switchroom 0.19.46 → 0.19.48

package/dist/host-control/main.js

@@ -10994,7 +10994,7 @@ var init_observation_scopes = __esm(() => {
 });
 
 // src/config/schema.ts
-var CodeRepoEntrySchema, AgentBindMountSchema, HttpDiffPollSchema, PollSpecSchema, TelegramMessageActionSchema, WebhookActionSchema, ActionSpecSchema, ScheduleEntrySchema, AgentSoulSchema, AgentToolsSchema, ObservationScopesSchema, ObservationScopeStrategySchema, AntiConfabulationDirectiveSchema, AgentMemorySchema, HookEntrySchema, AgentHooksSchema, SubagentSchema, SessionSchema, SessionContinuitySchema, webhookDispatchRule, TelegramChannelSchema, ChannelsSchema, TIMEZONE_REGEX, ApproverIdSchema, GoogleWorkspaceTierSchema, GoogleWorkspaceConfigSchema, LiteLLMConfigSchema, HindsightPerOpLlmSchema, HindsightConfigSchema, MicrosoftWorkspaceConfigSchema, NotionWorkspaceConfigSchema, AgentGoogleWorkspaceConfigSchema, MicrosoftAccountEmailSchema, MicrosoftToolTokenSchema, MicrosoftAccountBindingSchema, AgentMicrosoftWorkspaceConfigSchema, AgentNotionWorkspaceConfigSchema, ReactionsSchema, ReactionDispatchSchema, releaseBlockFields, ReleaseBlock, RootReleaseBlock, NetworkIsolationSchema, servesField, knowsField, profileFields, ProfileSchema, _omitExtends, defaultsFields, AgentDefaultsSchema, AgentSchema, TelegramConfigSchema, MemoryBackendConfigSchema, VaultConfigSchema, QuotaConfigSchema, AutoReleaseCheckSchema, HostControlConfigSchema, WebServiceConfigSchema, FleetHealthConfigSchema, HostdConfigSchema, CronEgressSchema, CronConfigSchema, UserSchema, SwitchroomConfigSchema;
+var CodeRepoEntrySchema, AgentBindMountSchema, HttpDiffPollSchema, PollSpecSchema, TelegramMessageActionSchema, WebhookActionSchema, ActionSpecSchema, ScheduleEntrySchema, AgentSoulSchema, AgentToolsSchema, ObservationScopesSchema, ObservationScopeStrategySchema, AntiConfabulationDirectiveSchema, AgentMemorySchema, HookEntrySchema, AgentHooksSchema, SubagentSchema, SessionSchema, SessionContinuitySchema, webhookDispatchRule, TelegramChannelSchema, ChannelsSchema, TIMEZONE_REGEX, ApproverIdSchema, GoogleWorkspaceTierSchema, GoogleServiceTokenSchema, GoogleWorkspaceConfigSchema, LiteLLMConfigSchema, HindsightPerOpLlmSchema, HindsightConfigSchema, MicrosoftWorkspaceConfigSchema, NotionWorkspaceConfigSchema, AgentGoogleWorkspaceConfigSchema, MicrosoftAccountEmailSchema, MicrosoftToolTokenSchema, MicrosoftAccountBindingSchema, AgentMicrosoftWorkspaceConfigSchema, AgentNotionWorkspaceConfigSchema, ReactionsSchema, ReactionDispatchSchema, releaseBlockFields, ReleaseBlock, RootReleaseBlock, NetworkIsolationSchema, servesField, knowsField, profileFields, ProfileSchema, _omitExtends, defaultsFields, AgentDefaultsSchema, AgentSchema, TelegramConfigSchema, MemoryBackendConfigSchema, VaultConfigSchema, QuotaConfigSchema, AutoReleaseCheckSchema, HostControlConfigSchema, WebServiceConfigSchema, FleetHealthConfigSchema, HostdConfigSchema, CronEgressSchema, CronConfigSchema, UserSchema, SwitchroomConfigSchema;
 var init_schema = __esm(() => {
   init_zod();
   init_observation_scopes();
@@ -11120,6 +11120,8 @@ var init_schema = __esm(() => {
     profile: exports_external.string().optional().describe("Memory profile bank this agent's curated memory defaults key off — " + "the built-in disposition + observations_mission in PROFILE_MEMORY_DEFAULTS. " + "Decouples the memory profile from `extends` (the filesystem persona " + "profile), so an agent on `extends: default` can opt into the `coding` " + "memory bundle via `memory.profile: coding` without inheriting the coding " + "persona. Resolution: memory.profile → extends → DEFAULT_PROFILE (see " + "resolveMemoryProfile). Unset ⇒ byte-identical to keying off `extends`."),
     bank_mission: exports_external.string().optional().describe("Bank-level mission statement used during recall to contextualize " + "results. NOTE: this is an alias for the Hindsight engine's " + "`reflect_mission` field (verified live: switchroom's bank_mission " + "lands in `config.reflect_mission`). Prefer `reflect_mission` going " + "forward; `bank_mission` is retained for back-compat. If both are " + "set, `reflect_mission` wins. Cascade: override."),
     reflect_mission: exports_external.string().optional().describe("Mission/context steering Hindsight Reflect operations (the bank's " + "'who am I / what matters' framing applied during recall). The " + "engine-accurate name for what `bank_mission` sets. Cascade: override."),
+    reflect_budget: exports_external.enum(["low", "mid", "high"]).optional().describe("Thinking/retrieval budget injected into reflect MCP calls when the " + "caller omits budget. Unset ⇒ shim default (mid). Explicit per-call " + "budget always wins. Higher = better recall on fuzzy queries, more " + "backend latency/compute. stdio-shim transport only (ignored under " + "memory.config.mcp_transport: http). Cascade: override (per-agent " + "wins over default)."),
+    reflect_max_tokens: exports_external.number().int().positive().max(8192).optional().describe("Token cap injected into reflect MCP calls when caller omits " + "max_tokens. Unset ⇒ shim default 1024. Values much above ~2048 risk " + "exceeding Claude Code's MCP output cap, silently dropping the " + "payload — raise deliberately. Explicit per-call values always win. " + "stdio-shim transport only. Cascade: override."),
     retain_mission: exports_external.string().optional().describe("Instructions for the fact extraction LLM during retain. Cascade: override."),
     mental_models: exports_external.array(exports_external.object({
       name: exports_external.string().min(1).describe("Stable model name (identity key for idempotent ensure). Two " + "declarations with the same name in one agent are rejected."),
@@ -11381,11 +11383,18 @@ var init_schema = __esm(() => {
     "extended",
     "complete"
   ]);
+  GoogleServiceTokenSchema = exports_external.enum([
+    "cal",
+    "drive",
+    "docs",
+    "sheets",
+    "slides"
+  ]);
   GoogleWorkspaceConfigSchema = exports_external.object({
     google_client_id: exports_external.string().min(1).describe("Google OAuth client ID (literal string or vault reference e.g. 'vault:google-oauth-client-id')"),
     google_client_secret: exports_external.string().min(1).describe("Google OAuth client secret (literal string or vault reference e.g. 'vault:google-oauth-client-secret')"),
     approvers: exports_external.array(ApproverIdSchema).min(1).describe("Array of numeric Telegram user IDs authorized to approve drive onboarding. " + "At least one must be specified."),
-    tier: GoogleWorkspaceTierSchema.optional().describe("RFC G Phase 1: which upstream MCP tier to expose. " + "core (default) = ~16 tools (Drive+Docs+Sheets+Calendar). " + "extended = ~40 tools (+Slides, Forms, Tasks, Chat). " + "complete = ~60+ tools (+Gmail; not recommended yet — see RFC G §5).")
+    tier: GoogleWorkspaceTierSchema.optional().describe("RFC G Phase 1: which upstream MCP tier to expose. " + "core (default) = ~16 tools (Drive+Docs+Sheets+Calendar tools; the " + "Calendar tools only authenticate when the account holds the opt-in " + "calendar scope — no tier mints it). " + "extended = ~40 tools (+Slides, Forms, Tasks, Chat). " + "complete = ~60+ tools (+Gmail; not recommended yet — see RFC G §5).")
   }).optional();
   LiteLLMConfigSchema = exports_external.object({
     enabled: exports_external.boolean().optional().describe("Opt-in toggle. When true, `switchroom apply` provisions a per-agent " + "LiteLLM virtual key and injects routing env into the container. " + "Default OFF."),
@@ -11418,7 +11427,7 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "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_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, 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_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains), 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_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, 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_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -11604,6 +11613,8 @@ var init_schema = __esm(() => {
       }).optional(),
       bank_mission: exports_external.string().optional(),
       reflect_mission: exports_external.string().optional(),
+      reflect_budget: exports_external.enum(["low", "mid", "high"]).optional(),
+      reflect_max_tokens: exports_external.number().int().positive().max(8192).optional(),
       retain_mission: exports_external.string().optional(),
       observations_mission: exports_external.string().optional(),
       disposition: exports_external.object({
@@ -11892,7 +11903,9 @@ var init_schema = __esm(() => {
     }).transform((v) => v.trim().toLowerCase()), exports_external.object({
       enabled_for: exports_external.array(exports_external.string().regex(/^[a-z0-9][a-z0-9_-]{0,50}$/, {
         message: "Agent name must match the standard agent-name pattern"
-      })).describe("Agent slugs that may read this account's vault slots " + "(`google:<account>:refresh_token` etc). Per-agent ACL is " + "enforced at the broker, not at the agent identity layer — " + "the agent still authenticates via socket-path-as-identity " + "per RFC D §4.1, broker just gates the cross-agent token share.")
+      })).describe("Agent slugs that may read this account's vault slots " + "(`google:<account>:refresh_token` etc). Per-agent ACL is " + "enforced at the broker, not at the agent identity layer — " + "the agent still authenticates via socket-path-as-identity " + "per RFC D §4.1, broker just gates the cross-agent token share."),
+      readonly: exports_external.boolean().optional().describe("Per-account read-only selection (v1 scope model). true → the " + "token minted by `auth google account add` carries ONLY " + ".readonly scope variants (zero write scopes), and the gdrive " + "MCP launcher passes upstream `--read-only` so write tools are " + "not exposed. Written by `account add --readonly`; re-read on " + "`--replace` so a re-consent cannot silently re-widen. Omitted " + "= legacy behaviour (tier-tied read-write document scopes)."),
+      services: exports_external.array(GoogleServiceTokenSchema).min(1).optional().describe("Per-account service selection (v1 scope model). Which Google " + "services the minted token covers AND the gdrive MCP exposes " + "(upstream `--tools`): cal, drive, docs, sheets, slides. " + "Written by `account add --services`; re-read on `--replace`. " + "Omitted = the tier's default services (drive,docs,sheets; " + "+slides at extended/complete).")
     })).optional().describe("RFC G Phase 2: per-Google-account ACL for vault slots holding " + "OAuth refresh tokens. Maps account email → list of agents " + "permitted to read that account's slots. Written by `switchroom " + "auth google enable|disable` (Phase 3); read by the broker on " + "every Google slot access. Replaces RFC D's per-agent vault slot " + "scope (which can't express 'two agents share one Google account')."),
     microsoft_accounts: exports_external.record(exports_external.string().regex(/^[^@\s:]+@[^@\s:]+\.[^@\s:]+$/, {
       message: "Account key must be a Microsoft account email like 'alice@outlook.com' or 'alice@contoso.com' (colons not allowed)"
@@ -21402,7 +21415,7 @@ function allocateAgentUid(name) {
 }
 
 // src/build-info.ts
-var VERSION = "0.19.46";
+var VERSION = "0.19.48";
 
 // src/setup/hindsight-recall-tunables.ts
 var RECALL_DEADLINE_HEADROOM_SECONDS = 2;
@@ -21473,6 +21486,9 @@ for (const name of SHARED_FRAGMENTS) {
   }
 }
 
+// src/memory/scaffold-integration.ts
+init_merge();
+
 // src/litellm/timeout-budget.ts
 var LITELLM_ROUTER_MARGIN_S = 10;
 var LITELLM_TIMEOUT_TIERS = {
@@ -21673,7 +21689,8 @@ var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_MCP_RECALL_BUDGET_MODE",
   "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT",
   "HINDSIGHT_API_RETAIN_WALL_TIMEOUT",
-  "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT"
+  "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT",
+  "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S"
 ]);
 var HINDSIGHT_WORKER_SLOT_TYPES = [
   "consolidation",

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

@@ -4434,7 +4434,7 @@ var init_observation_scopes = __esm(() => {
 });
 
 // src/config/schema.ts
-var CodeRepoEntrySchema, AgentBindMountSchema, HttpDiffPollSchema, PollSpecSchema, TelegramMessageActionSchema, WebhookActionSchema, ActionSpecSchema, ScheduleEntrySchema, AgentSoulSchema, AgentToolsSchema, ObservationScopesSchema, ObservationScopeStrategySchema, AntiConfabulationDirectiveSchema, AgentMemorySchema, HookEntrySchema, AgentHooksSchema, SubagentSchema, SessionSchema, SessionContinuitySchema, webhookDispatchRule, TelegramChannelSchema, ChannelsSchema, TIMEZONE_REGEX, ApproverIdSchema, GoogleWorkspaceTierSchema, GoogleWorkspaceConfigSchema, LiteLLMConfigSchema, HindsightPerOpLlmSchema, HindsightConfigSchema, MicrosoftWorkspaceConfigSchema, NotionWorkspaceConfigSchema, AgentGoogleWorkspaceConfigSchema, MicrosoftAccountEmailSchema, MicrosoftToolTokenSchema, MicrosoftAccountBindingSchema, AgentMicrosoftWorkspaceConfigSchema, AgentNotionWorkspaceConfigSchema, ReactionsSchema, ReactionDispatchSchema, releaseBlockFields, ReleaseBlock, RootReleaseBlock, NetworkIsolationSchema, servesField, knowsField, profileFields, ProfileSchema, _omitExtends, defaultsFields, AgentDefaultsSchema, AgentSchema, TelegramConfigSchema, MemoryBackendConfigSchema, VaultConfigSchema, QuotaConfigSchema, AutoReleaseCheckSchema, HostControlConfigSchema, WebServiceConfigSchema, FleetHealthConfigSchema, HostdConfigSchema, CronEgressSchema, CronConfigSchema, UserSchema, SwitchroomConfigSchema;
+var CodeRepoEntrySchema, AgentBindMountSchema, HttpDiffPollSchema, PollSpecSchema, TelegramMessageActionSchema, WebhookActionSchema, ActionSpecSchema, ScheduleEntrySchema, AgentSoulSchema, AgentToolsSchema, ObservationScopesSchema, ObservationScopeStrategySchema, AntiConfabulationDirectiveSchema, AgentMemorySchema, HookEntrySchema, AgentHooksSchema, SubagentSchema, SessionSchema, SessionContinuitySchema, webhookDispatchRule, TelegramChannelSchema, ChannelsSchema, TIMEZONE_REGEX, ApproverIdSchema, GoogleWorkspaceTierSchema, GoogleServiceTokenSchema, GoogleWorkspaceConfigSchema, LiteLLMConfigSchema, HindsightPerOpLlmSchema, HindsightConfigSchema, MicrosoftWorkspaceConfigSchema, NotionWorkspaceConfigSchema, AgentGoogleWorkspaceConfigSchema, MicrosoftAccountEmailSchema, MicrosoftToolTokenSchema, MicrosoftAccountBindingSchema, AgentMicrosoftWorkspaceConfigSchema, AgentNotionWorkspaceConfigSchema, ReactionsSchema, ReactionDispatchSchema, releaseBlockFields, ReleaseBlock, RootReleaseBlock, NetworkIsolationSchema, servesField, knowsField, profileFields, ProfileSchema, _omitExtends, defaultsFields, AgentDefaultsSchema, AgentSchema, TelegramConfigSchema, MemoryBackendConfigSchema, VaultConfigSchema, QuotaConfigSchema, AutoReleaseCheckSchema, HostControlConfigSchema, WebServiceConfigSchema, FleetHealthConfigSchema, HostdConfigSchema, CronEgressSchema, CronConfigSchema, UserSchema, SwitchroomConfigSchema;
 var init_schema = __esm(() => {
   init_zod();
   init_observation_scopes();
@@ -4560,6 +4560,8 @@ var init_schema = __esm(() => {
     profile: exports_external.string().optional().describe("Memory profile bank this agent's curated memory defaults key off — " + "the built-in disposition + observations_mission in PROFILE_MEMORY_DEFAULTS. " + "Decouples the memory profile from `extends` (the filesystem persona " + "profile), so an agent on `extends: default` can opt into the `coding` " + "memory bundle via `memory.profile: coding` without inheriting the coding " + "persona. Resolution: memory.profile → extends → DEFAULT_PROFILE (see " + "resolveMemoryProfile). Unset ⇒ byte-identical to keying off `extends`."),
     bank_mission: exports_external.string().optional().describe("Bank-level mission statement used during recall to contextualize " + "results. NOTE: this is an alias for the Hindsight engine's " + "`reflect_mission` field (verified live: switchroom's bank_mission " + "lands in `config.reflect_mission`). Prefer `reflect_mission` going " + "forward; `bank_mission` is retained for back-compat. If both are " + "set, `reflect_mission` wins. Cascade: override."),
     reflect_mission: exports_external.string().optional().describe("Mission/context steering Hindsight Reflect operations (the bank's " + "'who am I / what matters' framing applied during recall). The " + "engine-accurate name for what `bank_mission` sets. Cascade: override."),
+    reflect_budget: exports_external.enum(["low", "mid", "high"]).optional().describe("Thinking/retrieval budget injected into reflect MCP calls when the " + "caller omits budget. Unset ⇒ shim default (mid). Explicit per-call " + "budget always wins. Higher = better recall on fuzzy queries, more " + "backend latency/compute. stdio-shim transport only (ignored under " + "memory.config.mcp_transport: http). Cascade: override (per-agent " + "wins over default)."),
+    reflect_max_tokens: exports_external.number().int().positive().max(8192).optional().describe("Token cap injected into reflect MCP calls when caller omits " + "max_tokens. Unset ⇒ shim default 1024. Values much above ~2048 risk " + "exceeding Claude Code's MCP output cap, silently dropping the " + "payload — raise deliberately. Explicit per-call values always win. " + "stdio-shim transport only. Cascade: override."),
     retain_mission: exports_external.string().optional().describe("Instructions for the fact extraction LLM during retain. Cascade: override."),
     mental_models: exports_external.array(exports_external.object({
       name: exports_external.string().min(1).describe("Stable model name (identity key for idempotent ensure). Two " + "declarations with the same name in one agent are rejected."),
@@ -4821,11 +4823,18 @@ var init_schema = __esm(() => {
     "extended",
     "complete"
   ]);
+  GoogleServiceTokenSchema = exports_external.enum([
+    "cal",
+    "drive",
+    "docs",
+    "sheets",
+    "slides"
+  ]);
   GoogleWorkspaceConfigSchema = exports_external.object({
     google_client_id: exports_external.string().min(1).describe("Google OAuth client ID (literal string or vault reference e.g. 'vault:google-oauth-client-id')"),
     google_client_secret: exports_external.string().min(1).describe("Google OAuth client secret (literal string or vault reference e.g. 'vault:google-oauth-client-secret')"),
     approvers: exports_external.array(ApproverIdSchema).min(1).describe("Array of numeric Telegram user IDs authorized to approve drive onboarding. " + "At least one must be specified."),
-    tier: GoogleWorkspaceTierSchema.optional().describe("RFC G Phase 1: which upstream MCP tier to expose. " + "core (default) = ~16 tools (Drive+Docs+Sheets+Calendar). " + "extended = ~40 tools (+Slides, Forms, Tasks, Chat). " + "complete = ~60+ tools (+Gmail; not recommended yet — see RFC G §5).")
+    tier: GoogleWorkspaceTierSchema.optional().describe("RFC G Phase 1: which upstream MCP tier to expose. " + "core (default) = ~16 tools (Drive+Docs+Sheets+Calendar tools; the " + "Calendar tools only authenticate when the account holds the opt-in " + "calendar scope — no tier mints it). " + "extended = ~40 tools (+Slides, Forms, Tasks, Chat). " + "complete = ~60+ tools (+Gmail; not recommended yet — see RFC G §5).")
   }).optional();
   LiteLLMConfigSchema = exports_external.object({
     enabled: exports_external.boolean().optional().describe("Opt-in toggle. When true, `switchroom apply` provisions a per-agent " + "LiteLLM virtual key and injects routing env into the container. " + "Default OFF."),
@@ -4858,7 +4867,7 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "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_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, 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_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains), 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_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, 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_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -5044,6 +5053,8 @@ var init_schema = __esm(() => {
       }).optional(),
       bank_mission: exports_external.string().optional(),
       reflect_mission: exports_external.string().optional(),
+      reflect_budget: exports_external.enum(["low", "mid", "high"]).optional(),
+      reflect_max_tokens: exports_external.number().int().positive().max(8192).optional(),
       retain_mission: exports_external.string().optional(),
       observations_mission: exports_external.string().optional(),
       disposition: exports_external.object({
@@ -5332,7 +5343,9 @@ var init_schema = __esm(() => {
     }).transform((v) => v.trim().toLowerCase()), exports_external.object({
       enabled_for: exports_external.array(exports_external.string().regex(/^[a-z0-9][a-z0-9_-]{0,50}$/, {
         message: "Agent name must match the standard agent-name pattern"
-      })).describe("Agent slugs that may read this account's vault slots " + "(`google:<account>:refresh_token` etc). Per-agent ACL is " + "enforced at the broker, not at the agent identity layer — " + "the agent still authenticates via socket-path-as-identity " + "per RFC D §4.1, broker just gates the cross-agent token share.")
+      })).describe("Agent slugs that may read this account's vault slots " + "(`google:<account>:refresh_token` etc). Per-agent ACL is " + "enforced at the broker, not at the agent identity layer — " + "the agent still authenticates via socket-path-as-identity " + "per RFC D §4.1, broker just gates the cross-agent token share."),
+      readonly: exports_external.boolean().optional().describe("Per-account read-only selection (v1 scope model). true → the " + "token minted by `auth google account add` carries ONLY " + ".readonly scope variants (zero write scopes), and the gdrive " + "MCP launcher passes upstream `--read-only` so write tools are " + "not exposed. Written by `account add --readonly`; re-read on " + "`--replace` so a re-consent cannot silently re-widen. Omitted " + "= legacy behaviour (tier-tied read-write document scopes)."),
+      services: exports_external.array(GoogleServiceTokenSchema).min(1).optional().describe("Per-account service selection (v1 scope model). Which Google " + "services the minted token covers AND the gdrive MCP exposes " + "(upstream `--tools`): cal, drive, docs, sheets, slides. " + "Written by `account add --services`; re-read on `--replace`. " + "Omitted = the tier's default services (drive,docs,sheets; " + "+slides at extended/complete).")
     })).optional().describe("RFC G Phase 2: per-Google-account ACL for vault slots holding " + "OAuth refresh tokens. Maps account email → list of agents " + "permitted to read that account's slots. Written by `switchroom " + "auth google enable|disable` (Phase 3); read by the broker on " + "every Google slot access. Replaces RFC D's per-agent vault slot " + "scope (which can't express 'two agents share one Google account')."),
     microsoft_accounts: exports_external.record(exports_external.string().regex(/^[^@\s:]+@[^@\s:]+\.[^@\s:]+$/, {
       message: "Account key must be a Microsoft account email like 'alice@outlook.com' or 'alice@contoso.com' (colons not allowed)"
@@ -19368,6 +19381,9 @@ for (const name of SHARED_FRAGMENTS) {
   }
 }
 
+// src/memory/scaffold-integration.ts
+init_merge();
+
 // src/litellm/timeout-budget.ts
 var LITELLM_ROUTER_MARGIN_S = 10;
 var LITELLM_TIMEOUT_TIERS = {
@@ -19568,7 +19584,8 @@ var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_MCP_RECALL_BUDGET_MODE",
   "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT",
   "HINDSIGHT_API_RETAIN_WALL_TIMEOUT",
-  "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT"
+  "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT",
+  "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S"
 ]);
 var HINDSIGHT_WORKER_SLOT_TYPES = [
   "consolidation",

package/dist/vault/broker/server.js

@@ -4030,7 +4030,7 @@ var init_observation_scopes = __esm(() => {
 });
 
 // src/config/schema.ts
-var CodeRepoEntrySchema, AgentBindMountSchema, HttpDiffPollSchema, PollSpecSchema, TelegramMessageActionSchema, WebhookActionSchema, ActionSpecSchema, ScheduleEntrySchema, AgentSoulSchema, AgentToolsSchema, ObservationScopesSchema, ObservationScopeStrategySchema, AntiConfabulationDirectiveSchema, AgentMemorySchema, HookEntrySchema, AgentHooksSchema, SubagentSchema, SessionSchema, SessionContinuitySchema, webhookDispatchRule, TelegramChannelSchema, ChannelsSchema, TIMEZONE_REGEX, ApproverIdSchema, GoogleWorkspaceTierSchema, GoogleWorkspaceConfigSchema, LiteLLMConfigSchema, HindsightPerOpLlmSchema, HindsightConfigSchema, MicrosoftWorkspaceConfigSchema, NotionWorkspaceConfigSchema, AgentGoogleWorkspaceConfigSchema, MicrosoftAccountEmailSchema, MicrosoftToolTokenSchema, MicrosoftAccountBindingSchema, AgentMicrosoftWorkspaceConfigSchema, AgentNotionWorkspaceConfigSchema, ReactionsSchema, ReactionDispatchSchema, releaseBlockFields, ReleaseBlock, RootReleaseBlock, NetworkIsolationSchema, servesField, knowsField, profileFields, ProfileSchema, _omitExtends, defaultsFields, AgentDefaultsSchema, AgentSchema, TelegramConfigSchema, MemoryBackendConfigSchema, VaultConfigSchema, QuotaConfigSchema, AutoReleaseCheckSchema, HostControlConfigSchema, WebServiceConfigSchema, FleetHealthConfigSchema, HostdConfigSchema, CronEgressSchema, CronConfigSchema, UserSchema, SwitchroomConfigSchema;
+var CodeRepoEntrySchema, AgentBindMountSchema, HttpDiffPollSchema, PollSpecSchema, TelegramMessageActionSchema, WebhookActionSchema, ActionSpecSchema, ScheduleEntrySchema, AgentSoulSchema, AgentToolsSchema, ObservationScopesSchema, ObservationScopeStrategySchema, AntiConfabulationDirectiveSchema, AgentMemorySchema, HookEntrySchema, AgentHooksSchema, SubagentSchema, SessionSchema, SessionContinuitySchema, webhookDispatchRule, TelegramChannelSchema, ChannelsSchema, TIMEZONE_REGEX, ApproverIdSchema, GoogleWorkspaceTierSchema, GoogleServiceTokenSchema, GoogleWorkspaceConfigSchema, LiteLLMConfigSchema, HindsightPerOpLlmSchema, HindsightConfigSchema, MicrosoftWorkspaceConfigSchema, NotionWorkspaceConfigSchema, AgentGoogleWorkspaceConfigSchema, MicrosoftAccountEmailSchema, MicrosoftToolTokenSchema, MicrosoftAccountBindingSchema, AgentMicrosoftWorkspaceConfigSchema, AgentNotionWorkspaceConfigSchema, ReactionsSchema, ReactionDispatchSchema, releaseBlockFields, ReleaseBlock, RootReleaseBlock, NetworkIsolationSchema, servesField, knowsField, profileFields, ProfileSchema, _omitExtends, defaultsFields, AgentDefaultsSchema, AgentSchema, TelegramConfigSchema, MemoryBackendConfigSchema, VaultConfigSchema, QuotaConfigSchema, AutoReleaseCheckSchema, HostControlConfigSchema, WebServiceConfigSchema, FleetHealthConfigSchema, HostdConfigSchema, CronEgressSchema, CronConfigSchema, UserSchema, SwitchroomConfigSchema;
 var init_schema = __esm(() => {
   init_zod();
   init_observation_scopes();
@@ -4156,6 +4156,8 @@ var init_schema = __esm(() => {
     profile: exports_external.string().optional().describe("Memory profile bank this agent's curated memory defaults key off — " + "the built-in disposition + observations_mission in PROFILE_MEMORY_DEFAULTS. " + "Decouples the memory profile from `extends` (the filesystem persona " + "profile), so an agent on `extends: default` can opt into the `coding` " + "memory bundle via `memory.profile: coding` without inheriting the coding " + "persona. Resolution: memory.profile → extends → DEFAULT_PROFILE (see " + "resolveMemoryProfile). Unset ⇒ byte-identical to keying off `extends`."),
     bank_mission: exports_external.string().optional().describe("Bank-level mission statement used during recall to contextualize " + "results. NOTE: this is an alias for the Hindsight engine's " + "`reflect_mission` field (verified live: switchroom's bank_mission " + "lands in `config.reflect_mission`). Prefer `reflect_mission` going " + "forward; `bank_mission` is retained for back-compat. If both are " + "set, `reflect_mission` wins. Cascade: override."),
     reflect_mission: exports_external.string().optional().describe("Mission/context steering Hindsight Reflect operations (the bank's " + "'who am I / what matters' framing applied during recall). The " + "engine-accurate name for what `bank_mission` sets. Cascade: override."),
+    reflect_budget: exports_external.enum(["low", "mid", "high"]).optional().describe("Thinking/retrieval budget injected into reflect MCP calls when the " + "caller omits budget. Unset ⇒ shim default (mid). Explicit per-call " + "budget always wins. Higher = better recall on fuzzy queries, more " + "backend latency/compute. stdio-shim transport only (ignored under " + "memory.config.mcp_transport: http). Cascade: override (per-agent " + "wins over default)."),
+    reflect_max_tokens: exports_external.number().int().positive().max(8192).optional().describe("Token cap injected into reflect MCP calls when caller omits " + "max_tokens. Unset ⇒ shim default 1024. Values much above ~2048 risk " + "exceeding Claude Code's MCP output cap, silently dropping the " + "payload — raise deliberately. Explicit per-call values always win. " + "stdio-shim transport only. Cascade: override."),
     retain_mission: exports_external.string().optional().describe("Instructions for the fact extraction LLM during retain. Cascade: override."),
     mental_models: exports_external.array(exports_external.object({
       name: exports_external.string().min(1).describe("Stable model name (identity key for idempotent ensure). Two " + "declarations with the same name in one agent are rejected."),
@@ -4417,11 +4419,18 @@ var init_schema = __esm(() => {
     "extended",
     "complete"
   ]);
+  GoogleServiceTokenSchema = exports_external.enum([
+    "cal",
+    "drive",
+    "docs",
+    "sheets",
+    "slides"
+  ]);
   GoogleWorkspaceConfigSchema = exports_external.object({
     google_client_id: exports_external.string().min(1).describe("Google OAuth client ID (literal string or vault reference e.g. 'vault:google-oauth-client-id')"),
     google_client_secret: exports_external.string().min(1).describe("Google OAuth client secret (literal string or vault reference e.g. 'vault:google-oauth-client-secret')"),
     approvers: exports_external.array(ApproverIdSchema).min(1).describe("Array of numeric Telegram user IDs authorized to approve drive onboarding. " + "At least one must be specified."),
-    tier: GoogleWorkspaceTierSchema.optional().describe("RFC G Phase 1: which upstream MCP tier to expose. " + "core (default) = ~16 tools (Drive+Docs+Sheets+Calendar). " + "extended = ~40 tools (+Slides, Forms, Tasks, Chat). " + "complete = ~60+ tools (+Gmail; not recommended yet — see RFC G §5).")
+    tier: GoogleWorkspaceTierSchema.optional().describe("RFC G Phase 1: which upstream MCP tier to expose. " + "core (default) = ~16 tools (Drive+Docs+Sheets+Calendar tools; the " + "Calendar tools only authenticate when the account holds the opt-in " + "calendar scope — no tier mints it). " + "extended = ~40 tools (+Slides, Forms, Tasks, Chat). " + "complete = ~60+ tools (+Gmail; not recommended yet — see RFC G §5).")
   }).optional();
   LiteLLMConfigSchema = exports_external.object({
     enabled: exports_external.boolean().optional().describe("Opt-in toggle. When true, `switchroom apply` provisions a per-agent " + "LiteLLM virtual key and injects routing env into the container. " + "Default OFF."),
@@ -4454,7 +4463,7 @@ var init_schema = __esm(() => {
       reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
       consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
     }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-    env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "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_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, 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_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains), 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_RESERVED_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, GRAPH_SEED_MIN_SIMILARITY, " + "LLM_SUPPORTS_MAX_ITEMS, 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_RESERVED_SLOTS reserves out of; " + "unset means upstream's own default; and " + "HINDSIGHT_API_WORKER_RETAIN_RESERVED_SLOTS — the reserved slot FLOOR for the " + "retain (memory write) lane, carved from that same total; unset means " + "upstream's own 0, i.e. no floor and retain competes for the shared " + "pool; and " + "the pre-0.8.6 name HINDSIGHT_API_WORKER_<TYPE>_MAX_SLOTS is still " + "accepted for every operation type and normalised to " + "..._RESERVED_SLOTS on the way in, because setting both names for " + "one type is a hard boot failure in the engine and switchroom " + "therefore emits only the canonical name; and " + "HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY — the cosine-similarity floor a " + "candidate must reach to be returned by the semantic retrieval arm at " + "all (0..1); unset means upstream's own 0.3, and where it should sit " + "depends on the bank's embedding model and phrasing diversity, so " + "switchroom ships no opinion; and " + "HINDSIGHT_MCP_RECALL_BUDGET_MODE — the rollback knob for switchroom's " + "mcp-recall-token-budget image patch; `legacy` restores upstream's " + "exact pre-patch recall returns, anything else or unset is the " + "honest-envelope mode; and " + "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT — upstream's own per-op reflect " + "temperature knob, made live by switchroom's reflect-temperature image " + "patch; a float, or `none` to omit the kwarg (provider default, " + "upstream's accidental pre-patch behaviour); unset means the image's " + "baked default 0.1; and " + "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT — the background " + "half of switchroom's recall-admission split (#3660): of the " + "RECALL_MAX_CONCURRENT admission slots, at most this many may be held " + "by background consolidation recalls at once, so foreground per-turn " + "recall always keeps the remainder; must be >= 1 and strictly less " + "than RECALL_MAX_CONCURRENT or the engine refuses to boot; unset " + "means the image's derived default min(2, RECALL_MAX_CONCURRENT - 1); " + "1 biases hard toward the interactive lane while a consolidation " + "backlog drains; and " + "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S — the rollback/tuning knob for " + "switchroom's MM-refresh-debounce image patch: the minimum seconds " + "between consolidation-triggered refreshes of one mental model; unset " + "means the image's baked 3600, 0 restores upstream's " + "refresh-every-round behaviour; explicit and cron-scheduled refreshes " + "are never debounced), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
   });
   MicrosoftWorkspaceConfigSchema = exports_external.object({
     microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -4640,6 +4649,8 @@ var init_schema = __esm(() => {
       }).optional(),
       bank_mission: exports_external.string().optional(),
       reflect_mission: exports_external.string().optional(),
+      reflect_budget: exports_external.enum(["low", "mid", "high"]).optional(),
+      reflect_max_tokens: exports_external.number().int().positive().max(8192).optional(),
       retain_mission: exports_external.string().optional(),
       observations_mission: exports_external.string().optional(),
       disposition: exports_external.object({
@@ -4928,7 +4939,9 @@ var init_schema = __esm(() => {
     }).transform((v) => v.trim().toLowerCase()), exports_external.object({
       enabled_for: exports_external.array(exports_external.string().regex(/^[a-z0-9][a-z0-9_-]{0,50}$/, {
         message: "Agent name must match the standard agent-name pattern"
-      })).describe("Agent slugs that may read this account's vault slots " + "(`google:<account>:refresh_token` etc). Per-agent ACL is " + "enforced at the broker, not at the agent identity layer — " + "the agent still authenticates via socket-path-as-identity " + "per RFC D §4.1, broker just gates the cross-agent token share.")
+      })).describe("Agent slugs that may read this account's vault slots " + "(`google:<account>:refresh_token` etc). Per-agent ACL is " + "enforced at the broker, not at the agent identity layer — " + "the agent still authenticates via socket-path-as-identity " + "per RFC D §4.1, broker just gates the cross-agent token share."),
+      readonly: exports_external.boolean().optional().describe("Per-account read-only selection (v1 scope model). true → the " + "token minted by `auth google account add` carries ONLY " + ".readonly scope variants (zero write scopes), and the gdrive " + "MCP launcher passes upstream `--read-only` so write tools are " + "not exposed. Written by `account add --readonly`; re-read on " + "`--replace` so a re-consent cannot silently re-widen. Omitted " + "= legacy behaviour (tier-tied read-write document scopes)."),
+      services: exports_external.array(GoogleServiceTokenSchema).min(1).optional().describe("Per-account service selection (v1 scope model). Which Google " + "services the minted token covers AND the gdrive MCP exposes " + "(upstream `--tools`): cal, drive, docs, sheets, slides. " + "Written by `account add --services`; re-read on `--replace`. " + "Omitted = the tier's default services (drive,docs,sheets; " + "+slides at extended/complete).")
     })).optional().describe("RFC G Phase 2: per-Google-account ACL for vault slots holding " + "OAuth refresh tokens. Maps account email → list of agents " + "permitted to read that account's slots. Written by `switchroom " + "auth google enable|disable` (Phase 3); read by the broker on " + "every Google slot access. Replaces RFC D's per-agent vault slot " + "scope (which can't express 'two agents share one Google account')."),
     microsoft_accounts: exports_external.record(exports_external.string().regex(/^[^@\s:]+@[^@\s:]+\.[^@\s:]+$/, {
       message: "Account key must be a Microsoft account email like 'alice@outlook.com' or 'alice@contoso.com' (colons not allowed)"
@@ -19781,6 +19794,9 @@ for (const name of SHARED_FRAGMENTS) {
   }
 }
 
+// src/memory/scaffold-integration.ts
+init_merge();
+
 // src/litellm/timeout-budget.ts
 var LITELLM_ROUTER_MARGIN_S = 10;
 var LITELLM_TIMEOUT_TIERS = {
@@ -19981,7 +19997,8 @@ var HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS = new Set([
   "HINDSIGHT_MCP_RECALL_BUDGET_MODE",
   "HINDSIGHT_API_LLM_TEMPERATURE_REFLECT",
   "HINDSIGHT_API_RETAIN_WALL_TIMEOUT",
-  "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT"
+  "HINDSIGHT_API_CONSOLIDATION_RECALL_MAX_CONCURRENT",
+  "HINDSIGHT_MM_REFRESH_MIN_INTERVAL_S"
 ]);
 var HINDSIGHT_WORKER_SLOT_TYPES = [
   "consolidation",

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "switchroom",
   "//version": "NOT the release version — source of truth is the git tag, resolved by scripts/build.mjs:resolveVersion() (see CLAUDE.md > Standard release process). This field is stale by design and only the Layer-4 dev/non-tag fallback for build.mjs + src/cli/resolve-version.ts; do NOT bump it expecting a release to pick it up. npm-pack tarball naming needs a real version — do that as an UNCOMMITTED pack-time bump (see release step 6), never a committed one.",
-  "version": "0.19.46",
+  "version": "0.19.48",
   "description": "Run Claude Code 24/7 on your Claude Pro/Max subscription over Telegram. Open-source alternative to OpenClaw and NanoClaw — no API keys.",
   "type": "module",
   "bin": {

package/telegram-plugin/answer-ready-flush.ts

@@ -33,6 +33,61 @@ import { decideTurnFlush, type FlushDecisionInput } from './turn-flush-safety.js
  *  matching the repo's env-flag convention. */
 export const ANSWER_READY_FLUSH_MS = 1000
 
+/**
+ * Default STAGE window (ms) — the stage-don't-send fix for the live
+ * flush-then-late-reply duplicate (2026-08-02, klanker DM, msgs 25843/25844).
+ *
+ * The quiescence flush used to SEND the composed terminal answer the moment the
+ * ~1 s debounce fired. But quiescence is a heuristic, not a completion signal:
+ * the model routinely goes quiet for a few seconds BETWEEN its terminal
+ * narration and its `reply` tool call (observed gap: 6 s). The flush then posts
+ * a PROVISIONAL message that the supersede machinery must claw back — and the
+ * claw-back is deliberately conservative (#3429 content gate, #4176 sub-agent
+ * liveness hold), so a REWORDED own-answer with a background sub-agent live
+ * ships as a visible duplicate even on current main (the accepted residual
+ * documented in gateway/subagent-reply-authority.ts).
+ *
+ * Stage-don't-send removes the provisional send instead of reconciling it:
+ * when the quiescence debounce fires, the composed answer is STAGED (held on
+ * the still-live turn — the turn is NOT ended, so the typing indicator keeps
+ * running) and promoted to a real send only when the completion window closes
+ * with no reply:
+ *
+ *   - the model's `reply` lands → the staged text is DISCARDED (fire-time
+ *     re-verify declines on `replyCalled`) — exactly one message, the canonical
+ *     reply, on its normal path;
+ *   - the REAL turn_end lands first → the turn-flush branch at turn_end
+ *     delivers the captured answer (the pre-PR-A path) — promotion at the true
+ *     completion signal, usually seconds;
+ *   - neither arrives within this window (the KNOWN-UNRELIABLE turn_end case
+ *     PR A was built for) → the stage timer promotes through the same
+ *     synthetic-turn_end flush path as before.
+ *
+ * Worst-case added latency for a silent no-op turn whose turn_end never lands
+ * is this window (30 s) — still ~5× faster than the ~150 s orphaned-reply
+ * backstop PR A replaced, and it buys the hard guarantee that a composing
+ * model can no longer race the flush into a duplicate. Env-tunable via
+ * SWITCHROOM_ANSWER_STAGE_MS; 0 (or any non-positive value) disables staging
+ * and restores the legacy immediate flush.
+ */
+export const ANSWER_STAGE_MS = 30_000
+
+/**
+ * Resolve the stage window from the environment. Returns a positive integer
+ * ms, or 0 when explicitly disabled (non-positive value = legacy immediate
+ * flush). Unparseable values fail safe to the default.
+ */
+export function resolveAnswerStageMs(
+  env: Record<string, string | undefined>,
+): number {
+  const raw = env.SWITCHROOM_ANSWER_STAGE_MS
+  if (raw == null || raw.trim() === '') return ANSWER_STAGE_MS
+  const n = Number(raw)
+  if (!Number.isFinite(n)) return ANSWER_STAGE_MS
+  if (n <= 0) return 0
+  return Math.floor(n)
+}
+
 /**
  * Resolve the debounce window from the environment. Returns a positive integer
  * ms, or 0 when disabled (kill-switch) / unparseable / non-positive. The
@@ -117,6 +172,17 @@ export interface AnswerReadyFlushDeps<Turn> {
    * short-circuits — the exactly-once guarantee.
    */
   onFlush(turn: Turn): void
+  /**
+   * Stage-don't-send window (ms) — see {@link ANSWER_STAGE_MS}. When > 0 the
+   * quiescence expiry STAGES the answer (turn stays live, typing hold) and
+   * arms a promotion timer for this many ms instead of flushing immediately;
+   * the promotion re-verifies before delivering, so a `reply` that landed in
+   * the window discards the staged text, and a real turn_end that landed first
+   * delivered it through the turn-flush branch already (its
+   * `endCurrentTurnAtomic` → `clear` cancels the promotion). 0 / undefined =
+   * legacy immediate flush at quiescence.
+   */
+  stageWindowMs?: number
   /** Injectable timer primitives (default to the globals). */
   setTimeoutFn?: (fn: () => void, ms: number) => FlushTimerHandle
   clearTimeoutFn?: (handle: FlushTimerHandle) => void
@@ -139,9 +205,14 @@ export interface AnswerReadyFlushDeps<Turn> {
  * landed since the arm cancels the flush deterministically), then dispatches
  * exactly one `onFlush`.
  */
-export class AnswerReadyFlushController<Turn> {
+export class AnswerReadyFlushController<Turn extends object> {
   constructor(private readonly deps: AnswerReadyFlushDeps<Turn>) {}
 
+  /** Turns whose composed answer is currently STAGED (stage-don't-send): the
+   *  quiescence debounce fired, but delivery is held for the promotion window.
+   *  WeakSet so a superseded/ended turn atom can never leak an entry. */
+  private readonly staged = new WeakSet<Turn>()
+
   private get setTimeoutFn(): (fn: () => void, ms: number) => FlushTimerHandle {
     return this.deps.setTimeoutFn ?? ((fn, ms) => setTimeout(fn, ms))
   }
@@ -150,9 +221,11 @@ export class AnswerReadyFlushController<Turn> {
     return this.deps.clearTimeoutFn ?? ((h) => clearTimeout(h))
   }
 
-  /** Disarm the flush timer for a turn (idempotent). */
+  /** Disarm the flush timer for a turn (idempotent). Clears BOTH phases: a
+   *  pending quiescence debounce or a pending staged promotion. */
   clear(turn: Turn | null): void {
     if (turn == null) return
+    this.staged.delete(turn)
     const handle = this.deps.getTimerHandle(turn)
     if (handle != null) {
       this.clearTimeoutFn(handle)
@@ -160,6 +233,11 @@ export class AnswerReadyFlushController<Turn> {
     }
   }
 
+  /** True IFF `turn`'s composed answer is currently staged (test/diagnostics). */
+  isStaged(turn: Turn): boolean {
+    return this.staged.has(turn)
+  }
+
   /** (Re)arm the flush timer for the current turn — the debounce. */
   reset(): void {
     const turn = this.deps.getCurrentTurn()
@@ -171,7 +249,10 @@ export class AnswerReadyFlushController<Turn> {
     this.deps.setTimerHandle(turn, handle)
   }
 
-  /** Timer-expiry callback. Re-pins the turn, re-verifies quiescence, fires once. */
+  /** Timer-expiry callback — BOTH phases share it. Re-pins the turn,
+   *  re-verifies quiescence, then either STAGES (first fire, window > 0) or
+   *  delivers (legacy immediate mode, or the staged promotion firing after the
+   *  completion window closed with no reply). */
   private onExpiry(armedTurn: Turn): void {
     const live = this.deps.getCurrentTurn()
     // Rollover guard: a superseded turn's timer must not fire against a fresh atom.
@@ -179,9 +260,40 @@ export class AnswerReadyFlushController<Turn> {
     this.deps.setTimerHandle(live, null)
     // Fire-time re-verification: a tool that started (or a reply that landed)
     // since the arm deterministically cancels the flush even if the explicit
-    // disarm was somehow missed.
-    if (!shouldArmAnswerReadyFlush(this.deps.getArmInput(live))) return
-    this.deps.log?.('answer-ready quiescence flush — delivering composed terminal answer')
+    // disarm was somehow missed. On a staged promotion this is the DISCARD
+    // path: the reply landed inside the window, so the staged provisional text
+    // is dropped and the canonical reply is the one message the user sees.
+    if (!shouldArmAnswerReadyFlush(this.deps.getArmInput(live))) {
+      if (this.staged.delete(live)) {
+        this.deps.log?.(
+          'answer-ready stage discarded — reply/tool activity landed inside the completion window',
+        )
+      }
+      return
+    }
+    const stageMs = this.deps.stageWindowMs ?? 0
+    if (stageMs > 0 && !this.staged.has(live)) {
+      // STAGE-DON'T-SEND (see ANSWER_STAGE_MS): hold the composed answer on
+      // the still-live turn instead of posting a provisional message. The turn
+      // is NOT ended, so the typing indicator keeps running (the "typing
+      // hold") and a real turn_end landing first delivers through the normal
+      // turn-flush branch (endCurrentTurnAtomic's `clear` cancels this timer).
+      this.staged.add(live)
+      this.deps.log?.(
+        `answer-ready quiescence — staging composed terminal answer ` +
+          `(promotes in ${stageMs}ms unless the reply lands or the turn completes)`,
+      )
+      const handle = this.setTimeoutFn(() => this.onExpiry(live), stageMs)
+      this.deps.setTimerHandle(live, handle)
+      return
+    }
+    if (this.staged.delete(live)) {
+      this.deps.log?.(
+        'answer-ready stage promoted — completion window closed with no reply; delivering composed terminal answer',
+      )
+    } else {
+      this.deps.log?.('answer-ready quiescence flush — delivering composed terminal answer')
+    }
     this.deps.onFlush(live)
   }
 }