arkaos 2.64.0 → 2.66.0

package/VERSION

@@ -1 +1 @@
-2.64.0
+2.66.0

package/config/hooks/stop.sh

@@ -19,12 +19,25 @@ SESSION_ID=""
 TRANSCRIPT_PATH=""
 STOP_HOOK_ACTIVE=""
 CWD=""
+EFFORT_LEVEL=""
 if command -v jq &>/dev/null; then
   SESSION_ID=$(echo "$input" | jq -r '.session_id // ""' 2>/dev/null)
   TRANSCRIPT_PATH=$(echo "$input" | jq -r '.transcript_path // ""' 2>/dev/null)
   STOP_HOOK_ACTIVE=$(echo "$input" | jq -r '.stop_hook_active // ""' 2>/dev/null)
   CWD=$(echo "$input" | jq -r '.cwd // ""' 2>/dev/null)
+  # PR46 v2.65.0 — Claude Code W19 ships effort.level in hook stdin and
+  # $CLAUDE_EFFORT env var. Soft-block checks (kb-cite, meta-tag) only
+  # run at high|xhigh; hard enforcement runs regardless.
+  EFFORT_LEVEL=$(echo "$input" | jq -r '.effort.level // ""' 2>/dev/null)
 fi
+# Fallback to env var if stdin didn't carry it
+[ -z "$EFFORT_LEVEL" ] && EFFORT_LEVEL="${CLAUDE_EFFORT:-}"
+
+# Telemetry-only signal. Soft-block checks (kb_cite, meta_tag, sycophancy)
+# always run here because they're cheap and feed /arka compliance.
+# What is effort-gated is the NUDGE SURFACING in user-prompt-submit.sh
+# (whether the next turn sees a [arka:suggest] line). Record the level
+# on the telemetry row so we can later analyze suppression rates.
 
 # Prevent infinite loops when Stop hook was triggered by its own decision.
 if [ "$STOP_HOOK_ACTIVE" = "true" ]; then
@@ -60,6 +73,7 @@ SESSION_ID_VAL="$SESSION_ID" \
 TRANSCRIPT_PATH_VAL="$TRANSCRIPT_PATH" \
 CWD_VAL="$CWD" \
 ARKAOS_ROOT_VAL="$ARKAOS_ROOT" \
+EFFORT_LEVEL_VAL="$EFFORT_LEVEL" \
 python3 - <<'PY' 2>/dev/null
 import json
 import os
@@ -223,6 +237,9 @@ entry = {
     "kb_cite_topic_score": cite_topic_score,
     "meta_tag_check_passed": meta_passed,
     "meta_tag_check_reason": meta_reason,
+    # PR46 v2.65.0 — Claude Code effort level captured for later analysis
+    # of nudge-suppression rates. Unset / unknown values land as "".
+    "effort_level": os.environ.get("EFFORT_LEVEL_VAL", ""),
     "mode": "warn",
 }
 

package/config/hooks/user-prompt-submit.sh

@@ -84,10 +84,25 @@ mkdir -p "$CACHE_DIR" 2>/dev/null
 # ─── Extract user input from hook JSON ───────────────────────────────────
 user_input=""
 SESSION_ID=""
+EFFORT_LEVEL=""
 if command -v jq &>/dev/null; then
   user_input=$(echo "$input" | jq -r '.userInput // .message // ""' 2>/dev/null)
   SESSION_ID=$(echo "$input" | jq -r '.session_id // ""' 2>/dev/null)
+  # PR46 v2.65.0 — Claude Code W19 ships effort.level in hook stdin.
+  # Soft-block nudges (KB-first + meta-tag) are gated by effort: only
+  # surfaced at high|xhigh; low/medium skip the nudge to avoid forcing
+  # the model to comply with full contracts during cheap exploratory
+  # turns. Hard enforcement (PreToolUse flow_enforcer) runs regardless.
+  EFFORT_LEVEL=$(echo "$input" | jq -r '.effort.level // ""' 2>/dev/null)
 fi
+[ -z "$EFFORT_LEVEL" ] && EFFORT_LEVEL="${CLAUDE_EFFORT:-}"
+
+# Decide whether soft-block nudges surface to the next turn.
+_ARKA_SURFACE_NUDGES="true"
+case "${EFFORT_LEVEL:-high}" in
+  low|medium) _ARKA_SURFACE_NUDGES="false" ;;
+  *)          _ARKA_SURFACE_NUDGES="true"  ;;
+esac
 
 # ─── Flow marker cache invalidation (v2 — new turn, reset ALLOW cache) ──
 # Cheap, non-blocking, runs before Synapse so a stuck Python later cannot
@@ -376,7 +391,7 @@ fi
 # the suggestion to the model in this turn's additionalContext. One-shot:
 # the file is deleted after read so the nudge does not repeat across turns.
 _KB_CITE_NUDGE=""
-if [ -n "$SESSION_ID" ]; then
+if [ -n "$SESSION_ID" ] && [ "$_ARKA_SURFACE_NUDGES" = "true" ]; then
   _CITE_FILE="/tmp/arkaos-cite/${SESSION_ID}.json"
   if [ -f "$_CITE_FILE" ]; then
     if command -v jq &>/dev/null; then
@@ -397,7 +412,7 @@ fi
 # Mirror of the KB citation nudge but for the [arka:meta] one-liner
 # contract. One-shot; deleted after read.
 _META_TAG_NUDGE=""
-if [ -n "$SESSION_ID" ]; then
+if [ -n "$SESSION_ID" ] && [ "$_ARKA_SURFACE_NUDGES" = "true" ]; then
   _META_FILE="/tmp/arkaos-meta/${SESSION_ID}.json"
   if [ -f "$_META_FILE" ]; then
     if command -v jq &>/dev/null; then

package/core/runtime/llm_cost_telemetry.py

@@ -66,9 +66,15 @@ def record_cost(
     tokens_out: int,
     cached_tokens: int,
     estimated_cost_usd: float | None,
+    category: str = "",
 ) -> None:
     """Append one JSONL line describing an LLM call's cost.
 
+    `category` mirrors Claude Code v2.1.149's per-category usage
+    breakdown: ``"skill:<slug>"``, ``"subagent:<dept>"``,
+    ``"plugin:<id>"``, ``"mcp:<server>"``, or ``""`` for base usage.
+    Free-form string — the aggregator groups whatever it sees.
+
     Silently swallows all errors. Telemetry must never break a
     completion call. The caller decides whether to compute the cost via
     `core.runtime.pricing.estimate_cost_usd` or pass None.
@@ -87,6 +93,7 @@ def record_cost(
                 if estimated_cost_usd is not None
                 else None
             ),
+            "category": str(category or ""),
         }
         with _locked_append(_telemetry_path()) as fh:
             fh.write(json.dumps(entry, ensure_ascii=False) + "\n")
@@ -133,6 +140,7 @@ class CostSummary:
     call_count: int
     by_provider: dict[str, dict[str, Any]] = field(default_factory=dict)
     by_model: dict[str, dict[str, Any]] = field(default_factory=dict)
+    by_category: dict[str, dict[str, Any]] = field(default_factory=dict)
     by_session: list[dict[str, Any]] = field(default_factory=list)
     advisories: list[str] = field(default_factory=list)
     corrupt_line_count: int = 0
@@ -291,6 +299,7 @@ def summarise(
         call_count=finalised["call_count"],
         by_provider=_group(entries, "provider"),
         by_model=_group(entries, "model"),
+        by_category=_group(entries, "category"),
         by_session=sessions,
         advisories=_build_advisories(sessions, advisory_threshold_usd),
         corrupt_line_count=corrupt,

package/core/runtime/llm_cost_telemetry_cli.py

@@ -89,6 +89,12 @@ def _render_sessions(rows: list[dict[str, Any]], title: str) -> list[str]:
     return lines
 
 
+def _has_category_data(group: dict[str, dict[str, Any]]) -> bool:
+    # The summariser always returns at least the "" bucket for legacy
+    # rows. Hide the section unless ≥ 1 row has a non-empty key.
+    return any(k.strip() for k in group.keys())
+
+
 def _render_advisories(advisories: list[str]) -> list[str]:
     if not advisories:
         return []
@@ -101,6 +107,11 @@ def _format_summary(summary: CostSummary) -> str:
     parts.append("")
     parts.extend(_render_group("By provider", summary.by_provider))
     parts.extend(_render_group("By model", summary.by_model))
+    # Per-category breakdown (Claude Code v2.1.149+): skill, subagent,
+    # plugin, mcp-server. Renders only when at least one categorised
+    # entry exists so old telemetry doesn't show an empty section.
+    if _has_category_data(summary.by_category):
+        parts.extend(_render_group("By category", summary.by_category))
     parts.extend(_render_sessions(summary.by_session, "Top 10 sessions"))
     parts.extend(_render_advisories(summary.advisories))
     if summary.corrupt_line_count:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "arkaos",
-  "version": "2.64.0",
+  "version": "2.66.0",
   "description": "The Operating System for AI Agent Teams",
   "type": "module",
   "bin": {

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "arkaos-core"
-version = "2.64.0"
+version = "2.66.0"
 description = "Core engine for ArkaOS — The Operating System for AI Agent Teams"
 readme = "README.md"
 license = {text = "MIT"}