@rulemetric/hooks 0.11.0 → 0.12.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rulemetric/hooks",
-  "version": "0.11.0",
+  "version": "0.12.0",
   "publishConfig": {
     "access": "public"
   },

package/scripts/session-end.sh

@@ -121,10 +121,29 @@ fi
   "$AUTH_HEADER" "{\"sessionId\":\"$RW_SESSION_ID\"}" "session-end-link" 8 > /dev/null 2>&1) &
 
 # ── Auto-enrich from local session-meta ──
+#
+# 🚨 NOTHING BELOW THIS LINE CAN BE RELIED ON TO RUN. Claude Code cancels this
+# hook shortly after it starts ("SessionEnd hook … failed: Hook cancelled",
+# reproduced 2026-08-21 with `claude -p --debug`): the reimport above lands,
+# and the script is killed before it reaches even the `success` log at line 93.
+# Measured the same day: 0 `session-end` entries in $TMPDIR/rulemetric/hook.log
+# across 24 session-starts, while 59 of the last 100 claude_code sessions had
+# been reimported and 100 of 100 had ended_at.
+#
+# So enrichment is NO LONGER this block's responsibility. The API derives it
+# server-side at the finalization points this hook does reach — the reimport
+# handler (finalizeSessionIntelligence) and /close. See
+# apps/api/src/lib/sessions/derive-enrichment.ts. What is left here is the
+# PATCH /enrich path, which carries fields the server cannot derive (real
+# interruption counts, per-message hour histograms, response times) and which
+# the server's derivation refuses to overwrite. Treat it as a best-effort
+# upgrade of already-correct data, never as the thing that makes the dashboard
+# non-zero. Do not move new must-run work below here.
+#
 # For Claude Code: read local session-meta file when present; nothing writes
 # that dir anymore, so when it's missing fall back to deriving enrichment
-# from the reimported events (claude_code enrichment was silently dead
-# without this — with_meta dropped to ~0 from April on).
+# from the reimported events (redundant with the server-side derive, kept as
+# the path for an API old enough not to have it).
 # For other tools: derive enrichment from stored session events via API
 if [ "$HOOK_TOOL" = "claude_code" ]; then
   SESSION_META_FILE="$HOME/.claude/usage-data/session-meta/$HOOK_SESSION_ID.json"

package/scripts/session-start.sh

@@ -330,6 +330,60 @@ _rulemetric_suggest_instructions || true
 # traffic), so the exposure is BUFFERED by this harness session id and reconciled at
 # session close (reimport). All failures are silent — this block MUST NOT prevent
 # session start. Same 4s read / 3s write timeouts as the suggestion block above.
+
+# ── The review expectation ───────────────────────────────────────────────────
+# Human labeling is the only step of the loop no machine can take: an LLM
+# labeling anchors to calibrate an LLM judge is circular. Because every other
+# step runs unattended, this is the step that silently stops happening — and on
+# 2026-08-14..18 it did, while the evolution loop measured nightly and refused
+# to decide because the runs it needed sat unreviewed.
+#
+# A `labels_needed` notification already fires and is correctly deduped. It was
+# not dark, it was DROWNING: 20 of them against 522 usage alerts in one bell.
+# So the ask is placed where it cannot be scrolled past — the top of the session
+# the person is already starting.
+#
+# Prints NOTHING unless the loop is genuinely blocked (`blocking`), which the
+# API decides by requiring an ENROLLED target that is both short of the pilot
+# floor and has runs left to review. A nudge that fires when labeling would
+# change no verdict is the nudge that teaches you to ignore nudges.
+_rulemetric_review_gate() {
+  local api_url="${RULEMETRIC_API_URL:-}"
+  local auth_token="${RULEMETRIC_API_KEY:-${RULEMETRIC_ACCESS_TOKEN:-}}"
+
+  [ -n "$api_url" ] && [ -n "$auth_token" ] || return 0
+  command -v jq >/dev/null 2>&1 || return 0
+
+  local response
+  # Bounded like every other call here: a slow or absent API must cost a session
+  # start nothing it would not already cost.
+  response="$(curl -s --noproxy '*' --max-time 3 \
+    "${api_url}/api/loop/review-gate" \
+    -H "Authorization: Bearer ${auth_token}" 2>/dev/null)" || return 0
+
+  local blocking total
+  blocking="$(printf '%s' "$response" | jq -r '.blocking // false' 2>/dev/null)" || return 0
+  [ "$blocking" = "true" ] || return 0
+  total="$(printf '%s' "$response" | jq -r '.totalAwaitingReview // 0' 2>/dev/null)" || return 0
+  [ "$total" != "0" ] && [ -n "$total" ] || return 0
+
+  printf '\n---\n'
+  printf '[rulemetric] THE LOOP IS WAITING ON YOU — %s run(s) need human review.\n' "$total"
+  printf 'Nothing else can do this step, and until it is done the loop keeps measuring and keeps refusing to decide.\n'
+
+  # One line per target: what is unreviewed, how far the anchor set is from the
+  # pilot floor, and the cohort split — the gate is per cohort, so a total alone
+  # cannot say which arm is short.
+  printf '%s' "$response" | jq -r '
+    .targets[]? |
+    "  - \(.name)\n" +
+    "    \(.awaitingReview) unreviewed · \(.humanLabeled)/\(.pilotTarget) labels " +
+    "(\(.humanLabeledWith) with, \(.humanLabeledWithout) without) · " +
+    "\(.labelsToPilotFloor) more to reach the floor\n" +
+    "    Review: \(env.RULEMETRIC_WEB_URL // "http://localhost:5174")/eval-targets/\(.id)?tab=review"
+  ' 2>/dev/null || true
+}
+
 # Mirrors _rulemetric_suggest_instructions() deliberately.
 _rulemetric_inject_memories() {
   local api_url="${RULEMETRIC_API_URL:-}"
@@ -433,3 +487,7 @@ _rulemetric_inject_memories() {
 }
 
 _rulemetric_inject_memories || true
+
+# Last, and therefore closest to the prompt: the one item on this block that is
+# a request rather than context.
+_rulemetric_review_gate || true