arkaos 4.20.0 → 4.22.0

package/core/hooks/stop.py

@@ -13,9 +13,11 @@ Section order preserved exactly:
        [arka:meta], kb-cite, sycophancy, closing marker, skill proposer,
        kb=N reconciliation) → telemetry entry (mode: warn)
     6. Gate checkpoint persistence (core/workflow/gate_checkpoint.py)
-    7. Auto-documentor enqueue (flow-required + QG APPROVED + external
+    7. Stop-lint batch enqueue (detached scoped lint —
+       core/governance/stop_lint.py, warn-only telemetry)
+    8. Auto-documentor enqueue (flow-required + QG APPROVED + external
        research heuristic)
-    8. Belt-and-braces WF_MARKER removal
+    9. Belt-and-braces WF_MARKER removal
 
 The transcript is read from disk ONCE and shared across native-usage,
 flow validation, gate checkpoint, and the auto-doc external-research
@@ -24,13 +26,13 @@ heuristic (the old hook parsed it four times).
 
 from __future__ import annotations
 
+import contextlib
 import json
 import os
 import re
-from datetime import datetime, timezone
+from datetime import UTC, datetime
 from pathlib import Path
 
-from core.shared.temp_paths import arkaos_temp_dir
 from core.hooks._shared import (
     ensure_root_on_path,
     get_str,
@@ -39,6 +41,7 @@ from core.hooks._shared import (
     resolve_arkaos_root,
     safe_session_id,
 )
+from core.shared.temp_paths import arkaos_temp_dir
 
 _DISPATCH_RE = re.compile(
     r"\[arka:dispatch\][ \t]*[A-Za-z0-9_-]+[ \t]*->[ \t]*([A-Za-z0-9_-]+)",
@@ -164,8 +167,8 @@ def _flow_checks(
         from core.governance.phantom_action_check import (
             current_turn_assistant_texts,
         )
-        from core.workflow.gate_checkpoint import extract_latest_gate
         from core.workflow import plan_approval
+        from core.workflow.gate_checkpoint import extract_latest_gate
         turn = current_turn_assistant_texts(raw)
         scan = turn if turn is not None else [last]
         if session_id and extract_latest_gate(scan) == 2:
@@ -300,8 +303,31 @@ def _flow_checks(
     except Exception:
         pass
 
+    # Context-monitor tool-loop signal (warn-only): turn-boundary scan
+    # over the transcript sees every call, fast-pathed ones included.
+    loop_detected = False
+    loop_tool = ""
+    loop_repeats = 0
+    loop_pattern = ""
+    try:
+        from core.governance.tool_loop_check import check_tool_loops
+        lv = check_tool_loops(raw)
+        loop_detected = lv.detected
+        loop_tool = lv.tool
+        loop_repeats = lv.repeats
+        loop_pattern = lv.pattern
+        if lv.detected and safe_sid:
+            _write_tmp_state("arkaos-tool-loop", safe_sid, {
+                "tool": lv.tool,
+                "repeats": lv.repeats,
+                "pattern": lv.pattern,
+                "total_tool_uses": lv.total_tool_uses,
+            })
+    except Exception:
+        pass
+
     entry = {
-        "ts": datetime.now(timezone.utc).isoformat(),
+        "ts": datetime.now(UTC).isoformat(),
         "session_id": session_id,
         "cwd": cwd,
         "event": "stop-hook-flow-check",
@@ -326,6 +352,10 @@ def _flow_checks(
         "phantom_check_passed": phantom_passed,
         "phantom_check_reason": phantom_reason,
         "phantom_claims": phantom_claims,
+        "tool_loop_detected": loop_detected,
+        "tool_loop_tool": loop_tool,
+        "tool_loop_repeats": loop_repeats,
+        "tool_loop_pattern": loop_pattern,
         "effort_level": effort_level,
         "mode": "warn",
     }
@@ -336,16 +366,14 @@ def _flow_checks(
     except Exception:
         pass
 
-    try:
+    with contextlib.suppress(Exception):
         clear_flow_required(session_id)
-    except Exception:
-        pass
 
 
 def _auto_doc_enqueue(
     session_id: str, transcript_path: str, raw: str | None
 ) -> None:
-    """Section 7: fire-and-forget auto-documentor job enqueue."""
+    """Section 8: fire-and-forget auto-documentor job enqueue."""
     try:
         from core.jobs.auto_doc_worker import enqueue_job
         from core.workflow.flow_enforcer import _load_last_assistant_messages
@@ -387,10 +415,8 @@ def _auto_doc_enqueue(
     data = raw if raw is not None else ""
     if not any(marker in data for marker in _EXTERNAL_MARKERS):
         return
-    try:
+    with contextlib.suppress(Exception):
         enqueue_job(session_id, transcript_path, "APPROVED")
-    except Exception:
-        pass
 
 
 def _session_memory_enabled() -> bool:
@@ -426,7 +452,33 @@ def _enqueue_turn_capture(session_id: str, transcript_path: str, cwd: str) -> No
             env={**os.environ, "PYTHONPATH": repo},
             start_new_session=True,  # same pattern as _enqueue_cognition_capture
         )
-    except Exception:  # noqa: BLE001 — capture is best-effort, hook never breaks
+    except Exception:
+        pass
+
+
+def _enqueue_stop_lint(session_id: str, cwd: str) -> None:
+    """Section 7: detached scoped lint batch — deterministic evidence
+    lands in telemetry (stop-lint.jsonl), never on the 5s hook budget."""
+    if not cwd or not Path(cwd).is_dir():
+        return
+    try:
+        from core.governance.stop_lint import mode as stop_lint_mode
+        if stop_lint_mode() == "off":
+            return
+        repo = repo_path()
+        if not repo or not Path(repo).is_dir():
+            return
+        import subprocess
+        import sys as _sys
+        subprocess.Popen(
+            [_sys.executable, "-m", "core.governance.stop_lint",
+             cwd, session_id or ""],
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+            env={**os.environ, "PYTHONPATH": repo},
+            start_new_session=True,  # same pattern as _enqueue_turn_capture
+        )
+    except Exception:  # best-effort, hook never breaks
         pass
 
 
@@ -450,7 +502,7 @@ def _enqueue_routing_rebuild() -> None:
             env={**os.environ, "PYTHONPATH": repo},
             start_new_session=True,
         )
-    except Exception:  # noqa: BLE001 — rebuild is best-effort, hook never breaks
+    except Exception:
         pass
 
 
@@ -479,10 +531,8 @@ def main(stdin_json: dict | None = None) -> int:
 
     _native_usage(transcript_path, session_id, raw)
 
-    try:
+    with contextlib.suppress(Exception):
         _enqueue_turn_capture(session_id, transcript_path, cwd)
-    except Exception:  # noqa: BLE001 — never let capture break the hook
-        pass
 
     _enqueue_routing_rebuild()
 
@@ -494,13 +544,13 @@ def main(stdin_json: dict | None = None) -> int:
     if (Path(root) / "core" / "workflow" / "flow_enforcer.py").is_file():
         _flow_checks(session_id, transcript_path, cwd, effort_level, raw)
 
+    _enqueue_stop_lint(session_id, cwd)
+
     _auto_doc_enqueue(session_id, transcript_path, raw)
 
     # Belt-and-braces marker removal (session id already allowlisted).
-    try:
+    with contextlib.suppress(OSError):
         wf_marker.unlink(missing_ok=True)
-    except OSError:
-        pass
     return 0
 
 
@@ -510,4 +560,5 @@ if __name__ == "__main__":
     except SystemExit:
         raise
     except Exception:
-        raise SystemExit(0)  # Stop hook must never break the turn
+        # Stop hook must never break the turn.
+        raise SystemExit(0) from None

package/core/runtime/pricing.py

@@ -6,25 +6,59 @@ their public per-token pricing. It is used to compute an optional
 any model-selection logic — that is explicitly forbidden by the
 LLM-agnostic contract.
 
-TODO(pricing-snapshot 2026-04-20): values sourced from the public
-Anthropic (https://www.anthropic.com/pricing) and OpenAI
-(https://openai.com/api/pricing) pricing pages. Refresh when model
-families change. Unknown models return `None` from `estimate_cost_usd`
-and are logged with a null cost — never a guessed number.
+Snapshot 2026-07-18: values sourced from the public Anthropic pricing
+page (https://platform.claude.com/docs/en/about-claude/pricing) and
+OpenAI (https://openai.com/api/pricing). Refresh when model families
+change. Unknown models return `None` from `estimate_cost_usd` and are
+logged with a null cost — never a guessed number.
 """
 
 from __future__ import annotations
 
 
 # USD per 1M tokens. Only the `cache_read` and `cache_write` keys apply
-# to providers that expose prompt caching (currently Anthropic). Missing
-# keys fall back to 0 cost contribution rather than raising.
+# to providers that expose prompt caching (currently Anthropic; rates
+# are the 0.1x / 1.25x input multipliers for 5-minute cache writes).
+# Missing keys fall back to 0 cost contribution rather than raising.
 PRICING: dict[str, dict[str, float]] = {
+    "claude-fable-5": {
+        "input": 10.00,
+        "output": 50.00,
+        "cache_read": 1.00,
+        "cache_write": 12.50,
+    },
+    "claude-opus-4-8": {
+        "input": 5.00,
+        "output": 25.00,
+        "cache_read": 0.50,
+        "cache_write": 6.25,
+    },
     "claude-opus-4-7": {
-        "input": 15.00,
-        "output": 75.00,
-        "cache_read": 1.50,
-        "cache_write": 18.75,
+        "input": 5.00,
+        "output": 25.00,
+        "cache_read": 0.50,
+        "cache_write": 6.25,
+    },
+    "claude-opus-4-6": {
+        "input": 5.00,
+        "output": 25.00,
+        "cache_read": 0.50,
+        "cache_write": 6.25,
+    },
+    "claude-opus-4-5": {
+        "input": 5.00,
+        "output": 25.00,
+        "cache_read": 0.50,
+        "cache_write": 6.25,
+    },
+    # Introductory pricing ($2/$10) is in effect through 2026-08-31;
+    # from 2026-09-01 the standard rate is $3/$15 (cache $0.30/$3.75).
+    # Refresh this row when the introductory window closes.
+    "claude-sonnet-5": {
+        "input": 2.00,
+        "output": 10.00,
+        "cache_read": 0.20,
+        "cache_write": 2.50,
     },
     "claude-sonnet-4-6": {
         "input": 3.00,
@@ -33,19 +67,19 @@ PRICING: dict[str, dict[str, float]] = {
         "cache_write": 3.75,
     },
     "claude-haiku-4-5-20251001": {
-        "input": 0.80,
-        "output": 4.00,
-        "cache_read": 0.08,
-        "cache_write": 1.00,
+        "input": 1.00,
+        "output": 5.00,
+        "cache_read": 0.10,
+        "cache_write": 1.25,
     },
     # Generic (undated) alias of the dated haiku row above — native
     # transcripts sometimes report the tier id without the date suffix.
     # Same published price; NOT an invented number.
     "claude-haiku-4-5": {
-        "input": 0.80,
-        "output": 4.00,
-        "cache_read": 0.08,
-        "cache_write": 1.00,
+        "input": 1.00,
+        "output": 5.00,
+        "cache_read": 0.10,
+        "cache_write": 1.25,
     },
     "gpt-4": {
         "input": 30.00,