switchroom 0.13.21 → 0.13.23

package/dist/cli/switchroom.js

@@ -47331,8 +47331,8 @@ var {
 } = import__.default;
 
 // src/build-info.ts
-var VERSION = "0.13.21";
-var COMMIT_SHA = "4183c5df";
+var VERSION = "0.13.23";
+var COMMIT_SHA = "c4a29251";
 
 // src/cli/agent.ts
 init_source();
@@ -48699,6 +48699,8 @@ function applyHindsightSettingsOverrides(pluginDestPath) {
     return;
   }
   settings.retainEveryNTurns = 1;
+  settings.recallMaxMemories = 8;
+  settings.recallMinOverlap = 0.1;
   writeFileSync5(settingsPath, JSON.stringify(settings, null, 2) + `
 `, "utf-8");
 }
@@ -62230,6 +62232,13 @@ var HINDSIGHT_IMAGE = "ghcr.io/switchroom/switchroom-hindsight:latest";
 var HINDSIGHT_DEFAULT_MODEL = "claude-sonnet-4-6";
 var HINDSIGHT_DEFAULT_MCP_STATELESS = true;
 var HINDSIGHT_BROKER_SOCK_VOLUME = `auth-broker-${HINDSIGHT_CONSUMER_NAME}-sock`;
+var HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING = "true";
+var HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES = 150;
+var HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4;
+var HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT = 8;
+var HINDSIGHT_DEFAULT_MEM_LIMIT = "4g";
+var HINDSIGHT_DEFAULT_MEM_RESERVATION = "2g";
+var HINDSIGHT_DEFAULT_PIDS_LIMIT = 1000;
 function isPortFree(port) {
   return new Promise((resolve25) => {
     const server = createServer3();
@@ -62300,7 +62309,15 @@ function startHindsight(ports) {
     "-e",
     `HINDSIGHT_API_LLM_MODEL=${HINDSIGHT_DEFAULT_MODEL}`,
     "-e",
-    `HINDSIGHT_API_MCP_STATELESS=${HINDSIGHT_DEFAULT_MCP_STATELESS}`
+    `HINDSIGHT_API_MCP_STATELESS=${HINDSIGHT_DEFAULT_MCP_STATELESS}`,
+    "-e",
+    `HINDSIGHT_API_RERANKER_LOCAL_BUCKET_BATCHING=${HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING}`,
+    "-e",
+    `HINDSIGHT_API_RERANKER_MAX_CANDIDATES=${HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES}`,
+    "-e",
+    `HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT}`,
+    "-e",
+    `HINDSIGHT_API_RECALL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT}`
   ];
   const args = [
     "run",
@@ -62309,6 +62326,9 @@ function startHindsight(ports) {
     "switchroom-hindsight",
     "--restart",
     "unless-stopped",
+    `--memory=${HINDSIGHT_DEFAULT_MEM_LIMIT}`,
+    `--memory-reservation=${HINDSIGHT_DEFAULT_MEM_RESERVATION}`,
+    `--pids-limit=${HINDSIGHT_DEFAULT_PIDS_LIMIT}`,
     "-p",
     `127.0.0.1:${apiPort}:8888`,
     "-p",
@@ -62355,6 +62375,13 @@ function generateHindsightComposeSnippet() {
     "      - HINDSIGHT_API_LLM_PROVIDER=claude-code",
     `      - HINDSIGHT_API_LLM_MODEL=${HINDSIGHT_DEFAULT_MODEL}`,
     `      - HINDSIGHT_API_MCP_STATELESS=${HINDSIGHT_DEFAULT_MCP_STATELESS}`,
+    `      - HINDSIGHT_API_RERANKER_LOCAL_BUCKET_BATCHING=${HINDSIGHT_DEFAULT_RERANKER_BUCKET_BATCHING}`,
+    `      - HINDSIGHT_API_RERANKER_MAX_CANDIDATES=${HINDSIGHT_DEFAULT_RERANKER_MAX_CANDIDATES}`,
+    `      - HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT}`,
+    `      - HINDSIGHT_API_RECALL_MAX_CONCURRENT=${HINDSIGHT_DEFAULT_RECALL_MAX_CONCURRENT}`,
+    `    mem_limit: ${HINDSIGHT_DEFAULT_MEM_LIMIT}`,
+    `    mem_reservation: ${HINDSIGHT_DEFAULT_MEM_RESERVATION}`,
+    `    pids_limit: ${HINDSIGHT_DEFAULT_PIDS_LIMIT}`,
     "    volumes:",
     "      - switchroom-hindsight-data:/home/hindsight/.pg0",
     `      - ${HINDSIGHT_BROKER_SOCK_VOLUME}:/run/switchroom/auth-broker`,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "switchroom",
-  "version": "0.13.21",
+  "version": "0.13.23",
   "description": "Run Claude Code 24/7 on your Claude Pro/Max subscription over Telegram. Open-source alternative to OpenClaw and NanoClaw — no API keys.",
   "type": "module",
   "bin": {

package/profiles/_base/start.sh.hbs

@@ -207,14 +207,15 @@ export HINDSIGHT_AGENT_NAME="{{name}}"
 export HINDSIGHT_RECALL_MAX_MEMORIES={{hindsightRecallMaxMemories}}
 {{/if}}
 # Per-session recall cache (#424 phase 4.1). Identical (prompt, bank)
-# within a session reuses the prior recall result instead of round-
-# tripping to Hindsight. Defaults to 600s (10 min); operators can
-# override via memory.recall.cache_ttl_secs in switchroom.yaml.
-# Setting it to 0 disables caching for that agent.
+# within a session reuses the prior recall result. v0.13.22: the
+# unconditional default-600s export was dropped — the 2026-05-24
+# audit measured 0% hit rate across 430+ Telegram turns (cache key
+# is `(session_id, prompt, bank, extra_banks)`; each Telegram inbound
+# is unique). Operators who run Claude Code interactively (where
+# prompt-repeat happens) can still opt in via memory.recall.cache_ttl_secs
+# in switchroom.yaml. Setting it to 0 disables caching for that agent.
 {{#if (isNumber hindsightRecallCacheTtlSecs)}}
 export HINDSIGHT_RECALL_CACHE_TTL_SECS={{hindsightRecallCacheTtlSecs}}
-{{else}}
-export HINDSIGHT_RECALL_CACHE_TTL_SECS=600
 {{/if}}
 # Lexical-overlap relevance gate (#475). Drops memories whose Jaccard
 # overlap with the user's query is below this threshold (range 0.0–1.0).

package/telegram-plugin/dist/gateway/gateway.js

@@ -48315,10 +48315,10 @@ function sweepStaleTurnActiveMarker(stateDir, opts) {
 }
 
 // ../src/build-info.ts
-var VERSION = "0.13.21";
-var COMMIT_SHA = "4183c5df";
-var COMMIT_DATE = "2026-05-23T21:59:16Z";
-var LATEST_PR = 1692;
+var VERSION = "0.13.23";
+var COMMIT_SHA = "c4a29251";
+var COMMIT_DATE = "2026-05-23T23:54:40Z";
+var LATEST_PR = 1702;
 var COMMITS_AHEAD_OF_TAG = 0;
 
 // gateway/boot-version.ts

package/vendor/hindsight-memory/scripts/recall.py

@@ -556,7 +556,15 @@ def main():
             max_tokens=config.get("recallMaxTokens", 1024),
             budget=config.get("recallBudget", "mid"),
             types=config.get("recallTypes"),
-            timeout=10,
+            # 8s in-script timeout leaves 4s headroom inside the 12s
+            # UserPromptSubmit hook ceiling (see hooks.json:20) for cache
+            # write + block formatting. Tightened from 10s in switchroom
+            # v0.13.22: the 2026-05-24 audit showed 17-26% of turns
+            # breaching the 12s hook timeout on heavy agents (finn /
+            # gymbro / klanker), which dropped the recall entirely; an
+            # earlier-hard-timeout failure returns cleanly with no
+            # memories instead of blowing past the hook ceiling.
+            timeout=8,
         )
         results = response.get("results", [])
     except Exception as e:
@@ -577,7 +585,15 @@ def main():
                 max_tokens=config.get("recallMaxTokens", 1024),
                 budget=config.get("recallBudget", "mid"),
                 types=config.get("recallTypes"),
-                timeout=10,
+                # 8s in-script timeout leaves 4s headroom inside the 12s
+                # UserPromptSubmit hook ceiling (see hooks.json:20) for cache
+                # write + block formatting. Tightened from 10s in switchroom
+                # v0.13.22: the 2026-05-24 audit showed 17-26% of turns
+                # breaching the 12s hook timeout on heavy agents (finn /
+                # gymbro / klanker), which dropped the recall entirely; an
+                # earlier-hard-timeout failure returns cleanly with no
+                # memories instead of blowing past the hook ceiling.
+                timeout=8,
             )
             extra_results = extra_response.get("results", [])
             if extra_results: