switchroom 0.19.23 → 0.19.24

package/dist/cli/switchroom.js

@@ -2120,7 +2120,7 @@ var init_esm = __esm(() => {
 });
 
 // src/build-info.ts
-var VERSION = "0.19.23", COMMIT_SHA = "56e3c01e";
+var VERSION = "0.19.24", COMMIT_SHA = "0fa8ffb8";
 
 // src/cli/resolve-version.ts
 import { existsSync, readFileSync } from "node:fs";
@@ -13939,17 +13939,19 @@ var init_schema = __esm(() => {
     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.");
   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, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY \u2014 a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS \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).")
   });
   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)."),
@@ -15645,7 +15647,7 @@ function hindsightPerfEnv(caps, overrides = new Map) {
   }
   return out;
 }
-var HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION = 150, HINDSIGHT_DEFAULT_RECALL_MAX_CANDIDATES_PER_SOURCE, HINDSIGHT_DEFAULT_LINK_EXPANSION_PER_ENTITY_LIMIT = 50, HINDSIGHT_DEFAULT_LINK_EXPANSION_TIMEOUT_S = 2, HINDSIGHT_DEFAULT_LLM_REASONING_EFFORT = "low", HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT = 4, HINDSIGHT_DEFAULT_RETAIN_LLM_MAX_CONCURRENT = 1, HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_MAX_CONCURRENT = 1, HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16 = "true", HINDSIGHT_PERF_DEFAULTS_UNGATED, HINDSIGHT_PERF_DEFAULTS_GPU, HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM, HINDSIGHT_PERF_ENV_KEYS;
+var HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION = 150, HINDSIGHT_DEFAULT_RECALL_MAX_CANDIDATES_PER_SOURCE, HINDSIGHT_DEFAULT_LINK_EXPANSION_PER_ENTITY_LIMIT = 50, HINDSIGHT_DEFAULT_LINK_EXPANSION_TIMEOUT_S = 2, HINDSIGHT_DEFAULT_LLM_REASONING_EFFORT = "low", HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT = 4, HINDSIGHT_DEFAULT_RETAIN_LLM_MAX_CONCURRENT = 1, HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_MAX_CONCURRENT = 1, HINDSIGHT_DEFAULT_RERANKER_LOCAL_FP16 = "true", HINDSIGHT_DEFAULT_RERANKER_LOCAL_BATCH_SIZE = 128, HINDSIGHT_DEFAULT_LLM_STRICT_SCHEMA = "true", HINDSIGHT_DEFAULT_LLM_MAX_RETRIES = 2, HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM = 2, HINDSIGHT_PERF_DEFAULTS_UNGATED, HINDSIGHT_PERF_DEFAULTS_GPU, HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM, HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS, HINDSIGHT_PERF_ENV_KEYS;
 var init_hindsight_perf_defaults = __esm(() => {
   HINDSIGHT_DEFAULT_RECALL_MAX_CANDIDATES_PER_SOURCE = Math.ceil(HINDSIGHT_RERANKER_MAX_CANDIDATES_FOR_DERIVATION * 0.4);
   HINDSIGHT_PERF_DEFAULTS_UNGATED = [
@@ -15661,10 +15663,18 @@ var init_hindsight_perf_defaults = __esm(() => {
       "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_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)
+    ]
   ];
   HINDSIGHT_PERF_DEFAULTS_LOCAL_LLM = [
     ["HINDSIGHT_API_LLM_MAX_CONCURRENT", String(HINDSIGHT_DEFAULT_LLM_MAX_CONCURRENT)],
@@ -15675,13 +15685,21 @@ var init_hindsight_perf_defaults = __esm(() => {
     [
       "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)]
   ];
+  HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
+    "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY"
+  ]);
   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
@@ -15731,6 +15749,185 @@ var init_hindsight_pg_defaults = __esm(() => {
   HINDSIGHT_PG_ENV_KEYS = new Set(HINDSIGHT_PG_DEFAULTS.map(([k]) => k));
 });
 
+// src/setup/self-hosted-url.ts
+function isLoopbackHttpUrl(url) {
+  try {
+    const u = new URL(url.includes("://") ? url : `http://${url}`);
+    const h = (u.hostname || "").toLowerCase().replace(/^\[|\]$/g, "");
+    return h === "localhost" || h === "127.0.0.1" || h === "::1";
+  } catch {
+    return false;
+  }
+}
+function isSelfHostedHttpUrl(url) {
+  let host;
+  try {
+    const u = new URL(url.includes("://") ? url : `http://${url}`);
+    host = (u.hostname || "").toLowerCase().replace(/^\[|\]$/g, "");
+  } catch {
+    return false;
+  }
+  if (!host)
+    return false;
+  if (isLoopbackHttpUrl(url))
+    return true;
+  if (host === "host.docker.internal" || host === "gateway.docker.internal")
+    return true;
+  if (host.endsWith(".local") || host.endsWith(".internal"))
+    return true;
+  if (/^f[cd][0-9a-f]{2}:/.test(host) || /^fe[89ab][0-9a-f]:/.test(host))
+    return true;
+  const v4 = host.match(/^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.(\d{1,3})$/);
+  if (!v4)
+    return false;
+  const [a, b] = [Number(v4[1]), Number(v4[2])];
+  if (a === 10)
+    return true;
+  if (a === 127)
+    return true;
+  if (a === 192 && b === 168)
+    return true;
+  if (a === 172 && b >= 16 && b <= 31)
+    return true;
+  if (a === 169 && b === 254)
+    return true;
+  return false;
+}
+
+// src/setup/hindsight-context-budget.ts
+function clamp(value, floor, ceiling) {
+  return Math.min(ceiling, Math.max(floor, value));
+}
+function usableContextTokens(windowTokens) {
+  return windowTokens - Math.floor(windowTokens * HINDSIGHT_CONTEXT_SAFETY_FRACTION);
+}
+function defaultContextWindowForProvider(provider, baseUrl) {
+  const url = baseUrl?.trim();
+  if (url && isSelfHostedHttpUrl(url))
+    return HINDSIGHT_CONSERVATIVE_CONTEXT_WINDOW;
+  const p = provider.trim().toLowerCase();
+  if (p === "claude-code" || p === "anthropic")
+    return HINDSIGHT_CLAUDE_CONTEXT_WINDOW;
+  return HINDSIGHT_CONSERVATIVE_CONTEXT_WINDOW;
+}
+function resolveLaneContextWindow(lane, llm) {
+  const perOp = llm?.[lane];
+  const provider = perOp?.provider?.trim() || llm?.provider?.trim() || "claude-code";
+  const baseUrl = perOp?.base_url?.trim() || undefined;
+  const perOpWindow = perOp?.context_window;
+  if (typeof perOpWindow === "number" && perOpWindow > 0) {
+    return { windowTokens: Math.floor(perOpWindow), windowSource: "per-op", provider, baseUrl };
+  }
+  const globalWindow = llm?.context_window;
+  if (typeof globalWindow === "number" && globalWindow > 0) {
+    return { windowTokens: Math.floor(globalWindow), windowSource: "global", provider, baseUrl };
+  }
+  return {
+    windowTokens: defaultContextWindowForProvider(provider, baseUrl),
+    windowSource: "provider-default",
+    provider,
+    baseUrl
+  };
+}
+function resolveHindsightContextBudget(llm) {
+  const cons = resolveLaneContextWindow("consolidation", llm);
+  const ret = resolveLaneContextWindow("retain", llm);
+  const consMaxCompletion = clamp(Math.floor(cons.windowTokens / 4), HINDSIGHT_CONSOLIDATION_MAX_COMPLETION_FLOOR, HINDSIGHT_CONSOLIDATION_MAX_COMPLETION_CEILING);
+  const consUsable = usableContextTokens(cons.windowTokens);
+  const consPromptBudget = consUsable - consMaxCompletion;
+  const batchSize = clamp(Math.floor((consPromptBudget - HINDSIGHT_CONSOLIDATION_PROMPT_OVERHEAD_TOKENS) / HINDSIGHT_CONSOLIDATION_TOKENS_PER_FACT), HINDSIGHT_CONSOLIDATION_BATCH_SIZE_FLOOR, HINDSIGHT_CONSOLIDATION_BATCH_SIZE_CEILING);
+  const consPrompt = HINDSIGHT_CONSOLIDATION_PROMPT_OVERHEAD_TOKENS + batchSize * HINDSIGHT_CONSOLIDATION_TOKENS_PER_FACT;
+  const retMaxCompletion = clamp(Math.floor(ret.windowTokens * 3 / 16), HINDSIGHT_RETAIN_MAX_COMPLETION_FLOOR, HINDSIGHT_RETAIN_MAX_COMPLETION_CEILING);
+  const refl = resolveLaneContextWindow("reflect", llm);
+  const reflCompletionReserve = clamp(Math.floor(refl.windowTokens * 3 / 16), HINDSIGHT_RETAIN_MAX_COMPLETION_FLOOR, HINDSIGHT_RETAIN_MAX_COMPLETION_CEILING);
+  const reflUsable = usableContextTokens(refl.windowTokens);
+  const reflMaxContext = Math.max(HINDSIGHT_REFLECT_MAX_CONTEXT_FLOOR, Math.floor((reflUsable - reflCompletionReserve) / 1000) * 1000);
+  return {
+    consolidation: {
+      lane: "consolidation",
+      windowTokens: cons.windowTokens,
+      usableTokens: consUsable,
+      windowSource: cons.windowSource,
+      provider: cons.provider,
+      batchSize,
+      maxCompletionTokens: consMaxCompletion,
+      estimatedPromptTokens: consPrompt,
+      worstCaseTotalTokens: consPrompt + consMaxCompletion
+    },
+    reflect: {
+      lane: "reflect",
+      windowTokens: refl.windowTokens,
+      usableTokens: reflUsable,
+      windowSource: refl.windowSource,
+      provider: refl.provider,
+      maxContextTokens: reflMaxContext,
+      completionReserveTokens: reflCompletionReserve,
+      estimatedPromptTokens: reflMaxContext,
+      worstCaseTotalTokens: reflMaxContext + reflCompletionReserve
+    },
+    retain: {
+      lane: "retain",
+      windowTokens: ret.windowTokens,
+      usableTokens: usableContextTokens(ret.windowTokens),
+      windowSource: ret.windowSource,
+      provider: ret.provider,
+      maxCompletionTokens: retMaxCompletion,
+      estimatedPromptTokens: HINDSIGHT_RETAIN_PROMPT_ESTIMATE_TOKENS,
+      worstCaseTotalTokens: HINDSIGHT_RETAIN_PROMPT_ESTIMATE_TOKENS + retMaxCompletion
+    }
+  };
+}
+function laneCheck(b) {
+  return { ...b, completionTokens: b.maxCompletionTokens, completionLabel: "completion" };
+}
+function reflectLaneCheck(b) {
+  return {
+    ...b,
+    completionTokens: b.completionReserveTokens,
+    completionLabel: "completion reserve"
+  };
+}
+function laneFailure(b) {
+  if (b.worstCaseTotalTokens > b.usableTokens) {
+    return `hindsight ${b.lane}: worst-case ${b.worstCaseTotalTokens} tokens ` + `(${b.estimatedPromptTokens} prompt + ${b.completionTokens} ${b.completionLabel}) ` + `exceeds the ${b.usableTokens} usable tokens of ` + `the declared ${b.windowTokens}-token context window ` + `(${Math.round(HINDSIGHT_CONTEXT_SAFETY_FRACTION * 100)}% safety band held back; ` + `source: ${b.windowSource}, provider: ${b.provider}).`;
+  }
+  if (b.lane === "retain" && b.maxCompletionTokens !== undefined && b.maxCompletionTokens <= HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE) {
+    return `hindsight retain: derived max_completion_tokens ${b.maxCompletionTokens} is not ` + `greater than the upstream retain_chunk_size ${HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE}; ` + `the hindsight container refuses to boot with this combination.`;
+  }
+  return;
+}
+function assertHindsightContextBudgetFits(budget) {
+  for (const lane of [
+    laneCheck(budget.consolidation),
+    laneCheck(budget.retain),
+    reflectLaneCheck(budget.reflect)
+  ]) {
+    const failure = laneFailure(lane);
+    if (failure) {
+      throw new HindsightContextBudgetError(lane.lane, `${failure}
+` + `A local backend does NOT error on context overflow \u2014 llama.cpp context-shift ` + `silently drops the system prompt and returns HTTP 200 with conversational text ` + `instead of the required JSON, so this must be caught here.
+` + `Fix: raise \`hindsight.llm.context_window\` (or \`hindsight.llm.${lane.lane}.context_window\`) ` + `to the backend's real window, or point the lane at a larger-window model.`);
+    }
+  }
+}
+function resolveCheckedHindsightContextBudget(llm) {
+  const budget = resolveHindsightContextBudget(llm);
+  assertHindsightContextBudgetFits(budget);
+  return budget;
+}
+var HINDSIGHT_CLAUDE_CONTEXT_WINDOW = 200000, HINDSIGHT_CONSERVATIVE_CONTEXT_WINDOW = 32768, HINDSIGHT_CONTEXT_SAFETY_FRACTION = 0.2, HINDSIGHT_CONSOLIDATION_PROMPT_OVERHEAD_TOKENS = 2270, HINDSIGHT_CONSOLIDATION_TOKENS_PER_FACT = 2500, HINDSIGHT_RETAIN_PROMPT_ESTIMATE_TOKENS = 8000, HINDSIGHT_CONSOLIDATION_BATCH_SIZE_CEILING = 12, HINDSIGHT_CONSOLIDATION_BATCH_SIZE_FLOOR = 1, HINDSIGHT_CONSOLIDATION_MAX_COMPLETION_CEILING = 16384, HINDSIGHT_RETAIN_MAX_COMPLETION_CEILING = 16384, HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE = 3000, HINDSIGHT_CONSOLIDATION_MAX_COMPLETION_FLOOR = 1024, HINDSIGHT_RETAIN_MAX_COMPLETION_FLOOR, HINDSIGHT_REFLECT_MAX_CONTEXT_FLOOR = 2048, HindsightContextBudgetError;
+var init_hindsight_context_budget = __esm(() => {
+  HINDSIGHT_RETAIN_MAX_COMPLETION_FLOOR = HINDSIGHT_UPSTREAM_RETAIN_CHUNK_SIZE + 72;
+  HindsightContextBudgetError = class HindsightContextBudgetError extends Error {
+    lane;
+    constructor(lane, message) {
+      super(message);
+      this.name = "HindsightContextBudgetError";
+      this.lane = lane;
+    }
+  };
+});
+
 // src/util/atomic.ts
 var exports_atomic = {};
 __export(exports_atomic, {
@@ -15854,8 +16051,9 @@ function hindsightInteractiveLlmTimeoutSeconds() {
 function hindsightConsolidationLlmTimeoutSeconds() {
   return clientBudgetSeconds(LITELLM_TIMEOUT_TIERS.consolidation);
 }
-function hindsightLlmBudgetEnv() {
-  const maxCompletionTokens = HINDSIGHT_DEFAULT_RETAIN_MAX_COMPLETION_TOKENS;
+function hindsightLlmBudgetEnv(llm) {
+  const contextBudget = resolveCheckedHindsightContextBudget(llm);
+  const maxCompletionTokens = Math.min(HINDSIGHT_DEFAULT_RETAIN_MAX_COMPLETION_TOKENS, contextBudget.retain.maxCompletionTokens);
   const interactiveS = hindsightInteractiveLlmTimeoutSeconds();
   const retainS = hindsightRetainClientTimeoutSeconds();
   const consolidationS = hindsightConsolidationLlmTimeoutSeconds();
@@ -15885,7 +16083,16 @@ function hindsightLlmBudgetEnv() {
     ["HINDSIGHT_API_LLM_TIMEOUT", String(interactiveS)],
     ["HINDSIGHT_API_REFLECT_LLM_TIMEOUT", String(interactiveS)],
     ["HINDSIGHT_API_RETAIN_LLM_TIMEOUT", String(retainS)],
-    ["HINDSIGHT_API_CONSOLIDATION_LLM_TIMEOUT", String(consolidationS)]
+    ["HINDSIGHT_API_CONSOLIDATION_LLM_TIMEOUT", String(consolidationS)],
+    ["HINDSIGHT_API_CONSOLIDATION_LLM_BATCH_SIZE", String(contextBudget.consolidation.batchSize)],
+    [
+      "HINDSIGHT_API_CONSOLIDATION_MAX_COMPLETION_TOKENS",
+      String(contextBudget.consolidation.maxCompletionTokens)
+    ],
+    [
+      "HINDSIGHT_API_REFLECT_MAX_CONTEXT_TOKENS",
+      String(contextBudget.reflect.maxContextTokens)
+    ]
   ];
 }
 function isPortFree(port) {
@@ -16014,15 +16221,6 @@ function resolveHindsightLlm(llm, litellm) {
   const model = llm?.model?.trim() || (litellm ? litellm.model ?? HINDSIGHT_DEFAULT_LITELLM_MODEL : HINDSIGHT_DEFAULT_MODEL);
   return { provider, model };
 }
-function isLoopbackHttpUrl(url) {
-  try {
-    const u = new URL(url.includes("://") ? url : `http://${url}`);
-    const h = (u.hostname || "").toLowerCase().replace(/^\[|\]$/g, "");
-    return h === "localhost" || h === "127.0.0.1" || h === "::1";
-  } catch {
-    return false;
-  }
-}
 function collectHindsightLlmBaseUrls(llm, litellm) {
   const out = [];
   const seen = new Set;
@@ -16045,40 +16243,6 @@ function hindsightNeedsHostNetwork(llm, litellm) {
     return true;
   return collectHindsightLlmBaseUrls(llm).some(isLoopbackHttpUrl);
 }
-function isSelfHostedHttpUrl(url) {
-  let host;
-  try {
-    const u = new URL(url.includes("://") ? url : `http://${url}`);
-    host = (u.hostname || "").toLowerCase().replace(/^\[|\]$/g, "");
-  } catch {
-    return false;
-  }
-  if (!host)
-    return false;
-  if (isLoopbackHttpUrl(url))
-    return true;
-  if (host === "host.docker.internal" || host === "gateway.docker.internal")
-    return true;
-  if (host.endsWith(".local") || host.endsWith(".internal"))
-    return true;
-  if (/^f[cd][0-9a-f]{2}:/.test(host) || /^fe[89ab][0-9a-f]:/.test(host))
-    return true;
-  const v4 = host.match(/^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.(\d{1,3})$/);
-  if (!v4)
-    return false;
-  const [a, b] = [Number(v4[1]), Number(v4[2])];
-  if (a === 10)
-    return true;
-  if (a === 127)
-    return true;
-  if (a === 192 && b === 168)
-    return true;
-  if (a === 172 && b >= 16 && b <= 31)
-    return true;
-  if (a === 169 && b === 254)
-    return true;
-  return false;
-}
 function hindsightLocalLlmEnabled(llm, litellm, override) {
   if (override !== undefined)
     return override;
@@ -16140,13 +16304,11 @@ function startHindsight(ports, litellm, imageTag, llm, mirrorDir, gpu, perf) {
     `HINDSIGHT_API_RECALL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT}`,
     "-e",
     `HINDSIGHT_API_REFLECT_WALL_TIMEOUT=${HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S}`,
-    ...hindsightLlmBudgetEnv().flatMap(([k, v]) => ["-e", `${k}=${v}`]),
-    "-e",
-    `HINDSIGHT_API_CONSOLIDATION_LLM_BATCH_SIZE=${HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE}`,
+    ...hindsightLlmBudgetEnv(llm).flatMap(([k, v]) => ["-e", `${k}=${v}`]),
     "-e",
     `HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS=${HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS}`,
     "-e",
-    `HINDSIGHT_API_CONSOLIDATION_LLM_PARALLELISM=${HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM}`,
+    `HINDSIGHT_API_WORKER_CONSOLIDATION_SLOT_LIMIT=${HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT}`,
     "-e",
     `HINDSIGHT_API_CONSOLIDATION_MAX_MEMORIES_PER_ROUND=${HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND}`,
     "-e",
@@ -16375,10 +16537,9 @@ function generateHindsightComposeSnippet(llm, mirrorDir, litellm, gpu, perf) {
     `      - HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT}`,
     `      - HINDSIGHT_API_RECALL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT}`,
     `      - HINDSIGHT_API_REFLECT_WALL_TIMEOUT=${HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S}`,
-    ...hindsightLlmBudgetEnv().map(([k, v]) => `      - ${k}=${v}`),
-    `      - HINDSIGHT_API_CONSOLIDATION_LLM_BATCH_SIZE=${HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE}`,
+    ...hindsightLlmBudgetEnv(llm).map(([k, v]) => `      - ${k}=${v}`),
     `      - HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS=${HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS}`,
-    `      - HINDSIGHT_API_CONSOLIDATION_LLM_PARALLELISM=${HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM}`,
+    `      - HINDSIGHT_API_WORKER_CONSOLIDATION_SLOT_LIMIT=${HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT}`,
     `      - HINDSIGHT_API_CONSOLIDATION_MAX_MEMORIES_PER_ROUND=${HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND}`,
     `      - HINDSIGHT_API_WORKER_ID=${HINDSIGHT_DEFAULT_WORKER_ID}`,
     ...hindsightPerfEnvPairs(llm, litellm, gpu, perf).map(([k, v]) => `      - ${k}=${v}`),
@@ -16481,7 +16642,7 @@ async function ensureHindsightConsumer(configPath, account, uid = HINDSIGHT_DEFA
   atomicWriteFileSync2(configPath, tail, mode);
   return { added: true, reason: "added" };
 }
-var HINDSIGHT_DEFAULT_API_PORT = 18888, HINDSIGHT_DEFAULT_UI_PORT = 9999, HINDSIGHT_DEFAULT_MCP_URL, HINDSIGHT_DEFAULT_API_BASE_URL, HINDSIGHT_DEFAULT_MAX_OBSERVATIONS_PER_SCOPE = 1000, HINDSIGHT_CONSUMER_NAME = "hindsight", HINDSIGHT_DEFAULT_WORKER_ID = "switchroom-hindsight", HINDSIGHT_DATA_VOLUME = "switchroom-hindsight-data", HINDSIGHT_DEFAULT_UID = 11000, HINDSIGHT_IMAGE_REPO = "ghcr.io/switchroom/switchroom-hindsight", HINDSIGHT_IMAGE, HINDSIGHT_DEFAULT_MODEL = "claude-sonnet-5", HINDSIGHT_DEFAULT_LITELLM_MODEL = "openrouter/google/gemini-3.1-flash-lite", HINDSIGHT_DEFAULT_MCP_STATELESS = true, HINDSIGHT_BROKER_SOCK_VOLUME, HINDSIGHT_CREDS_MIRROR_VOLUME, HINDSIGHT_CRED_DIR = "/run/claude-creds", HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING = "true", HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES = 150, HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4, HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT = 8, HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S = 600, HINDSIGHT_RETAIN_CHUNK_SIZE = 3000, HINDSIGHT_DEFAULT_RETAIN_MAX_COMPLETION_TOKENS = 16384, HINDSIGHT_RETAIN_MIN_TOKENS_PER_SECOND = 80.6, HINDSIGHT_RETAIN_CLIENT_DEADLINE_S, HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE = 12, HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS = 1, HINDSIGHT_DEFAULT_CONSOLIDATION_LLM_PARALLELISM = 2, HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND = 100, HINDSIGHT_DEFAULT_MEM_LIMIT = "8g", HINDSIGHT_DEFAULT_MEM_RESERVATION = "4g", HINDSIGHT_DEFAULT_PIDS_LIMIT = 1000, HINDSIGHT_DEFAULT_SHM_SIZE = "2g", HINDSIGHT_HEALTHCHECK_PY = 'import urllib.request,sys; sys.exit(0 if urllib.request.urlopen("http://localhost:8888/health",timeout=4).getcode()==200 else 1)', HINDSIGHT_HEALTHCHECK_CMD, DOCKER_PROBE_TIMEOUT_MS, defaultDockerProbe = (args) => {
+var HINDSIGHT_DEFAULT_API_PORT = 18888, HINDSIGHT_DEFAULT_UI_PORT = 9999, HINDSIGHT_DEFAULT_MCP_URL, HINDSIGHT_DEFAULT_API_BASE_URL, HINDSIGHT_DEFAULT_MAX_OBSERVATIONS_PER_SCOPE = 1000, HINDSIGHT_CONSUMER_NAME = "hindsight", HINDSIGHT_DEFAULT_WORKER_ID = "switchroom-hindsight", HINDSIGHT_DATA_VOLUME = "switchroom-hindsight-data", HINDSIGHT_DEFAULT_UID = 11000, HINDSIGHT_IMAGE_REPO = "ghcr.io/switchroom/switchroom-hindsight", HINDSIGHT_IMAGE, HINDSIGHT_DEFAULT_MODEL = "claude-sonnet-5", HINDSIGHT_DEFAULT_LITELLM_MODEL = "openrouter/google/gemini-3.1-flash-lite", HINDSIGHT_DEFAULT_MCP_STATELESS = true, HINDSIGHT_BROKER_SOCK_VOLUME, HINDSIGHT_CREDS_MIRROR_VOLUME, HINDSIGHT_CRED_DIR = "/run/claude-creds", HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING = "true", HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES = 150, HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4, HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT = 8, HINDSIGHT_DEFAULT_REFLECT_WALL_TIMEOUT_S = 600, HINDSIGHT_RETAIN_CHUNK_SIZE = 3000, HINDSIGHT_DEFAULT_RETAIN_MAX_COMPLETION_TOKENS = 16384, HINDSIGHT_RETAIN_MIN_TOKENS_PER_SECOND = 80.6, HINDSIGHT_RETAIN_CLIENT_DEADLINE_S, HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_SLOTS = 1, HINDSIGHT_DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND = 500, HINDSIGHT_DEFAULT_CONSOLIDATION_SLOT_LIMIT = 6, HINDSIGHT_DEFAULT_MEM_LIMIT = "8g", HINDSIGHT_DEFAULT_MEM_RESERVATION = "4g", HINDSIGHT_DEFAULT_PIDS_LIMIT = 1000, HINDSIGHT_DEFAULT_SHM_SIZE = "2g", HINDSIGHT_HEALTHCHECK_PY = 'import urllib.request,sys; sys.exit(0 if urllib.request.urlopen("http://localhost:8888/health",timeout=4).getcode()==200 else 1)', HINDSIGHT_HEALTHCHECK_CMD, DOCKER_PROBE_TIMEOUT_MS, defaultDockerProbe = (args) => {
   try {
     return execFileSync("docker", args, {
       stdio: "pipe",
@@ -16499,6 +16660,8 @@ var init_hindsight = __esm(() => {
   init_host_capabilities();
   init_hindsight_perf_defaults();
   init_hindsight_pg_defaults();
+  init_hindsight_context_budget();
+  init_hindsight_perf_defaults();
   HINDSIGHT_DEFAULT_MCP_URL = `http://127.0.0.1:${HINDSIGHT_DEFAULT_API_PORT}/mcp/`;
   HINDSIGHT_DEFAULT_API_BASE_URL = HINDSIGHT_DEFAULT_MCP_URL.replace(/\/mcp\/?$/, "");
   HINDSIGHT_IMAGE = `${HINDSIGHT_IMAGE_REPO}:latest`;
@@ -28932,7 +29095,7 @@ function buildSettingsHooksBlock(p) {
 ` + `  - Question with a short answer \u2192 just reply with the answer.
 ` + "  - Complex tool-driven work \u2192 go straight to the tools (the " + "compose-area preview is the ambient liveness signal), then reply " + 'once with the answer or a genuine mid-work pivot ("halfway ' + 'through \u2014 found an unexpected issue, want me to continue?"). Not ' + `"still working".
 
-` + "NARRATE YOUR INTENT (your reasoning is NOT shown to the user). The " + "user cannot see your thinking \u2014 only your tool activity and your " + "replies. So before you fire a burst of tools or work silently for a " + "while, first write ONE short plain-language line of intent as " + "ordinary working text (NOT a reply tool call): name what you are " + 'about to do and why \u2014 "Now checking the gateway logs to find why ' + 'the turn stalled". The framework mirrors that line into the live ' + "compose-area preview the user watches; because it is plain text and " + "not a reply, it is never a chat message and never pings. One intent " + "line before a silent stretch, in plain words, never raw tool names " + '("calling Bash", "Read(x)") and never a debug dump. And when you run ' + "a Bash command, always give it a plain-English `description` naming " + "the goal \u2014 that description, not the raw command, is what the user " + "sees. This intent narration is the ONE thing to keep saying; it is " + "NOT the placeholder ack banned above (that is a chat reply, which " + `you still skip).
+` + "LABEL THE WORK AS YOU DO IT. The user watches a live preview built " + "from your tool activity, so what you put in a tool call is what they " + "read. Two output requirements. FIRST: every Bash and tool call gets a " + 'plain-English `description` naming the goal \u2014 "Checking the gateway ' + 'logs to find why the turn stalled", never raw tool names ("calling ' + 'Bash", "Read(x)") and never a debug dump. That description, not the ' + "raw command, is what the user sees. SECOND: before a long stretch of " + "work, post ONE short status line saying what you are starting \u2014 one " + "line per stretch, in plain words. Emit it as the `description` on the " + "first tool call of the stretch, or as a status line through the " + "progress tool where you have one; do NOT leave it sitting as unsent " + "transcript text (see the reply rule below). This labelling is the ONE " + "thing to keep doing; it is NOT the placeholder ack banned above (that " + `is a chat reply, which you still skip).
 
 ` + 'Do NOT send a trailing confirmation after your answer \u2014 no "Done.", ' + '"Sent.", "Hope that helps." as a separate message once you have ' + "already replied. Your answer is the last thing the user should " + `see; a follow-up "Done." is dead-air clutter (and the user's ` + `device already pinged on the answer). Stop after the answer.
 
@@ -28942,7 +29105,7 @@ function buildSettingsHooksBlock(p) {
 
 ` + "TURN-END SHAPE: the human reads your turn-end on a phone. Build it " + "for a thumb, not a terminal. LEAD WITH THE ONE DECISION OR ANSWER " + "in the first line \u2014 the single thing the user has to know. A routine " + "turn-end is a FEW SHORT LINES, not a multi-section report; if you " + "find yourself writing more than ~6 lines or stacking labelled " + "sections, you are dumping, not answering. ONE IDEA PER MESSAGE: " + "split genuinely separate topics into separate replies rather than " + "piling sections into one bubble. Push detail BELOW the answer or " + 'leave it out \u2014 offer "want the detail?" instead of pre-emptively ' + "dumping logs, full file contents, or a section per thread you " + "touched. RESERVE the long multi-section structured message for when " + 'the operator EXPLICITLY asked to go deep ("full breakdown", "walk me ' + 'through it", "show everything"). Default to short; expand on ' + "request. A wall of text the user has to re-read to find the one fact " + `that matters is the failure this exists to prevent.
 
-` + 'CRITICAL: "answer" means a call to the reply tool ' + "(mcp__switchroom-telegram__reply). " + "Your terminal/transcript text is NEVER delivered to Telegram \u2014 the " + "user sees only what you send through the reply tool. After a long " + "tool sequence (scheduling, multi-step research, sub-agent handback), " + "do not let your closing narration stand as the answer: end the turn " + "by passing that narration to the reply tool. No reply tool call = the " + "user got nothing, however much text you wrote. Call the reply tool as " + "your FIRST action when you have the answer \u2014 do not write it out as " + "transcript text first and call reply afterward: a framework backstop " + "flushes unsent text after a delay and then your real reply lands late " + "and out of order.</turn-pacing>";
+` + 'CRITICAL: "answer" means a call to the reply tool ' + "(mcp__switchroom-telegram__reply). " + "Your terminal/transcript text is NEVER delivered to Telegram \u2014 the " + "user sees only what you send through the reply tool. After a long " + "tool sequence (scheduling, multi-step research, sub-agent handback), " + "do not let your closing narration stand as the answer: end the turn " + "by passing that narration to the reply tool. No reply tool call = the " + "user got nothing, however much text you wrote. Call the reply tool as " + "your FIRST action when you have the answer \u2014 do not write it out as " + "transcript text first and call reply afterward: a framework backstop " + "flushes unsent text after a delay and then your real reply lands late " + "and out of order. This is why the work-labelling above rides on tool " + "`description`s rather than loose prose: a tool description reaches the " + "preview immediately and is never subject to that flush, so there is " + "exactly one rule \u2014 user-facing text goes through a tool call, always." + "</turn-pacing>";
   const switchroomUserPromptSubmit = [
     ...useHotReloadStable ? [
       {
@@ -30000,22 +30163,19 @@ messaging a capable colleague \u2014 not a tool emitting output. Five beats:
    compose the full answer. This holds even for a pure-thinking
    answer: if it will run to a paragraph, ack first. It is the line
    between a colleague and a black box.
-2. **Then go quiet and work \u2014 but leave a trail of intent.** Heads-down
+2. **Then go quiet and work \u2014 but label what you are doing.** Heads-down
    is correct: do NOT narrate every tool call, and a typing indicator
-   runs automatically (you do not maintain it). One thing you SHOULD do,
-   because your private reasoning is never shown to the user: before a
-   burst of tool calls or a long silent stretch, drop ONE short
-   plain-language line of intent in your own working text \u2014 *what* you
-   are about to do and *why* ("Now checking the gateway logs to find why
-   the turn stalled"). Write it as ordinary text, NOT a \`reply\` call, so
-   it never becomes a chat message and never pings \u2014 the framework
-   mirrors that line into the live preview the user watches while you
-   work. One line per stretch, plain words, never raw tool names
-   ("calling Bash") or debug dumps. It is the difference between the user
-   seeing "checking the logs to find the stall" and seeing a black box.
-   When a step is a \`Bash\` command, always give it a plain-English
-   \`description\` naming the goal \u2014 that description, not the raw command,
-   is what the user sees.
+   runs automatically (you do not maintain it). What you SHOULD do is
+   make the work legible through the tool calls themselves: always give a
+   \`Bash\` (or any tool) call a plain-English \`description\` naming the
+   goal \u2014 "Checking the gateway logs to find why the turn stalled". That
+   description, not the raw command, is what the user sees in the live
+   preview. Before a long silent stretch, put one short status line there
+   too, saying what you are starting. One line per stretch, plain words,
+   never raw tool names ("calling Bash") or debug dumps. It is the
+   difference between the user seeing "checking the logs to find the
+   stall" and seeing a black box. Keep it in the tool call \u2014 loose
+   transcript prose is not delivered and can be flushed late (see beat 5).
 3. **Surface meaningful progress** at genuine inflection points \u2014 a
    hard step finished, a blocker, a pivot, dispatching a sub-agent, a
    notably slow wait, a finding worth knowing now. One short \`reply\`,

package/dist/host-control/main.js

@@ -19434,17 +19434,19 @@ 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, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
 });
 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)."),
@@ -21121,7 +21123,7 @@ function allocateAgentUid(name) {
 }
 
 // src/build-info.ts
-var VERSION = "0.19.23";
+var VERSION = "0.19.24";
 
 // src/cli/agent-config.ts
 import { join as join2 } from "node:path";
@@ -21224,6 +21226,10 @@ 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_PERF_DEFAULTS_UNGATED = [
   [
     "HINDSIGHT_API_RECALL_MAX_CANDIDATES_PER_SOURCE",
@@ -21237,10 +21243,18 @@ 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)
+  ]
 ];
 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)],
@@ -21251,13 +21265,21 @@ 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"
+]);
 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;
@@ -21280,6 +21302,10 @@ var HINDSIGHT_PG_DEFAULTS = [
 ];
 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;
 var HINDSIGHT_DEFAULT_MCP_URL = `http://127.0.0.1:${HINDSIGHT_DEFAULT_API_PORT}/mcp/`;