memstack-skill-loader 4.18.0__tar.gz → 4.19.0__tar.gz

{memstack_skill_loader-4.18.0/src/memstack_skill_loader.egg-info → memstack_skill_loader-4.19.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memstack-skill-loader
-Version: 4.18.0
+Version: 4.19.0
 Summary: MCP server that vector-indexes MemStack Pro skills for on-demand loading
 Project-URL: Homepage, https://memstack.pro
 Project-URL: Documentation, https://memstack.pro/docs

{memstack_skill_loader-4.18.0 → memstack_skill_loader-4.19.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "memstack-skill-loader"
-version = "4.18.0"
+version = "4.19.0"
 description = "MCP server that vector-indexes MemStack Pro skills for on-demand loading"
 readme = "README.md"
 requires-python = ">=3.10"

{memstack_skill_loader-4.18.0 → memstack_skill_loader-4.19.0}/src/memstack_skill_loader/__init__.py

@@ -1,3 +1,3 @@
 """MemStack Skill Loader — MCP server for semantic skill search."""
 
-__version__ = "4.18.0"
+__version__ = "4.19.0"

{memstack_skill_loader-4.18.0 → memstack_skill_loader-4.19.0}/src/memstack_skill_loader/agent_runner.py

@@ -20,19 +20,14 @@ from .stats import log_agent_invocation, log_api_cost
 
 
 # ---------------------------------------------------------------------------
-# Anthropic model pricing (per million tokens, May 2026)
+# Anthropic model pricing
 # ---------------------------------------------------------------------------
+# The rate card lives in models.py, which is also what the dashboard reads, so
+# there is one table rather than two that can drift. models.MODEL_PRICING is
+# prefix-matched, so a dated id such as claude-haiku-4-5-20251001 resolves to
+# its family rate instead of falling through to the unknown-model default.
+from .models import MODEL_PRICING, price_for as _price_for  # noqa: F401
 
-MODEL_PRICING = {
-    "claude-fable-5": {"input": 10.0, "output": 50.0},
-    "claude-opus-5": {"input": 5.0, "output": 25.0},
-    "claude-sonnet-5": {"input": 3.0, "output": 15.0},
-    "claude-sonnet-4-6": {"input": 3.0, "output": 15.0},
-    "claude-opus-4-6": {"input": 5.0, "output": 25.0},
-    "claude-opus-4-7": {"input": 5.0, "output": 25.0},
-    "claude-opus-4-8": {"input": 5.0, "output": 25.0},
-    "claude-haiku-4-5": {"input": 1.0, "output": 5.0},
-}
 # Cache read: 90% discount on input rate
 # Cache creation: 25% premium on input rate
 
@@ -44,8 +39,15 @@ def _calculate_api_cost(
     cache_read_tokens: int = 0,
     cache_creation_tokens: int = 0,
 ) -> float:
-    """Calculate USD cost from token counts and model pricing."""
-    pricing = MODEL_PRICING.get(model, MODEL_PRICING["claude-sonnet-4-6"])
+    """Calculate USD cost from token counts and model pricing.
+
+    Returns 0.0 for an id with no list rate (a non-model placeholder such as
+    "<synthetic>"): there is no rate to multiply by, and inventing one at the
+    Opus rate is what this avoids.
+    """
+    pricing = _price_for(model)
+    if not pricing["has_rate"]:
+        return 0.0
     input_rate = pricing["input"]
     output_rate = pricing["output"]
 
@@ -71,6 +73,9 @@ MAX_ITERATIONS = 9999
 
 ANTHROPIC_BASE_URL = os.environ.get("ANTHROPIC_BASE_URL", "")
 API_KEY_FILE = Path.home() / ".memstack" / "api_key"
+# Holds the Claude Code SUBSCRIPTION token (sk-ant-oat01...), NOT an API key.
+# The file name says "oauth" for install-compatibility only; do not rename it.
+# An api03 key filed here is sent as a Bearer token and silently never works.
 OAUTH_TOKEN_FILE = Path.home() / ".memstack" / "oauth_token"
 API_DEFAULT_MODEL = "claude-sonnet-4-6"
 API_MAX_TOKENS = 16000
@@ -94,7 +99,12 @@ def _load_api_key() -> str | None:
     return None
 
 def _load_oauth_token() -> str | None:
-    """Get the OAuth token from environment or stored config file."""
+    """Get the Claude Code SUBSCRIPTION token (sk-ant-oat01...).
+
+    Reads CLAUDE_CODE_OAUTH_TOKEN, then ~/.memstack/oauth_token. That file holds
+    the subscription token from `claude setup-token`, not an API key, despite
+    the historic file name.
+    """
     token = os.environ.get("CLAUDE_CODE_OAUTH_TOKEN", "").strip()
     if token:
         return token
@@ -108,6 +118,54 @@ def _load_oauth_token() -> str | None:
     return None
 
 
+def _live_api_key():
+    """(has_live_api_key, dead_rows) for the api-execution-mode fallback gate.
+
+    Validity, not presence. A key counts only when a probe of GET /v1/models no
+    older than 24h says the credential is live; the probe runs here when the
+    status file is absent or stale. ANTHROPIC_BASE_URL is a proxy address, not a
+    credential, and never counts. The old gate let it stand in for one, so a
+    machine with the TokenStack proxy exported ran "api" mode with no key at all.
+
+    Fails open on any error: an unreachable probe returns no live key, which
+    downgrades to subscription rather than running api mode blind.
+    """
+    try:
+        from . import models as model_registry
+        return model_registry.live_api_key_status()
+    except Exception as exc:
+        return False, [{"label": "probe", "kind_label": "",
+                        "status": "unchecked", "error": str(exc)}]
+
+
+def _dead_key_reason(dead_rows) -> str:
+    """One sentence naming which key was dead. Never any part of a value."""
+    if not dead_rows:
+        return "no API key credential found (env ANTHROPIC_API_KEY or ~/.memstack/api_key)"
+    parts = []
+    for r in dead_rows:
+        label = r.get("label", "?")
+        kind = r.get("kind_label", "")
+        code = r.get("http_status")
+        if r.get("status") == "dead":
+            detail = f"dead (HTTP {code})" if code else "dead"
+        else:
+            detail = "unchecked" + (f": {r.get('error', '')}" if r.get("error") else "")
+        parts.append(f"{label}" + (f" [{kind}]" if kind else "") + f": {detail}")
+    return "; ".join(parts)
+
+
+def _append_agent_log(path: Path, line: str) -> None:
+    """Append one timestamped runner line to an agent log. Never raises."""
+    try:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        ts = time.strftime("%Y-%m-%dT%H:%M:%S")
+        with open(path, "a", encoding="utf-8") as f:
+            f.write(f"[{ts}] [agent-runner] {line}\n")
+    except Exception:
+        pass
+
+
 SYSTEM_PROMPTS = {
     "manager": (
         "You are a Manager agent. You receive a task and project context (directory listing "
@@ -1071,6 +1129,8 @@ def _build_env(strip_api_key: bool = False, inject_api_key: str | None = None) -
     # Set strip_api_key=False to revert.
     if strip_api_key:
         env.pop("ANTHROPIC_API_KEY", None)
+        # Subscription token (sk-ant-oat01) from ~/.memstack/oauth_token: that
+        # file holds the Claude Code subscription token, not an API key.
         oauth_token = _load_oauth_token()
         if oauth_token:
             env["ANTHROPIC_AUTH_TOKEN"] = oauth_token
@@ -1094,11 +1154,17 @@ def _build_env(strip_api_key: bool = False, inject_api_key: str | None = None) -
     return env
 
 
-def _parse_stream_json(raw: str) -> tuple[str, int, int, float, int, str]:
+def _parse_stream_json(raw: str) -> tuple[str, int, int, float, int, str, int, int]:
     """Extract text and token usage from claude --output-format stream-json.
 
-    Returns (text, input_tokens, output_tokens, cost_usd, context_tokens, model).
+    Returns (text, input_tokens, output_tokens, cost_usd, context_tokens, model,
+    cache_read_tokens, cache_creation_tokens).
     context_tokens is the last iteration's total tokens (accurate context window fill).
+
+    The two cache populations are returned separately because the CLI bills for
+    them inside total_cost_usd while leaving them out of usage.input_tokens.
+    Dropping them here is what made the Burn Report's cost columns unprovable
+    from its token columns.
     """
     text_parts: list[str] = []
     tool_parts: list[str] = []
@@ -1108,6 +1174,8 @@ def _parse_stream_json(raw: str) -> tuple[str, int, int, float, int, str]:
     cost_usd = 0.0
     context_tokens = 0
     stream_model = ""
+    cache_read_tokens = 0
+    cache_creation_tokens = 0
     last_assistant_usage: dict = {}
     for line in raw.splitlines():
         line = line.strip()
@@ -1139,6 +1207,8 @@ def _parse_stream_json(raw: str) -> tuple[str, int, int, float, int, str]:
             usage = obj.get("usage", {})
             input_tokens = usage.get("input_tokens", 0)
             output_tokens = usage.get("output_tokens", 0)
+            cache_read_tokens = usage.get("cache_read_input_tokens", 0) or 0
+            cache_creation_tokens = usage.get("cache_creation_input_tokens", 0) or 0
             cost_usd = obj.get("total_cost_usd", 0.0) or 0.0
     if last_assistant_usage:
         context_tokens = (
@@ -1156,7 +1226,8 @@ def _parse_stream_json(raw: str) -> tuple[str, int, int, float, int, str]:
         text = "[Agent completed with tool calls but no text summary]\n\n" + "\n\n".join(tool_parts)
     else:
         text = ""
-    return text, input_tokens, output_tokens, cost_usd, context_tokens, stream_model
+    return (text, input_tokens, output_tokens, cost_usd, context_tokens,
+            stream_model, cache_read_tokens, cache_creation_tokens)
 
 
 def _extract_text_from_stream_line(line: str) -> Optional[str]:
@@ -1574,7 +1645,8 @@ def _invoke_agent(name: str, prompt: str, working_dir: str, log_path: Optional[P
 
     raw_stdout = stdout_data or ""
     stderr = stderr_data or ""
-    output, input_tokens, output_tokens, cost_usd, context_tokens, stream_model = _parse_stream_json(raw_stdout)
+    (output, input_tokens, output_tokens, cost_usd, context_tokens, stream_model,
+     cache_read_tokens, cache_creation_tokens) = _parse_stream_json(raw_stdout)
     output = output.strip()
     ts = time.strftime("%H:%M:%S")
 
@@ -1615,6 +1687,8 @@ def _invoke_agent(name: str, prompt: str, working_dir: str, log_path: Optional[P
             input_tokens=input_tokens, output_tokens=output_tokens,
             actual_cost_usd=cost_usd, execution_mode=exec_mode,
             project=_project,
+            cache_read_tokens=cache_read_tokens,
+            cache_creation_tokens=cache_creation_tokens,
         )
     except Exception:
         pass
@@ -2091,11 +2165,18 @@ def _orchestrate(session: Session) -> None:
         # Load execution modes and custom display names for this run
         agent_modes = load_agent_modes()
         display_names = _load_agent_display_names()
-        has_api_key = bool(_load_api_key() or ANTHROPIC_BASE_URL)
+        has_api_key, dead_keys = _live_api_key()
         for _ag in ("manager", "builder", "reviewer"):
             if agent_modes[_ag] == "api" and not has_api_key:
                 agent_modes[_ag] = "subscription"
-                print(f"[agent-runner] {_ag}: mode=api but no API key found, falling back to subscription", file=sys.stderr)
+                reason = _dead_key_reason(dead_keys)
+                print(f"[agent-runner] {_ag}: mode=api but no live API key, "
+                      f"falling back to subscription ({reason})", file=sys.stderr)
+                _append_agent_log(
+                    session_log_dir / f"{_ag}.log",
+                    f"mode=api requested but no live API key; "
+                    f"falling back to subscription. {reason}",
+                )
 
         # Step 1: Manager analyzes the task
         session.agents["manager"]["status"] = "busy"