switchroom 0.19.26 → 0.19.27

package/dist/host-control/main.js

@@ -19228,10 +19228,37 @@ var AgentMemorySchema = exports_external.object({
     request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
     own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
     additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+    min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
+    min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
     types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
     additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
     sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
     skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
+    budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+    max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
+    prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
+    context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
+    roles: exports_external.array(exports_external.string().min(1)).min(1).optional().describe("Transcript roles the multi-turn composition may draw from. " + 'Default ["user", "assistant"]. Set ["user"] to compose ' + "the query from the human's words only — worth trying when an " + "agent's own verbose replies are dominating the query terms. No " + "effect while `context_turns` is 1."),
+    prompt_preamble: exports_external.string().min(1).optional().describe("The banner rendered above injected memories. The agent reads " + "this line as the instruction for how to treat the block, so it " + "is a behaviour knob, not cosmetics. Default tells the model to " + "prioritise recent memories on conflict and ignore irrelevant " + "ones. Override to tighten that framing for a specialised agent."),
+    tags: exports_external.array(exports_external.string().min(1)).optional().describe("Restrict recall to memories carrying these tags. Default [] = " + "no filter (match everything). This is a HARD filter applied " + "server-side — a memory without the tags cannot surface at any " + "score — so it is for a genuinely scoped agent, not for " + "ranking. Use `tag_weights` when you want a preference rather " + "than an exclusion."),
+    tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional().describe('How `tags` combine. "any" (default) = at least one; ' + '"all" = every tag. The `_strict` forms additionally require ' + "the memory to actually carry the tags rather than merely rank " + "for them. No effect while `tags` and `tag_groups` are empty."),
+    tag_groups: exports_external.union([
+      exports_external.array(exports_external.array(exports_external.string().min(1))),
+      exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+    ]).optional().describe("Tag filtering with grouping — either an OR-of-ANDs list " + '([["a","b"],["c"]] = (a AND b) OR c) or a named ' + "{group: [tags]} map. Default unset (no grouping). Use when a " + "flat `tags` + `tags_match` cannot express the scope you need."),
+    tag_weights: exports_external.record(exports_external.string(), exports_external.number().min(0)).optional().describe("Per-tag multipliers applied to `scores.final` just before the " + "final sort — a DEMOTION/PROMOTION, never a drop, so a " + "down-weighted memory still surfaces when it is the only " + "relevant hit. MERGED over switchroom's seed " + '({"sidechain": 0.8}, which ranks delegated sub-agent ' + "process-memories just under first-party ones), so setting one " + "unrelated weight does not silently undo it; pass " + "`sidechain: 1.0` to neutralise the seed. Reach for this when " + "recall_log shows one class of memory crowding the block."),
+    additional_bank_filters: exports_external.record(exports_external.string(), exports_external.object({
+      tags: exports_external.array(exports_external.string().min(1)).optional(),
+      tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional(),
+      tag_groups: exports_external.union([
+        exports_external.array(exports_external.array(exports_external.string().min(1))),
+        exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+      ]).optional()
+    }).strict()).optional().describe("Per-bank overrides of the tag filters above, keyed by bank id " + "(applies to `additional_banks` AND to sender banks). Default " + "{} = every extra bank inherits the global filters. Use it to " + "scope a shared bank — e.g. recall only `profile`-tagged " + "memories from the operator's profile bank while leaving the " + "agent's own bank unfiltered."),
+    transcript_fallback: exports_external.boolean().optional().describe("When every bank returns zero results AND no bank hit its " + "deadline, grep the current session's transcript tail for turns " + "matching the query and inject them as a clearly-labelled " + "lower-confidence block. Default true — it covers the window " + "between an abrupt kill and the next boot reconciliation, where " + "the fact layer was never told about the lost turns. Set false " + "if you never want un-consolidated transcript text in context."),
+    transcript_tail_bytes: exports_external.number().int().min(0).optional().describe("Bytes of the session transcript read from the tail for the " + "multi-turn query composition. Default 262144 (256 KiB), which " + "keeps the per-turn read O(1) on a session log that can grow to " + "many MB. 0 = read the whole file (the pre-bound behaviour, and " + "the rollback lever if a composition ever needs older turns)."),
+    max_query_chars: exports_external.number().int().min(1).optional().describe("Character bound on the composed recall query, applied before " + "`query_max_tokens` shapes it. Default 800. Truncation preserves " + "the latest turn and drops the oldest context first. Lower it " + "for an agent whose turns are long pasted payloads."),
+    parallel: exports_external.boolean().optional().describe("Run the directives fetch and every bank recall concurrently " + "under one shared deadline, so total latency is the SLOWEST slot " + "rather than their SUM. Default true. false restores the serial " + "path — the rollback lever if the parallel path ever misbehaves; " + "expect multi-bank recall latency to add up."),
     topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
   }).optional().describe("Auto-recall tuning knobs"),
   retain: exports_external.object({
@@ -19444,6 +19471,7 @@ var HindsightPerOpLlmSchema = exports_external.object({
   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({
+  gpu: exports_external.boolean().optional().describe("Force GPU passthrough for the hindsight container on (`true`) or off " + "(`false`), overriding host autodetection in BOTH directions. Absent " + "(the default) → autodetect from the persisted host-capabilities " + "verdict (`~/.switchroom/host-capabilities.json`), which enables " + "`--gpus all` only when that file proves BOTH a GPU and the nvidia " + "container toolkit. Set `true` when that verdict is wrong or unreadable " + "and you know the host has a working toolkit — switchroom cannot verify " + "it for you, and `docker run --gpus all` hard-fails container create on " + "a host without one. Set `false` to pin the container to CPU on a GPU " + "host. This is also the declarative opt-out for the recreate-time GPU " + "drop guard (`switchroom memory setup --recreate` refuses to silently " + "turn a GPU container into a CPU one). `--gpu`/`--no-gpu` on `memory " + "setup` override this for a single run."),
   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."),
@@ -19595,6 +19623,33 @@ var profileFields = {
       request_timeout_seconds: exports_external.number().int().min(1).optional(),
       own_bank_min_slots: exports_external.number().int().min(0).optional(),
       additional_bank_min_slots: exports_external.number().int().min(0).optional(),
+      min_score: exports_external.number().min(0).optional(),
+      min_score_scope: exports_external.enum(["degraded", "all"]).optional(),
+      budget: exports_external.enum(["low", "mid", "high"]).optional(),
+      max_tokens: exports_external.number().int().min(1).optional(),
+      prefer_observations: exports_external.boolean().optional(),
+      context_turns: exports_external.number().int().min(1).optional(),
+      roles: exports_external.array(exports_external.string().min(1)).min(1).optional(),
+      prompt_preamble: exports_external.string().min(1).optional(),
+      tags: exports_external.array(exports_external.string().min(1)).optional(),
+      tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional(),
+      tag_groups: exports_external.union([
+        exports_external.array(exports_external.array(exports_external.string().min(1))),
+        exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+      ]).optional(),
+      tag_weights: exports_external.record(exports_external.string(), exports_external.number().min(0)).optional(),
+      additional_bank_filters: exports_external.record(exports_external.string(), exports_external.object({
+        tags: exports_external.array(exports_external.string().min(1)).optional(),
+        tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional(),
+        tag_groups: exports_external.union([
+          exports_external.array(exports_external.array(exports_external.string().min(1))),
+          exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+        ]).optional()
+      }).strict()).optional(),
+      transcript_fallback: exports_external.boolean().optional(),
+      transcript_tail_bytes: exports_external.number().int().min(0).optional(),
+      max_query_chars: exports_external.number().int().min(1).optional(),
+      parallel: exports_external.boolean().optional(),
       additional_banks: exports_external.array(exports_external.string()).optional(),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional()
     }).optional()
@@ -21137,7 +21192,7 @@ function allocateAgentUid(name) {
 }
 
 // src/build-info.ts
-var VERSION = "0.19.26";
+var VERSION = "0.19.27";
 
 // src/setup/hindsight-recall-tunables.ts
 var RECALL_DEADLINE_HEADROOM_SECONDS = 2;
@@ -21233,6 +21288,9 @@ function assertPositive(value, label) {
   }
 }
 
+// src/setup/host-capabilities.ts
+var _warnedReads = new Set;
+
 // src/setup/hindsight-pg-defaults.ts
 var HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION = 8 * 1024;
 var HINDSIGHT_PG_APP_ANON_MIB = 2560;
@@ -21395,7 +21453,42 @@ var HINDSIGHT_HEALTHCHECK_CMD = `python3 -c '${HINDSIGHT_HEALTHCHECK_PY}'`;
 var DOCKER_PROBE_TIMEOUT_MS = 60 * 1000;
 
 // src/memory/hindsight.ts
-var DEFAULT_RETAIN_MISSION = "Extract durable facts that will still be true and useful weeks from now: " + "user preferences and standing rules, ongoing projects and recurring " + "commitments, technical and architectural decisions with their rationale, " + "and people/tool relationships. A preference revealed by a request is " + "durable — record the preference (what the user likes, wants, or always " + `does), not the request itself.
+var DEFAULT_RETAIN_MISSION = `Extract durable facts that will still be true and useful weeks from now: user preferences and standing rules, ongoing projects and recurring commitments, technical and architectural decisions with their rationale, and people/tool relationships. A preference revealed by a request is durable — record the preference (what the user likes, wants, or always does), not the request itself.
+` + `
+` + `A TOOL RESULT IS NOT A FACT. Before extracting, ask: is the subject of this
+` + `candidate a file path, a command/process/agent/session id, a temp directory, or
+` + `the location where some output was written? If yes, drop it — it is transcript
+` + `exhaust, not memory.
+` + `
+` + `NEVER extract:
+` + `- Tool results verbatim or paraphrased. Concretely, never produce a fact whose
+` + `  text resembles any of these: "File created successfully at /path/to/file",
+` + `  "A background command with ID bctz4yskm is running, and its output will be
+` + `  written to /tmp/...", "Async agent a745598ba84e71df1 was launched successfully
+` + `  and is running in the background", "User executed a Bash command to sleep for
+` + `  200 seconds", "The assistant used grep to locate 'truncateSync' in src/foo.ts".
+` + `- Anything mentioning a path under /tmp, a scratchpad directory, or a .tmp file.
+` + `- Agent tool-use traces or narration of what the assistant did (e.g. "the
+` + `  assistant used X to query Y", "ran a search", "sent the message").
+` + `- In-flight workflow/process narration (a sub-task started, paused, or is still
+` + `  running) — retain the outcome only once the task completes or a decision is made.
+` + `- Operation, request, batch, agent, command or session IDs, UUIDs, hashes, or error codes.
+` + `- Slash commands the user typed and their effects (e.g. "User issued /clear to
+` + `  reset assistant state").
+` + `- Hindsight's own errors, retries, backlogs, or internal state — the memory
+` + `  system's self-reports are not memories.
+` + `- Restatements of the user's current request or the task in progress.
+` + `- Transient state (unread counts, build status, what is running right now) unless
+` + `  the fact is explicitly dated, in which case record it as a dated observation.
+` + `- Greetings, acknowledgements, and routine operational chatter.
+` + `
+` + `If a candidate fact matches an exclusion, drop it rather than rewording it. If
+` + "nothing durable remains, return an empty facts list.";
+var SUPERSEDED_RETAIN_MISSIONS = [
+  "Extract technical decisions, architectural choices, user preferences, project context, and people/tool relationships. Ignore routine greetings and transient operational details.",
+  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise.",
+  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise, " + "including in-flight workflow/process narration (a sub-task started, " + "paused, or is still running) — only retain the outcome once a task " + "actually completes or a decision is made.",
+  "Extract durable facts that will still be true and useful weeks from now: " + "user preferences and standing rules, ongoing projects and recurring " + "commitments, technical and architectural decisions with their rationale, " + "and people/tool relationships. A preference revealed by a request is " + "durable — record the preference (what the user likes, wants, or always " + `does), not the request itself.
 
 ` + `NEVER extract:
 ` + "- Agent tool-use traces or narration of what the assistant did (e.g. " + `"the assistant used X to query Y", "ran a search", "sent the message").
@@ -21406,11 +21499,7 @@ var DEFAULT_RETAIN_MISSION = "Extract durable facts that will still be true and
 ` + "- Transient state (unread counts, build status, what is running right now) " + "unless the fact is explicitly dated, in which case record it as a dated " + `observation.
 ` + `- Greetings, acknowledgements, and routine operational chatter.
 
-` + "If a candidate fact matches an exclusion, drop it rather than rewording " + "it. If nothing durable remains, return an empty facts list.";
-var SUPERSEDED_RETAIN_MISSIONS = [
-  "Extract technical decisions, architectural choices, user preferences, project context, and people/tool relationships. Ignore routine greetings and transient operational details.",
-  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise.",
-  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise, " + "including in-flight workflow/process narration (a sub-task started, " + "paused, or is still running) — only retain the outcome once a task " + "actually completes or a decision is made."
+` + "If a candidate fact matches an exclusion, drop it rather than rewording " + "it. If nothing durable remains, return an empty facts list."
 ];
 var PROFILE_MEMORY_DEFAULTS = {
   "health-coach": {
@@ -21492,6 +21581,27 @@ var SCRYPT_MAXMEM = 128 * 1024 * 1024;
 init_client();
 var materializedDirs = new Set;
 
+// src/setup/hindsight-recall-passthrough.ts
+var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
+var HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT = "Relevant memories from past conversations (prioritize recent when " + "conflicting). Only use memories that are directly useful to continue " + "this conversation; ignore the rest:";
+var RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
+  budget: "low",
+  maxTokens: 1024,
+  preferObservations: true,
+  contextTurns: 2,
+  roles: ["user", "assistant"],
+  promptPreamble: HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT,
+  tags: [],
+  tagsMatch: "any",
+  tagGroups: {},
+  tagWeights: HINDSIGHT_RECALL_TAG_WEIGHT_SEED,
+  additionalBankFilters: {},
+  transcriptFallback: true,
+  transcriptTailBytes: 262144,
+  maxQueryChars: 800,
+  parallel: true
+});
+
 // src/agents/scaffold.ts
 var REPO_ROOT = resolve5(import.meta.dirname, "../..");
 var CLAUDE_MD_YOURS_PLACEHOLDER = "This space is yours. Add per-agent rules, exceptions, or context the " + "Switchroom template doesn't capture. Everything above the marker line is " + "regenerated on every apply; this section is preserved.";

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

@@ -4171,10 +4171,37 @@ var init_schema = __esm(() => {
       request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Even parallelised, each bank carries its own deadline " + "so ONE hung bank returns empty instead of consuming the shared " + "deadline and starving its siblings. The plugin default is 12 " + "(raised from a hardcoded 8 in #3757, which fired on 96.8% of " + "one agent's own-bank recalls). Switchroom defaults it to the " + "effective `parallel_deadline_seconds` instead — 10 at the " + "shipped ceiling — because the shared fan-out deadline is " + "already the tighter outer guard, so a per-bank value above it " + "can never bind. An explicitly configured value above the " + "effective deadline is clamped down to it, and the clamp is " + "reported."),
       own_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the agent's " + "own bank when recall fans out to more than one bank. The merged " + "set is sorted globally by relevance and head-sliced, which is " + "winner-take-all across banks: when both banks return more " + "candidates than the cap, one bank's score distribution can fill " + "every slot and the agent gets a dossier about its operator with " + "none of its own session memory. A floor, not a quota: at most " + "this many slots, only if the own bank returned that many, and " + "only up to HALF the cap shared with `additional_bank_min_slots` " + "— the rest is always won on pure relevance, so composition still " + "moves with the scores. Fixes score-based crowd-out only; a " + "timed-out bank returns no candidates and reservation is a no-op " + "there. 0 disables (default). Switchroom-managed agents use 2 " + "against the fleet-deployed cap of 6."),
       additional_bank_min_slots: exports_external.number().int().min(0).optional().describe("Slots inside `max_memories` reserved as a FLOOR for the " + "additional (profile / shared / sender) banks. Symmetric with " + "`own_bank_min_slots` — same floor-not-quota semantics, and the " + "two share the same half-of-cap reservation budget. When they sum " + "above that budget the own-bank floor is honoured first. 0 " + "disables (default). Switchroom-managed agents use 1 against the " + "fleet-deployed cap of 6. Observe `injected_own_bank_count` / " + "`injected_additional_bank_count` via " + "`switchroom memory recall-log`."),
+      min_score: exports_external.number().min(0).optional().describe("Absolute floor on a memory's engine relevance score " + "(`scores.final`) for it to be injected. 0 disables (default, " + "and the shipped fleet behaviour). Exists for one measured " + "failure: when the agent's own bank times out, recall still " + "injects side-bank residue under the banner 'Relevant memories " + "from past conversations' — 98.4% of degraded turns have a best " + "injected score below 0.01, against 28.4% of healthy ones. Six " + "noise memories are worse than none, because the agent cannot " + "tell them apart. Below-floor results are dropped BEFORE " + "rendering, and when the floor empties the set the turn says so " + "rather than going silent. Do NOT read this as a general " + "precision control: `scores.final` is not calibrated across " + "queries, and #3761 measured that an unconditional 0.01 floor " + "empties ~28% of HEALTHY recalls — which is why " + "`min_score_scope` defaults to degraded turns only. Observe " + "`dropped_below_min_score` via `switchroom memory recall-log`."),
+      min_score_scope: exports_external.enum(["degraded", "all"]).optional().describe('Which turns `min_score` binds on. "degraded" (default) — only ' + "turns where the agent's OWN bank timed out or was unreachable, " + "the population where a below-floor score actually predicts " + "noise and where the agent already receives the degraded-recall " + 'disclosure. "all" — every turn; only for an operator who has ' + "measured their own bank's score distribution, since it " + "re-creates the empty-recall failure of #3541 at any floor " + "calibrated on degraded data. No effect while `min_score` is 0."),
       types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
       additional_banks: exports_external.array(exports_external.string()).optional().describe("Extra Hindsight banks to recall from on every turn, merged into " + "the agent's own bank results — e.g. a shared operator/household " + "profile bank authored via `switchroom memory profile`. Each is " + "recalled with the `request_timeout_seconds` per-bank timeout " + "(defaults to the effective `parallel_deadline_seconds`, 10s at " + "the shipped ceiling) and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
       sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Per-speaker recall routing: a map of Telegram sender → extra " + "recall bank. When a message arrives, the agent also recalls the " + "speaker's bank (matched by Telegram username — a leading @ is " + "optional — or numeric user_id), merged " + "into its own results — so each trusted user gets their own " + "profile context. Additive recall scoping within the single " + "tenant: never an access boundary (who may drive an agent stays " + "the per-agent user assignment in `access.allowFrom`). Author the " + "banks via `switchroom memory profile`."),
       skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
+      budget: exports_external.enum(["low", "mid", "high"]).optional().describe('How hard Hindsight searches. "low" (switchroom default) = ' + 'vector retrieval only, ~1-2s. "mid" adds the LLM rerank pass ' + "and measured ~5s of hook latency on real fleet turns — the " + "second-largest contributor to perceived dead air after model " + 'TTFT. "high" is thorough and slower still. Raise it for an ' + "agent whose recall quality matters more than its reply latency " + "(a research or audit role); leave it at low for chat."),
+      max_tokens: exports_external.number().int().min(1).optional().describe("Token budget for the injected memory block. Default 1024. This " + "is the TOKEN bound; `max_memories` is the separate COUNT bound " + "and the tighter of the two wins. Raise it only alongside " + "`max_memories` — on its own it buys nothing once the count cap " + "binds."),
+      prefer_observations: exports_external.boolean().optional().describe("Bias recall toward the synthesized `observation` tier, " + "backfilling the slots freed by superseded raw facts for denser " + "coverage inside the same budget. Default true. Set false to " + "rank raw `world`/`experience` facts on equal footing — useful " + "when auditing what the consolidation engine actually stored, or " + "if a bank's observations are stale."),
+      context_turns: exports_external.number().int().min(1).optional().describe("How many recent human turns are composed into the recall query. " + 'Default 2, so a bare follow-up ("and the port?") embeds with ' + "its antecedent instead of recalling on the pronoun alone. 1 = " + "the latest turn only. Raising it costs BM25 terms, which is the " + "real recall cost driver — `query_max_tokens` still bounds the " + "result, so a large value mostly shifts which terms survive."),
+      roles: exports_external.array(exports_external.string().min(1)).min(1).optional().describe("Transcript roles the multi-turn composition may draw from. " + 'Default ["user", "assistant"]. Set ["user"] to compose ' + "the query from the human's words only — worth trying when an " + "agent's own verbose replies are dominating the query terms. No " + "effect while `context_turns` is 1."),
+      prompt_preamble: exports_external.string().min(1).optional().describe("The banner rendered above injected memories. The agent reads " + "this line as the instruction for how to treat the block, so it " + "is a behaviour knob, not cosmetics. Default tells the model to " + "prioritise recent memories on conflict and ignore irrelevant " + "ones. Override to tighten that framing for a specialised agent."),
+      tags: exports_external.array(exports_external.string().min(1)).optional().describe("Restrict recall to memories carrying these tags. Default [] = " + "no filter (match everything). This is a HARD filter applied " + "server-side — a memory without the tags cannot surface at any " + "score — so it is for a genuinely scoped agent, not for " + "ranking. Use `tag_weights` when you want a preference rather " + "than an exclusion."),
+      tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional().describe('How `tags` combine. "any" (default) = at least one; ' + '"all" = every tag. The `_strict` forms additionally require ' + "the memory to actually carry the tags rather than merely rank " + "for them. No effect while `tags` and `tag_groups` are empty."),
+      tag_groups: exports_external.union([
+        exports_external.array(exports_external.array(exports_external.string().min(1))),
+        exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+      ]).optional().describe("Tag filtering with grouping — either an OR-of-ANDs list " + '([["a","b"],["c"]] = (a AND b) OR c) or a named ' + "{group: [tags]} map. Default unset (no grouping). Use when a " + "flat `tags` + `tags_match` cannot express the scope you need."),
+      tag_weights: exports_external.record(exports_external.string(), exports_external.number().min(0)).optional().describe("Per-tag multipliers applied to `scores.final` just before the " + "final sort — a DEMOTION/PROMOTION, never a drop, so a " + "down-weighted memory still surfaces when it is the only " + "relevant hit. MERGED over switchroom's seed " + '({"sidechain": 0.8}, which ranks delegated sub-agent ' + "process-memories just under first-party ones), so setting one " + "unrelated weight does not silently undo it; pass " + "`sidechain: 1.0` to neutralise the seed. Reach for this when " + "recall_log shows one class of memory crowding the block."),
+      additional_bank_filters: exports_external.record(exports_external.string(), exports_external.object({
+        tags: exports_external.array(exports_external.string().min(1)).optional(),
+        tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional(),
+        tag_groups: exports_external.union([
+          exports_external.array(exports_external.array(exports_external.string().min(1))),
+          exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+        ]).optional()
+      }).strict()).optional().describe("Per-bank overrides of the tag filters above, keyed by bank id " + "(applies to `additional_banks` AND to sender banks). Default " + "{} = every extra bank inherits the global filters. Use it to " + "scope a shared bank — e.g. recall only `profile`-tagged " + "memories from the operator's profile bank while leaving the " + "agent's own bank unfiltered."),
+      transcript_fallback: exports_external.boolean().optional().describe("When every bank returns zero results AND no bank hit its " + "deadline, grep the current session's transcript tail for turns " + "matching the query and inject them as a clearly-labelled " + "lower-confidence block. Default true — it covers the window " + "between an abrupt kill and the next boot reconciliation, where " + "the fact layer was never told about the lost turns. Set false " + "if you never want un-consolidated transcript text in context."),
+      transcript_tail_bytes: exports_external.number().int().min(0).optional().describe("Bytes of the session transcript read from the tail for the " + "multi-turn query composition. Default 262144 (256 KiB), which " + "keeps the per-turn read O(1) on a session log that can grow to " + "many MB. 0 = read the whole file (the pre-bound behaviour, and " + "the rollback lever if a composition ever needs older turns)."),
+      max_query_chars: exports_external.number().int().min(1).optional().describe("Character bound on the composed recall query, applied before " + "`query_max_tokens` shapes it. Default 800. Truncation preserves " + "the latest turn and drops the oldest context first. Lower it " + "for an agent whose turns are long pasted payloads."),
+      parallel: exports_external.boolean().optional().describe("Run the directives fetch and every bank recall concurrently " + "under one shared deadline, so total latency is the SLOWEST slot " + "rather than their SUM. Default true. false restores the serial " + "path — the rollback lever if the parallel path ever misbehaves; " + "expect multi-bank recall latency to add up."),
       topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
     }).optional().describe("Auto-recall tuning knobs"),
     retain: exports_external.object({
@@ -4387,6 +4414,7 @@ var init_schema = __esm(() => {
     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.");
   HindsightConfigSchema = exports_external.object({
+    gpu: exports_external.boolean().optional().describe("Force GPU passthrough for the hindsight container on (`true`) or off " + "(`false`), overriding host autodetection in BOTH directions. Absent " + "(the default) → autodetect from the persisted host-capabilities " + "verdict (`~/.switchroom/host-capabilities.json`), which enables " + "`--gpus all` only when that file proves BOTH a GPU and the nvidia " + "container toolkit. Set `true` when that verdict is wrong or unreadable " + "and you know the host has a working toolkit — switchroom cannot verify " + "it for you, and `docker run --gpus all` hard-fails container create on " + "a host without one. Set `false` to pin the container to CPU on a GPU " + "host. This is also the declarative opt-out for the recreate-time GPU " + "drop guard (`switchroom memory setup --recreate` refuses to silently " + "turn a GPU container into a CPU one). `--gpu`/`--no-gpu` on `memory " + "setup` override this for a single run."),
     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."),
@@ -4538,6 +4566,33 @@ var init_schema = __esm(() => {
         request_timeout_seconds: exports_external.number().int().min(1).optional(),
         own_bank_min_slots: exports_external.number().int().min(0).optional(),
         additional_bank_min_slots: exports_external.number().int().min(0).optional(),
+        min_score: exports_external.number().min(0).optional(),
+        min_score_scope: exports_external.enum(["degraded", "all"]).optional(),
+        budget: exports_external.enum(["low", "mid", "high"]).optional(),
+        max_tokens: exports_external.number().int().min(1).optional(),
+        prefer_observations: exports_external.boolean().optional(),
+        context_turns: exports_external.number().int().min(1).optional(),
+        roles: exports_external.array(exports_external.string().min(1)).min(1).optional(),
+        prompt_preamble: exports_external.string().min(1).optional(),
+        tags: exports_external.array(exports_external.string().min(1)).optional(),
+        tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional(),
+        tag_groups: exports_external.union([
+          exports_external.array(exports_external.array(exports_external.string().min(1))),
+          exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+        ]).optional(),
+        tag_weights: exports_external.record(exports_external.string(), exports_external.number().min(0)).optional(),
+        additional_bank_filters: exports_external.record(exports_external.string(), exports_external.object({
+          tags: exports_external.array(exports_external.string().min(1)).optional(),
+          tags_match: exports_external.enum(["any", "all", "any_strict", "all_strict"]).optional(),
+          tag_groups: exports_external.union([
+            exports_external.array(exports_external.array(exports_external.string().min(1))),
+            exports_external.record(exports_external.string(), exports_external.array(exports_external.string().min(1)))
+          ]).optional()
+        }).strict()).optional(),
+        transcript_fallback: exports_external.boolean().optional(),
+        transcript_tail_bytes: exports_external.number().int().min(0).optional(),
+        max_query_chars: exports_external.number().int().min(1).optional(),
+        parallel: exports_external.boolean().optional(),
         additional_banks: exports_external.array(exports_external.string()).optional(),
         sender_banks: exports_external.record(exports_external.string(), exports_external.string()).optional()
       }).optional()
@@ -19211,6 +19266,7 @@ function assertPositive(value, label) {
 
 // src/setup/host-capabilities.ts
 init_paths();
+var _warnedReads = new Set;
 
 // src/setup/hindsight-pg-defaults.ts
 var HINDSIGHT_PG_MEM_LIMIT_MIB_FOR_DERIVATION = 8 * 1024;
@@ -19374,7 +19430,42 @@ var HINDSIGHT_HEALTHCHECK_CMD = `python3 -c '${HINDSIGHT_HEALTHCHECK_PY}'`;
 var DOCKER_PROBE_TIMEOUT_MS = 60 * 1000;
 
 // src/memory/hindsight.ts
-var DEFAULT_RETAIN_MISSION = "Extract durable facts that will still be true and useful weeks from now: " + "user preferences and standing rules, ongoing projects and recurring " + "commitments, technical and architectural decisions with their rationale, " + "and people/tool relationships. A preference revealed by a request is " + "durable — record the preference (what the user likes, wants, or always " + `does), not the request itself.
+var DEFAULT_RETAIN_MISSION = `Extract durable facts that will still be true and useful weeks from now: user preferences and standing rules, ongoing projects and recurring commitments, technical and architectural decisions with their rationale, and people/tool relationships. A preference revealed by a request is durable — record the preference (what the user likes, wants, or always does), not the request itself.
+` + `
+` + `A TOOL RESULT IS NOT A FACT. Before extracting, ask: is the subject of this
+` + `candidate a file path, a command/process/agent/session id, a temp directory, or
+` + `the location where some output was written? If yes, drop it — it is transcript
+` + `exhaust, not memory.
+` + `
+` + `NEVER extract:
+` + `- Tool results verbatim or paraphrased. Concretely, never produce a fact whose
+` + `  text resembles any of these: "File created successfully at /path/to/file",
+` + `  "A background command with ID bctz4yskm is running, and its output will be
+` + `  written to /tmp/...", "Async agent a745598ba84e71df1 was launched successfully
+` + `  and is running in the background", "User executed a Bash command to sleep for
+` + `  200 seconds", "The assistant used grep to locate 'truncateSync' in src/foo.ts".
+` + `- Anything mentioning a path under /tmp, a scratchpad directory, or a .tmp file.
+` + `- Agent tool-use traces or narration of what the assistant did (e.g. "the
+` + `  assistant used X to query Y", "ran a search", "sent the message").
+` + `- In-flight workflow/process narration (a sub-task started, paused, or is still
+` + `  running) — retain the outcome only once the task completes or a decision is made.
+` + `- Operation, request, batch, agent, command or session IDs, UUIDs, hashes, or error codes.
+` + `- Slash commands the user typed and their effects (e.g. "User issued /clear to
+` + `  reset assistant state").
+` + `- Hindsight's own errors, retries, backlogs, or internal state — the memory
+` + `  system's self-reports are not memories.
+` + `- Restatements of the user's current request or the task in progress.
+` + `- Transient state (unread counts, build status, what is running right now) unless
+` + `  the fact is explicitly dated, in which case record it as a dated observation.
+` + `- Greetings, acknowledgements, and routine operational chatter.
+` + `
+` + `If a candidate fact matches an exclusion, drop it rather than rewording it. If
+` + "nothing durable remains, return an empty facts list.";
+var SUPERSEDED_RETAIN_MISSIONS = [
+  "Extract technical decisions, architectural choices, user preferences, project context, and people/tool relationships. Ignore routine greetings and transient operational details.",
+  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise.",
+  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise, " + "including in-flight workflow/process narration (a sub-task started, " + "paused, or is still running) — only retain the outcome once a task " + "actually completes or a decision is made.",
+  "Extract durable facts that will still be true and useful weeks from now: " + "user preferences and standing rules, ongoing projects and recurring " + "commitments, technical and architectural decisions with their rationale, " + "and people/tool relationships. A preference revealed by a request is " + "durable — record the preference (what the user likes, wants, or always " + `does), not the request itself.
 
 ` + `NEVER extract:
 ` + "- Agent tool-use traces or narration of what the assistant did (e.g. " + `"the assistant used X to query Y", "ran a search", "sent the message").
@@ -19385,11 +19476,7 @@ var DEFAULT_RETAIN_MISSION = "Extract durable facts that will still be true and
 ` + "- Transient state (unread counts, build status, what is running right now) " + "unless the fact is explicitly dated, in which case record it as a dated " + `observation.
 ` + `- Greetings, acknowledgements, and routine operational chatter.
 
-` + "If a candidate fact matches an exclusion, drop it rather than rewording " + "it. If nothing durable remains, return an empty facts list.";
-var SUPERSEDED_RETAIN_MISSIONS = [
-  "Extract technical decisions, architectural choices, user preferences, project context, and people/tool relationships. Ignore routine greetings and transient operational details.",
-  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise.",
-  "Extract user preferences, ongoing projects, recurring commitments, " + "important context, and durable facts that should help across future " + "conversations. Skip one-off chatter and temporary task noise, " + "including in-flight workflow/process narration (a sub-task started, " + "paused, or is still running) — only retain the outcome once a task " + "actually completes or a decision is made."
+` + "If a candidate fact matches an exclusion, drop it rather than rewording " + "it. If nothing durable remains, return an empty facts list."
 ];
 var PROFILE_MEMORY_DEFAULTS = {
   "health-coach": {
@@ -19439,6 +19526,27 @@ var materializedDirs = new Set;
 // src/setup/onboarding.ts
 init_paths();
 
+// src/setup/hindsight-recall-passthrough.ts
+var HINDSIGHT_RECALL_TAG_WEIGHT_SEED = Object.freeze({ sidechain: 0.8 });
+var HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT = "Relevant memories from past conversations (prioritize recent when " + "conflicting). Only use memories that are directly useful to continue " + "this conversation; ignore the rest:";
+var RECALL_PASSTHROUGH_DEFAULTS = Object.freeze({
+  budget: "low",
+  maxTokens: 1024,
+  preferObservations: true,
+  contextTurns: 2,
+  roles: ["user", "assistant"],
+  promptPreamble: HINDSIGHT_RECALL_PROMPT_PREAMBLE_DEFAULT,
+  tags: [],
+  tagsMatch: "any",
+  tagGroups: {},
+  tagWeights: HINDSIGHT_RECALL_TAG_WEIGHT_SEED,
+  additionalBankFilters: {},
+  transcriptFallback: true,
+  transcriptTailBytes: 262144,
+  maxQueryChars: 800,
+  parallel: true
+});
+
 // src/repos/bare-clone.ts
 init_paths();