chore: config coherence + constant-time auth (GLM review P0s)

apps/orchestrator/core/auth.py

@@ -2,11 +2,23 @@
 
 from __future__ import annotations
 
+import secrets
+
 from fastapi import HTTPException, Request, Security
 from fastapi.security import APIKeyHeader
 
 from core.config import settings
 
+
+def _token_matches(token: str | None) -> bool:
+    """Constant-time compare so the key can't be recovered a byte at a time.
+
+    A plain `==` short-circuits on the first differing byte, leaking the shared
+    key's prefix through response timing to anyone who can reach the endpoint.
+    `compare_digest` runs in time independent of where the mismatch is.
+    """
+    return secrets.compare_digest(token or "", settings.api_key)
+
 _api_key_header = APIKeyHeader(name="X-API-Key", auto_error=False)
 
 
@@ -24,7 +36,7 @@ async def require_api_key(
         if auth.lower().startswith("bearer "):
             token = auth[7:].strip()
 
-    if token != settings.api_key:
+    if not _token_matches(token):
         raise HTTPException(status_code=401, detail="Invalid or missing API key")
 
 
@@ -39,4 +51,4 @@ def verify_ws_token(query_token: str | None, request: Request) -> bool:
         auth = request.headers.get("Authorization", "")
         if auth.lower().startswith("bearer "):
             token = auth[7:].strip()
-    return token == settings.api_key
+    return _token_matches(token)

apps/orchestrator/core/config.py

@@ -7,72 +7,41 @@
 
 
 class Settings(BaseSettings):
+    """Runtime configuration.
+
+    Two groups, and the split is deliberate. The **SIEM flight recorder** is the
+    product: capture, canonical schema, detection, DLP, tool policy, and the
+    forward sinks. Everything a security engineer touches lives in that block.
+
+    The **optional governed runtime** below it (Obsidian vault + LangGraph
+    skills, Gemini/Ollama, the Telegram/Gmail channel drivers) is a separate,
+    off-by-default lineage documented in ``docs/advanced-governed-runtime.md``.
+    It is not required for IDE-hook or MCP capture. It is kept, not dead, but a
+    cloner evaluating the recorder can ignore it entirely.
+    """
+
     model_config = SettingsConfigDict(
         env_prefix="AGENTMETRY_",
         env_file=_ORCHESTRATOR_ROOT / ".env",
         env_file_encoding="utf-8",
         extra="ignore",
     )
 
-    vault_path: Path = Path(__file__).resolve().parents[3] / "vault"
-    qdrant_url: str = "http://localhost:6333"
-    ollama_base_url: str = "http://localhost:11434"
-    ollama_model: str = "llama3.2"
-    embedding_model: str = "nomic-embed-text"
-    llm_provider: str = "gemini"  # gemini | ollama | mock
-    allow_mock: bool = False  # permit mock fallback when no real provider is available
-    gemini_api_key: str = Field(
+    # ------------------------------------------------------------------ SIEM
+    # Operator identity + optional API key protecting ingest/tail/export.
+    operator_id: str = Field(
         default="",
         validation_alias=AliasChoices(
-            "GEMINI_API_KEY",
-            "GOOGLE_API_KEY",
+            "OPERATOR_ID",
+            "AGENTMETRY_OPERATOR_ID",
         ),
     )
-    gemini_model: str = "gemini-2.5-flash-lite"
-    gemini_embedding_model: str = "gemini-embedding-2"
-    embedding_dimensions: int = 768
-    collection_name: str = "agent_memory"
-    database_url: str = "sqlite:///./data/telemetry.db"
-    postgres_url: str = "postgresql://agentmetry:agentmetry@localhost:5432/agentic_os"
-    use_postgres: bool = False
-    approval_threshold: float = 0.9
-    cost_alert_threshold: float = 1.0
     api_key: str = Field(
         default="",
         validation_alias=AliasChoices("AGENTMETRY_API_KEY"),
     )
-    context_window_tokens: int = 1_048_576
-    gemini_health_cache_seconds: int = 300
-    gemini_health_probe: bool = False
-    gemini_embed_min_interval_seconds: float = 0.7
-    # Pacing: 60 / RPM. Flash-lite free tier ≈10 RPM → 6s; 2.5-flash ≈5 RPM → 13s.
-    gemini_flash_min_interval_seconds: float = 6.5
-    gemini_flash_daily_limit: int = 20
-    gemini_flash_interactive_reserve: int = 8
-    kernel_background_run_limit: int = 2
-    sandbox_tier1_allowed: str = "git"  # comma-separated binaries runnable in the jail
-    # Telegram channel adapter — ships disabled, like the gmail/search drivers.
-    channel_telegram_enabled: bool = False
-    telegram_bot_token: str = Field(
-        default="",
-        validation_alias=AliasChoices(
-            "TELEGRAM_BOT_TOKEN",
-        ),
-    )
-    telegram_allowed_chat_ids: str = ""  # comma-separated; empty refuses to start
-    startup_vault_index: bool = True
-    startup_index_skip_unchanged: bool = True
-    active_loop_archive_days: int = 7
-    active_loop_auto_archive: bool = True
-    gmail_send_enabled: bool = False  # Phase 4-E — requires explicit unlock
-    # Agentmetry — operator identity + SIEM-ready JSONL forwarder
-    operator_id: str = Field(
-        default="",
-        validation_alias=AliasChoices(
-            "OPERATOR_ID",
-            "AGENTMETRY_OPERATOR_ID",
-        ),
-    )
+
+    # Canonical JSONL trail + query index (system of record for the hook path).
     audit_export_enabled: bool = Field(
         default=True,
         validation_alias=AliasChoices(
@@ -82,28 +51,10 @@ class Settings(BaseSettings):
     audit_export_path: Path = _ORCHESTRATOR_ROOT / "data" / "audit-forward.jsonl"
     audit_db_path: Path = _ORCHESTRATOR_ROOT / "data" / "audit.db"
     detection_live_db_path: Path = _ORCHESTRATOR_ROOT / "data" / "detection_live.db"
-    # Local policy checks (core/audit/policy.py). Off by default: the built-in
-    # ruleset is a hardcoded starting point, and it annotates only, it cannot
-    # block. Real prevention lives in the hook (DLP block mode).
-    policy_enabled: bool = Field(
-        default=False,
-        validation_alias=AliasChoices("AGENTMETRY_POLICY_ENABLED"),
-    )
-    # off-hours-activity detection. Off by default: "unusual hours" is only a
-    # signal once an operator says which hours are usual, and scheduled jobs
-    # legitimately run at night.
-    detect_off_hours: bool = Field(
-        default=False,
-        validation_alias=AliasChoices("AGENTMETRY_DETECT_OFF_HOURS"),
-    )
-    business_hours: str = Field(
-        default="09-18",  # local start-end, 24h
-        validation_alias=AliasChoices("AGENTMETRY_BUSINESS_HOURS"),
-    )
-    business_tz: str = Field(
-        default="UTC",  # IANA name, e.g. Europe/Athens
-        validation_alias=AliasChoices("AGENTMETRY_BUSINESS_TZ"),
-    )
+    audit_ingest_enabled: bool = True
+    audit_ingest_url: str = "http://127.0.0.1:8000"
+
+    # Forward sinks (all optional; file is the default and never a cloud ledger).
     audit_sink: str = Field(
         default="file",
         validation_alias=AliasChoices("AGENTMETRY_AUDIT_SINK"),
@@ -146,21 +97,21 @@ class Settings(BaseSettings):
     audit_splunk_index: str = "main"
     audit_splunk_sourcetype: str = "agentmetry:json"
     audit_splunk_verify_tls: bool = True
-    audit_ingest_enabled: bool = True
-    audit_ingest_url: str = "http://127.0.0.1:8000"
     audit_alert_webhook_url: str = Field(
         default="",
         validation_alias=AliasChoices(
             "AGENTMETRY_AUDIT_ALERT_WEBHOOK_URL",
         ),
     )
-    # Semantic DLP
+
+    # Semantic DLP — regex scan of tool arguments at the hook boundary.
     dlp_mode: str = Field(
         default="log",
         validation_alias=AliasChoices("AGENTMETRY_DLP_MODE"),
     )
     dlp_rules_path: Path = _ORCHESTRATOR_ROOT.parent.parent / "policies" / "dlp" / "manifest.yaml"
     dlp_pii: bool = True
+
     # Tool allow/deny policy — enforced at the hook boundary (like DLP block mode).
     tool_policy_mode: str = Field(
         default="log",
@@ -170,6 +121,87 @@ class Settings(BaseSettings):
         _ORCHESTRATOR_ROOT.parent.parent / "policies" / "tool" / "manifest.yaml"
     )
 
+    # Post-ingest policy checks (core/audit/policy.py). Off by default: the
+    # built-in ruleset is a hardcoded starting point, and it annotates only, it
+    # cannot block. Real prevention lives in the hook (DLP block mode).
+    policy_enabled: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("AGENTMETRY_POLICY_ENABLED"),
+    )
+    # off-hours-activity detection. Off by default: "unusual hours" is only a
+    # signal once an operator says which hours are usual, and scheduled jobs
+    # legitimately run at night.
+    detect_off_hours: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("AGENTMETRY_DETECT_OFF_HOURS"),
+    )
+    business_hours: str = Field(
+        default="09-18",  # local start-end, 24h
+        validation_alias=AliasChoices("AGENTMETRY_BUSINESS_HOURS"),
+    )
+    business_tz: str = Field(
+        default="UTC",  # IANA name, e.g. Europe/Athens
+        validation_alias=AliasChoices("AGENTMETRY_BUSINESS_TZ"),
+    )
+
+    # ------------------------------------ optional governed runtime (advanced)
+    # Obsidian vault + LangGraph skills, model providers, and channel drivers.
+    # Off by default and not required for SIEM capture. See
+    # docs/advanced-governed-runtime.md. Kept for governed-agent demos; a
+    # recorder-only deployment can leave this whole block at its defaults.
+    vault_path: Path = Path(__file__).resolve().parents[3] / "vault"
+    startup_vault_index: bool = True
+    startup_index_skip_unchanged: bool = True
+    active_loop_archive_days: int = 7
+    active_loop_auto_archive: bool = True
+    kernel_background_run_limit: int = 2
+    sandbox_tier1_allowed: str = "git"  # comma-separated binaries runnable in the jail
+    approval_threshold: float = 0.9
+    cost_alert_threshold: float = 1.0
+    context_window_tokens: int = 1_048_576
+
+    # Model providers (governed runtime only).
+    llm_provider: str = "gemini"  # gemini | ollama | mock
+    allow_mock: bool = False  # permit mock fallback when no real provider is available
+    qdrant_url: str = "http://localhost:6333"
+    ollama_base_url: str = "http://localhost:11434"
+    ollama_model: str = "llama3.2"
+    embedding_model: str = "nomic-embed-text"
+    gemini_api_key: str = Field(
+        default="",
+        validation_alias=AliasChoices(
+            "GEMINI_API_KEY",
+            "GOOGLE_API_KEY",
+        ),
+    )
+    gemini_model: str = "gemini-2.5-flash-lite"
+    gemini_embedding_model: str = "gemini-embedding-2"
+    embedding_dimensions: int = 768
+    collection_name: str = "agent_memory"
+    gemini_health_cache_seconds: int = 300
+    gemini_health_probe: bool = False
+    gemini_embed_min_interval_seconds: float = 0.7
+    # Pacing: 60 / RPM. Flash-lite free tier ≈10 RPM → 6s; 2.5-flash ≈5 RPM → 13s.
+    gemini_flash_min_interval_seconds: float = 6.5
+    gemini_flash_daily_limit: int = 20
+    gemini_flash_interactive_reserve: int = 8
+
+    # Runtime telemetry store (separate from the audit trail above).
+    database_url: str = "sqlite:///./data/telemetry.db"
+    postgres_url: str = "postgresql://agentmetry:agentmetry@localhost:5432/agentic_os"
+    use_postgres: bool = False
+
+    # Channel drivers — ship disabled, require explicit unlock.
+    channel_telegram_enabled: bool = False
+    telegram_bot_token: str = Field(
+        default="",
+        validation_alias=AliasChoices(
+            "TELEGRAM_BOT_TOKEN",
+        ),
+    )
+    telegram_allowed_chat_ids: str = ""  # comma-separated; empty refuses to start
+    gmail_send_enabled: bool = False  # Phase 4-E — requires explicit unlock
+
 
 settings = Settings()