) " + "is the tighter outer guard in the default configuration, so " + "this is a per-request safety net. Was a hardcoded 8 in the " + "plugin before #3757."),
+    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)."),
     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`."),
@@ -16962,17 +16966,19 @@ var HindsightPerOpLlmSchema = exports_external.object({
   model: exports_external.string().min(1).optional().describe("Per-op model (upstream `HINDSIGHT_API_<OP>_LLM_MODEL`). Absent → " + "inherit the global `hindsight.llm.model`."),
   provider: exports_external.string().min(1).optional().describe("Per-op provider (upstream `HINDSIGHT_API_<OP>_LLM_PROVIDER`). " + "Absent → inherit the global `hindsight.llm.provider`."),
   base_url: exports_external.string().min(1).optional().describe("Per-op base URL (upstream `HINDSIGHT_API_<OP>_LLM_BASE_URL`). " + "Optional passthrough; absent → inherit the global."),
-  api_key: exports_external.string().min(1).optional().describe("Per-op API key (upstream `HINDSIGHT_API_<OP>_LLM_API_KEY`). Literal " + "or `vault:` reference. Optional passthrough; absent → inherit global.")
+  api_key: exports_external.string().min(1).optional().describe("Per-op API key (upstream `HINDSIGHT_API_<OP>_LLM_API_KEY`). Literal " + "or `vault:` reference. Optional passthrough; absent → inherit global."),
+  context_window: exports_external.number().int().positive().optional().describe("Context window (tokens) of the backend serving THIS op. NOT an " + "upstream env var — switchroom derives the op's token budget " + "(consolidation batch size / max-completion caps / reflect " + "max-context cap) from it so a single call can never overflow the " + "window. Absent → inherit " + "`hindsight.llm.context_window`, else a per-provider default " + "(conservative for non-`claude-code` providers, which usually mean " + "a local llama.cpp/Ollama slot; a self-hosted `base_url` — loopback, " + "RFC1918, `.local`/`.internal` — forces the conservative default too, " + "regardless of the provider NAME, since the endpoint is where the " + "traffic actually terminates). All three lanes (`retain`, " + "`reflect`, `consolidation`) are budgeted independently.")
 }).describe("Per-operation LLM override. Every field optional; an unset field (or " + "an omitted op block) inherits the global `hindsight.llm.*`, which is " + "already the engine's fallback — switchroom emits only the vars set.");
 var HindsightConfigSchema = exports_external.object({
   llm: exports_external.object({
     provider: exports_external.string().min(1).optional().describe("Hindsight LLM provider (upstream `HINDSIGHT_API_LLM_PROVIDER`). " + "Defaults to `claude-code` (subscription-honest, broker-fed OAuth). " + "Any litellm-routable provider the upstream image supports is valid. " + "Serves as the GLOBAL default for every op absent a per-op override."),
     model: exports_external.string().min(1).optional().describe("Hindsight LLM model (upstream `HINDSIGHT_API_LLM_MODEL`). Defaults " + "to HINDSIGHT_DEFAULT_MODEL. Any model your LiteLLM proxy can route " + "is valid, e.g. `openrouter/z-ai/glm-5.2` when routing through the " + "fleet proxy. With provider=claude-code this value is ALSO exported " + "as `ANTHROPIC_MODEL` to the claude subprocess. Serves as the GLOBAL " + "default for every op absent a per-op override."),
+    context_window: exports_external.number().int().positive().optional().describe("GLOBAL context window (tokens) of the backend serving hindsight's " + "LLM ops — the declared size switchroom derives every token budget " + "from. Set this to the real window of whatever you point " + "`hindsight.llm` at (e.g. 32768 for a llama.cpp slot launched with " + "`-c 65536 -np 2`, 131072 for a large-window OpenRouter model). " + "Absent → a per-provider default: 200000 for `claude-code`, a " + "conservative 32768 for everything else. Overflowing a local " + "backend's window does NOT error — llama.cpp context-shift silently " + "drops the system prompt and the model answers conversationally " + "with HTTP 200 — so this value is what makes the failure " + "detectable at setup time instead of never."),
     retain: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `retain` LLM op (memory ingestion). Emits " + "`HINDSIGHT_API_RETAIN_LLM_*`. Absent → uses the global model/provider."),
     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, " + "LLM_MAX_CONCURRENT, RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), 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)."),
@@ -17108,7 +17114,11 @@ 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(),
+      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()
@@ -17142,7 +17152,8 @@ 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]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional().describe("Size of the agent container's RAM-backed `/tmp` (Docker " + "`tmpfs: - /tmp:size=<this>,mode=1777`). The container root FS is " + "read-only, so /tmp is the only scratch space the agent and its " + "sub-agents get for repo clones and toolchain caches (bunx/npx/pip) " + "— a fan-out of several sub-agents exhausts the 1g default. A tmpfs " + "only consumes host RAM for the pages actually written, so raising " + "the ceiling costs nothing until it is used; those pages are still " + "charged to the container's `memory` cap, so raise both together. " + "Format: '1g', '4g', '512m'. Default when unset at every cascade " + "layer: '1g'.")
   }).optional().describe("Per-agent resource limits. Cascades through defaults → profile → " + "per-agent with per-field merge (agent wins on each field independently). " + "Any field left unset at every layer falls back to the hard-coded " + "per-profile defaults in src/agents/compose.ts."),
   experimental: exports_external.object({
     legacy_pty: exports_external.boolean().optional().describe("Opt out of the default tmux supervisor (#725) and run the agent under " + "the legacy PTY supervisor instead. Default: false (tmux is the default)."),
@@ -17235,7 +17246,8 @@ var AgentSchema = exports_external.object({
     memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).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)
@@ -18394,6 +18406,27 @@ 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);
@@ -18404,6 +18437,19 @@ var HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT = 4;
 var HINDSIGHT_DEFAULT_RETAIN_LLM_MAX_CONCURRENT = 1;
 var HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_MAX_CONCURRENT = 1;
 var HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16 = "true";
+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",
@@ -18417,10 +18463,54 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
     "HINDSIGHT_API_LINK_EXPANSION_TIMEOUT",
     String(HINDSIGHT_DEFAULT_LINK_EXPANSION_TIMEOUT_S)
   ],
-  ["HINDSIGHT_API_LLM_REASONING_EFFORT", HINDSIGHT_DEFAULT_LLM_REASONING_EFFORT]
+  ["HINDSIGHT_API_LLM_REASONING_EFFORT", HINDSIGHT_DEFAULT_LLM_REASONING_EFFORT],
+  [
+    "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 = [
-  ["HINDSIGHT_API_RERANKER_LOCAL_FP16", HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16]
+  ["HINDSIGHT_API_RERANKER_LOCAL_FP16", HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16],
+  [
+    "HINDSIGHT_API_RERANKER_LOCAL_BATCH_SIZE",
+    String(HINDSIGHT_DEFAULT_RERANKER_LOCAL_BATCH_SIZE)
+  ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM = [
   ["HINDSIGHT_API_LLM_MAX_CONCURRENT", String(HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT)],
@@ -18431,34 +18521,26 @@ var HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM = [
   [
     "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_CONCURRENT",
     String(HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_MAX_CONCURRENT)
-  ]
+  ],
+  ["HINDSIGHT_API_LLM_STRICT_SCHEMA", HINDSIGHT_DEFAULT_LLM_STRICT_SCHEMA],
+  ["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_CE_DECISIVE_RELATIVE_GAP"
+]);
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
-  ...HINDSIGHT_PERF_DEFAULTS_UNGATED,
-  ...HINDSIGHT_PERF_DEFAULTS_GPU,
-  ...HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM
-].map(([k]) => k));
+  ...[
+    ...HINDSIGHT_PERF_DEFAULTS_UNGATED,
+    ...HINDSIGHT_PERF_DEFAULTS_GPU,
+    ...HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM
+  ].map(([k]) => k),
+  ...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;
 
 // src/setup/hindsight.ts
 var HINDSIGHT_DEFAULT_API_PORT = 18888;
@@ -21715,9 +21797,11 @@ class AuthBroker {
       last_seen_at: this.consumerLastSeen[c.name] ?? null
     }));
     const active_overage_serving = this.isActiveOverageServing(this.callerAccount(identity));
+    const serving = this.accountWithFailover(auth.active) ?? auth.active ?? "";
     this.audit({ op: "list-state", identity, ok: true });
     socket.write(encodeSuccess(id, {
       active: auth.active ?? "",
+      serving,
       fallback_order: auth.fallback_order ?? [],
       accounts,
       agents,

package/dist/cli/autoaccept-poll.js

@@ -5061,7 +5061,6 @@ class AuthBrokerUnreachableError extends Error {
     this.name = "AuthBrokerUnreachableError";
   }
 }
-
 class AuthBrokerClient {
   socketPath;
   timeoutMs;

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

@@ -4289,6 +4289,7 @@ var exports_client = {};
 __export(exports_client, {
   withAuthBrokerClient: () => withAuthBrokerClient,
   resolveAuthBrokerSocketPath: () => resolveAuthBrokerSocketPath,
+  effectiveServingLabel: () => effectiveServingLabel,
   AuthBrokerUnreachableError: () => AuthBrokerUnreachableError,
   AuthBrokerError: () => AuthBrokerError,
   AuthBrokerClient: () => AuthBrokerClient
@@ -4316,6 +4317,10 @@ function resolveAuthBrokerSocketPath(opts) {
     return env;
   return operatorSocketPath(opts?.home);
 }
+function effectiveServingLabel(state) {
+  const serving = state.serving;
+  return typeof serving === "string" && serving.length > 0 ? serving : state.active;
+}
 
 class AuthBrokerClient {
   socketPath;

package/dist/cli/ms-365-write-pretool.mjs

@@ -4291,6 +4291,7 @@ var exports_client = {};
 __export(exports_client, {
   withAuthBrokerClient: () => withAuthBrokerClient,
   resolveAuthBrokerSocketPath: () => resolveAuthBrokerSocketPath,
+  effectiveServingLabel: () => effectiveServingLabel,
   AuthBrokerUnreachableError: () => AuthBrokerUnreachableError,
   AuthBrokerError: () => AuthBrokerError,
   AuthBrokerClient: () => AuthBrokerClient
@@ -4318,6 +4319,10 @@ function resolveAuthBrokerSocketPath(opts) {
     return env;
   return operatorSocketPath(opts?.home);
 }
+function effectiveServingLabel(state) {
+  const serving = state.serving;
+  return typeof serving === "string" && serving.length > 0 ? serving : state.active;
+}
 
 class AuthBrokerClient {
   socketPath;

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

@@ -11866,7 +11866,11 @@ 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\u20131.0] 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 \u2014 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 \u2014 current " + "behaviour). Use 0.10; values at or above 0.20 measurably " + "starve recall \u2014 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`."),
+    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. Plugin default is 12, matching the UserPromptSubmit " + "hook ceiling; the shared `recallParallelDeadlineSeconds` (10) " + "is the tighter outer guard in the default configuration, so " + "this is a per-request safety net. Was a hardcoded 8 in the " + "plugin before #3757."),
+    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`."),
     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 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)."),
     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`."),
@@ -12079,17 +12083,19 @@ var HindsightPerOpLlmSchema = exports_external.object({
   model: exports_external.string().min(1).optional().describe("Per-op model (upstream `HINDSIGHT_API_<OP>_LLM_MODEL`). Absent \u2192 " + "inherit the global `hindsight.llm.model`."),
   provider: exports_external.string().min(1).optional().describe("Per-op provider (upstream `HINDSIGHT_API_<OP>_LLM_PROVIDER`). " + "Absent \u2192 inherit the global `hindsight.llm.provider`."),
   base_url: exports_external.string().min(1).optional().describe("Per-op base URL (upstream `HINDSIGHT_API_<OP>_LLM_BASE_URL`). " + "Optional passthrough; absent \u2192 inherit the global."),
-  api_key: exports_external.string().min(1).optional().describe("Per-op API key (upstream `HINDSIGHT_API_<OP>_LLM_API_KEY`). Literal " + "or `vault:` reference. Optional passthrough; absent \u2192 inherit global.")
+  api_key: exports_external.string().min(1).optional().describe("Per-op API key (upstream `HINDSIGHT_API_<OP>_LLM_API_KEY`). Literal " + "or `vault:` reference. Optional passthrough; absent \u2192 inherit global."),
+  context_window: exports_external.number().int().positive().optional().describe("Context window (tokens) of the backend serving THIS op. NOT an " + "upstream env var \u2014 switchroom derives the op's token budget " + "(consolidation batch size / max-completion caps / reflect " + "max-context cap) from it so a single call can never overflow the " + "window. Absent \u2192 inherit " + "`hindsight.llm.context_window`, else a per-provider default " + "(conservative for non-`claude-code` providers, which usually mean " + "a local llama.cpp/Ollama slot; a self-hosted `base_url` \u2014 loopback, " + "RFC1918, `.local`/`.internal` \u2014 forces the conservative default too, " + "regardless of the provider NAME, since the endpoint is where the " + "traffic actually terminates). All three lanes (`retain`, " + "`reflect`, `consolidation`) are budgeted independently.")
 }).describe("Per-operation LLM override. Every field optional; an unset field (or " + "an omitted op block) inherits the global `hindsight.llm.*`, which is " + "already the engine's fallback \u2014 switchroom emits only the vars set.");
 var HindsightConfigSchema = exports_external.object({
   llm: exports_external.object({
     provider: exports_external.string().min(1).optional().describe("Hindsight LLM provider (upstream `HINDSIGHT_API_LLM_PROVIDER`). " + "Defaults to `claude-code` (subscription-honest, broker-fed OAuth). " + "Any litellm-routable provider the upstream image supports is valid. " + "Serves as the GLOBAL default for every op absent a per-op override."),
     model: exports_external.string().min(1).optional().describe("Hindsight LLM model (upstream `HINDSIGHT_API_LLM_MODEL`). Defaults " + "to HINDSIGHT_DEFAULT_MODEL. Any model your LiteLLM proxy can route " + "is valid, e.g. `openrouter/z-ai/glm-5.2` when routing through the " + "fleet proxy. With provider=claude-code this value is ALSO exported " + "as `ANTHROPIC_MODEL` to the claude subprocess. Serves as the GLOBAL " + "default for every op absent a per-op override."),
+    context_window: exports_external.number().int().positive().optional().describe("GLOBAL context window (tokens) of the backend serving hindsight's " + "LLM ops \u2014 the declared size switchroom derives every token budget " + "from. Set this to the real window of whatever you point " + "`hindsight.llm` at (e.g. 32768 for a llama.cpp slot launched with " + "`-c 65536 -np 2`, 131072 for a large-window OpenRouter model). " + "Absent \u2192 a per-provider default: 200000 for `claude-code`, a " + "conservative 32768 for everything else. Overflowing a local " + "backend's window does NOT error \u2014 llama.cpp context-shift silently " + "drops the system prompt and the model answers conversationally " + "with HTTP 200 \u2014 so this value is what makes the failure " + "detectable at setup time instead of never."),
     retain: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `retain` LLM op (memory ingestion). Emits " + "`HINDSIGHT_API_RETAIN_LLM_*`. Absent \u2192 uses the global model/provider."),
     reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent \u2192 uses global."),
     consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent \u2192 global.")
   }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-  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, " + "LLM_MAX_CONCURRENT, RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS \u2014 a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED \u2014 a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+  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 \u2014 a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP \u2014 the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS \u2014 a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED \u2014 a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
 });
 var MicrosoftWorkspaceConfigSchema = exports_external.object({
   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 \u2014 omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -12225,7 +12231,11 @@ 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(),
+      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()
@@ -12259,7 +12269,8 @@ var profileFields = {
     memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "memory must be a Docker size string like '6g', '512m', '1.5g'").optional().describe("Hard memory cap (Docker `mem_limit` \u2192 cgroup memory.max). When the " + "container exceeds this, the kernel OOM-kills processes in the cgroup. " + "Format: '6g', '1.5g', '512m'. When unset at every cascade layer the " + "compose generator falls back to the hard-coded per-profile defaults " + "in src/agents/compose.ts (klanker 6g, coding 2g, conversational 1.5g, " + "lightweight 1g, default 1.5g)."),
     memory_reservation: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/, "memory_reservation must be a Docker size string like '4g', '256m'").optional().describe("Soft memory floor (Docker `mem_reservation` \u2192 cgroup memory.low). " + "Under host-wide memory pressure, the kernel protects at least this " + "much from being reclaimed from the cgroup. Must be \u2264 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]?$/, "tmp_size must be a Docker size string like '1g', '4g', '512m'").refine((v) => parseFloat(v) > 0, "tmp_size must be greater than zero").optional().describe("Size of the agent container's RAM-backed `/tmp` (Docker " + "`tmpfs: - /tmp:size=<this>,mode=1777`). The container root FS is " + "read-only, so /tmp is the only scratch space the agent and its " + "sub-agents get for repo clones and toolchain caches (bunx/npx/pip) " + "\u2014 a fan-out of several sub-agents exhausts the 1g default. A tmpfs " + "only consumes host RAM for the pages actually written, so raising " + "the ceiling costs nothing until it is used; those pages are still " + "charged to the container's `memory` cap, so raise both together. " + "Format: '1g', '4g', '512m'. Default when unset at every cascade " + "layer: '1g'.")
   }).optional().describe("Per-agent resource limits. Cascades through defaults \u2192 profile \u2192 " + "per-agent with per-field merge (agent wins on each field independently). " + "Any field left unset at every layer falls back to the hard-coded " + "per-profile defaults in src/agents/compose.ts."),
   experimental: exports_external.object({
     legacy_pty: exports_external.boolean().optional().describe("Opt out of the default tmux supervisor (#725) and run the agent under " + "the legacy PTY supervisor instead. Default: false (tmux is the default)."),
@@ -12352,7 +12363,8 @@ var AgentSchema = exports_external.object({
     memory: exports_external.string().regex(/^\d+(\.\d+)?[kmgKMG]?$/).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)