* fix(llm): forward provider_params.timeout to the OpenAI-compatible embedding client
#832 and #903 added a configurable request timeout for the LLM registry
and the Gemini embedding client respectively, but the OpenAI-compatible
embedding client (src/embedding_client.py) was never wired up. It
constructed AsyncOpenAI with no timeout at all, so a stalled socket
against a slow or contended OpenAI-compatible backend (e.g. a
self-hosted embedding model under load) wedges the deriver worker's
event loop indefinitely — the exact failure #785/#903 describe, just
via a code path #903 didn't cover.
EmbeddingModelConfig now carries provider_params through from
resolve_embedding_model_config, mirroring how resolve_model_config
already does it for ModelConfig, and the OpenAI branch of
_EmbeddingClient.__init__ extracts `timeout` via the existing
request_timeout_from_extra_params helper. Unset stays unset — no
existing behavior changes.
Reproduced and verified against a real self-hosted deployment (local
Ollama backend under load): before this fix, a single stuck embedding
call blocked all deriver queue processing for 20+ minutes with no
error logged, twice in one session.
* fix(embedding): use first-class timeout on embedding model config
provider_params is the LLM per-request escape hatch; embedding timeouts are
client-construction knobs and belong next to max_batch_size. Wire the field
for OpenAI and Gemini, omit the OpenAI kwarg when unset so the SDK default
stays, and keep Gemini's 10-minute floor when unset.
* test(embedding): live coverage for first-class embedding timeout
Exercise EmbeddingModelConfig.timeout on one representative OpenAI and
Gemini model: configured timeout lands on the SDK client, and a near-zero
timeout aborts before the provider answers.
---------
Co-authored-by: Aakash Kattelu <aakash@plasticlabs.ai>
* test(embedding): add reproducer for missing encoding_format on openai paths
The openai SDK defaults encoding_format to base64 when it is not passed. OpenAI-compatible providers that don't support base64 embeddings (e.g. OpenRouter with nvidia/nemotron-3-embed-1b:free) return HTTP 200 with empty data, and every embedding call fails with 'No embedding data received'.
* fix(embedding): request float encoding_format on openai embedding calls
The openai SDK defaults encoding_format to base64 when the caller does not pass one. OpenAI-compatible providers that don't support base64 embeddings (e.g. OpenRouter hosting nvidia/nemotron-3-embed-1b:free) answer HTTP 200 with empty embedding data, and every embedding call fails with 'No embedding data received', breaking conclusions, semantic search, and the deriver. Pass encoding_format='float' explicitly on both the single-query and batch call paths.
* test(embedding): cover openai-compatible providers in the live embedding matrix
The existing openai family runs against real OpenAI, which serves base64
embeddings happily, so the matrix passes with or without the #932 fix. Adds an
`openai_compatible_embedding` family (openai transport, third-party base_url)
so the matrix can reach a provider that rejects base64. Empty default_models
keeps it skipped unless LIVE_EMBEDDING_OPENAI_COMPATIBLE_MODELS is set.
Also adds test_live_openai_float_encoding_matches_base64, which pins the other
direction: switching the wire format must not move vectors on real OpenAI.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(embedding): keep an explicit embedding-count check on the openai paths
Passing `encoding_format` disables the openai SDK's own empty-data guard, so a
provider answering 200 with missing embeddings surfaced as `IndexError: list
index out of range` on the single path and `zip() argument 2 is shorter than
argument 1` on the batch path. The latter is also #745's signature, which would
have left it with two unrelated causes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* docs(live-llm): correct the openai-compatible embedding matrix env docs
The documented default dimensions said 2048 after the family moved to 3072, and
LIVE_EMBEDDING_OPENAI_COMPATIBLE_SEND_DIMENSIONS was missing entirely. Also
points the example and the coverage note at a model that is actually reachable,
and records that OpenRouter load-balances, so the base64 failure is per-attempt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* feat(embedding): resolve openai encoding_format by mode instead of pinning float
Requesting float unconditionally costs ~3.6x the response bytes of base64 and up
to +83% latency on a 500-item batch, which the default deployment on real OpenAI
pays for nothing: only third-party OpenAI-compatible providers reject base64.
Adds EMBEDDING_MODEL_CONFIG__ENCODING_FORMAT_MODE, mirroring dimensions_mode.
`auto` keeps base64 when no base_url override is set or it points at
api.openai.com, and picks float elsewhere. The format is still always sent
explicitly, since the SDK otherwise injects base64 on its own.
Also corrects the _validate_embedding_count docstring, which said "fewer" where
the guard is an inequality.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(embedding): request base64 embeddings by omission, not by name
The openai SDK decodes a base64 response only when it injected the default
itself; naming any format makes it skip the decode and hand back the raw string,
which then fails the dimension check with "Expected 1536, got 8192". base64 mode
therefore has to omit the kwarg rather than pass it.
The unit fake returned float lists whatever was asked for, so it could not catch
this. It now mirrors the SDK and returns a base64 string for a named base64
request, which fails against the previous commit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Aakash Kattelu <aakash@plasticlabs.ai>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* Fix Gemini batch embedding for gemini-embedding-2* models
The google-genai SDK treats embed_content(contents=[list_of_strings]) as a
single multi-part document for gemini-embedding-2* models, silently returning
exactly 1 embedding regardless of input count. This caused a zip(...,
strict=True) ValueError in _process_batch.
Wrap each text in genai_types.Content(parts=[genai_types.Part(text=...)])
so the SDK treats each string as a separate content item. This matches the
workaround used by pydantic-ai (#4873) and graphiti (#1474).
Upstream SDK issue: googleapis/python-genai#2523Fixesplastic-labs/honcho#744
* test(embedding): live embedding coverage for every Gemini and OpenAI model
Adds tests/live_llm/test_live_embeddings.py plus an env-driven embedding
matrix alongside the existing LLM one. Covers single embed, batched embed,
batch-vs-single alignment, chunk-to-id mapping, and the batch-split path.
Only a live call catches the gemini-embedding-2* collapse: the SDK folds a
list of bare strings into one document and returns a single embedding.
Reverting the Content wrapping fails all four batch tests for
gemini-embedding-2-preview and gemini-embedding-2 with the reported
`zip() argument 2 is shorter than argument 1`, while gemini-embedding-001
and text-embedding-3-small stay green.
Also makes the concatenation in the conclusions semantic-search validation
message explicit, so the repo-wide basedpyright pre-push hook passes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(embedding): drop the preview twin from the default embedding matrix
gemini-embedding-2 is the GA release of gemini-embedding-2-preview and
behaves identically, so running both by default doubles the Gemini cost for
no extra coverage. The preview stays reachable through
LIVE_EMBEDDING_GEMINI_MODELS.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Aakash Kattelu <aakash@plasticlabs.ai>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix: Add JSON repair for truncated LLM responses across all providers and Gemini thinking budget support
LengthFinishReasonError from OpenAI-compatible providers (custom, openai, groq) was crashing the deriver
with 14k+ occurrences in production. The vLLM path already had repair logic but it was gated on
provider=="vllm", unreachable when routing through litellm as a custom provider.
- Extract shared _repair_response_model_json() helper for all providers
- Catch LengthFinishReasonError in OpenAI/custom parse() path and repair truncated JSON
- Add repair fallback to Anthropic and Gemini response_model paths
- Add repair fallback to Groq response_model path
- Pass thinking_budget_tokens to Gemini 2.5 models via thinking_config
- Add 14 tests covering repair paths for all providers and Gemini thinking budget
Fixes HONCHO-YC
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: live llm integration tests
* feat: Consistent Model Config Protocol
* fix: migrate the remaining app callers off the legacy llm_settings path
* fix: Docs and regression tests
* fix: refactor llm runtime path to model-config-only API
* fix: refactor config to nested model-config source of truth
* fix: refactor llm streaming and tool dispatch through backends
* fix: cut over llm config to nested model_config only
* fix: collapse vllm and custom into openai_compatible transport
* feat: refactor llm config to explicit transports and bare model ids
* feat: (embed) Add configurability for embedding model
* fix: tests for embedding provider
* fix: Address Review Comments
* fix: (llm) remove Groq backend and per-vendor base URLs
* chore: move llm tests
* fix: (llm) address review findings — config regressions, backend bugs, dead code
* fix: address backend end silly errors
* chore: (docs) update configuration and self-hosting guides
* chore: fix tests
* fix: address code rabbit comments
* fix: add validation to the dream settings
* fix: further address code rabbit comments
* fix: Address Code Rabbit Comments
* fix: Another round of code rabbit
* fix: Address Code Rabbit Nits
* fix: tests
* refactor: rename thinking validator to reflect transport scope
_validate_anthropic_thinking_minimum only enforces the >=1024 rule for
Anthropic and no-ops for other transports, so the name was misleading
now that it's shared across ConfiguredModelSettings, FallbackModelSettings,
and ModelConfig. Renamed to _validate_thinking_constraints with a docstring
clarifying per-transport behavior. No logic change.
* fix(config): drop transport-specific thinking params when env override changes transport
_fill_defaults_for_nested_field previously preserved the default MODEL_CONFIG's
thinking_budget_tokens/thinking_effort across a transport override. This leaked
Gemini-family defaults (e.g. thinking_budget_tokens=1024) into OpenAI-transport
overrides, and the OpenAI backend then correctly rejected the unsupported param
at call time (OpenAI uses reasoning.effort, not a token budget).
The helper now strips thinking_budget_tokens and thinking_effort from the
default dict when the env override supplies a transport different from the
default's. Explicit thinking params in the override are preserved.
* fix(config): apply thinking-param strip to dialectic level merge too
DialecticSettings._merge_level_defaults does its own inline MODEL_CONFIG
merge (parallel to _fill_defaults_for_nested_field), so the previous fix
missed dialectic-level overrides. E.g. flipping
DIALECTIC_LEVELS__minimal__MODEL_CONFIG__TRANSPORT from gemini (default)
to openai still leaked the default thinking_budget_tokens=0 into the
openai config, which the OpenAI backend then rejected at call time.
The level-merge path now applies the same 'strip transport-specific
thinking params when transport changes' rule as the generic helper.
Added a regression test exercising the merge validator directly.
* refactor(llm): wire ModelConfig knobs through, prune clients.py migration leftovers
Three connected fixes to finish carving the LLM stack out of src/utils/clients.py
and into src/llm/:
1. Propagate ModelConfig tuning knobs into backend calls.
honcho_llm_call_inner built extra_params from only {json_mode, verbosity},
silently dropping top_p, top_k, frequency_penalty, presence_penalty, seed,
and operator-supplied provider_params from any ModelConfig. Thread the
selected config through ProviderSelection and merge
build_config_extra_params(selected_config) into extra_params; per-call
kwargs still win over provider_params defaults. Makes
_build_config_extra_params public as build_config_extra_params so
clients.py and request_builder.py share one translation. Adds
TestModelConfigExtraParamsPropagation covering OpenAI/Anthropic knob
propagation, provider_params passthrough, and per-call override
precedence.
2. Drop dead extract_openai_* duplicates in clients.py.
extract_openai_reasoning_content, extract_openai_reasoning_details, and
extract_openai_cache_tokens had no callers outside their own definitions
— the live implementations live in src/llm/backends/openai.py. -103
lines from clients.py.
3. Unify on ModelTransport, delete SupportedProviders.
The "google" vs "gemini" split forced a _provider_for_model_config
translation shim in two places. Replace all SupportedProviders usages
with ModelTransport, rename CLIENTS["google"] → CLIENTS["gemini"],
update provider branches + LLMError labels + reasoning-trace entries
accordingly. Trace JSONL now writes "provider": "gemini" instead of
"google" — consistent with the broader env-var rename cutover.
Also tidies up pre-existing basedpyright findings in tests/llm/test_model_config.py
(pydantic before-validator dict inputs + descriptor-proxy call).
ruff: clean. basedpyright: 0 errors, 0 warnings. Tests: 153/153 pass across
tests/utils/test_clients.py, tests/utils/test_length_finish_reason.py,
tests/llm/, tests/dialectic/, tests/deriver/.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(llm): finish the src/utils/clients.py → src/llm/ migration
honcho_llm_call_inner now delegates to request_builder.execute_completion
and execute_stream instead of re-implementing backend call scaffolding
inline. The new _effective_config_for_call helper carries per-call kwargs
(temperature, stop_seqs, thinking_budget_tokens, reasoning_effort) onto
the selected ModelConfig — or synthesizes a minimal config for the
test-only callers that pass provider+model directly. max_output_tokens
is zeroed on the effective config to preserve the current
"per-call max_tokens wins" semantic; honoring ModelConfig.max_output_tokens
is a separable correctness concern.
Side effect of routing through the new path: ConfiguredModelSettings'
thinking_budget_tokens validator now fires on synthesized configs.
test_anthropic_thinking_budget was asserting that a sub-1024 budget
propagated to Anthropic — bumped to 1024 to match what Anthropic actually
accepts.
Unified client construction. Promoted the cached client factories in
src/llm/__init__.py (get_anthropic_client, get_openai_client,
get_gemini_client, get_{anthropic,openai,gemini}_override_client) to
public API and added them to __all__. Promoted
credentials._default_transport_api_key → default_transport_api_key.
Deleted the duplicate _build_client and _default_credentials_for_provider
from clients.py; _client_for_model_config now falls through to the
public factories. CLIENTS dict and _get_backend_for_provider stay as the
mockable seam for the ~50 patch.dict(CLIENTS, {...}) test call sites.
Wired operator-configurable Gemini cached-content reuse end-to-end.
PromptCachePolicy moved from src/llm/caching.py into src/config.py so
ModelConfig can reference it as a field without a circular import;
caching.py re-exports the name for existing imports. Added
cache_policy: PromptCachePolicy | None on ConfiguredModelSettings,
FallbackModelSettings, ResolvedFallbackConfig, and ModelConfig.
resolve_model_config, _resolve_fallback_config, and
_select_model_config_for_attempt copy the field through.
honcho_llm_call_inner passes effective_config.cache_policy into
execute_completion / execute_stream, so operators opt in via
e.g. DERIVER_MODEL_CONFIG__CACHE_POLICY__MODE=gemini_cached_content
and the selection actually fires instead of sitting on a dead path.
New regression test test_cache_policy_reaches_gemini_backend asserts the
PromptCachePolicy object reaches the Gemini backend's extra_params.
ruff + basedpyright: clean. Tests: 154/154 pass across
tests/utils/test_clients.py, tests/utils/test_length_finish_reason.py,
tests/llm/, tests/dialectic/, tests/deriver/.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(llm): move all LLM orchestration into src/llm/ and delete clients.py
The 1624-line src/utils/clients.py has been carved up into focused modules
under src/llm/ and deleted. There is now one golden path for LLM
orchestration and no dual entrypoint.
New module layout:
src/llm/
__init__.py thin stable re-export surface
api.py public honcho_llm_call with retry + fallback + tool
loop delegation
executor.py honcho_llm_call_inner (single-call executor); bridges
to request_builder.execute_completion / execute_stream
tool_loop.py execute_tool_loop + stream_final_response, plus
assistant-tool-message and tool-result formatting
runtime.py AttemptPlan dataclass (replaces the loose
ProviderSelection NamedTuple), effective_config_for_call,
plan_attempt, per-retry temperature bump, attempt
ContextVar
registry.py single owner of CLIENTS dict + cached default and
override SDK-client factories + backend/history-adapter
selection + high-level get_backend(config)
conversation.py count_message_tokens, tool-aware message grouping,
truncate_messages_to_fit
types.py HonchoLLMCallResponse, HonchoLLMCallStreamChunk,
StreamingResponseWithMetadata, IterationData,
IterationCallback, ReasoningEffortType, VerbosityType,
ProviderClient
request_builder.py low-level request assembly (ModelConfig → backend
complete/stream); no longer owns credential resolution
credentials.py default_transport_api_key, resolve_credentials
caching.py gemini_cache_store; re-exports PromptCachePolicy
from src.config
backend.py Protocol + normalized result types
history_adapters.py provider-specific assistant/tool message shapes
structured_output.py
backends/ AnthropicBackend, OpenAIBackend, GeminiBackend
handle_streaming_response had no production callers; it is deleted. The
three tests that used it now drive honcho_llm_call_inner(stream=True,
client_override=...) directly, which exercises the same code path the
public API uses.
Dead credential passthrough removed. The ProviderBackend Protocol and
all three concrete backends no longer accept api_key / api_base — those
are baked into the underlying SDK client at registry construction time
and were being del'd everywhere they appeared. request_builder also
stops resolving and forwarding them.
Client construction is unified. The cached default-client factories
(get_anthropic_client, get_openai_client, get_gemini_client) and override
factories (get_*_override_client) are promoted to public API; the
module-level CLIENTS dict populates from them and remains the
patch.dict(CLIENTS, {...}) mocking seam tests rely on. Old duplicate
helpers (_build_client, _default_credentials_for_provider) are gone.
default_transport_api_key is promoted to public.
Application imports now come from src.llm (dreamer, dialectic, deriver,
summarizer, telemetry-adjacent tests). No code imports from
src.utils.clients anywhere in the repo.
ruff: clean. basedpyright: 0 errors, 0 warnings. Tests: 1013/1013 pass
across the entire non-infra test suite (excluding tests/unified,
tests/bench, tests/live_llm, tests/alembic).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(llm): sanitize tool schemas for Gemini's function_declarations validator
Gemini's native-transport function-declarations validator accepts a narrow
subset of JSON-Schema / OpenAPI: type, format, description, nullable, enum,
properties, required, items, minItems, maxItems, minimum, maximum, title.
Anything else — additionalProperties, allOf, if/then/else, $ref, anyOf,
oneOf, $defs, patternProperties — triggers an INVALID_ARGUMENT 400 at call
time.
Our agent tool schemas in src/utils/agent_tools.py use several of those
(additionalProperties: false, allOf + if/then conditionals) because they
were authored for OpenAI strict-mode + Anthropic, which need the richer
vocabulary. GeminiBackend._convert_tools was passing them straight through.
Add _sanitize_schema(): walks the parameters tree and drops unsupported
keywords while preserving semantics for the keywords that hold user data
(properties maps field-name → sub-schema; required / enum are lists of
literals; items is a single sub-schema). Other backends are untouched and
continue to receive the full strict schemas.
Regression tests:
- test_gemini_sanitize_schema_strips_unsupported_keywords: confirms
additionalProperties, allOf + if/then, and $defs are stripped at nested
levels while legitimate fields survive.
- test_gemini_convert_tools_sanitizes_parameters_schema: end-to-end
_convert_tools output has no forbidden keys.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: fix tool calling syntax for gemini
* refactor(llm): normalize defaults, widen OpenAI reasoning-model routing
* chore: fix test
* fix(llm): address post-migration review feedback
* fix(llm): gemini robustness + dreamer specialist ergonomics
* chore: addres review comments
* chore: (docs) unrelease changelog addition
* chore: (docs) merge commit changes
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Erosika <eri@plasticlabs.ai>
Joe-Kneeland
Vansh Sharma
Niyaz Almufti
Vineeth Voruganti