switchroom 0.19.26 → 0.19.28

package/bin/git-agent-attribution-hook.sh

@@ -0,0 +1,144 @@
+#!/bin/sh
+# git-agent-attribution-hook.sh — deterministic `Switchroom-Agent:` /
+# `Switchroom-Model:` commit trailers for work done inside a switchroom
+# agent container.
+#
+# WHY THIS IS A HOOK AND NOT A RULE IN CLAUDE.md
+# ----------------------------------------------
+# Every PR and commit in the fleet's repos is authored by the operator's
+# GitHub account, because agents commit with the operator's git identity.
+# That is deliberate and stays. The cost is that `git log` cannot answer
+# "which agent, on which model, produced this" — the only machine marker
+# is the generic `Co-authored-by: Claude ...` trailer, which names no
+# agent.
+#
+# A prompt rule ("always add these trailers") is model-dependent: it is
+# forgotten under compaction, dropped by a sub-agent that never read the
+# rule, and silently absent forever after. This hook makes the trailers a
+# property of the ENVIRONMENT instead: any `git commit` run inside an
+# agent container gets them, from `$SWITCHROOM_AGENT_NAME` and the live
+# session model, with no model cooperation at all.
+#
+# WHY `prepare-commit-msg` AND NOT `commit-msg`
+# ---------------------------------------------
+# `git commit --no-verify` bypasses `pre-commit` and `commit-msg`. It does
+# NOT bypass `prepare-commit-msg` (verified against git 2.47). Agents reach
+# for `--no-verify` routinely to get past a slow or broken repo hook, so a
+# `commit-msg` implementation would be one habitual flag away from silent
+# non-attribution. `prepare-commit-msg` also runs for `git merge` and
+# `git commit --amend`, and it puts the trailers in the message BEFORE the
+# editor, so they are visible rather than injected behind the author's back.
+#
+# ACTIVATION
+# ----------
+# `docker/Dockerfile.agent` installs this file (read-only, in the image, at
+# a path no agent can rewrite) as
+# `/opt/switchroom/git-hooks/prepare-commit-msg`, and the scaffolded
+# `start.sh` sets `git config --global core.hooksPath
+# /opt/switchroom/git-hooks` on every boot. Global config means it applies
+# to EVERY repo the agent clones, including worktrees, without per-repo
+# setup — which is the point: a new clone must not be able to start out
+# un-attributed.
+#
+# REPO-LOCAL HOOK CHAINING
+# ------------------------
+# `core.hooksPath` REPLACES `.git/hooks` wholesale, so a repo that installs
+# its own `prepare-commit-msg` into `.git/hooks` (the `pre-commit`
+# framework does exactly this) would silently stop running it. We therefore
+# delegate to `$GIT_DIR/hooks/prepare-commit-msg` FIRST, propagate a
+# non-zero exit from it, and only then append our trailers — so our
+# trailers are the last word and cannot be clobbered by the repo hook
+# rewriting the message. (husky and friends set `core.hooksPath` LOCALLY,
+# and local config beats global, so those repos never reach this file.)
+#
+# NOT ADDED HERE: `Co-authored-by:`
+# ---------------------------------
+# Deliberate. `Co-authored-by: Claude ...` is the independent "a model
+# wrote this" marker that `scripts/check-agent-attribution-trailers.mjs`
+# keys off to decide whether the Switchroom-* trailers are REQUIRED. If
+# this hook emitted it too, then disabling the hook would remove the
+# marker and the evidence of its own absence in the same stroke, and CI
+# could never go red. Keeping the two producers separate (model writes
+# Co-authored-by, environment writes Switchroom-*) is what gives the CI
+# check teeth.
+#
+# Exit code: always 0 unless a chained repo hook failed. Attribution must
+# never be the reason a commit cannot be made.
+
+set -u
+
+MSG_FILE="${1-}"
+COMMIT_SOURCE="${2-}"
+
+[ -n "$MSG_FILE" ] || exit 0
+[ -f "$MSG_FILE" ] || exit 0
+
+# --- 1. Chain to the repo-local hook this one displaced -------------------
+_git_dir="$(git rev-parse --git-dir 2>/dev/null || true)"
+if [ -n "$_git_dir" ] && [ -x "$_git_dir/hooks/prepare-commit-msg" ]; then
+  "$_git_dir/hooks/prepare-commit-msg" "$@" || exit $?
+fi
+
+# --- 2. Are we in an agent container? -------------------------------------
+# `$SWITCHROOM_AGENT_NAME` is set by compose for every agent container and
+# by nothing else. Absent ⇒ a human at a laptop, or a stray copy of this
+# script: do nothing at all. A human commit must never be stamped with an
+# agent name it did not come from.
+AGENT="${SWITCHROOM_AGENT_NAME:-}"
+[ -n "$AGENT" ] || exit 0
+
+# `SWITCHROOM_ATTRIBUTION_DISABLE=1` is the documented, greppable escape
+# hatch for the rare case where a trailer genuinely must not be written
+# (e.g. replaying an upstream patch verbatim). It is intentionally an env
+# var and not a config knob, so its use is confined to one command.
+[ "${SWITCHROOM_ATTRIBUTION_DISABLE:-0}" = "1" ] && exit 0
+
+# --- 3. Resolve the model -------------------------------------------------
+# Chain, most specific first:
+#   1. SWITCHROOM_SESSION_MODEL — exported by start.sh immediately before
+#      `exec claude`, from the SAME `$_EFFECTIVE_MODEL` the session is
+#      actually launched on (so a `/model` override is reflected).
+#   2. SWITCHROOM_AGENT_MODEL   — the configured default, written into the
+#      container env by compose. Covers shells that did not inherit
+#      start.sh's export (a bare `docker exec`).
+#   3. "unknown" — never empty: a present-but-unknown trailer still proves
+#      the hook ran, and an empty trailer value would read as a bug.
+MODEL="${SWITCHROOM_SESSION_MODEL:-}"
+[ -n "$MODEL" ] || MODEL="${SWITCHROOM_AGENT_MODEL:-}"
+[ -n "$MODEL" ] || MODEL="unknown"
+
+# --- 4. Sanitize ----------------------------------------------------------
+# A git trailer value is single-line by definition. Both inputs are
+# operator-controlled rather than attacker-controlled, but they land
+# verbatim in an argv passed to `git interpret-trailers`, so clamp them to
+# a conservative charset and length regardless: anything else is replaced,
+# not silently dropped, so a malformed value is visible in the log rather
+# than mistaken for a different agent.
+_sanitize() {
+  printf '%s' "$1" \
+    | tr -d '\n\r' \
+    | sed 's/[^A-Za-z0-9._@:+/-]/_/g' \
+    | cut -c1-100
+}
+AGENT="$(_sanitize "$AGENT")"
+MODEL="$(_sanitize "$MODEL")"
+[ -n "$AGENT" ] || exit 0
+[ -n "$MODEL" ] || MODEL="unknown"
+
+# --- 5. Write the trailers ------------------------------------------------
+# `--if-exists doNothing` makes this idempotent: `--amend`, a rebase that
+# re-runs the hook, and a `-c` reuse all leave a single pair of trailers
+# rather than stacking duplicates.
+git interpret-trailers \
+  --in-place \
+  --if-exists doNothing \
+  --trailer "Switchroom-Agent: $AGENT" \
+  --trailer "Switchroom-Model: $MODEL" \
+  "$MSG_FILE" 2>/dev/null || true
+
+# COMMIT_SOURCE is unused today but is part of the hook contract; naming it
+# keeps `set -u` honest and documents that we deliberately stamp every
+# source (message, template, merge, squash, commit/amend).
+: "${COMMIT_SOURCE}"
+
+exit 0

package/dist/agent-scheduler/index.js

@@ -11114,10 +11114,37 @@ 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({
@@ -11320,7 +11347,10 @@ var LiteLLMConfigSchema = exports_external.object({
   admin_key: exports_external.string().optional().describe("LiteLLM master/admin key used at apply time to provision the team + " + "virtual key. Supports a vault reference (e.g. " + "'vault:litellm/master-key') — resolution happens at apply time via " + "the vault-broker. Never injected into the agent container."),
   team: exports_external.string().optional().describe("LiteLLM team alias the per-agent key is created under. Defaults to " + "'switchroom' (applied in code, not as a schema default)."),
   small_fast_model: exports_external.string().optional().describe("Model id exported as ANTHROPIC_SMALL_FAST_MODEL for the claude CLI's " + "background/fast lane, e.g. 'claude-haiku-4-5-20251001'."),
-  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins).")
+  tags: exports_external.record(exports_external.string(), exports_external.string()).optional().describe("Extra key/value metadata tags attached to the provisioned LiteLLM " + "virtual key. Merged per-key across cascade layers (agent wins)."),
+  max_budget: exports_external.number().positive().optional().describe("HARD spend cap in USD for this agent's virtual key over one " + "`budget_duration` window. LiteLLM refuses the request once the key's " + "tracked spend exceeds it, so a runaway loop costs at most this much " + "before it is stopped. Defaults to " + "DEFAULT_KEY_MAX_BUDGET_USD (see src/litellm/budget.ts) — deliberately " + "conservative; raise it per-agent rather than removing it. Set 0 or " + "omit `budget_duration` at your own risk: an uncapped key is only as " + "bounded as the upstream account balance."),
+  soft_budget: exports_external.number().positive().optional().describe("ADVISORY spend threshold in USD. LiteLLM keeps serving past it and " + "raises a budget alert instead. Must be < max_budget. NOTE: LiteLLM " + "accepts soft_budget only on POST /key/generate (GenerateKeyRequest); " + "UpdateKeyRequest does NOT carry it, so changing this value only takes " + "effect on a key that is (re)generated, not on an existing one."),
+  budget_duration: exports_external.string().regex(/^\d+(s|m|h|d|mo)$/, "budget_duration must be a LiteLLM duration like '30d', '24h', '1mo'").optional().describe("Rolling window the budget resets on, in LiteLLM duration syntax " + "('30d', '24h', '1mo'). Defaults to DEFAULT_KEY_BUDGET_DURATION. " + "WITHOUT a duration LiteLLM treats max_budget as a LIFETIME cap that " + "never resets — the key silently dies for good once it is hit.")
 }).optional().describe("LiteLLM routing config — opt-in per-agent virtual-key auto-provisioning " + "+ routing env. Default OFF. See LiteLLMConfigSchema doc for the full flow.");
 var HindsightPerOpLlmSchema = exports_external.object({
   model: exports_external.string().min(1).optional().describe("Per-op model (upstream `HINDSIGHT_API_<OP>_LLM_MODEL`). Absent → " + "inherit the global `hindsight.llm.model`."),
@@ -11330,6 +11360,7 @@ 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."),
@@ -11338,7 +11369,7 @@ var HindsightConfigSchema = exports_external.object({
     reflect: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `reflect` LLM op (synthesis / mental-model " + "refresh). Emits `HINDSIGHT_API_REFLECT_LLM_*`. Absent → uses global."),
     consolidation: HindsightPerOpLlmSchema.optional().describe("Per-op override for the `consolidation` LLM op (background memory " + "merge). Emits `HINDSIGHT_API_CONSOLIDATION_LLM_*`. Absent → global.")
   }).optional().describe("LLM knob for the hindsight container. The flat `provider`/`model` set " + "the global default (backward-compatible); optional `retain`/`reflect`/" + "`consolidation` blocks override individual ops. All fields optional; " + "unset fields fall back to the hard-coded defaults."),
-  env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_MAX_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
+  env: exports_external.record(exports_external.union([exports_external.string(), exports_external.number(), exports_external.boolean()])).optional().describe("Operator overrides for switchroom's capability-gated Hindsight " + "performance defaults. Only the keys switchroom actually manages are " + "honoured (`HINDSIGHT_PERF_ENV_KEYS` in " + "src/setup/hindsight-perf-defaults.ts: RERANKER_LOCAL_FP16, " + "RERANKER_LOCAL_BATCH_SIZE, LLM_MAX_CONCURRENT, " + "RETAIN/CONSOLIDATION_LLM_MAX_CONCURRENT, LLM_STRICT_SCHEMA, " + "LLM_MAX_RETRIES, CONSOLIDATION_LLM_PARALLELISM, " + "MAX_OBSERVATIONS_PER_SCOPE, " + "RECALL_MAX_CANDIDATES_PER_SOURCE, LINK_EXPANSION_PER_ENTITY_LIMIT, " + "LINK_EXPANSION_TIMEOUT, LLM_REASONING_EFFORT, " + "RERANKER_LOCAL_BUCKET_BATCHING, RERANKER_MAX_CANDIDATES, " + "RERANKER_LOCAL_MAX_CONCURRENT, RECALL_MAX_CONCURRENT, " + "REFLECT_WALL_TIMEOUT, WORKER_CONSOLIDATION_MAX_SLOTS, " + "WORKER_CONSOLIDATION_SLOT_LIMIT, " + "CONSOLIDATION_MAX_MEMORIES_PER_ROUND, RECENCY_DECAY_FUNCTION, " + "RECENCY_DECAY_HALFLIFE_DAYS — switchroom defaults recall's recency " + "curve to `exponential` with a 30-day half-life so a fact retained " + "today outranks a stale one, instead of upstream's near-flat " + "linear/365-day window), the override-only keys " + "switchroom manages but ships NO default for " + "(`HINDSIGHT_PERF_OVERRIDE_ONLY_KEYS`: " + "HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY — a per-deployment " + "`bank-pattern:priority,...` map; unset means upstream's flat " + "created_at FIFO across banks; and " + "HINDSIGHT_CE_DECISIVE_RELATIVE_GAP — the rollback knob for " + "switchroom's CE-saturation damping patch, a float; >= ~0.65 backs the " + "damping out entirely, unset means the patch's own derived gap; and " + "HINDSIGHT_API_RECENCY_DECAY_LINEAR_WINDOW_DAYS — only read when the " + "decay function is `linear`, so switchroom ships no default for it but " + "still honours an operator who flips the function back; and " + "HINDSIGHT_API_WORKER_MAX_SLOTS — the worker poller's TOTAL in-flight " + "task budget, the pool WORKER_CONSOLIDATION_MAX_SLOTS reserves out of; " + "unset means upstream's own default), plus the " + "embedded-PostgreSQL (pg0) sizing keys switchroom manages in " + "src/setup/hindsight-pg-defaults.ts (`HINDSIGHT_PG_ENV_KEYS`: " + "SWITCHROOM_HINDSIGHT_PG_EFFECTIVE_CACHE_SIZE, " + "SWITCHROOM_HINDSIGHT_PG_SHARED_BUFFERS — a postgres size string such " + "as `4GB`, or the sentinel `off` to leave pg0's own default for that " + "one knob). A value set here " + "REPLACES switchroom's default and is emitted even when the gating " + "capability is absent, so an operator can always force a knob. Other " + "`HINDSIGHT_API_*` keys are deliberately IGNORED — a blanket " + "passthrough would collide with the vars startHindsight() derives " + "itself (HINDSIGHT_API_PORT, the retain token/deadline budget).")
 });
 var MicrosoftWorkspaceConfigSchema = exports_external.object({
   microsoft_client_id: exports_external.string().min(1).optional().describe("Microsoft OAuth application (client) ID from Entra portal " + "(literal string or vault reference e.g. " + "'vault:microsoft-oauth-client-id'). OPTIONAL — omit it to use " + "switchroom's shipped default Microsoft app (zero-config). " + "Set it only to bring your own Entra app (BYO)."),
@@ -11481,6 +11512,33 @@ 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()