switchroom 0.19.27 → 0.19.28

package/dist/agent-scheduler/index.js

@@ -11347,7 +11347,10 @@ var LiteLLMConfigSchema = exports_external.object({
   admin_key: exports_external.string().optional().describe("LiteLLM master/admin key used at apply time to provision the team + " + "virtual key. Supports a vault reference (e.g. " + "'vault:litellm/master-key') — resolution happens at apply time via " + "the vault-broker. Never injected into the agent container."),
   team: exports_external.string().optional().describe("LiteLLM team alias the per-agent key is created under. Defaults to " + "'switchroom' (applied in code, not as a schema default)."),
   small_fast_model: exports_external.string().optional().describe("Model id exported as ANTHROPIC_SMALL_FAST_MODEL for the claude CLI's " + "background/fast lane, e.g. 'claude-haiku-4-5-20251001'."),
-  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins).")
+  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins)."),
+  max_budget: exports_external.number().positive().optional().describe("HARD spend cap in USD for this agent's virtual key over one " + "`budget_duration` window. LiteLLM refuses the request once the key's " + "tracked spend exceeds it, so a runaway loop costs at most this much " + "before it is stopped. Defaults to " + "DEFAULT_KEY_MAX_BUDGET_USD (see src/litellm/budget.ts) — deliberately " + "conservative; raise it per-agent rather than removing it. Set 0 or " + "omit `budget_duration` at your own risk: an uncapped key is only as " + "bounded as the upstream account balance."),
+  soft_budget: exports_external.number().positive().optional().describe("ADVISORY spend threshold in USD. LiteLLM keeps serving past it and " + "raises a budget alert instead. Must be < max_budget. NOTE: LiteLLM " + "accepts soft_budget only on POST /key/generate (GenerateKeyRequest); " + "UpdateKeyRequest does NOT carry it, so changing this value only takes " + "effect on a key that is (re)generated, not on an existing one."),
+  budget_duration: exports_external.string().regex(/^\d+(s|m|h|d|mo)$/, "budget_duration must be a LiteLLM duration like '30d', '24h', '1mo'").optional().describe("Rolling window the budget resets on, in LiteLLM duration syntax " + "('30d', '24h', '1mo'). Defaults to DEFAULT_KEY_BUDGET_DURATION. " + "WITHOUT a duration LiteLLM treats max_budget as a LIFETIME cap that " + "never resets — the key silently dies for good once it is hit.")
 }).optional().describe("LiteLLM routing config — opt-in per-agent virtual-key auto-provisioning " + "+ routing env. Default OFF. See LiteLLMConfigSchema doc for the full flow.");
 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`."),
@@ -11366,7 +11369,7 @@ 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, " + "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).")
+  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, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_MAX_SLOTS reserves out of; " + "unset means upstream's own default), 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)."),

package/dist/auth-broker/index.js

@@ -16989,7 +16989,10 @@ var LiteLLMConfigSchema = exports_external.object({
   admin_key: exports_external.string().optional().describe("LiteLLM master/admin key used at apply time to provision the team + " + "virtual key. Supports a vault reference (e.g. " + "'vault:litellm/master-key') — resolution happens at apply time via " + "the vault-broker. Never injected into the agent container."),
   team: exports_external.string().optional().describe("LiteLLM team alias the per-agent key is created under. Defaults to " + "'switchroom' (applied in code, not as a schema default)."),
   small_fast_model: exports_external.string().optional().describe("Model id exported as ANTHROPIC_SMALL_FAST_MODEL for the claude CLI's " + "background/fast lane, e.g. 'claude-haiku-4-5-20251001'."),
-  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins).")
+  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins)."),
+  max_budget: exports_external.number().positive().optional().describe("HARD spend cap in USD for this agent's virtual key over one " + "`budget_duration` window. LiteLLM refuses the request once the key's " + "tracked spend exceeds it, so a runaway loop costs at most this much " + "before it is stopped. Defaults to " + "DEFAULT_KEY_MAX_BUDGET_USD (see src/litellm/budget.ts) — deliberately " + "conservative; raise it per-agent rather than removing it. Set 0 or " + "omit `budget_duration` at your own risk: an uncapped key is only as " + "bounded as the upstream account balance."),
+  soft_budget: exports_external.number().positive().optional().describe("ADVISORY spend threshold in USD. LiteLLM keeps serving past it and " + "raises a budget alert instead. Must be < max_budget. NOTE: LiteLLM " + "accepts soft_budget only on POST /key/generate (GenerateKeyRequest); " + "UpdateKeyRequest does NOT carry it, so changing this value only takes " + "effect on a key that is (re)generated, not on an existing one."),
+  budget_duration: exports_external.string().regex(/^\d+(s|m|h|d|mo)$/, "budget_duration must be a LiteLLM duration like '30d', '24h', '1mo'").optional().describe("Rolling window the budget resets on, in LiteLLM duration syntax " + "('30d', '24h', '1mo'). Defaults to DEFAULT_KEY_BUDGET_DURATION. " + "WITHOUT a duration LiteLLM treats max_budget as a LIFETIME cap that " + "never resets — the key silently dies for good once it is hit.")
 }).optional().describe("LiteLLM routing config — opt-in per-agent virtual-key auto-provisioning " + "+ routing env. Default OFF. See LiteLLMConfigSchema doc for the full flow.");
 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`."),
@@ -17008,7 +17011,7 @@ 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, " + "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).")
+  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, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_MAX_SLOTS reserves out of; " + "unset means upstream's own default), 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)."),
@@ -18500,7 +18503,13 @@ var HINDSIGHT_DEFAULT_LINK_EXPANSION_TIMEOUT_S = 2;
 var HINDSIGHT_DEFAULT_LLM_REASONING_EFFORT = "low";
 var HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT = 4;
 var HINDSIGHT_DEFAULT_RETAIN_LLM_MAX_CONCURRENT = 1;
-var HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_MAX_CONCURRENT = 1;
+function hindsightConsolidationLlmMaxConcurrentDefault(globalMaxConcurrent = HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT, retainMaxConcurrent = HINDSIGHT_DEFAULT_RETAIN_LLM_MAX_CONCURRENT) {
+  const globalCap = Number.isFinite(globalMaxConcurrent) && globalMaxConcurrent >= 1 ? Math.floor(globalMaxConcurrent) : HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT;
+  const retainCap = Number.isFinite(retainMaxConcurrent) && retainMaxConcurrent >= 0 ? Math.floor(retainMaxConcurrent) : HINDSIGHT_DEFAULT_RETAIN_LLM_MAX_CONCURRENT;
+  const headroomBound = Math.max(1, globalCap - 1);
+  return Math.min(headroomBound, Math.max(1, globalCap - retainCap - 1));
+}
+var HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_MAX_CONCURRENT = hindsightConsolidationLlmMaxConcurrentDefault();
 var HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16 = "true";
 var HINDSIGHT_DEFAULT_RERANKER_LOCAL_BATCH_SIZE = 128;
 var HINDSIGHT_DEFAULT_LLM_STRICT_SCHEMA = "true";
@@ -18515,6 +18524,8 @@ 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_DEFAULT_RECENCY_DECAY_FUNCTION = "exponential";
+var HINDSIGHT_DEFAULT_RECENCY_DECAY_HALFLIFE_DAYS = 30;
 var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -18568,6 +18579,14 @@ var HINDSIGHT_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_CONSOLIDATION_MAX_MEMORIES_PER_ROUND",
     String(HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND)
+  ],
+  [
+    "HINDSIGHT_API_RECENCY_DECAY_FUNCTION",
+    HINDSIGHT_DEFAULT_RECENCY_DECAY_FUNCTION
+  ],
+  [
+    "HINDSIGHT_API_RECENCY_DECAY_HALFLIFE_DAYS",
+    String(HINDSIGHT_DEFAULT_RECENCY_DECAY_HALFLIFE_DAYS)
   ]
 ];
 var HINDSIGHT_PERF_DEFAULTS_GPU = [
@@ -18592,7 +18611,9 @@ var HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM = [
 ];
 var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY",
-  "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP"
+  "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP",
+  "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS",
+  "HINDSIGHT_API_WORKER_MAX_SLOTS"
 ]);
 var HINDSIGHT_PERF_ENV_KEYS = new Set([
   ...[
@@ -18658,6 +18679,16 @@ var DEFAULT_RETAIN_MISSION = `Extract durable facts that will still be true and
 ` + `- Hindsight's own errors, retries, backlogs, or internal state — the memory
 ` + `  system's self-reports are not memories.
 ` + `- Restatements of the user's current request or the task in progress.
+` + `- Volatile state written as a timeless assertion. A version, count, size,
+` + `  backlog, status, or any "X is running Y" / "X is at Y" / "X is currently Y"
+` + `  claim is true only at the instant it was said. Concretely, never produce a
+` + `  fact whose text resembles any of these: "Switchroom fleet is running image
+` + `  version v0.18.19", "The switchroom repo is at /path/to/fleet, version
+` + `  v0.19.5", "Bank overlord has 43155 pending consolidations", "The build is
+` + `  currently green". If the claim is worth keeping, put the date INSIDE the
+` + `  fact text ("As of 2026-07-19 the fleet was running v0.18.19"); if you
+` + `  cannot date it, drop it. An undated one is recalled forever as though it
+` + `  were still true, which is worse than not remembering it at all.
 ` + `- Transient state (unread counts, build status, what is running right now) unless
 ` + `  the fact is explicitly dated, in which case record it as a dated observation.
 ` + `- Greetings, acknowledgements, and routine operational chatter.
@@ -18679,18 +18710,108 @@ var SUPERSEDED_RETAIN_MISSIONS = [
 ` + "- Transient state (unread counts, build status, what is running right now) " + "unless the fact is explicitly dated, in which case record it as a dated " + `observation.
 ` + `- Greetings, acknowledgements, and routine operational chatter.
 
-` + "If a candidate fact matches an exclusion, drop it rather than rewording " + "it. If nothing durable remains, return an empty facts list."
+` + "If a candidate fact matches an exclusion, drop it rather than rewording " + "it. If nothing durable remains, return an empty facts list.",
+  `Extract durable facts that will still be true and useful weeks from now: user preferences and standing rules, ongoing projects and recurring commitments, technical and architectural decisions with their rationale, and people/tool relationships. A preference revealed by a request is durable — record the preference (what the user likes, wants, or always does), not the request itself.
+` + `
+` + `A TOOL RESULT IS NOT A FACT. Before extracting, ask: is the subject of this
+` + `candidate a file path, a command/process/agent/session id, a temp directory, or
+` + `the location where some output was written? If yes, drop it — it is transcript
+` + `exhaust, not memory.
+` + `
+` + `NEVER extract:
+` + `- Tool results verbatim or paraphrased. Concretely, never produce a fact whose
+` + `  text resembles any of these: "File created successfully at /path/to/file",
+` + `  "A background command with ID bctz4yskm is running, and its output will be
+` + `  written to /tmp/...", "Async agent a745598ba84e71df1 was launched successfully
+` + `  and is running in the background", "User executed a Bash command to sleep for
+` + `  200 seconds", "The assistant used grep to locate 'truncateSync' in src/foo.ts".
+` + `- Anything mentioning a path under /tmp, a scratchpad directory, or a .tmp file.
+` + `- Agent tool-use traces or narration of what the assistant did (e.g. "the
+` + `  assistant used X to query Y", "ran a search", "sent the message").
+` + `- In-flight workflow/process narration (a sub-task started, paused, or is still
+` + `  running) — retain the outcome only once the task completes or a decision is made.
+` + `- Operation, request, batch, agent, command or session IDs, UUIDs, hashes, or error codes.
+` + `- Slash commands the user typed and their effects (e.g. "User issued /clear to
+` + `  reset assistant state").
+` + `- Hindsight's own errors, retries, backlogs, or internal state — the memory
+` + `  system's self-reports are not memories.
+` + `- Restatements of the user's current request or the task in progress.
+` + `- Transient state (unread counts, build status, what is running right now) unless
+` + `  the fact is explicitly dated, in which case record it as a dated observation.
+` + `- Greetings, acknowledgements, and routine operational chatter.
+` + `
+` + `If a candidate fact matches an exclusion, drop it rather than rewording it. If
+` + "nothing durable remains, return an empty facts list."
+];
+var DEFAULT_OBSERVATIONS_MISSION = `Synthesise durable, standing knowledge about the people, projects, and systems this agent works with: preferences and standing rules, roles and relationships, skills and recurring patterns, technical and operational decisions with their rationale, and the state of long-running work once it lands.
+` + `
+` + `The test is durability, not notability: an observation must still be worth reading weeks from now. A single dated event belongs in an observation only when it establishes or changes a standing fact.
+` + `
+` + `Do NOT synthesise observations from:
+` + `- Transcript exhaust — tool calls and their results, file paths, temp or scratchpad directories, where some output was written, or narration of what an assistant did.
+` + `- Identifiers with no standing meaning: session, agent, request, batch or command IDs, UUIDs, hashes, error codes.
+` + `- In-flight process narration — a task started, paused, or still running. Record the outcome once it lands, not the running state.
+` + `- The memory system's own errors, retries, backlogs, or internal state.
+` + `- Transient state (what is running right now, unread counts, build status) unless the fact is explicitly dated, in which case record it as dated.
+` + `
+` + "If the new facts contain nothing durable, record nothing rather than synthesising a weak observation.";
+var SUPERSEDED_OBSERVATIONS_MISSIONS = [
+  "Synthesise the person's wellbeing patterns, motivations, and emotional " + "context — how habits, setbacks, and encouragement connect over time."
 ];
 var PROFILE_MEMORY_DEFAULTS = {
   "health-coach": {
     disposition: { skepticism: 2, literalism: 2, empathy: 5 },
-    observations_mission: "Synthesise the person's wellbeing patterns, motivations, and emotional " + "context — how habits, setbacks, and encouragement connect over time."
+    observations_mission: `You consolidate the memory of a health and fitness coach working with one person. This bank records how that person actually lives and trains.
+` + `
+` + `Synthesise into durable observations:
+` + `- Goals, targets, and the plan currently in force, with the reasoning behind each.
+` + `- Training, nutrition, sleep, and alcohol patterns as they hold over weeks — what the person reliably does, not what they did once.
+` + `- Constraints that shape the plan: injuries, medical guidance, schedule, equipment, foods and sessions they refuse.
+` + `- Motivations, and what actually helps or backfires when they slip.
+` + `- Trends in the numbers: direction and range over time, not any single reading.
+` + `- Corrections to earlier beliefs, recorded explicitly as corrections.
+` + `
+` + `A single day's log, weight reading, or session is evidence, not an observation. Record one only when it establishes or changes a standing pattern, target, or constraint — then fold it into the observation for that pattern and say what changed and roughly when.
+` + `
+` + `Granularity: one observation per habit, target, constraint, or trend. Aggregate repeated daily evidence into the observation for that pattern rather than creating one per day, and keep training, nutrition, and sleep as separate observations.
+` + `
+` + "Word observations as the person's own pattern and framing, never as a verdict on them."
   },
   "executive-assistant": {
-    disposition: { skepticism: 4, literalism: 4, empathy: 3 }
+    disposition: { skepticism: 4, literalism: 4, empathy: 3 },
+    observations_mission: `You consolidate the memory of an executive assistant working for one person. This bank records that person's commitments, people, and standing arrangements.
+` + `
+` + `Synthesise into durable observations:
+` + `- Standing rules and preferences: how they want things scheduled, written, and filed, and when they want to be interrupted.
+` + `- People and organisations, and the relationship: who they are, what they are involved in, how to reach them.
+` + `- Recurring commitments and routines, and the constraints around them.
+` + `- Obligations and their state: what was promised to whom, the deadline, and what is still outstanding.
+` + `- Decisions made and decisions deferred, with the reasoning and the trade accepted.
+` + `- Corrections to earlier beliefs, recorded explicitly as corrections.
+` + `
+` + `Live commitment state IS durable knowledge here, not ephemeral chatter. An outstanding obligation, a travel window, an unanswered request, a "currently X" arrangement — these are precisely what this agent must recall later. Keep them, and embed the date inside the observation text so a later reader can judge staleness. Do not drop them as ephemeral.
+` + `
+` + `Granularity: one observation per person, arrangement, or obligation. Aggregate repeated mentions into that observation rather than creating siblings; never merge two different people or two different commitments into one.
+` + `
+` + "Do not synthesise from transcript exhaust: tool calls and their results, message or event identifiers, or narration of what the assistant did."
   },
   coding: {
-    disposition: { skepticism: 4, literalism: 5, empathy: 2 }
+    disposition: { skepticism: 4, literalism: 5, empathy: 2 },
+    observations_mission: `You consolidate the memory of a software-engineering agent. This bank records real work on real codebases.
+` + `
+` + `Synthesise into durable observations:
+` + `- Architecture and design decisions, each with its rationale and the trade accepted.
+` + `- Root causes, with the evidence chain, and negative results — what was ruled out matters as much as what was found.
+` + `- How the repository works: build, test, and lint commands, conventions, CI gates, and where things live.
+` + `- Outcomes of code work: issue and PR numbers, what changed, whether it merged, what review found.
+` + `- The user's standing rules, preferences, and corrections to this agent's behaviour.
+` + `- Corrections to earlier beliefs, recorded explicitly as corrections.
+` + `
+` + `Repository and service state IS durable knowledge here, not ephemeral chatter. Versions, a failing gate, an open PR, a measured number, a "currently X" claim — these are precisely what this agent must recall later. Keep them, and embed the date inside the observation text so a later reader can judge staleness. Do not drop them as ephemeral.
+` + `
+` + `Granularity: one observation per distinct decision, cause, convention, or work item. Aggregate repeated evidence about the same one into that observation rather than creating siblings; never merge two separate decisions into a single summary.
+` + `
+` + "Do not synthesise from transcript exhaust: tool calls and their results, scratch paths, session or request identifiers, or narration of what the agent did. Prefer specific and falsifiable — naming the file, the number, the commit, or the decision beats summarising the topic."
   }
 };
 

package/dist/cli/autoaccept-poll.js

@@ -295,6 +295,182 @@ function queryPendingPermission(agentName, opts = {}) {
   });
 }
 
+// telegram-plugin/flood-circuit-breaker.ts
+import {
+  existsSync,
+  readFileSync as readFileSync2,
+  writeFileSync as writeFileSync2,
+  mkdirSync as mkdirSync2,
+  chmodSync as chmodSync2,
+  unlinkSync as unlinkSync2,
+  renameSync
+} from "node:fs";
+import { dirname as dirname2, join as join3 } from "node:path";
+
+// telegram-plugin/flood-429-ledger.ts
+import { readFileSync, writeFileSync, mkdirSync, chmodSync, unlinkSync } from "node:fs";
+import { dirname, join as join2 } from "node:path";
+var FLOOD_429_LEDGER_FILE = "429-ledger.json";
+var FLOOD_429_LEDGER_RETENTION_MS = 30 * 24 * 60 * 60 * 1000;
+var PENALTY_WINDOW_MS = 7 * 24 * 60 * 60 * 1000;
+var TRIVIAL_PRESSURE_WINDOW_MS = 24 * 60 * 60 * 1000;
+function flood429LedgerPath(stateDir) {
+  return join2(stateDir, FLOOD_429_LEDGER_FILE);
+}
+function parseEpisode(raw) {
+  if (typeof raw !== "object" || raw === null)
+    return null;
+  const r = raw;
+  const num = (v) => typeof v === "number" && Number.isFinite(v) ? v : null;
+  const firstTs = num(r.firstTs);
+  const untilTs = num(r.untilTs);
+  const peak = num(r.peakRetryAfterSec);
+  if (firstTs === null || untilTs === null || peak === null)
+    return null;
+  return {
+    firstTs,
+    lastTs: num(r.lastTs) ?? firstTs,
+    peakRetryAfterSec: peak,
+    untilTs,
+    count: num(r.count) ?? 1
+  };
+}
+function readFlood429LedgerResult(path, readFile = (p) => readFileSync(p, "utf-8")) {
+  let text;
+  try {
+    text = readFile(path);
+  } catch (err) {
+    const code = err?.code;
+    if (code === "ENOENT" || code === "ENOTDIR")
+      return { status: "absent", episodes: [] };
+    return {
+      status: "unreadable",
+      episodes: [],
+      error: `${code ?? "EUNKNOWN"}: ${err?.message ?? String(err)}`
+    };
+  }
+  let parsed;
+  try {
+    parsed = JSON.parse(text);
+  } catch (err) {
+    return { status: "corrupt", episodes: [], error: err?.message ?? String(err) };
+  }
+  if (!Array.isArray(parsed)) {
+    return { status: "corrupt", episodes: [], error: "ledger is not a JSON array" };
+  }
+  const out = [];
+  let dropped = 0;
+  for (const raw of parsed) {
+    const ep = parseEpisode(raw);
+    if (ep)
+      out.push(ep);
+    else
+      dropped += 1;
+  }
+  if (dropped > 0 && out.length === 0) {
+    return { status: "corrupt", episodes: [], error: `${dropped} unparseable entr(ies)` };
+  }
+  return { status: "ok", episodes: out };
+}
+
+// telegram-plugin/flood-circuit-breaker.ts
+var blindLoggedAt = new Map;
+var FLOOD_WINDOWS_FILE = "flood-windows.json";
+function floodWindowsPath(stateDir) {
+  return join3(stateDir, FLOOD_WINDOWS_FILE);
+}
+var FLOOD_WINDOWS_CORRUPT_SUPPRESS_MS = 5 * 60000;
+function readFloodWindows(path, now, log = (l) => process.stderr.write(l)) {
+  if (!existsSync(path))
+    return [];
+  const failSafe = (why) => {
+    log(`telegram gateway: flood-breaker: scoped-windows file ${path} is ${why} \u2014 ` + `failing SAFE: opening a conservative ${FLOOD_WINDOWS_CORRUPT_SUPPRESS_MS}ms global ` + `window rather than booting into a possible open ban with no windows (issue #3106)
+`);
+    return [
+      {
+        scopeKey: "global",
+        untilTs: now + FLOOD_WINDOWS_CORRUPT_SUPPRESS_MS,
+        retryAfterSrc: `failsafe:${why}`,
+        observedAt: now
+      }
+    ];
+  };
+  let text;
+  try {
+    text = readFileSync2(path, "utf-8");
+  } catch (err) {
+    const code = err?.code;
+    if (code === "ENOENT" || code === "ENOTDIR")
+      return [];
+    return failSafe(`unreadable (${code ?? "error"})`);
+  }
+  let raw;
+  try {
+    raw = JSON.parse(text);
+  } catch {
+    return failSafe("corrupt (invalid JSON)");
+  }
+  if (!Array.isArray(raw))
+    return failSafe("corrupt (not an array)");
+  const out = [];
+  for (const r of raw) {
+    const rec = r;
+    if (typeof rec.scopeKey !== "string" || typeof rec.untilTs !== "number")
+      continue;
+    if (rec.untilTs <= now)
+      continue;
+    out.push({
+      scopeKey: rec.scopeKey,
+      untilTs: rec.untilTs,
+      retryAfterSrc: typeof rec.retryAfterSrc === "string" ? rec.retryAfterSrc : "unknown",
+      observedAt: typeof rec.observedAt === "number" ? rec.observedAt : now,
+      ...typeof rec.alertedAt === "number" ? { alertedAt: rec.alertedAt } : {}
+    });
+  }
+  return out;
+}
+
+// src/agents/flood-pressure-probe.ts
+var FLOOD_PRESSURE_GRACE_MS = 120000;
+function scopeBlocksCardDelivery(scopeKey) {
+  return !scopeKey.startsWith("msg-edit:");
+}
+var NO_PRESSURE = { active: false, reason: "" };
+function resolveTelegramStateDir() {
+  return process.env.TELEGRAM_STATE_DIR ?? "/state/agent/telegram";
+}
+function readFloodPressure(now, opts = {}) {
+  const stateDir = opts.stateDir ?? resolveTelegramStateDir();
+  const graceMs = opts.graceMs ?? FLOOD_PRESSURE_GRACE_MS;
+  try {
+    const windows = readFloodWindows(floodWindowsPath(stateDir), now - graceMs, () => {});
+    for (const w of windows) {
+      if (!scopeBlocksCardDelivery(w.scopeKey))
+        continue;
+      const msLeft = w.untilTs - now;
+      return {
+        active: true,
+        reason: msLeft > 0 ? `open flood window ${w.scopeKey} (${Math.ceil(msLeft / 1000)}s left, src=${w.retryAfterSrc})` : `flood window ${w.scopeKey} closed ${Math.ceil(-msLeft / 1000)}s ago (within ${Math.round(graceMs / 1000)}s grace)`
+      };
+    }
+    const ledger = readFlood429LedgerResult(flood429LedgerPath(stateDir));
+    if (ledger.status === "ok") {
+      for (const ep of ledger.episodes) {
+        const lastEvidenceTs = Math.max(ep.lastTs, ep.untilTs);
+        if (lastEvidenceTs >= now - graceMs) {
+          return {
+            active: true,
+            reason: `429 episode ${ep.count}x peak=${ep.peakRetryAfterSec}s last seen ` + `${Math.max(0, Math.round((now - ep.lastTs) / 1000))}s ago ` + `(within ${Math.round(graceMs / 1000)}s grace)`
+          };
+        }
+      }
+    }
+    return NO_PRESSURE;
+  } catch {
+    return NO_PRESSURE;
+  }
+}
+
 // src/agents/wedge-watchdog.ts
 var WEDGE_FOOTER_SIGNATURE = /(?=[\s\S]*[Ee]sc(?:ape)?[^\n]*cancel)(?=[\s\S]*(?:to select|to navigate|\u2191\/\u2193))/;
 var RATE_LIMIT_MENU_SIGNATURE = /(?=[\s\S]*Stop and wait for)(?=[\s\S]*(?:usage credits|Upgrade your plan|\/rate-limit-options))/;
@@ -438,6 +614,8 @@ var DEFAULT_COOLDOWN_MS = 60000;
 var DEFAULT_CONFIRM_MODAL_POLLS = 3;
 var DEFAULT_MANIFEST_STALL_POLLS = 60;
 var DEFAULT_PERMISSION_PROMPT_POLLS = 3;
+var DEFAULT_PERMISSION_CARDLESS_POLLS = 24;
+var DEFAULT_PERMISSION_FLOOD_MAX_POLLS = 120;
 function envInt(name, fallback) {
   const raw = process.env[name];
   if (raw === undefined || raw.trim() === "")
@@ -467,6 +645,9 @@ async function runWedgeWatchdog(opts) {
   const confirmModalPolls = opts.confirmModalPolls ?? envInt("SWITCHROOM_WEDGE_CONFIRM_POLLS", DEFAULT_CONFIRM_MODAL_POLLS);
   const permissionPromptSignature = opts.permissionPromptSignature === null ? null : opts.permissionPromptSignature ?? PERMISSION_PROMPT_SIGNATURE;
   const permissionPromptPolls = opts.permissionPromptPolls ?? envInt("SWITCHROOM_WEDGE_PERMISSION_POLLS", DEFAULT_PERMISSION_PROMPT_POLLS);
+  const permissionCardlessPolls = Math.max(permissionPromptPolls, opts.permissionCardlessPolls ?? envInt("SWITCHROOM_WEDGE_PERMISSION_CARDLESS_POLLS", DEFAULT_PERMISSION_CARDLESS_POLLS));
+  const floodPressure = opts.floodPressure === null ? null : opts.floodPressure ?? ((n) => readFloodPressure(n));
+  const permissionFloodMaxPolls = Math.max(permissionCardlessPolls, opts.permissionFloodMaxPolls ?? envInt("SWITCHROOM_WEDGE_PERMISSION_FLOOD_MAX_POLLS", DEFAULT_PERMISSION_FLOOD_MAX_POLLS));
   const manifestStallSignature = opts.manifestStallSignature === null ? null : opts.manifestStallSignature ?? MANIFESTING_SIGNATURE;
   const manifestStallPolls = opts.manifestStallPolls ?? envInt("SWITCHROOM_WEDGE_MANIFEST_POLLS", DEFAULT_MANIFEST_STALL_POLLS);
   const requestRestart = opts.requestRestart;
@@ -485,6 +666,8 @@ async function runWedgeWatchdog(opts) {
   let confirmModalFires = 0;
   let permissionPromptFires = 0;
   let permissionPromptDeferrals = 0;
+  let permissionPromptFloodHolds = 0;
+  let permissionPromptCardlessHolds = 0;
   let restartEscalations = 0;
   let polls = 0;
   let confirmModalPresent = 0;
@@ -493,6 +676,7 @@ async function runWedgeWatchdog(opts) {
   let lastManifestKey = null;
   let confirmCooldownUntil = 0;
   let permissionCooldownUntil = 0;
+  let lastPermissionHoldLogAt = Number.NEGATIVE_INFINITY;
   while (polls < maxPolls) {
     polls++;
     let text = "";
@@ -602,17 +786,37 @@ async function runWedgeWatchdog(opts) {
           permissionPromptDeferrals++;
           permissionCooldownUntil = now() + cooldownMs;
         } else {
-          const reason = queryPendingPermission2 == null ? "card-aware check disabled" : status && status.ok ? "gateway reports no pending permission (card-less prompt)" : `gateway unreachable/no answer within budget${status && !status.ok ? ` (${status.reason})` : ""}`;
-          console.error(`[wedge-watchdog] ${opts.agentName}: dismissing stuck per-tool ` + `permission prompt (Esc == decline == safe DENY) after ` + `${permissionPromptPresent} polls present ` + `(~${Math.round(permissionPromptPresent * pollIntervalMs / 1000)}s) \u2014 ${reason}`);
-          try {
-            send(opts.agentName, ["Escape"]);
-          } catch (err) {
-            console.error(`[wedge-watchdog] ${opts.agentName}: send threw: ${err.message}`);
+          const held = now();
+          const pressure = floodPressure ? floodPressure(held) : null;
+          const logHold = (line) => {
+            if (held - lastPermissionHoldLogAt < cooldownMs)
+              return;
+            lastPermissionHoldLogAt = held;
+            console.error(line);
+          };
+          const floodHold = pressure?.active === true && permissionPromptPresent < permissionFloodMaxPolls;
+          const cardless = status != null && status.ok && !status.pending;
+          const cardlessHold = !floodHold && cardless && permissionPromptPresent < permissionCardlessPolls;
+          if (floodHold) {
+            permissionPromptFloodHolds++;
+            logHold(`[wedge-watchdog] ${opts.agentName}: permission prompt present ` + `${permissionPromptPresent} polls ` + `(~${Math.round(permissionPromptPresent * pollIntervalMs / 1000)}s) but ` + `Telegram is flood-throttling this agent (${pressure.reason}) \u2014 ` + `an approval card may be deferred, NOT sending Esc ` + `(ceiling ${permissionFloodMaxPolls} polls)`);
+          } else if (cardlessHold) {
+            permissionPromptCardlessHolds++;
+            logHold(`[wedge-watchdog] ${opts.agentName}: gateway reports no pending ` + `permission after ${permissionPromptPresent} polls ` + `(~${Math.round(permissionPromptPresent * pollIntervalMs / 1000)}s) \u2014 ` + `holding Esc until the card-less streak of ${permissionCardlessPolls} polls ` + `(~${Math.round(permissionCardlessPolls * pollIntervalMs / 1000)}s), ` + `in case the card is merely late`);
+          } else {
+            const reason = queryPendingPermission2 == null ? "card-aware check disabled" : status && status.ok ? `gateway reports no pending permission (card-less prompt) for ${permissionCardlessPolls}+ polls` : `gateway unreachable/no answer within budget${status && !status.ok ? ` (${status.reason})` : ""}`;
+            const floodNote = pressure?.active ? ` [flood ceiling ${permissionFloodMaxPolls} polls reached while ${pressure.reason}]` : "";
+            console.error(`[wedge-watchdog] ${opts.agentName}: dismissing stuck per-tool ` + `permission prompt (Esc == decline == safe DENY) after ` + `${permissionPromptPresent} polls present ` + `(~${Math.round(permissionPromptPresent * pollIntervalMs / 1000)}s) \u2014 ${reason}${floodNote}`);
+            try {
+              send(opts.agentName, ["Escape"]);
+            } catch (err) {
+              console.error(`[wedge-watchdog] ${opts.agentName}: send threw: ${err.message}`);
+            }
+            fires++;
+            permissionPromptFires++;
+            permissionCooldownUntil = now() + cooldownMs;
+            permissionPromptPresent = 0;
           }
-          fires++;
-          permissionPromptFires++;
-          permissionCooldownUntil = now() + cooldownMs;
-          permissionPromptPresent = 0;
         }
       }
     } else if (isConfirmModal) {
@@ -667,6 +871,8 @@ async function runWedgeWatchdog(opts) {
     confirmModalFires,
     permissionPromptFires,
     permissionPromptDeferrals,
+    permissionPromptFloodHolds,
+    permissionPromptCardlessHolds,
     restartEscalations,
     polls,
     reason: "max-polls"
@@ -675,10 +881,10 @@ async function runWedgeWatchdog(opts) {
 
 // src/agents/rate-limit-signal.ts
 import { createConnection as createConnection2 } from "node:net";
-import { join as join2 } from "node:path";
+import { join as join4 } from "node:path";
 function resolveGatewaySocketPath2() {
   const stateDir = process.env.TELEGRAM_STATE_DIR ?? "/state/agent/telegram";
-  return process.env.SWITCHROOM_GATEWAY_SOCKET ?? join2(stateDir, "gateway.sock");
+  return process.env.SWITCHROOM_GATEWAY_SOCKET ?? join4(stateDir, "gateway.sock");
 }
 function signalQuotaWall(agentName, resetAt, opts = {}) {
   const socketPath = opts.socketPath ?? resolveGatewaySocketPath2();
@@ -744,7 +950,7 @@ function signalQuotaWall(agentName, resetAt, opts = {}) {
 import * as net from "node:net";
 import { homedir } from "node:os";
 import { randomUUID as randomUUID2 } from "node:crypto";
-import { join as join3 } from "node:path";
+import { join as join5 } from "node:path";
 
 // node_modules/.bun/zod@3.25.76/node_modules/zod/v3/external.js
 var exports_external = {};
@@ -5031,7 +5237,7 @@ function reviveDate(v) {
   return Number.isNaN(d.getTime()) ? null : d;
 }
 function operatorSocketPath(home = homedir()) {
-  return join3(home, ".switchroom", "state", "auth-broker-operator", "sock");
+  return join5(home, ".switchroom", "state", "auth-broker-operator", "sock");
 }
 function resolveAuthBrokerSocketPath(opts) {
   if (opts?.socket)
@@ -5406,8 +5612,9 @@ async function main() {
   const rateLimitDetect = process.env.SWITCHROOM_RATE_LIMIT_DETECT !== "0";
   const overageSelect = rateLimitDetect && process.env.SWITCHROOM_RATE_LIMIT_OVERAGE !== "0";
   const permissionCardAware = process.env.SWITCHROOM_PERMISSION_CARD_AWARE !== "0";
+  const permissionFloodAware = process.env.SWITCHROOM_WEDGE_FLOOD_AWARE !== "0";
   try {
-    console.error(`[autoaccept-poll] ${agentName}: entering wedge-watchdog (continuous)` + (rateLimitDetect ? " +rate-limit-detect" : " (rate-limit-detect OFF)") + (overageSelect ? " +overage-carveout" : "") + (permissionCardAware ? " +permission-card-aware" : " (permission-card-aware OFF)"));
+    console.error(`[autoaccept-poll] ${agentName}: entering wedge-watchdog (continuous)` + (rateLimitDetect ? " +rate-limit-detect" : " (rate-limit-detect OFF)") + (overageSelect ? " +overage-carveout" : "") + (permissionCardAware ? " +permission-card-aware" : " (permission-card-aware OFF)") + (permissionFloodAware ? " +permission-flood-aware" : " (permission-flood-aware OFF)"));
     const res = await runWedgeWatchdog({
       agentName,
       rateLimitSignature: rateLimitDetect ? undefined : null,
@@ -5416,9 +5623,10 @@ async function main() {
       } : undefined,
       overageDecision: overageSelect ? () => queryActiveOverageServing() : undefined,
       requestRestart: requestWedgeRestart,
-      queryPendingPermission: permissionCardAware ? undefined : null
+      queryPendingPermission: permissionCardAware ? undefined : null,
+      floodPressure: permissionFloodAware ? undefined : null
     });
-    console.error(`[autoaccept-poll] ${agentName}: wedge-watchdog returned reason=${res.reason} fires=${res.fires} rateLimitFires=${res.rateLimitFires} overageCreditSelections=${res.overageCreditSelections} confirmModalFires=${res.confirmModalFires} permissionPromptFires=${res.permissionPromptFires} permissionPromptDeferrals=${res.permissionPromptDeferrals} restartEscalations=${res.restartEscalations}`);
+    console.error(`[autoaccept-poll] ${agentName}: wedge-watchdog returned reason=${res.reason} fires=${res.fires} rateLimitFires=${res.rateLimitFires} overageCreditSelections=${res.overageCreditSelections} confirmModalFires=${res.confirmModalFires} permissionPromptFires=${res.permissionPromptFires} permissionPromptDeferrals=${res.permissionPromptDeferrals} permissionPromptFloodHolds=${res.permissionPromptFloodHolds} permissionPromptCardlessHolds=${res.permissionPromptCardlessHolds} restartEscalations=${res.restartEscalations}`);
   } catch (err) {
     console.error(`[autoaccept-poll] ${agentName}: wedge-watchdog unexpected throw: ${err.message}`);
   }

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

@@ -12106,7 +12106,10 @@ var LiteLLMConfigSchema = exports_external.object({
   admin_key: exports_external.string().optional().describe("LiteLLM master/admin key used at apply time to provision the team + " + "virtual key. Supports a vault reference (e.g. " + "'vault:litellm/master-key') \u2014 resolution happens at apply time via " + "the vault-broker. Never injected into the agent container."),
   team: exports_external.string().optional().describe("LiteLLM team alias the per-agent key is created under. Defaults to " + "'switchroom' (applied in code, not as a schema default)."),
   small_fast_model: exports_external.string().optional().describe("Model id exported as ANTHROPIC_SMALL_FAST_MODEL for the claude CLI's " + "background/fast lane, e.g. 'claude-haiku-4-5-20251001'."),
-  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins).")
+  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins)."),
+  max_budget: exports_external.number().positive().optional().describe("HARD spend cap in USD for this agent's virtual key over one " + "`budget_duration` window. LiteLLM refuses the request once the key's " + "tracked spend exceeds it, so a runaway loop costs at most this much " + "before it is stopped. Defaults to " + "DEFAULT_KEY_MAX_BUDGET_USD (see src/litellm/budget.ts) \u2014 deliberately " + "conservative; raise it per-agent rather than removing it. Set 0 or " + "omit `budget_duration` at your own risk: an uncapped key is only as " + "bounded as the upstream account balance."),
+  soft_budget: exports_external.number().positive().optional().describe("ADVISORY spend threshold in USD. LiteLLM keeps serving past it and " + "raises a budget alert instead. Must be < max_budget. NOTE: LiteLLM " + "accepts soft_budget only on POST /key/generate (GenerateKeyRequest); " + "UpdateKeyRequest does NOT carry it, so changing this value only takes " + "effect on a key that is (re)generated, not on an existing one."),
+  budget_duration: exports_external.string().regex(/^\d+(s|m|h|d|mo)$/, "budget_duration must be a LiteLLM duration like '30d', '24h', '1mo'").optional().describe("Rolling window the budget resets on, in LiteLLM duration syntax " + "('30d', '24h', '1mo'). Defaults to DEFAULT_KEY_BUDGET_DURATION. " + "WITHOUT a duration LiteLLM treats max_budget as a LIFETIME cap that " + "never resets \u2014 the key silently dies for good once it is hit.")
 }).optional().describe("LiteLLM routing config \u2014 opt-in per-agent virtual-key auto-provisioning " + "+ routing env. Default OFF. See LiteLLMConfigSchema doc for the full flow.");
 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`."),
@@ -12125,7 +12128,7 @@ 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 \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, " + "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).")
+  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, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS \u2014 switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY \u2014 a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP \u2014 the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS \u2014 only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS \u2014 the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_MAX_SLOTS reserves out of; " + "unset means upstream's own default), 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)."),