diff --git a/CHANGELOG.md b/CHANGELOG.md index ea463b0dd..d51bfa558 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -3,6 +3,10 @@ ## [Unreleased] +### Fixed + +- **Custom provider context-length probes now include the provider API key during session hydration and streaming fallbacks (#4059).** Static session loads and fallback usage/session-save paths now reuse the matched custom provider credentials when querying `/v1/models`, so authenticated custom endpoints do not fall back to the default 256K window and clobber a larger persisted context length / compression threshold. + ## [v0.51.369] — 2026-06-12 — Release MH (WebUI streaming honors runtime target model/base_url) ### Fixed diff --git a/api/routes.py b/api/routes.py index 520c91e15..8b247cdfc 100644 --- a/api/routes.py +++ b/api/routes.py @@ -2276,7 +2276,7 @@ def _model_matches_configured_default( class _ContextLengthLookupInputs: - __slots__ = ("config_context_length", "custom_providers", "base_url", "provider") + __slots__ = ("config_context_length", "custom_providers", "base_url", "provider", "api_key") def __init__( self, @@ -2285,11 +2285,13 @@ class _ContextLengthLookupInputs: custom_providers: list | None = None, base_url: str = "", provider: str = "", + api_key: str = "", ) -> None: self.config_context_length = config_context_length self.custom_providers = custom_providers self.base_url = base_url self.provider = provider + self.api_key = api_key def _positive_context_length(value) -> int | None: @@ -2379,6 +2381,38 @@ def _providers_match_for_context(config_key: object, requested_provider: str) -> ) +def _custom_provider_api_key_for_context(entry: dict, provider: str) -> str: + """Resolve the API key for a matched ``custom_providers`` entry. + + Static session hydration/update routes already have a per-profile config + snapshot. Resolve from the matched entry instead of re-reading global config, + while preserving the same literal, ``${ENV_VAR}``, ``key_env``, and + sanitized-env shapes used by streaming/provider resolution. + """ + raw_api_key = entry.get("api_key") + if raw_api_key is not None: + api_key_text = str(raw_api_key).strip() + if api_key_text.startswith("${") and api_key_text.endswith("}") and len(api_key_text) > 3: + resolved = os.getenv(api_key_text[2:-1], "").strip() + if resolved: + return resolved + elif api_key_text: + return api_key_text + + key_env = str(entry.get("key_env") or "").strip() + if key_env: + resolved = os.getenv(key_env, "").strip() + if resolved: + return resolved + + try: + from api.config import _lookup_custom_api_key_env + + return _lookup_custom_api_key_env(provider) or "" + except Exception: + return "" + + def _context_length_lookup_inputs_for_model( model: str | None, provider: str | None = None, @@ -2438,6 +2472,7 @@ def _context_length_lookup_inputs_for_model( break custom_context_length = None + effective_api_key = "" if custom_providers: target_base = effective_base_url.rstrip("/") model_candidates = set(_model_lookup_candidates(bare_model or model_for_lookup)) @@ -2467,6 +2502,7 @@ def _context_length_lookup_inputs_for_model( effective_provider = entry_slug if not effective_base_url and entry_base: effective_base_url = entry_base + effective_api_key = _custom_provider_api_key_for_context(entry, effective_provider or entry_slug) custom_context_length = _models_config_context_length(models_cfg, bare_model or model_for_lookup) break @@ -2489,6 +2525,7 @@ def _context_length_lookup_inputs_for_model( custom_providers=custom_providers, base_url=effective_base_url, provider=effective_provider, + api_key=effective_api_key, ) @@ -3002,6 +3039,7 @@ def _resolve_context_length_for_session_model( return _get_cl( model_for_lookup, _ctx_lookup.base_url, + api_key=_ctx_lookup.api_key, config_context_length=_ctx_lookup.config_context_length, provider=_ctx_lookup.provider or provider or "", custom_providers=_ctx_lookup.custom_providers, diff --git a/api/streaming.py b/api/streaming.py index d1ed35836..2ac461d96 100644 --- a/api/streaming.py +++ b/api/streaming.py @@ -7294,11 +7294,13 @@ def _run_agent_streaming( ) _cfg_ctx_len = _ctx_lookup.config_context_length _cfg_custom_providers = _ctx_lookup.custom_providers + _cfg_api_key = _ctx_lookup.api_key or getattr(agent, 'api_key', '') or resolved_api_key or '' _cfg_base_url = _ctx_lookup.base_url or _cfg_base_url _cfg_provider = _ctx_lookup.provider or resolved_provider or '' _resolved_cl = get_model_context_length( getattr(agent, 'model', resolved_model or '') or '', _cfg_base_url, + api_key=_cfg_api_key, config_context_length=_cfg_ctx_len, provider=_cfg_provider, custom_providers=_cfg_custom_providers, @@ -7552,12 +7554,14 @@ def _run_agent_streaming( ) _cfg_ctx_len = _ctx_lookup.config_context_length _cfg_custom_providers = _ctx_lookup.custom_providers + _cfg_api_key = _ctx_lookup.api_key or getattr(agent, 'api_key', '') or resolved_api_key or '' _cfg_base_url = _ctx_lookup.base_url _cfg_provider = _ctx_lookup.provider or resolved_provider or '' try: _fb_cl = _get_cl( getattr(agent, 'model', resolved_model or '') or '', _cfg_base_url, + api_key=_cfg_api_key, config_context_length=_cfg_ctx_len, provider=_cfg_provider, custom_providers=_cfg_custom_providers, diff --git a/tests/test_issue1896_context_length_fallback_args.py b/tests/test_issue1896_context_length_fallback_args.py index 2ea48c4e2..67dce8257 100644 --- a/tests/test_issue1896_context_length_fallback_args.py +++ b/tests/test_issue1896_context_length_fallback_args.py @@ -233,3 +233,133 @@ def test_routes_session_load_fallback_passes_config_overrides(): "session-load fallback must catch TypeError to support older " "hermes-agent builds without the new kwargs." ) + + +def test_context_lookup_returns_custom_provider_api_key_from_entry(): + """#4059: static session hydration must carry custom-provider API keys. + + A named ``custom_providers`` entry can require auth for its ``/v1/models`` + endpoint. The route-side lookup helper already identifies the matching + provider/base/model; it must also return that entry's API key so + ``get_model_context_length`` does not probe anonymously and fall back to + 256K. + """ + from api.routes import _context_length_lookup_inputs_for_model + + lookup = _context_length_lookup_inputs_for_model( + "custom-model-id", + "custom:llm-proxy", + cfg={ + "custom_providers": [ + { + "name": "llm-proxy", + "base_url": "https://llm.example.test/v1", + "api_key": "sk-test-entry", + "model": "custom-model-id", + } + ] + }, + ) + + assert lookup.provider == "custom:llm-proxy" + assert lookup.base_url == "https://llm.example.test/v1" + assert lookup.api_key == "sk-test-entry" + + +def test_context_lookup_resolves_custom_provider_api_key_env_template(monkeypatch): + """#4059: ``${ENV_VAR}`` custom-provider keys resolve before metadata probes.""" + from api.routes import _context_length_lookup_inputs_for_model + + monkeypatch.setenv("ISSUE_4059_CONTEXT_KEY", "env-template-key") + + lookup = _context_length_lookup_inputs_for_model( + "custom-model-id", + "custom:llm-proxy", + cfg={ + "custom_providers": [ + { + "name": "llm-proxy", + "base_url": "https://llm.example.test/v1", + "api_key": "${ISSUE_4059_CONTEXT_KEY}", + "model": "custom-model-id", + } + ] + }, + ) + + assert lookup.api_key == "env-template-key" + + +def test_context_lookup_resolves_custom_provider_key_env(monkeypatch): + """#4059: ``key_env`` custom-provider keys resolve before metadata probes.""" + from api.routes import _context_length_lookup_inputs_for_model + + monkeypatch.setenv("ISSUE_4059_KEY_ENV", "key-env-value") + + lookup = _context_length_lookup_inputs_for_model( + "custom-model-id", + "custom:llm-proxy", + cfg={ + "custom_providers": [ + { + "name": "llm-proxy", + "base_url": "https://llm.example.test/v1", + "key_env": "ISSUE_4059_KEY_ENV", + "model": "custom-model-id", + } + ] + }, + ) + + assert lookup.api_key == "key-env-value" + + +def test_routes_session_model_resolver_passes_custom_provider_api_key(monkeypatch): + """#4059: ``_resolve_context_length_for_session_model`` passes api_key. + + This is the session-load/static-update path that was clobbering persisted + 500K context metadata back to the unauthenticated 256K fallback. + """ + from api import config as cfg_mod + from api import routes + import agent.model_metadata as metadata + + seen = {} + + monkeypatch.setattr( + cfg_mod, + "get_config", + lambda: { + "custom_providers": [ + { + "name": "llm-proxy", + "base_url": "https://llm.example.test/v1", + "api_key": "sk-test-route", + "model": "custom-model-id", + } + ] + }, + ) + + def fake_get_model_context_length(model, base_url, **kwargs): + seen.update(model=model, base_url=base_url, kwargs=kwargs) + return 500_000 if kwargs.get("api_key") == "sk-test-route" else 256_000 + + monkeypatch.setattr(metadata, "get_model_context_length", fake_get_model_context_length) + + assert routes._resolve_context_length_for_session_model( + "custom-model-id", + "custom:llm-proxy", + ) == 500_000 + assert seen["kwargs"]["api_key"] == "sk-test-route" + + +def test_streaming_context_length_fallbacks_pass_api_key(): + """#4059: both streaming fallback probes must pass the resolved api_key.""" + blocks = _both_callsites() + for i, block in enumerate(blocks): + assert "api_key=" in block, ( + f"Callsite #{i+1} is missing api_key=. Authenticated custom " + f"provider /v1/models probes then fail and fall back to 256K. " + f"See #4059.\n\nBlock:\n{block}" + ) diff --git a/tests/test_issue3256_context_length_default_only_guard.py b/tests/test_issue3256_context_length_default_only_guard.py index fab27abad..164d21935 100644 --- a/tests/test_issue3256_context_length_default_only_guard.py +++ b/tests/test_issue3256_context_length_default_only_guard.py @@ -20,8 +20,9 @@ def _install_fake_get_model_context_length(monkeypatch, recorder): """Install a fake get_model_context_length into a stand-in agent.model_metadata.""" mod = types.ModuleType("agent.model_metadata") - def _fake(model, base_url="", config_context_length=None, provider="", custom_providers=None): + def _fake(model, base_url="", api_key="", config_context_length=None, provider="", custom_providers=None): recorder["model"] = model + recorder["api_key"] = api_key recorder["config_context_length"] = config_context_length # Pretend the real per-model metadata window is 1,000,000 unless the # caller forced a config cap, in which case honor the cap (mirrors the