arkaos 2.75.0 → 2.77.0

package/VERSION

@@ -1 +1 @@
-2.75.0
+2.77.0

package/config/hooks/stop.sh

@@ -219,6 +219,38 @@ try:
 except Exception:
     pass
 
+# PR59 v2.76.0 — Closing-marker soft block. Telemetry analysis showed
+# 0% [arka:phase:13]/[arka:trivial] rate on flow-required turns. Persist
+# result to /tmp/arkaos-closing/<session>.json so the next
+# UserPromptSubmit can surface a nudge if missing.
+closing_check_passed = True
+closing_check_reason = "trivial"
+closing_check_suggestion: str | None = None
+try:
+    from core.governance.closing_marker_check import check_closing_marker
+    cmr = check_closing_marker(last)
+    closing_check_passed = cmr.passed
+    closing_check_reason = cmr.reason
+    closing_check_suggestion = cmr.suggestion
+    if safe_sid:
+        prev_umask = os.umask(0o077)
+        try:
+            closing_dir = Path("/tmp/arkaos-closing")
+            closing_dir.mkdir(parents=True, exist_ok=True)
+            closing_path = closing_dir / f"{safe_sid}.json"
+            closing_path.write_text(
+                json.dumps({
+                    "passed": cmr.passed,
+                    "reason": cmr.reason,
+                    "suggestion": cmr.suggestion,
+                }),
+                encoding="utf-8",
+            )
+        finally:
+            os.umask(prev_umask)
+except Exception:
+    pass
+
 entry = {
     "ts": datetime.now(timezone.utc).isoformat(),
     "session_id": session_id,
@@ -237,6 +269,9 @@ entry = {
     "kb_cite_topic_score": cite_topic_score,
     "meta_tag_check_passed": meta_passed,
     "meta_tag_check_reason": meta_reason,
+    # PR59 v2.76.0 — Closing-marker soft-block telemetry.
+    "closing_marker_check_passed": closing_check_passed,
+    "closing_marker_check_reason": closing_check_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", ""),

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

@@ -426,6 +426,24 @@ if [ -n "$SESSION_ID" ] && [ "$_ARKA_SURFACE_NUDGES" = "true" ]; then
   fi
 fi
 
+# ─── Closing-marker nudge (PR59 v2.76.0) ─────────────────────────────────
+# Mirror of meta-tag nudge but for [arka:phase:13] / [arka:trivial]
+# closing markers. One-shot; deleted after read.
+_CLOSING_MARKER_NUDGE=""
+if [ -n "$SESSION_ID" ] && [ "$_ARKA_SURFACE_NUDGES" = "true" ]; then
+  _CLOSING_FILE="/tmp/arkaos-closing/${SESSION_ID}.json"
+  if [ -f "$_CLOSING_FILE" ]; then
+    if command -v jq &>/dev/null; then
+      _CLOSING_PASSED=$(jq -r '.passed' "$_CLOSING_FILE" 2>/dev/null)
+      _CLOSING_SUGGEST=$(jq -r '.suggestion // ""' "$_CLOSING_FILE" 2>/dev/null)
+      if [ "$_CLOSING_PASSED" = "false" ] && [ -n "$_CLOSING_SUGGEST" ] && [ "$_CLOSING_SUGGEST" != "null" ]; then
+        _CLOSING_MARKER_NUDGE="[arka:suggest] ${_CLOSING_SUGGEST}"
+      fi
+    fi
+    rm -f "$_CLOSING_FILE" 2>/dev/null
+  fi
+fi
+
 # ─── Output ──────────────────────────────────────────────────────────────
 _OUT_CONTEXT="${_ARKA_GREETING:-}${_SYNC_NOTICE:-}${_ROUTE_REMINDER}${_WORKFLOW_DIRECTIVE} $python_result"
 [ -n "$_HYGIENE" ] && _OUT_CONTEXT="$_OUT_CONTEXT $_HYGIENE"
@@ -433,6 +451,8 @@ _OUT_CONTEXT="${_ARKA_GREETING:-}${_SYNC_NOTICE:-}${_ROUTE_REMINDER}${_WORKFLOW_
 $_KB_CITE_NUDGE"
 [ -n "$_META_TAG_NUDGE" ] && _OUT_CONTEXT="$_OUT_CONTEXT
 $_META_TAG_NUDGE"
+[ -n "$_CLOSING_MARKER_NUDGE" ] && _OUT_CONTEXT="$_OUT_CONTEXT
+$_CLOSING_MARKER_NUDGE"
 [ -n "$_ARKA_CONTEXT_HITS" ] && _OUT_CONTEXT="$_OUT_CONTEXT
 $_ARKA_CONTEXT_HITS"
 # Escape for JSON

package/core/governance/closing_marker_check.py

@@ -0,0 +1,65 @@
+"""[arka:phase:13] / [arka:trivial] closing-marker soft-block (PR59 v2.76.0).
+
+Response-side classifier. Inspects the closing assistant message of a
+flow-required turn for the mandatory closure marker — either
+``[arka:phase:13]`` (full flow completed) or ``[arka:trivial]``
+(trivial bypass). Mirrors the contract of
+``core.governance.meta_tag_check`` (PR30 v2.49.0) and
+``core.governance.kb_cite_check`` (PR18 v2.40.0).
+
+Telemetry analysis from the May 24-25 continuous-build session showed
+**0% closing-marker rate** on every flow-required turn (5/5 rows
+without ``[arka:phase:13]`` or ``[arka:trivial]``). PR59 surfaces the
+gap to the next-turn nudge layer so the model is reminded to close
+each flow-required turn with an explicit marker.
+
+Soft-block contract — never raises. Hooks consume ClosingMarkerResult
+and decide whether to surface a suggestion.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+
+
+_PHASE13_RE: re.Pattern[str] = re.compile(r"\[arka:phase:13\]", re.IGNORECASE)
+_TRIVIAL_RE: re.Pattern[str] = re.compile(r"\[arka:trivial\]", re.IGNORECASE)
+_TRIVIAL_WORD_THRESHOLD: int = 15
+_SUGGESTION_TEXT: str = (
+    "Closing marker missing — end every flow-required turn with "
+    "`[arka:phase:13] <label>` (full canonical flow) or "
+    "`[arka:trivial] <reason>` (single-file edit < 10 lines, "
+    "imperative verb). Without the marker, telemetry can't confirm "
+    "the turn closed cleanly."
+)
+
+
+@dataclass(frozen=True)
+class ClosingMarkerResult:
+    """Verdict of a closing-marker check. Immutable; safe to log as JSON."""
+
+    passed: bool
+    reason: str
+    suggestion: str | None
+
+
+def check_closing_marker(response_text: str) -> ClosingMarkerResult:
+    """Classify whether a response carries a closing flow marker.
+
+    Order matters: a SHORT response *with* a marker still counts as
+    `present` — the trivial-length bypass only short-circuits when no
+    marker is found.
+    """
+    text = response_text or ""
+    if _PHASE13_RE.search(text):
+        return ClosingMarkerResult(True, "phase13", None)
+    if _TRIVIAL_RE.search(text):
+        return ClosingMarkerResult(True, "trivial", None)
+    if _is_trivial_length(text):
+        return ClosingMarkerResult(True, "trivial-length", None)
+    return ClosingMarkerResult(False, "missing", _SUGGESTION_TEXT)
+
+
+def _is_trivial_length(text: str) -> bool:
+    return len(text.split()) < _TRIVIAL_WORD_THRESHOLD

package/core/runtime/codex_cli.py

@@ -76,10 +76,16 @@ class CodexCliAdapter(RuntimeAdapter):
         raise NotImplementedError("Use Codex CLI's native content search")
 
     def headless_supported(self) -> bool:
-        # Codex CLI headless invocation syntax is not stable as of
-        # 2026-04-20. Until verified, we surface unsupported and let
-        # SubagentProvider fall back to AnthropicDirect or stub.
-        return False
+        # Auto-detect: headless is supported iff the `codex` binary is
+        # on PATH. When the operator installs Codex CLI later, this
+        # lights up without any code change (the headless_complete()
+        # method below already gates on shutil.which() too, so a missing
+        # binary will raise cleanly).
+        #
+        # Note: even when the binary is present, headless_complete()
+        # still raises until the invocation syntax is verified locally.
+        # See TODO(llm-agnostic) below for the verification checklist.
+        return shutil.which("codex") is not None
 
     def headless_complete(
         self,
@@ -96,14 +102,15 @@ class CodexCliAdapter(RuntimeAdapter):
             )
         # TODO(llm-agnostic): Implement real headless completion.
         #
-        # Status as of 2026-04-20: Codex CLI is NOT installed on the
-        # development machine, so actual invocation syntax could not
-        # be verified. Until a local install is available, refuse
-        # rather than ship guessed arguments.
+        # Status as of 2026-05-25 (PR60): Codex CLI still not verified
+        # in any ArkaOS dev environment. headless_supported() now
+        # auto-detects the binary on PATH so this lights up the moment
+        # someone installs it — but the actual subprocess call below
+        # still needs syntax verification before we can stop refusing.
         #
         # Verification checklist for whoever picks this up:
         #   1. Install:   npm install -g @openai/codex-cli
-        #   2. Discover:  codex --help   (confirm non-interactive flag)
+        #   2. Discover:  codex --help    (confirm non-interactive flag)
         #   3. Pattern:   likely `codex exec "<prompt>"` or
         #                 `codex --prompt "<prompt>" --format json`
         #   4. Wire the subprocess call (mirror the Gemini adapter —
@@ -113,9 +120,11 @@ class CodexCliAdapter(RuntimeAdapter):
         # SubagentProvider cleanly falls back to anthropic-direct or
         # stub when this raises, so the chain keeps working.
         raise NotImplementedError(
-            "Codex CLI headless mode requires local `codex` CLI. "
-            "Install: `npm install -g @openai/codex-cli` (verified 2026-04-20). "
-            "Verify syntax: `codex --help`. "
-            "See TODO(llm-agnostic) in this file. "
+            "Codex CLI headless mode requires verified invocation syntax. "
+            "The `codex` binary is on PATH but ArkaOS has not validated "
+            "the non-interactive call shape locally. "
+            "Verification steps: `codex --help`, then update "
+            "core/runtime/codex_cli.py::headless_complete to call the "
+            "discovered subprocess shape. "
             "SubagentProvider will cleanly fall back to anthropic-direct or stub."
         )

package/core/runtime/llm_provider.py

@@ -362,6 +362,19 @@ def get_llm_provider(config_path: Path | None = None) -> LLMProvider:
     return last if last is not None else StubProvider()
 
 
+def _current_category() -> str:
+    """Resolve the per-call category from the environment.
+
+    PR60 v2.77.0 — orchestration layers can set
+    ``ARKA_CALL_CATEGORY=skill:<slug>`` /
+    ``subagent:<dept>`` / ``plugin:<id>`` / ``mcp:<server>`` before
+    invoking the provider so `/arka costs --by-category` (PR47) can
+    attribute spend. Returns "" when unset, which lands in the base
+    bucket (backward-compatible).
+    """
+    return os.environ.get("ARKA_CALL_CATEGORY", "").strip()
+
+
 def _log_fallback(preferred: str, selected: str, reason: str = "") -> None:
     # Piggy-back on the cost telemetry file: zero-token, provider-only row.
     # Downstream can group by provider to spot degraded chains.
@@ -373,6 +386,7 @@ def _log_fallback(preferred: str, selected: str, reason: str = "") -> None:
         tokens_out=0,
         cached_tokens=0,
         estimated_cost_usd=None,
+        category=_current_category(),
     )
 
 
@@ -391,4 +405,5 @@ def _record(session_id: str, provider: str, response: LLMResponse) -> None:
         tokens_out=response.tokens_out,
         cached_tokens=response.cached_tokens,
         estimated_cost_usd=cost,
+        category=_current_category(),
     )

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "arkaos",
-  "version": "2.75.0",
+  "version": "2.77.0",
   "description": "The Operating System for AI Agent Teams",
   "type": "module",
   "bin": {

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "arkaos-core"
-version = "2.75.0"
+version = "2.77.0"
 description = "Core engine for ArkaOS — The Operating System for AI Agent Teams"
 readme = "README.md"
 license = {text = "MIT"}