switchroom 0.19.25 → 0.19.27

package/dist/vault/broker/server.js

@@ -4164,15 +4164,44 @@ var init_schema = __esm(() => {
     recall: exports_external.object({
       max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
+      hook_timeout_seconds: exports_external.number().int().min(1).optional().describe("Ceiling (seconds) Claude Code gives the UserPromptSubmit recall " + "hook before killing it. Stamped into the installed plugin's " + "hooks/hooks.json, so it survives `switchroom apply` reinstalling " + "the plugin. Default 12. Raising it lets slow banks finish at the " + "cost of pre-turn dead air; `parallel_deadline_seconds` and " + "`request_timeout_seconds` are both kept under it. A value below " + "3s is raised to 3s and reported: the fan-out deadline must be at " + "least 1s AND still leave 2s of post-deadline headroom, so a " + "lower ceiling admits no usable envelope at all."),
+      parallel_deadline_seconds: exports_external.number().int().min(1).optional().describe("Shared deadline (seconds) for the whole parallel multi-bank " + "recall fan-out. Slots unfinished when it elapses are abandoned " + "and reported as timed out. Defaults to `hook_timeout_seconds` " + "minus 2s of headroom for block formatting, cache write and " + "stdout flush, so a straggler bank can never push the hook past " + "its ceiling. Set explicitly to override that derivation; a value " + "that would leave less than 2s under the hook ceiling — including " + "one set EQUAL to it — is clamped back to `hook_timeout_seconds` " + "minus 2, and the clamp is reported. Equality is not allowed: at " + "zero headroom the hook is killed mid-write and the turn loses " + "both the memories and the recall_log row explaining why."),
       query_max_tokens: exports_external.number().int().min(0).optional().describe("Cap on the number of DISTINCT BM25 terms the recall hook may " + "put on the wire. `recallMaxQueryChars` bounds characters, which " + "is not the cost driver: Hindsight OR-joins every query token " + "into one tsquery and Postgres native FTS ranks the entire " + "matched set before the top-60 heapsort, so cost tracks TERMS. " + "An 800-char composed query is ~96 distinct terms and matched " + "119,510 of 135,565 rows on the `overlord` bank (14.0s for the " + "3-arm " + "BM25 UNION), past the per-bank timeout — 96.8% of that agent's " + "own-bank recalls returned nothing. Plugin default is 24 " + "(measured 48,433 rows / 2.7s on the same bank). Terms are " + "chosen recency-first (the latest turn beats prior context), " + "then by selectivity. 0 disables shaping (rollback lever)."),
       query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional().describe("Extra terms dropped from the BM25 recall query, on top of the " + "built-in English stopword list. For BANK-SPECIFIC " + "high-document-frequency words a generic stoplist cannot know " + "about: on `overlord`, `switchroom` matches 20% of the bank and " + "`agent` another 20%, purely because that is what the corpus is " + "about, and each such term drags tens of thousands of rows into " + "the ranking. Defaults to []."),
-      request_timeout_seconds: exports_external.number().int().min(1).optional().describe("Per-bank HTTP read timeout, in seconds, for one recall " + "request. Plugin default is 12, matching the UserPromptSubmit " + "hook ceiling; the shared `recallParallelDeadlineSeconds` (10) " + "is the tighter outer guard in the default configuration, so " + "this is a per-request safety net. Was a hardcoded 8 in the " + "plugin before #3757."),
+      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 an 8s timeout and is non-fatal on failure. Stays " + "within the single tenant: all banks are the operator's data, in " + "the operator's Hindsight instance (see the `single-tenant` " + "invariant). Defaults to [] (no extra banks)."),
+      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({
@@ -4385,6 +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."),
@@ -4529,11 +4559,40 @@ var init_schema = __esm(() => {
       recall: exports_external.object({
         max_memories: exports_external.number().int().min(0).optional(),
         cache_ttl_secs: exports_external.number().int().min(0).optional(),
+        hook_timeout_seconds: exports_external.number().int().min(1).optional(),
+        parallel_deadline_seconds: exports_external.number().int().min(1).optional(),
         query_max_tokens: exports_external.number().int().min(0).optional(),
         query_stop_terms: exports_external.array(exports_external.string().min(1).regex(/^[\w./-]+$/)).optional(),
         request_timeout_seconds: exports_external.number().int().min(1).optional(),
         own_bank_min_slots: exports_external.number().int().min(0).optional(),
         additional_bank_min_slots: exports_external.number().int().min(0).optional(),
+        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()
@@ -18591,6 +18650,10 @@ function allocateAgentUid(name) {
   return AGENT_UID_MIN + u32 % range;
 }
 
+// src/setup/hindsight-recall-tunables.ts
+var RECALL_DEADLINE_HEADROOM_SECONDS = 2;
+var MIN_RECALL_HOOK_TIMEOUT_SECONDS = RECALL_DEADLINE_HEADROOM_SECONDS + 1;
+
 // src/agents/scaffold.ts
 init_schema();
 
@@ -18876,6 +18939,7 @@ 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;
@@ -19039,7 +19103,42 @@ 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").
@@ -19050,11 +19149,7 @@ 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": {
@@ -19859,6 +19954,27 @@ 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();
 

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": "0.19.25",
+  "version": "0.19.27",
   "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": {
@@ -28,11 +28,12 @@
     "test:vitest": "vitest run",
     "test:bun": "bun test telegram-plugin/tests/agent-state-dir-preload.test.ts telegram-plugin/tests/catch-all-forwarded-history.test.ts src/vault/grants.test.ts src/vault/grants-db.test.ts src/vault/write-grants.test.ts src/vault/broker/server-grants.test.ts src/vault/broker/server-write-grants.test.ts src/vault/broker/server-scope-persist.test.ts src/vault/broker/server-tokenless-scope.test.ts src/vault/broker/server-mint-grant-passphrase-attest.test.ts src/vault/broker/server-passphrase-attest.test.ts src/vault/broker/server-mint-grant-posture-attest.test.ts src/vault/broker/server-admin-only-keys.test.ts src/vault/broker/client-token.test.ts src/vault/broker/server-unlock.test.ts src/vault/broker/auto-unlock.test.ts src/vault/broker/drift-detection.test.ts tests/vault-broker-passphrase.test.ts src/cli/vault-get-broker.test.ts src/vault/resolver-via-broker.test.ts src/vault/broker/scope.test.ts src/vault/broker/server.test.ts src/litellm/provision-apply-e2e.test.ts src/drive/disconnect.test.ts src/drive/grants.test.ts src/drive/oauth.test.ts src/drive/onboarding.test.ts src/drive/reconciler.test.ts src/drive/vault-slots.test.ts src/drive/wrapper.test.ts src/vault/approvals/kernel.test.ts src/vault/approvals/approval-origin.test.ts src/vault/approvals/self-approval-bypass.test.ts src/vault/approvals/schema-idempotent.test.ts src/vault/broker/server-approvals.test.ts telegram-plugin/tests/boot-probes.test.ts telegram-plugin/tests/boot-version-string.test.ts telegram-plugin/tests/history.test.ts telegram-plugin/tests/cross-turn-card-gate.test.ts telegram-plugin/tests/emission-authority-open-gate.test.ts telegram-plugin/tests/emission-authority-ping-gate.test.ts telegram-plugin/tests/emission-authority-card-drain-gate.test.ts telegram-plugin/tests/per-topic-current-turn.test.ts telegram-plugin/tests/history-reaper.test.ts telegram-plugin/tests/ipc-server-client.test.ts telegram-plugin/tests/ipc-server-race.test.ts telegram-plugin/tests/ipc-server-query-pending-permission.test.ts telegram-plugin/tests/ipc-server-check-pre-approved.test.ts telegram-plugin/tests/gateway-bridge.test.ts telegram-plugin/tests/gateway-startup-mutex.test.ts telegram-plugin/tests/gateway-clean-shutdown-marker.test.ts telegram-plugin/tests/boot-card-dedupe.test.ts telegram-plugin/tests/boot-card-reason.test.ts telegram-plugin/tests/progress-update.test.ts telegram-plugin/tests/quota-cache.test.ts telegram-plugin/tests/silent-reply-guard.test.ts telegram-plugin/tests/unhandled-rejection-policy.test.ts telegram-plugin/tests/registry-turns.test.ts telegram-plugin/registry/subagents.test.ts telegram-plugin/registry/subagents-bugs.test.ts telegram-plugin/tests/subagent-watcher-parent-turn-key.test.ts telegram-plugin/tests/subagent-nested-dispatch.test.ts telegram-plugin/tests/nested-worker-visibility-harness.test.ts telegram-plugin/tests/turns-writer.test.ts telegram-plugin/tests/resume-inbound-builder.test.ts telegram-plugin/tests/subagent-tracker-hooks.test.ts telegram-plugin/tests/resolve-calling-subagent.test.ts telegram-plugin/tests/gateway-update-placeholder-dispatch.test.ts telegram-plugin/tests/status-query-telemetry.test.ts telegram-plugin/tests/reaction-trigger.test.ts telegram-plugin/tests/reaction-trigger-flow.test.ts telegram-plugin/tests/subagent-watcher-workflow-visibility.test.ts telegram-plugin/uat/load-env.test.ts telegram-plugin/uat/feed-matcher.test.ts telegram-plugin/uat/uat-driver.test.ts telegram-plugin/gateway/webhook-ingest-server.test.ts telegram-plugin/tests/skill-proposal-card.test.ts",
     "test:watch": "vitest",
-    "lint": "tsc --noEmit && node scripts/check-plugin-references.mjs && bash scripts/check-bot-api-wrapping.sh && node scripts/check-bun-test-imports.mjs && node scripts/check-bun-module-mock-scope.mjs && node scripts/check-no-pii-secrets.mjs && node scripts/check-vault-test-hermeticity.mjs && node scripts/check-auth-test-hermeticity.mjs && node scripts/check-agent-state-dir-hermeticity.mjs && node scripts/check-no-broadcast-delivery.mjs && node scripts/check-stale-tool-descriptions.mjs && node scripts/check-mcp-instructions-budget.mjs && node scripts/check-web-subscription-honest.mjs && node scripts/check-no-unpinned-npx-playwright.mjs && node scripts/check-gateway-line-ratchet.mjs && node scripts/check-litellm-config-guard.mjs && node scripts/check-release-asset-names.mjs",
+    "lint": "tsc --noEmit && node scripts/check-plugin-references.mjs && bash scripts/check-bot-api-wrapping.sh && node scripts/check-bun-test-imports.mjs && node scripts/check-test-runner-coverage.mjs && node scripts/check-bun-module-mock-scope.mjs && node scripts/check-no-pii-secrets.mjs && node scripts/check-vault-test-hermeticity.mjs && node scripts/check-auth-test-hermeticity.mjs && node scripts/check-agent-state-dir-hermeticity.mjs && node scripts/check-no-broadcast-delivery.mjs && node scripts/check-stale-tool-descriptions.mjs && node scripts/check-mcp-instructions-budget.mjs && node scripts/check-web-subscription-honest.mjs && node scripts/check-no-unpinned-npx-playwright.mjs && node scripts/check-gateway-line-ratchet.mjs && node scripts/check-retry-flood-hooks.mjs && node scripts/check-status-pin-single-path.mjs && node scripts/check-litellm-config-guard.mjs && node scripts/check-release-asset-names.mjs && node scripts/check-agent-attribution-trailers.mjs",
     "lint:tsc": "tsc --noEmit",
     "lint:plugin-references": "node scripts/check-plugin-references.mjs",
     "lint:bot-api-wrapping": "bash scripts/check-bot-api-wrapping.sh",
     "lint:bun-test-imports": "node scripts/check-bun-test-imports.mjs",
+    "lint:test-runner-coverage": "node scripts/check-test-runner-coverage.mjs",
     "lint:bun-module-mock-scope": "node scripts/check-bun-module-mock-scope.mjs",
     "lint:no-pii": "node scripts/check-no-pii-secrets.mjs",
     "lint:auth-test-hermeticity": "node scripts/check-auth-test-hermeticity.mjs",
@@ -40,9 +41,12 @@
     "lint:web-subscription-honest": "node scripts/check-web-subscription-honest.mjs",
     "lint:no-broadcast-delivery": "node scripts/check-no-broadcast-delivery.mjs",
     "lint:gateway-line-ratchet": "node scripts/check-gateway-line-ratchet.mjs",
+    "lint:retry-flood-hooks": "node scripts/check-retry-flood-hooks.mjs",
+    "lint:status-pin-single-path": "node scripts/check-status-pin-single-path.mjs",
     "lint:mcp-instructions-budget": "node scripts/check-mcp-instructions-budget.mjs",
     "lint:litellm-config-guard": "node scripts/check-litellm-config-guard.mjs",
     "lint:release-asset-contract": "node scripts/check-release-asset-names.mjs",
+    "lint:agent-attribution-trailers": "node scripts/check-agent-attribution-trailers.mjs",
     "prepublishOnly": "npm run build && npm run lint && npm test"
   },
   "dependencies": {

package/profiles/_base/cron-session.sh.hbs

@@ -74,6 +74,12 @@ export CLAUDE_CONFIG_DIR="$CRON_CONFIG_DIR"
 # (honouring any value already in the env). See start.sh.hbs for rationale.
 export CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH="${CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH:-2}"
 
+# Pin sub-agent fan-out width below the CLI's own default of 20 — the
+# deterministic half of the dev-protocol's "max 15 parallel sub-agents".
+# Normally inherited from start.sh's env; set here as a belt-and-braces default
+# (honouring any value already in the env). See start.sh.hbs for rationale.
+export CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS="${CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS:-15}"
+
 # Defer the (now full) MCP schema set via tool-search so the cron session keeps
 # its low-context cost — the heavy servers load on demand; switchroom-telegram
 # + hindsight stay alwaysLoad. Normally inherited from start.sh's env; set here
@@ -171,6 +177,14 @@ fi
 # claude in a detached pane that registers fine; a human can still
 # `tmux -L "$CRON_SOCKET" attach -t "$CRON_NAME"` later. The `-x/-y` give the
 # pane a size since there's no TTY to derive one from.
+#
+# Commit-attribution carrier for the model half of the trailer pair. The Tier-1
+# cron session runs a DIFFERENT model from the main session, so it exports its
+# own value — otherwise a commit made during a cron fire would be stamped with
+# the main session's model. Mirrors the export in start.sh; consumed by
+# bin/git-agent-attribution-hook.sh. ({{{cronModelQ}}} is already
+# shell-single-quoted by the scaffold, so it is assigned BARE.)
+export SWITCHROOM_SESSION_MODEL={{{cronModelQ}}}
 tmux -L "$CRON_SOCKET" \
   new-session -d -s "$CRON_NAME" -x 400 -y 50 \
   claude \

package/profiles/_base/start.sh.hbs

@@ -594,6 +594,30 @@ if [ ! -e "$HOME/.switchroom-config" ] || [ -L "$HOME/.switchroom-config" ]; the
 fi
 {{/if}}
 
+# --- Commit attribution: global git hooks path ---------------------------
+#
+# Every commit an agent makes is AUTHORED by the operator's GitHub account
+# (deliberate, and staying that way), which leaves `git log` unable to say
+# WHICH agent and WHICH model produced a change. The image ships a
+# `prepare-commit-msg` hook that stamps `Switchroom-Agent:` /
+# `Switchroom-Model:` trailers from $SWITCHROOM_AGENT_NAME and the launched
+# model — see bin/git-agent-attribution-hook.sh for the full rationale
+# (including why prepare-commit-msg, which `--no-verify` does NOT bypass).
+#
+# Set GLOBALLY, not per-repo: an agent clones repos on demand and a per-repo
+# install would mean every new clone starts out unattributed until someone
+# remembered to run a setup step. Repo-local `core.hooksPath` (husky et al)
+# still wins, and the hook chains to any displaced `.git/hooks` copy.
+#
+# Idempotent, and non-fatal: attribution must never be the reason a boot
+# fails. `git config --global` writes $HOME/.gitconfig, which lives on the
+# /state/agent bind mount and so persists — re-asserting it each boot is what
+# repairs an agent whose config was hand-edited.
+if command -v git >/dev/null 2>&1 && [ -x /opt/switchroom/git-hooks/prepare-commit-msg ]; then
+  git config --global core.hooksPath /opt/switchroom/git-hooks 2>/dev/null \
+    || echo "attribution: WARNING — could not set core.hooksPath; commits from this agent will not carry Switchroom-Agent/Switchroom-Model trailers and will fail the repo's attribution lint" >&2
+fi
+
 export NVM_DIR="$HOME/.nvm"
 [ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"
 export PATH="$HOME/.bun/bin:$PATH"
@@ -685,10 +709,19 @@ unset CLAUDE_CODE_OAUTH_TOKEN
 # out its own sub-agents), so we restore the pre-2.1.217 behaviour. Depth 2
 # = main → sub-agent → grandchild, which covers the one extra level the
 # fleet actually uses; there is no evidence of deeper nesting, so we don't
-# raise it further. (The companion 2.1.217 change — CLAUDE_CODE_MAX_
-# CONCURRENT_SUBAGENTS defaulting to 20 — is already compatible: our fleet
-# self-caps at 15 parallel sub-agents, so the 20 default is left untouched.)
+# raise it further.
 export CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH="${CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH:-2}"
+# Fan-out width. The CLI's own default is 20 (2.1.217's companion change);
+# switchroom deliberately pins it LOWER, at 15, because the dev-protocol
+# fragment every agent boots with ("max 15 parallel sub-agents",
+# profiles/_shared/dev-protocol.md.hbs) is prompt text — a model can ignore
+# it, and nothing else caps the width. This export is the deterministic
+# mechanism behind that sentence: past 15 the CLI refuses the launch instead
+# of relying on the agent to count. 15 rather than 20 keeps a fully fanned-out
+# worker inside the per-container CPU/RAM envelope and leaves headroom for the
+# nested level that SPAWN_DEPTH=2 above permits. Override-able for a one-off
+# wide job; raise both this and the fragment together or they disagree.
+export CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS="${CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS:-15}"
 export TELEGRAM_STATE_DIR="{{agentDir}}/telegram"
 # SWITCHROOM_AGENT_NAME is the canonical "which agent am I" identifier the
 # telegram-plugin reads to detect self-restart commands. We can't rely on
@@ -737,6 +770,33 @@ export HINDSIGHT_RECALL_MAX_MEMORIES={{hindsightRecallMaxMemories}}
 {{#if (isNumber hindsightRecallCacheTtlSecs)}}
 export HINDSIGHT_RECALL_CACHE_TTL_SECS={{hindsightRecallCacheTtlSecs}}
 {{/if}}
+# Recall latency envelope (memory.recall.parallel_deadline_seconds /
+# .request_timeout_seconds cascade). Exported UNCONDITIONALLY, unlike the
+# override-only knobs above, and that is deliberate.
+#
+# The plugin's config load order is: built-in defaults -> plugin settings.json
+# -> ~/.hindsight/claude-code.json -> environment. Switchroom stamps these two
+# into settings.json as well, but claude-code.json sits BETWEEN the two and
+# SURVIVES `switchroom apply`. A hand-written value left in that file (the
+# documented pre-promotion workaround — the live fleet had exactly this,
+# pinning recallParallelDeadlineSeconds to a stale 16) would therefore silently
+# defeat a settings.json-only stamp. Env is the top of the chain, so exporting
+# the resolved value unconditionally is what actually makes switchroom.yaml
+# authoritative. Both values come from the same resolver
+# (src/setup/hindsight-recall-tunables.ts) fed the same CASCADED agent config —
+# scaffold/reconcile thread their resolved `agentConfig` into
+# installHindsightPlugin instead of letting it re-derive — so the two channels
+# cannot diverge. Sharing a resolver was NOT enough on its own: the stamp used
+# to re-derive from `agents[name] ?? {}`, and `{}` carries no `extends`, so it
+# silently dropped the profile tier on one side only.
+# Pinned by tests/scaffold.recall-envelope-carrier.test.ts, which asserts on the
+# RENDERED start.sh — a typo'd handlebars variable renders empty rather than
+# failing, and an empty export hands authority straight back to claude-code.json.
+#
+# The third knob in this envelope — the UserPromptSubmit hook ceiling — has no
+# env channel at all (Claude Code does not expand env in hooks.json `timeout`),
+# so it is stamped into the deployed hooks/hooks.json instead.
+export HINDSIGHT_RECALL_PARALLEL_DEADLINE_SECONDS={{hindsightRecallParallelDeadlineSeconds}}
 # BM25 term budget for the recall query (#3757). The hook composes a query
 # from the last 2 turns and Hindsight OR-joins EVERY token into one
 # to_tsquery; Postgres native FTS cannot top-k from the GIN index, so it
@@ -789,6 +849,78 @@ export HINDSIGHT_RECALL_REQUEST_TIMEOUT_SECONDS={{hindsightRecallRequestTimeoutS
 # Observe `injected_own_bank_count` via `switchroom memory recall-log <agent>`.
 export HINDSIGHT_RECALL_OWN_BANK_MIN_SLOTS={{hindsightRecallOwnBankMinSlots}}
 export HINDSIGHT_RECALL_ADDITIONAL_BANK_MIN_SLOTS={{hindsightRecallAdditionalBankMinSlots}}
+# Absolute relevance floor on an injected memory's engine score (#3837).
+# 0 = OFF, and 0 is what switchroom ships: at 0 recall.py drops nothing and the
+# injected set is exactly what it was before #3837.
+# It exists for the measured failure where the agent's own bank times out and
+# recall injects side-bank residue under the "relevant memories" banner —
+# 98.4% of degraded turns carry a best injected score below 0.01 against 28.4%
+# of healthy ones, and the agent cannot tell noise from recall. SCOPE decides
+# which turns a non-zero floor binds on: `degraded` (default) only when the own
+# bank timed out / was unreachable, `all` on every turn. `all` is NOT a
+# recommended fleet default — an unconditional 0.01 floor empties ~28% of
+# HEALTHY recalls (#3761), because `scores.final` is not calibrated across
+# queries. EXPORTED UNCONDITIONALLY for the #3774 reason above: a stale
+# hand-edited ~/.hindsight/claude-code.json must not be able to switch a recall
+# filter ON that switchroom ships OFF.
+# Observe `dropped_below_min_score` via `switchroom memory recall-log <agent>`.
+export HINDSIGHT_RECALL_MIN_SCORE={{hindsightRecallMinScore}}
+export HINDSIGHT_RECALL_MIN_SCORE_SCOPE={{hindsightRecallMinScoreScope}}
+# ── The remaining recall knobs (#3841) ───────────────────────────────────────
+# Everything else `recall.py` reads out of the plugin config, resolved from
+# `memory.recall.*` in switchroom.yaml and exported UNCONDITIONALLY for the
+# #3774 reason spelled out above: `~/.hindsight/claude-code.json` loads AFTER
+# the plugin's settings.json and survives an apply, so a knob switchroom does
+# not export is a knob a stale hand-edit owns. Every default below is the value
+# the fleet already ran before this block existed — an operator who sets none of
+# these gets byte-identical recall — and each is pinned against the vendored
+# plugin by tests/scaffold.recall-passthrough.test.ts so the two cannot drift.
+# Structural values are single-quoted JSON; `_cast_env` parses them and, unlike
+# an empty export, actually ASSIGNS (an empty one is skipped and hands authority
+# straight back to claude-code.json).
+#
+# How hard Hindsight searches: low (vector only, ~1-2s) | mid (+LLM rerank,
+# ~5s measured) | high. Switchroom ships `low` — mid was the second-largest
+# contributor to perceived dead air after model TTFT.
+export HINDSIGHT_RECALL_BUDGET={{hindsightRecallPass.budget}}
+# Token budget for the injected block. The count cap (…MAX_MEMORIES) is the
+# other bound and the tighter of the two wins.
+export HINDSIGHT_RECALL_MAX_TOKENS={{hindsightRecallPass.maxTokens}}
+# Bias toward the synthesized `observation` tier, backfilling slots freed by
+# superseded raw facts. On by default (memory.recall.prefer_observations).
+export HINDSIGHT_RECALL_PREFER_OBSERVATIONS={{hindsightRecallPass.preferObservations}}
+# Recent human turns composed into the query, and the transcript roles they may
+# be drawn from. 2 turns so a bare follow-up ("and the port?") embeds with its
+# antecedent instead of recalling on the pronoun alone.
+export HINDSIGHT_RECALL_CONTEXT_TURNS={{hindsightRecallPass.contextTurns}}
+export HINDSIGHT_RECALL_ROLES={{{hindsightRecallPass.rolesQ}}}
+# Character bound on the composed query, applied BEFORE …QUERY_MAX_TOKENS
+# shapes it; truncation keeps the latest turn and drops oldest context first.
+export HINDSIGHT_RECALL_MAX_QUERY_CHARS={{hindsightRecallPass.maxQueryChars}}
+# Tail bytes read from the session transcript for that composition — the read
+# stays O(1) on a session log that can grow to many MB. 0 = whole file.
+export HINDSIGHT_RECALL_TRANSCRIPT_TAIL_BYTES={{hindsightRecallPass.transcriptTailBytes}}
+# The banner above injected memories. The agent reads this line as the
+# instruction for how to treat the block, so it is behaviour, not cosmetics.
+export HINDSIGHT_RECALL_PROMPT_PREAMBLE={{{hindsightRecallPass.promptPreambleQ}}}
+# Tag filters (upstream 962140eef port). Dormant by default and exported at
+# their no-op values — `[]` / `{}` fold to None in recall.py, so nothing is
+# filtered unless an operator sets memory.recall.tags / .tag_groups.
+export HINDSIGHT_RECALL_TAGS={{{hindsightRecallPass.tagsQ}}}
+export HINDSIGHT_RECALL_TAGS_MATCH={{hindsightRecallPass.tagsMatch}}
+export HINDSIGHT_RECALL_TAG_GROUPS={{{hindsightRecallPass.tagGroupsQ}}}
+export HINDSIGHT_RECALL_ADDITIONAL_BANK_FILTERS={{{hindsightRecallPass.additionalBankFiltersQ}}}
+# Per-tag score multipliers — a DOWN-RANK, never a drop. Carries switchroom's
+# `sidechain: 0.8` seed (delegated sub-agent memories rank just under
+# first-party ones), merged with memory.recall.tag_weights.
+export HINDSIGHT_RECALL_TAG_WEIGHTS={{{hindsightRecallPass.tagWeightsQ}}}
+# Bounded transcript-grep fallback: only when EVERY bank returned zero AND no
+# bank hit its deadline. Covers the window between an abrupt kill and the next
+# boot reconciliation, where the fact layer never heard about the lost turns.
+export HINDSIGHT_RECALL_TRANSCRIPT_FALLBACK={{hindsightRecallPass.transcriptFallback}}
+# Parallel multi-bank fan-out: total latency is the SLOWEST slot, not the SUM.
+# false is the serial rollback lever.
+export HINDSIGHT_RECALL_PARALLEL={{hindsightRecallPass.parallel}}
 # Recall fact types (memory.recall.types cascade). Switchroom default is
 # world,experience,observation (the synthesized `observation` tier is ON
 # by default, set in the plugin settings.json). Export only when the
@@ -1772,7 +1904,7 @@ esac
 # tells the operator once.
 #
 # NB: the CONFIGURED default resolution is untouched (#3186 constraint) —
-# the fleet `thinking_effort: low` pin (see #1978 /
+# the configured `thinking_effort` (see #1978 /
 # src/config/thinking-effort-risk.ts) resolves exactly as before.
 _EFFECTIVE_EFFORT={{#if thinkingEffort}}'{{thinkingEffort}}'{{else}}''{{/if}}
 if [ -f "{{agentDir}}/.session-effort" ]; then
@@ -1857,6 +1989,15 @@ esac
 # welcome card honest. Overwrite (not consumed) — the gateway reads it at boot.
 printf '%s\n' "$_EFFECTIVE_MODEL" > "{{agentDir}}/.active-session-model" 2>/dev/null || true
 
+# Commit-attribution carrier for the model half of the trailer pair. Exported
+# here, from the SAME $_EFFECTIVE_MODEL the session is about to launch on, so a
+# `/model` override is reflected in the trailer rather than the configured
+# default. `exec claude` below inherits it, and so does every shell claude
+# spawns — which is where `git commit` actually runs.
+# Consumer: bin/git-agent-attribution-hook.sh (falls back to the static
+# SWITCHROOM_AGENT_MODEL that compose.ts writes, then to "unknown").
+export SWITCHROOM_SESSION_MODEL="$_EFFECTIVE_MODEL"
+
 # --- Routing-mode observability (.routing-mode, 2026-07-17 boot-race incident) ---
 #
 # Deterministic record of WHERE this boot actually landed, written immediately