diff --git a/litellm/proxy/guardrails/guardrail_hooks/straiker/straiker.py b/litellm/proxy/guardrails/guardrail_hooks/straiker/straiker.py index 9231525d9b8..eae93971b2f 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/straiker/straiker.py +++ b/litellm/proxy/guardrails/guardrail_hooks/straiker/straiker.py @@ -72,14 +72,10 @@ V3_DERIVED_SESSION_PREFIX: Final = "litellm-" V3_AGENT_HEADER: Final = "x-s6r-agent" V3_RESPONSE_PHASE: Final = "response-sync" V3_BLOCK_DECISIONS: Final = frozenset({"block", "deny"}) -# Provider body fields, for every surface the proxy fronts. An allowlist rather than a -# denylist: the hook sees the client body merged with proxy state (`deployment` carries -# the resolved credential, `proxy_server_request` the client's Authorization header), and -# a field this list does not know is not relayed. Detection reads messages, system, tools, -# input and instructions; the rest travels so Straiker records the turn as the client sent it. +# An allowlist: the hook's request dict merges the client body with proxy state (`deployment` +# carries the resolved credential), so only fields named here are relayed. _V3_PROVIDER_BODY_KEYS: Final = frozenset( { - # OpenAI chat completions "model", "messages", "tools", @@ -110,12 +106,10 @@ _V3_PROVIDER_BODY_KEYS: Final = frozenset( "store", "service_tier", "web_search_options", - # OpenAI text completions "prompt", "suffix", "echo", "best_of", - # Anthropic messages "system", "stop_sequences", "top_k", @@ -124,7 +118,6 @@ _V3_PROVIDER_BODY_KEYS: Final = frozenset( "mcp_servers", "context_management", "output_format", - # OpenAI responses "input", "instructions", "previous_response_id", @@ -135,19 +128,14 @@ _V3_PROVIDER_BODY_KEYS: Final = frozenset( "max_output_tokens", "background", "conversation", - # Conversation grouping a client may state itself "session_id", } ) -# Fields on a `tools` or `mcp_servers` entry that carry a credential for the model's own remote -# calls (an OpenAI `mcp` tool's `headers` and `authorization`, Anthropic's `authorization_token`). -# Detection reads tool names, descriptions and schemas, never these. They sit directly on the -# entry, so the scrub is one level deep on purpose: a function schema that defines a `token` or -# `headers` property lives under `function.parameters` and is relayed exactly as sent. +# The scrub of these is one level deep on purpose: a function schema that defines a `token` or +# `headers` property lives under `function.parameters` and must be relayed as sent. _V3_CREDENTIAL_FIELDS: Final = frozenset({"authorization_token", "authorization", "headers"}) _V3_REDACTED_VALUE: Final = "[redacted]" _V3_REDACTED_KEYS: Final = frozenset({"tools", "mcp_servers"}) -# The identity fields Straiker's LiteLLM adapter reads from `metadata`, most specific first. _V3_IDENTITY_METADATA_KEYS: Final = ( "user_api_key_end_user_id", "user_api_key_user_email", @@ -672,7 +660,7 @@ def _v3_session_id( if conversation == "\0": return None seed: Final = f"{_v3_user(envelope) or ''}\0{conversation}" - return V3_DERIVED_SESSION_PREFIX + hashlib.md5(seed.encode("utf-8"), usedforsecurity=False).hexdigest() + return V3_DERIVED_SESSION_PREFIX + hashlib.sha256(seed.encode("utf-8")).hexdigest()[:32] _V3_PREAMBLE_ROLES: Final = frozenset({"system", "developer"}) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_straiker.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_straiker.py index 4720de78839..2fb4026a0b6 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_straiker.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_straiker.py @@ -1252,19 +1252,15 @@ async def test_v3_request_phase_relays_the_provider_body_and_nothing_else(): assert g.async_handler.post.call_args.args[0] == "https://test.straiker.ai/api/v3/detect" payload = _posted_payload(g) - # provider body, relayed assert payload["messages"] == data["messages"] assert payload["tools"] == data["tools"] assert payload["model"] == "claude-haiku-4-5-20251001" - # the body is the provider body and nothing pre-digested: Straiker parses it itself for flat in ("prompt", "app_response", "source", "user_name", "straiker_phase"): assert flat not in payload, flat - # identity and session, the way the unified Kong plugin sends them assert payload["original"] == {"processed": {"Meta": {"user": "alice.chen@example.com"}}} assert payload["metadata"] == {"user_api_key_end_user_id": "alice.chen@example.com"} # the client's Claude Code session header outranks LiteLLM's own session id (Kong precedence) assert payload["session_id"] == "cc-sess-9" - # nothing the proxy added serialized = json.dumps(payload) for leaked in ( "deployment", @@ -1519,12 +1515,9 @@ def test_v3_session_follows_kong_precedence(): ) data = _v3_request_data() - # 1. the client's own Claude Code session header wins assert _v3_session_id(envelope_with("meta-sess"), data, _v3_request_body(data)) == "cc-sess-9" - # 2. then LiteLLM's resolved session data["proxy_server_request"] = {"headers": {}} assert _v3_session_id(envelope_with("meta-sess"), data, _v3_request_body(data)) == "meta-sess" - # 3. then a hash of system + first message, stable across the conversation's replays a = _v3_session_id(envelope_with(None), data, _v3_request_body(data)) data2 = _v3_request_data() data2["proxy_server_request"] = {"headers": {}} @@ -1534,7 +1527,6 @@ def test_v3_session_follows_kong_precedence(): ] b = _v3_session_id(envelope_with(None), data2, _v3_request_body(data2)) assert a == b and a.startswith("litellm-") and len(a) == len("litellm-") + 32 - # 4. nothing to hash: no session assert _v3_session_id(envelope_with(None), {"proxy_server_request": {"headers": {}}}, {}) is None