At end of drain the pump enqueued the sentinel first and picked the
billing mode from client_detached afterward, so a client that consumed
the sentinel and tore the relay down before the pump resumed (possible
whenever the sentinel enqueue hit a full queue) had its fully delivered
response billed through the teardown path, skipping the proxy's
post-response hook. Bill or park before the sentinel goes out, and let
an unconsumed sentinel fall back to dispatching the parked billing.
Shadow eval jobs previously targeted only virtual keys, so deployments on
pure JWT auth (which present no key at all) could never sample their
traffic. Jobs now carry a typed (target_type, target_id) pair covering
keys, teams, and users; sampling matches the identity every request
resolves to at auth time, so team and user jobs cover JWT traffic with
no client changes.
Resolves LIT-6578
* feat(complexity_router): escalate oversized prompts to a tier that fits before dispatch
The classifier scores complexity and never prompt size, so a long agentic
session whose newest ask is trivial classifies SIMPLE onto a small-window
tier and the provider rejects it with a context-window 400 that nothing
retries. The gate runs after classification on every decision path
(classify tail and session-affinity pin), estimates prompt tokens
including the out-of-band carriers (top-level system, tools,
instructions), and when the decided tier provably cannot hold the prompt
moves the request to the lowest configured tier with a model whose
declared window fits, restricting the pick to fitting models when the
decided tier can keep it. Models with no resolvable window are never
escalated away from or onto, escalated decisions are never written as
session pins, and the decision records context_escalated plus the
original tier in spend logs.
Resolves LIT-6503
* fix(complexity_router): judge groups by smallest window, bound skips by bytes, filter adaptive picks
Review-round rework, one mechanism per finding. A group is judged by its
smallest resolvable deployment window, since the core router picks within
a group with no fit check. The counting skip is gated on UTF-8 byte
length, which BPE token counts can never exceed, so token-dense scripts
cannot slip past it; only a real tokenizer count ever moves a request and
a failed count leaves the placement alone. The fit facts now filter every
adaptive phase including cold start and the tier fallbacks. Window
questions adopt the declared provider and never resolve authenticating
providers, and a router instance without get_model_list degrades the gate
to a no-op. Tests rebuilt on real Router instances resolving deployment
model_info end to end, plus a full-path test through
async_get_available_deployment
A guardrail modify_response verdict on a streaming request only produced a
proper replacement on /v1/messages: the chat completions and Responses API
translations had no build_block_sse_chunks, so the ModifyResponseException
re-raised and surfaced as an in-stream 500 error frame (or a whole-request
500 in buffered mode) instead of the documented 200 replacement.
Implement build_block_sse_chunks for both OpenAI translations: chat emits a
content delta plus a finish_reason content_filter chunk with real usage;
Responses emits the typed event sequence (standalone via
build_synthetic_response_events pre-stream, or an output-item continuation
under the in-progress response id mid-stream) ending in response.completed.
GigaChat reports prompt_tokens and total_tokens after subtracting cached
tokens (the docs example is prompt_tokens=1, precached_prompt_tokens=37,
total_tokens=5, so the fields are disjoint, not a subset). Map to the
OpenAI convention by adding precached_prompt_tokens back onto prompt and
total while still surfacing it as prompt_tokens_details.cached_tokens.
The base added router_metadata to SpendLogsMetadata in #39001 without
updating this fixture, and its CI run never executed logging_testing,
so the job now fails on every branch merged with current staging.
precached_prompt_tokens is a subset of prompt_tokens (OpenAI cached_tokens
semantics), so map it to prompt_tokens_details.cached_tokens instead of
adding it on top of prompt/total. Emit stream usage from any final chunk
carrying it rather than only finish_reason stop, which dropped tokens for
function_call and length streams. Merge auth metadata into a new dict in
the gigachat router handler instead of mutating the shared parsed-body
cache in place.
* fix(otel): emit cache token counts on OTel v2 LLM spans
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): trim comment in LLMUsage adapter
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): drop casts in LLMUsage cache token adapter
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(deps): bump restrictedpython to 8.3 for GHSA-ffg3-p8fm-mjx2
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The advisory was modified 2026-08-31 and flags mlflow 3.13.0 through
3.15.2 with no fixed release published, so every osv-scan run fails
with nothing to bump. Same treatment as the existing diskcache entry.
`ChatCompletionFileObject` is in the union `_count_content_list` accepts, but
`file` was missing from its match, so every local count of a Responses
`input_file` raised `Invalid content item type: file`. On
/v1/responses/input_tokens that surfaced as an opaque 500 whenever the model's
provider counting API refused the block and the local tokenizer took over.
Count it the way the module already counts the same thing in Anthropic's
dialect: the filename like a document title, the inline bytes through the
image pricer.
Since the requires-python cap moved to <3.15, uv resolved the project
python to 3.14, downloaded a managed interpreter under
/root/.local/share/uv that the runtime stage never receives, and every
layer-cache-miss image build broke: first at uvloop's cp314 sdist
configure step, then, with file/make added, at the runtime stage where
the copied venv's python symlink dangles and prisma imports fall through
to the system python. UV_PYTHON_DOWNLOADS=0 (already the convention in
migrations/backend/gateway) roots the venv on the apk python3.
The wolfi-base digest bump is required alongside it: the pinned 08-22
base ships glibc-2.43 while the current apk python-3.13 needs
GLIBC_2.44, and wolfi version-names glibc packages so apk upgrade
cannot cross that boundary.
With the venv on system 3.13 every dependency installs from wheels
again, so the file and make packages added for the sdist build are
reverted.
The Responses-to-chat transform dropped the filename OpenAI requires next to
file_data, so a request carrying an inline PDF counted 13 tokens instead of 36
and a real completion through the chat bridge got a 400.
- sync llm_passthrough_route: read and close an error-status streaming
response before mapping it, so upstream 4xx/5xx surface as the provider
error instead of httpx.ResponseNotRead
- AsyncPassthroughStreamingResponse: expose aiter_bytes() and carry
_hidden_params so the router attaches headers in place instead of
wrapping the stream in HiddenParamsAsyncIteratorWrapper, which 500'd
every streaming azure router-model passthrough request
- logging: swap the passthrough httpx result for the transformed
ModelResponse/EmbeddingResponse when firing success callbacks
- get_llm_provider: resolve gigachat from its api base and drop the dead
gigachat_models elif branch
- constants: register the gigachat api base in openai_compatible_endpoints
