chore(deps): refresh memory, liter-llm, and knowledge pins

config/prometheus-exec-binary.json

@@ -6,6 +6,6 @@
   "sourcePath": "crates/prometheus-exec/target/release/prometheus-exec",
   "installPath": "~/.local/bin/prometheus-exec",
   "expectedVersion": "prometheus-exec 1.7.0",
-  "expectedBuildSha256": "a67de3b6c4f41a471e00bcfb683de5afa3164ddbf96af6b1ed4e434d9d079b56",
+  "expectedBuildSha256": "97c6b5bdb853f2db9aaabe918d9e09beecf6ac2141705e0b9181e0e19fe5958f",
   "installReceipt": "~/.prometheus/install/manifests/prometheus-exec.json"
 }

dist/plugins/claude/prometheus-skill-pack/shared/harnesses/generated/release-manifest.json

@@ -52,7 +52,7 @@
     },
     {
       "path": "shared/scripts/enqueue-learning-job.py",
-      "sha256": "ead5ceb1c815c154500415b00c80722e18bd4df86e1f0ced8887c861e27ce28c",
+      "sha256": "fff1f6132b40bdeb82033528de3e52611cd6ca90adfeeea52dd73930ace61e05",
       "mode": "0755"
     },
     {
@@ -316,5 +316,5 @@
       "mode": "0755"
     }
   ],
-  "bundleId": "1e45626629cdce87b6cced49acc1bb5efbfbf0076b6eeb8fbd5d4fa25a202d68"
+  "bundleId": "7cf81b78dcdc1cce7734d4d6cdc3db0629a10edf622ffd91bee1e9437e97f391"
 }

dist/plugins/claude/prometheus-skill-pack/shared/launchagents/ai.prometheus.surreal-memory-native.plist

@@ -55,8 +55,22 @@
     <string>local</string>
     <key>LOCAL_EMBEDDING_MODEL</key>
     <string>BAAI/bge-small-en-v1.5</string>
+    <key>LOCAL_EMBEDDING_MODEL_REVISION</key>
+    <string>5c38ec7c405ec4b44b94cc5a9bb96e735b38267a</string>
+    <key>LOCAL_EMBEDDING_DIMENSIONS</key>
+    <string>384</string>
+    <key>LOCAL_EMBEDDING_BACKEND</key>
+    <string>__LOCAL_EMBEDDING_BACKEND__</string>
+    <key>LOCAL_EMBEDDING_DEVICE</key>
+    <string>__LOCAL_EMBEDDING_DEVICE__</string>
+    <key>LOCAL_EMBEDDING_EXECUTOR</key>
+    <string>__SURREAL_MLX_EXECUTOR__</string>
     <key>MODEL_CACHE_DIR</key>
     <string>__PROMETHEUS_HOME__/.cache/huggingface</string>
+    <key>HF_HUB_CACHE</key>
+    <string>__PROMETHEUS_HOME__/.cache/huggingface/hub</string>
+    <key>SURREAL_EXECUTOR_STARTUP_MS</key>
+    <string>120000</string>
 
     <!-- Retry / reconnection config -->
     <key>SURREAL_MAX_CONNECT_RETRIES</key>

dist/plugins/claude/prometheus-skill-pack/shared/scripts/enqueue-learning-job.py

@@ -69,9 +69,25 @@ def main() -> int:
         payload, "session_id", "sessionId", "conversation_id", "conversationId"
     ) or os.environ.get("CLAUDE_SESSION_ID") or os.environ.get("CODEX_THREAD_ID") or "unknown"
     transcript_path = first(payload, "transcript_path", "transcriptPath")
+    # Transcript length is the turn cursor: it advances once per completed turn
+    # and is identical for the twin hooks that fire for the same turn.
+    try:
+        transcript_bytes = str(Path(transcript_path).stat().st_size) if transcript_path else ""
+    except OSError:
+        transcript_bytes = ""
     payload_digest = hashlib.sha256(raw).hexdigest()
+    # ~keep The turn identity deliberately EXCLUDES `event` and `payload_digest`.
+    # Claude Code fires both `stop` and `subagent-executor-karpathy-learning`
+    # (`executor_complete`) for the same turn — measured 549 vs 548 across all
+    # recorded sessions, i.e. ~1:1 on ordinary turns with no subagent involved —
+    # and their payloads are structurally identical (same keys, sessionId,
+    # transcriptPath), differing only in the event name. Including either field
+    # made every turn unique, so the queue emitted one wiki record per hook per
+    # turn. Keying on the transcript position instead collapses the twins and
+    # still distinguishes genuinely different turns, because `transcript_bytes`
+    # grows monotonically as the session advances.
     identity = "\0".join(
-        (harness, event, session_id, str(project_root), transcript_path, payload_digest)
+        (harness, session_id, str(project_root), transcript_path, transcript_bytes)
     ).encode()
     event_id = hashlib.sha256(identity).hexdigest()
 

dist/plugins/claude/prometheus-skill-pack/shared/scripts/tests/test-karpathy-dispatch.sh

@@ -102,5 +102,38 @@ else
   fail "duplicate delivery is idempotent across durable states"
 fi
 
+# Revert line: restore `event` or `payload_digest` to the `identity` tuple in
+# enqueue-learning-job.py to make this test fail.
+# ~keep Claude Code fires stop AND executor_complete for the SAME turn with
+# structurally identical payloads (measured ~1:1, 549 vs 548, across all
+# recorded sessions). Both must collapse to one job, or the worker writes two
+# near-identical wiki records per turn.
+mkdir -p "$TMP/twin"
+printf 'line one\n' > "$TMP/twin-transcript.jsonl"
+twin_payload="$(printf '{"session_id":"twin-session","cwd":"%s","transcript_path":"%s"}' \
+  "$TMP/project" "$TMP/twin-transcript.jsonl")"
+for ev in stop executor_complete; do
+  printf '%s' "$twin_payload" | HOME="$TMP/home" PATH="$TMP/bin:$PATH" \
+    PROMETHEUS_LEARNING_QUEUE="$TMP/twin" "$DISPATCH" "$ev" claude-code
+done
+twin_jobs="$(find "$TMP/twin/pending" -type f -name '*.json' | wc -l | tr -d ' ')"
+if [ "$twin_jobs" = 1 ]; then
+  pass "stop and executor_complete for one turn collapse to a single job"
+else
+  fail "stop and executor_complete for one turn collapse to a single job (got $twin_jobs)"
+fi
+
+# Revert line: drop `transcript_bytes` from the `identity` tuple to make this
+# test fail (all turns would collapse into one job).
+printf 'line two\n' >> "$TMP/twin-transcript.jsonl"
+printf '%s' "$twin_payload" | HOME="$TMP/home" PATH="$TMP/bin:$PATH" \
+  PROMETHEUS_LEARNING_QUEUE="$TMP/twin" "$DISPATCH" stop claude-code
+next_jobs="$(find "$TMP/twin/pending" -type f -name '*.json' | wc -l | tr -d ' ')"
+if [ "$next_jobs" = 2 ]; then
+  pass "a genuinely later turn still enqueues its own job"
+else
+  fail "a genuinely later turn still enqueues its own job (got $next_jobs)"
+fi
+
 printf '\nPASS=%d FAIL=%d\n' "$PASS" "$FAIL"
 [ "$FAIL" -eq 0 ]