* refactor(rust): rename litellm-core to litellm-inference
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(rust): expose inference base API for format crates
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(rust): move shared inference test helpers behind test-support
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Yujong Lee <yujong@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(ui): wait for the classifier model popup before picking its option
Base UI exposes its select option asynchronously, and the helper waits for the option and its positioner to become clickable before selection
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: isolate two router tests from work leaked by earlier tests
Collect earlier tests' garbage before warning capture so their unawaited coroutines cannot be attributed to the target's warning assertion
Filter success-event callbacks by the request's litellm_call_id so queued logging work cannot replace the current request's captured messages
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>
* refactor: remove fresh tech debt from the 2026-10-05 window
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(lens): drop the review models left unused by the dead review helpers
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>
* fix(router): price a model group from the deployments that serve it
A model group's info read composed the group's own model_group_alias
entry into its deployments, a hop the router never takes: an alias is
resolved exactly once at request time, so a group reached as an alias
target is served by its own deployments. For the chain X -> T -> U the
price read for X included U's deployments too, and the free-model
budget waiver refused a free request to X on an over-budget key, while
GET /model_group/info reported U's providers and price for X.
The group info read now prices a group from the deployments routing
serves it with: the ones named after it, the routing group of that
name, or the wildcard route matching it when neither exists. The
budget waiver, GET /model_group/info, the rate limiters, and the
response headers all read the same set as routing. get_model_list
keeps its behavior for every other caller.
* test(router): give the paid fixtures explicit per-token prices
* test(router): call the routed-group read by name so the router coverage gate sees it
* test(integration): audit the alias chain budget waiver on every route, shape, and outage
Thirty-four cells under the management group prove an over-budget key is served through an alias chain entry at the price of the deployment that serves it, on chat, responses, and messages, sync and streamed, through the OpenAI and Anthropic SDKs and raw httpx on both replicas, with the chain middle, the reverse chain, a cost-map priced middle, a ghost middle, wildcard and routing-group targets, malformed and hostile inputs, a cached reply, a repointed alias, a provider failure, and two chaos bursts (a killed worker, a scripted outage)
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* perf(types): defer pydantic schema builds via shared LiteLLMBaseModel
Add LiteLLMBaseModel with defer_build driven by DEFER_PYDANTIC_BUILD (default true) and move litellm and enterprise pydantic models onto it
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(types): build deferred models created by a parent validator; keep lens worker models litellm-free
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(types): link pydantic issue on deferred-build rebuild hook
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: nate <nate@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(vertex_ai): add regional endpoint uplift to gemini-3.1-flash-image
Google prices Gemini 3.1 Flash Image at 1.1x on non-global endpoints for
input, text output and image output, but the cost map row had no
regional_endpoint_uplift_multiplier, so regional calls billed at the
global rate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(integration): cover regional uplift spend for gemini-3.1-flash-image
* test(integration): read the proxy salt from the environment in the uplift spend cells
---------
Co-authored-by: Nate Armstrong <narmstrong@Nates-MacBook-Pro.local>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(responses): honor caller stream flag when provider forces SSE
The Responses handlers decided whether to hand back a streaming iterator
from the provider payload's `stream` field, which chatgpt sets
unconditionally because the Codex backend only serves SSE. A caller that
sent `stream: false` therefore received a raw SSE stream on /v1/responses,
and the chat-completions bridge failed with "Unknown items in responses
API response: []" once its recovery path lost the raw SSE it reads from
Transport streaming still follows the provider payload; only the caller's
own `stream` value now decides the response shape. When the provider
forces SSE for a non-streaming caller the body is read and aggregated
through the existing path
* test(chatgpt): inject the authenticator into the responses config so handler tests never log in
* fix(responses): treat an extra_body stream flag as the caller's own and drop a redundant comment
* test(integration): audit the chatgpt caller stream flag across responses, chat and messages
---------
Co-authored-by: SeongWoon Cho <coffee@soylatte.kr>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(chatgpt,github_copilot): refuse device-code login when an event loop is running
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(github_copilot): drop stray whitespace change
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(chatgpt): bound token refresh timeout and drop placeholder assignment
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(chatgpt,github_copilot): keep the token file path out of the event-loop 401 message
* fix(auth): refuse device-code login from worker threads too
/v1/messages runs its handler in an executor thread, where the
running-loop check never fires, so a chatgpt or github_copilot model
still started the interactive device-code login there and the request
hung for up to 15 minutes. The guard now also requires the main thread,
so the login only runs where a human can actually answer it.
* test(chatgpt): keep authenticator tests out of the real token directory
* fix(chatgpt): keep the 5 second connect timeout and the operator's request_timeout on the token refresh call
* test(integration): cover the device-code login guard on the proxy and the SDK
---------
Co-authored-by: mateo <mateo@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(proxy): cap batch file records, daily batch uploads, and per-file downloads
Adds three opt-in limits for batch jobs, each settable in general_settings as a
per-key default and overridable in key or team metadata by a proxy admin:
max_batch_file_records rejects a purpose=batch upload with more request lines
than allowed with a 413 before it reaches the provider.
max_batch_file_uploads_per_day counts accepted batch uploads per key and per
team in a UTC day and returns 429 with Retry-After once the count is used.
max_file_downloads_per_minute counts GET /v1/files/{id}/content per key and per
team for each file in a one-minute window and returns 429 with Retry-After.
The existing admin-only guard for batch_enqueued_token_limit now covers all four
metadata keys.
* fix(proxy): keep file usage counters in their own store and gate batch limits on user creation
* fix(proxy): count keyless JWT callers per user, require positive file caps, and take the upload slot after request validation
* refactor(proxy): end file usage cap describers with an explicit return after the match
* fix(proxy): keep file usage counters when more than 200 are live without Redis
The file usage counter store used a default in-memory cache, which holds 200
entries and evicts the one that expires soonest. Without Redis, a caller got a
fresh per-file download allowance after touching about 200 other file ids in
the same minute, and a key got a fresh daily upload allowance once about 200
other keys had uploaded that day. The store now tracks up to 20,000 live
counters per worker, the same bound the login throttle uses
* test(proxy): move the file usage cap tests into the directory the proxy shard runs
Main's shard coverage check found tests/unit/proxy/openai_files_endpoints
claimed by no shard, so its tests would not run in CI. The file moves next to
the other files endpoint tests in tests/unit/proxy/openai_files_endpoint, which
the proxy-endpoints shard already runs
* fix(proxy): declare the file usage counters as rate limit calls
Main's redis producer gate requires every module that writes a shared cache to name its key family, and the file usage counters wrote theirs without one.
* test(files): audit batch file usage caps across processes, Redis outages, and config reloads
* test(files): guard the chaos cells against minute boundaries and open Redis breakers
Two chaos cells each failed once in the audit run. The restart check ran
three sequential downloads with no guard against straddling a UTC minute,
and the exact-cap probe after a Redis outage ran while both workers' Redis
circuit breakers were still open (60 s default recovery), so it counted in
per-process memory and the two workers split the cap
Every burst now carries a window guard, the chaos fixture lowers the
breaker recovery to 2 s, and the post-outage check drives a fresh key to its
cap through a one-worker sibling proxy and then expects the two-worker
candidate to refuse the whole burst, which only the shared Redis count can
produce, polled until the breakers close
* test(files): release the held uploads when the killed-worker cell fails early
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(vertex_ai): apply regional endpoint uplift on image generation cost path
completion_cost had vertex_location but never passed it to the image
generation cost router, so a regional_endpoint_uplift_multiplier on a
Vertex image row would be ignored. No image row carries the multiplier
yet, so nothing is misbilled today. Pass the location through to the
Vertex image calculator for both the token-based price and the
per-image fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(vertex_ai): cover regional image cost through the proxy logging path
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(images): hand image_edit's vertex_location to its cost resolver
* test(integration): audit Vertex image regional uplift billing
* test(integration): require every spend row after a proxy restart
* test(integration): reject extra spend rows after a proxy restart
---------
Co-authored-by: Nate Armstrong <narmstrong@Nates-MacBook-Pro.local>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(proxy): judge the free-model budget waiver by the group an alias routes to
A hidden model_group_alias can reuse the name of a real model_name. The
router serves that name from the alias target, but two of the three checks
behind the free-model budget waiver read the deployments of both the alias
target and the real model that shares the name.
So an over-budget key was served through an alias whose name belongs to a
model with an explicit $0 price when the target is $0 only by the cost map,
which the same key is refused on by its own name. The mirror case refused a
free target because the shadowed name belongs to a PTU-priced deployment.
Resolve the alias once and have the explicit-cost and PTU checks read the
routed group, the same group the price check already reads.
* test(proxy): cover the plain alias form of a shadowed free model name
The same wrong verdict exists for a plain string alias, so the unpriced-target case now runs for both alias shapes.
* fix(proxy): judge an alias chain's budget waiver by the deployments it is served from
The explicit-price and PTU checks read the alias target through
Router.get_model_list(), which follows a second alias hop when the target is
itself an alias key. The router never takes that hop, so an alias chain was
judged by a deployment the request never reaches. The checks now take the
deployments named after the routed group, or the wildcard deployment serving
it when none carries its name.
* fix(proxy): refuse the budget waiver when an alias chain is served by a priced wildcard route
* test(integration): cover the shadowing alias budget gate end to end
Adds the audit cells for a hidden alias whose name shadows an explicitly
free group: streamed SDK refusals on chat, responses and messages,
embeddings, the free wildcard and mixed-group paths, per-model budgets,
JWT and custom auth callers, cache hits, alias removal under traffic,
provider failure and fallback, a concurrent outage burst, and a worker
kill on an owned two-worker proxy
* test(integration): cite the cost-map rows the shadowing alias cells rely on
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(ci): namespace claude session ids in tracing seeds and allowlist /v1/logs on backend
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): allowlist /v1/logs on the gateway alongside /v1/traces
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>
* fix(gemini): stop replaying thinking block signatures to Gemini
A thinking block's signature has no provenance, and LiteLLM never fills it
from a Gemini response (Google signs text and functionCall parts, which ride
provider_specific_fields and the tool call id), so a Claude signature replayed
through a mixed model group reached Gemini as a thoughtSignature and Google
answered 400 Invalid thought signature on every later Gemini-served turn. The
same replay also sent the thinking text a second time as a plain text part.
The thinking text now goes out once, as the thought part built from
reasoning_content, and no part is built from thinking_blocks
* test(gemini): type the parts helper and split its comprehension
* test(gemini): cover thinking signature replay on the integration rig
Two integration files from the audit of the foreign thought signature fix: 62 wire cells asserting the model turn Google receives on chat, messages and responses across gemini and vertex_ai, streaming and not, SDK and httpx clients, the sad shapes of thinking_blocks, context caching through cachedContents, and 3 chaos cells (a concurrent burst across endpoints, upstream stream drops, a worker SIGKILL mid burst)
* test(gemini): read the integration salt from the environment
The wire test decrypted Responses ids with a literal salt; tests/integration/_support/process.py boots the proxy with LITELLM_SALT_KEY when it is set, so the test now reads the same variable with the same default
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(prometheus): cap series per metric for every labeled metric
Add prometheus_metrics_max_series_per_metric: per metric and per worker process, the first N label
sets keep a series of their own. Counters and histograms record every later label set on one series
whose labels are all "other", so totals stay exact, and gauges skip it. The cap holds with multiple
workers because it never needs to remove a series.
Add prometheus_metrics_ttl_seconds: a series idle for that long is removed and its slot is freed.
The prometheus client cannot remove a series in multi-process mode, so the TTL is ignored there with
a startup warning.
Both settings are off by default. The end_user caps are unchanged.
* fix(prometheus): share the series cap across workers of one proxy instance
Workers writing to one PROMETHEUS_MULTIPROC_DIR now agree on which label
sets get a series through an append-only admissions file per metric, so a
merged scrape stays at the cap plus `other` instead of growing with every
worker and every worker restart. The two fallback counters now pass their
label names as a keyword so the cap and prometheus_exclude_labels apply to
them, admission and child creation happen under one lock, the test fixture
restores the shared registry, and the `other` label value lives in
constants.py.
* test(prometheus): check emitted labels instead of wrapper types, close the admission match
The exclude-labels test now emits through the spend and provider budget
metrics and checks the scrape keeps all their labels. The admission match
arms end in assert_never so the match is exhaustive.
* fix(prometheus): return the exhaustive-match fallback so every admission arm returns
* fix(prometheus): pick the series tracker with isinstance so every path of _admits returns
* fix(prometheus): skip an admissions line a worker could only write part of
* fix(prometheus): frame each admissions record with newlines so a cut-off record cannot swallow the next
A record a worker could only write part of used to merge with the next worker's record, and both were skipped for one request. Each record is now written between two newlines, so the fragment is a line of its own. The clock fixture in the series tests starts from a constant instead of reading the real clock
* fix(prometheus): ignore a non-positive series cap or TTL with a warning instead of failing the logger
A cap or TTL of 0 or less raised at logger init. The proxy logs that as a non-blocking error and keeps serving, so the result was a running proxy with no Prometheus metrics at all. The setting is now ignored with a startup warning naming it, the same rule the end_user cap already follows for a non-positive value
* fix(prometheus): start the series cap over on a one-worker restart and audit it live
A proxy with one worker and an operator-set PROMETHEUS_MULTIPROC_DIR now drops litellm's admission files at boot, so a restart frees every slot there the way it already does with several workers. A cap or TTL that is not a number greater than 0 (a bool, a non-numeric string, an empty value) is ignored with the startup warning instead of breaking the logger
The integration cells drive the cap on every endpoint through the OpenAI and Anthropic SDKs and raw httpx, streaming and not, plus gauges, cache hits, failures, both workers of one instance, the TTL on one worker and its warning on two, ignored settings, excluded labels on the fallback counters, a null cap, /config/update, a concurrent burst scraped mid-flight, a provider outage, a killed worker, and restarts with one and two workers
* fix(prometheus): wipe an operator-set multiprocess directory on a one-worker boot too
* fix(prometheus): leave the multiprocess directory alone on a setup-only run
A run with --skip_server_startup starts no worker, so it no longer creates or
wipes PROMETHEUS_MULTIPROC_DIR. Wiping there deleted the samples of a proxy
already running against the same directory
* fix(prometheus): free the capped series slots when a gateway or backend container restarts
The component image entrypoint starts uvicorn without the proxy CLI and wiped only the .db sample files at container start, so the admitted-series files of the previous container survived an in-place restart. Every label set seen after the restart was then counted on `other` once the previous container had filled the cap
* test(prometheus): cover a setup-only run and a gateway image restart under the cap
Two integration cells from the audit: a `--skip_server_startup` run pointed at a live
two-worker proxy's operator directory leaves its samples alone, and the gateway image
(`docker/component_entrypoint.sh` running `python -m gateway.launch`) restarted on a kept
PROMETHEUS_MULTIPROC_DIR starts the cap over. The burst cells now wait for every counter
they assert on, since the request and failure counters of one call increment at different
points of the logging callback
* test(prometheus): prove the cap reaches the fallback counters in the X1 cell
* fix(prometheus): ignore a cleanup interval that is not a number of at least 0
A string or negative prometheus_metrics_cleanup_interval_seconds reached the
series tracker unvalidated, so the first labeled emit with a TTL on raised
TypeError inside the callback and recorded no series. The interval is now
validated the way the cap and the TTL are: an invalid value is ignored with a
warning and the default 60 seconds applies. The I2 integration cell drives a
string interval through a live proxy and reads the warning from its log
* test(prometheus): give the restart cells the boot budget of their siblings
C4 and C5 boot two proxies each and hit the file's 240 s budget on a loaded
box; C3 and D1 already carry 420 s
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(lens): restate response contract during model repair
* fix(lens): separate instructions and recover rejected results
* fix(lens): correct loop type annotations and checks
* fix(lens): preserve access to prior findings after compaction
* fix(lens): cap result retries and preserve partial completion
* fix(responses): honor request cache controls on chat completions bridged to the Responses API
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): assert spend and cache-hit status on bridged no-cache rows
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): unskip LIT-9196 openai_responses basic translation cases
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): restore azure attribution check on bridged no-cache rows
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): exercise bridged cache controls through a real local cache instead of patched responses
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* Add Reka as an OpenAI-compatible provider
* Add Reka supported endpoints
* fix: remove stray fragment after reka entry in provider_endpoints_support.json
Co-authored-by: Cursor <cursoragent@cursor.com>
* test(reka): register provider identity and cover routing, credentials, and bridged endpoints
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(lens): show investigation findings for agent traces
Replace child tool-error counts in the trace table with distinct investigation findings. Keep unassessed traces separate from completed clean investigations and use the root status for failure filters and timeline counts
* fix(lens): stabilize findings updates and repair UI checks
* fix(lens): allow viewer findings reads and index trace lookups
* test(lens): cover viewer findings reads with Postgres
* fix(router): retry a /v1/messages stream the provider drops before the first content chunk
A /v1/messages stream that the upstream closed before any content reached the
client answered an error event after a single attempt, so the router's
num_retries never applied to that drop. The pre-content failure is now retried
within the model group before the fallback chain runs, with the budget resolved
the way a failure raised before the stream opened resolves it: a retry policy
that names the error class, then the request's num_retries, then the
deployment's, then the router's. A drop after content reached the client keeps
surfacing the provider's error after one attempt.
Fixes#44238
* fix(router): hand a retry's non-retriable error to the fallback chain and type the retry helpers
A retry that failed before its stream opened with an error no retry covers raised straight to the
client, skipping a fallback the first attempt would have used. assert_never now comes from
typing_extensions so the router imports on Python 3.10, and the retry helpers read their kwargs
through typed narrowing instead of Mapping[str, Any]
* fix(router): cast the untyped router fallback defaults the stream retry gate reads
The retry gate passed the router's fallback attributes, declared without element types, to the
typed request override helper, which basedpyright counted as new unknown-argument errors
* fix(router): consult context_window_fallbacks when a retried /v1/messages stream overflows
A retry attempt raising ContextWindowExceededError reached the fallback chain inside its
mid-stream envelope, so only the regular fallbacks list matched. The fallback attempt now
unwraps it the way it unwraps a content policy error. The new router helpers are covered for
the router code coverage check with two direct-call tests and named covering tests
* fix(router): retry a 408 raised by a /v1/messages retry and honor deployment num_retries before the stream opens
* fix(router): attribute a retried /v1/messages stream to the deployment that served it and bound the retry-policy hold
* fix(router): retry /v1/messages error frames under their retry-policy class and keep the first drop's committed budget
An `event: error` frame that arrives before the first content delta now raises the exception class the pre-stream mapping gives an HTTP answer with the same status (429 RateLimitError, 500 and 529 InternalServerError, 503 ServiceUnavailableError, 504 Timeout), so a retry policy's per-class budget governs it the way it governs the error before the stream opened. The status the client sees is unchanged
A retry that lands on a sibling deployment keeps the budget the first drop committed to, read back from the request's attempted_retries and max_retries, instead of recomputing it from the new deployment's num_retries, matching the pre-stream retry loop
* refactor(anthropic): keep the error-frame exception mapping under llms and type the retry test helper
The status-to-exception mapping an `event: error` frame gets before the retry policy is consulted now lives next to the Anthropic error status map in llms/anthropic/common_utils.py, with its own unit test, and the two-deployment retry test helper takes explicit typed parameters instead of a bare dict and untyped kwargs
* refactor(anthropic): map an error frame's status with explicit returns on every path
* fix(router): map stream error frames through the pre-stream exception mapping
An overloaded `event: error` frame on a /v1/messages stream now raises the InternalServerError a 529 answer maps to, built by exception_type from the frame's own body, so one retry policy class governs the error before and after the first byte; a failed fallback after such a frame answers 500 like every other litellm path instead of the frame map's 503
A model_group_retry_policy that does not parse (a non-integer budget, an entry that is not a mapping) no longer fails every healthy stream of that group before its first attempt: the stream runs with no policy and the plain num_retries budget, with a warning naming the group
* fix(router): forward an error frame nothing can take over for as the provider sent it
A pre-content error frame whose class the retry policy grants no retry, with no fallback configured, raised an HTTP error only on the first attempt while the same frame after exhausted retries reached the client verbatim. Both now pass through as sent, the way the merge base forwarded every frame.
* test(integration): audit /v1/messages pre-content retry across routes and budgets
Adds the /audit cells for the pre-content stream retry: the native Anthropic route
(drops and error frames before content, HTTP rejections before the stream opens, SDK
sync and async, after-content and non-retriable controls, budget exhaustion, cache
twin, spend row and headers), the chat and responses bridges, the generic routes
(responses, chat, vllm pass-through, Gemini generateContent, fine-tuning jobs list),
owned two-worker proxies for router-level budgets, retry policies and fallbacks, and
two chaos cells (a worker killed mid burst, an outage on every first attempt). Shared
helpers for scripted Anthropic SSE upstreams and OpenAI-compatible wire replies live
in tests/integration/_support
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(proxy): let a listed team alias win over a same-named key alias in the customer model check
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(proxy): check each requested name in a plain loop in can_customer_access_model
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): only let a listed team alias skip the customer check when its target is live
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(proxy): move the per-name customer alias check into a local function
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(spend-tracking): stop caching failed spend-log metadata lookups as confirmed misses
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(spend-tracking): share the short-lived miss cache write
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(spend): cover key alias recovery after spend log lookup failures across usage routes
* test(spend): bound outage alias lookups per miss window instead of a fixed count
* fix(spend-tracking): treat any spend-log lookup failure as a short-lived miss
The Prisma client raises a plain AttributeError when the database drops
the connection mid-query, so the PrismaError catch let it through and
the whole usage call answered 500. Any failure now keeps the 30 second
backoff only, and the integration proxy patches its test entitlement at
import so uvicorn's spawned workers inherit it
* test(integration): audit spend-log metadata recovery under timeouts and dropped connections
Cover the daily activity routes, the usage AI chat, the Vantage and
CloudZero dry runs and exports under a locked spend-log table and under
a database connection dropped mid-lookup, on a two-worker proxy, with
the recovery after the outage asserted through the proxy's own miss TTL.
Add a dropped_connection_relay that closes only the connection whose
bytes carry a trigger, so a cell can drop the one connection the
recovery query runs on while the rest of the pool keeps serving. Rewrite
the sweep and JWT cells for the merged main: the export route reads
metadata by SQL join and never calls the recovery, the search routes
answer key rows and find deleted keys by alias, and the daily-spend
owner recovery names the user while the alias stays blank. The sweep
cell now times out a second lookup under the same lock, which pins the
keys blank on the merge base and recovers on this branch.
* test(integration): match a dropped-connection trigger split across two reads
The dropped-connection relay checked each TCP read on its own, so a SQL
marker that straddled two reads never tripped it and the outage cells
would run without the outage they meant to exercise. Carry the tail of
the previous read into the next check, as the held-statement relay
already does, and pin that with a unit test that splits the trigger
across two writes.
* test(integration): scan relay triggers through an in-process helper
The dropped-connection relay now matches its SQL trigger through a TriggerScanner that carries the previous read's tail, and the unit test exercises that scanner directly instead of opening loopback sockets, which tests/unit forbids. The relay's end to end behavior stays covered by the integration cells
---------
Co-authored-by: gabriele <gabriele@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(lens): add per-trace review models to jobs and progress
* feat(lens): append worker reviews to the job, capped, and count every review
* feat(lens): report a review with reasoning for each screened trace
* chore(ui): regenerate api types for lens job reviews
* feat(lens): type job reviews and fill them in lens fixtures
* feat(lens): add live review playback model
* feat(lens): pick the analysis model and slow single-review pacing
* feat(lens): add sample reviews for previewing the live run
* feat(lens): add live run layout with queue, reading trace and conclusions
* feat(lens): show the live run on investigations and open it from run now
* feat(lens): stream large review backlogs at 150ms or less and list newest first
* fix(lens): show the live run only for real reviews and keep fixtures test-only
* refactor(lens): restyle the live run as the native progress panel
* fix(lens): retry contended investigation updates with jittered backoff
* feat(lens): add a reading ticker line and replay for finished runs
* feat(lens): collapse the live run to an ambient line with show work
* fix(ui): crop the cerebras logo viewBox to its mark so it reads at icon size
* feat(lens): format review span previews as readable messages
* feat(lens): derive strip status, honest issue counts and drawer focus from a job
* feat(lens): track active jobs before their first review
* feat(lens): add a live trace results drawer with readable spans
* feat(lens): put the live strip under the progress bar and drop the inline panel
* feat(lens): add an ambient live strip that opens the drawer
* fix(lens): wait out provider rate limits and retry model calls four times
* style(lens): format repository contention tests
* feat(lens): read review spans as a conversation timeline
Turns spans into the user's ask, tool calls with args and results, and the agent's reply, dropping system prompts. Also handles a preview cut that lands inside the Output header.
* fix(lens): list recorded agents in the run now dialog
The run now agent field used a native datalist, whose suggestions do not show inside the modal dialog, so the agent list looked empty even though /lens/agents returned names. Use the same Combobox as investigation setup.
* feat(lens): pace live playback so each trace stays readable
Every trace now stays up for at least 1.5s. A backlog is cleared by skipping to the newest few instead of flickering through them. Conclusions count traces per check and kind, and new helpers cover share bars, group filters and flashes.
* feat(lens): keep the live run ambient until View run is clicked
The drawer no longer opens on Run or when entering a running investigation. LiveRun takes reviews as a prop so it can move to a dedicated reviews endpoint.
* feat(lens): show the live run as a two-pane trace and conclusions view
Left pane: the trace being reviewed as a readable timeline, followed by Lens's reasoning and the verdict. Right pane: ranked conclusion groups with share bars, plus a trace list you can filter.
* fix(lens): run several investigations per worker and poll every two seconds
* feat(lens): add worker slot and poll interval settings
* feat(lens): add list summaries and an incremental review filter
* perf(lens): strip reviews and run attributes from the lens list and serve reviews separately
* test(lens): cover list summaries, review polling and review access
* feat(lens): explain why a queued investigation is waiting
Works out whether no worker is connected, the worker is busy (with its running investigations and an estimated start time), or it is just being picked up.
* feat(lens): show the queue reason and what the worker is doing in the live strip
The progress header and the strip replace "Queued for your worker" with the concrete reason. While waiting, the strip lists the busy worker's investigations; click one to open it.
* feat(lens): add a review page model carrying the total reviewed count
* fix(lens): page live reviews by index so out-of-order reviews are never skipped
* feat(lens): take an index cursor on the reviews endpoint
* test(lens): cover index cursors across out-of-order and rolled-over reviews
* chore(ui): regenerate api types for the lens reviews endpoint
* feat(lens): page job reviews by index cursor
Adds api.reviews for GET /lens/{id}/runs/{job}/reviews?after=N, with a demo implementation. appendPage adds pages in arrival order and keeps the latest 200. liveJob now keys off reviewed, since the list no longer carries reviews.
* feat(ui): add a lens reviews query that polls the index cursor while live
* fix(lens): feed the live run from the reviews endpoint and keep View run open
LiveRun now gets its reviews from useJobReviews instead of the list, which no longer carries them. View run stays clickable while a run is queued or running, and before the first trace the opened view says what the worker is doing.
* fix(lens): split live conclusions into issues and patterns
A check could show up twice with the same label, once as an issue and once as a pattern.
* fix(lens): group live conclusions by check with short labels
There is now one group per check_id: issue traces are the main count and pattern traces a secondary note, so there are no duplicate red and grey cards. A long instruction falls back to the humanized check id. Adds briefReasoning and traceRows for the simplified trace list, and drops helpers nothing uses.
* feat(lens): simplify View run to traces and conclusions
The left pane is the trace list. A soft highlighter carrying the provider and model slides to the trace being reviewed, and clicking a row shows just Lens's reasoning and verdicts. The right pane keeps one conclusion card per check.
* refactor(lens): drop client-side replay in favour of real in-flight rows
Removes the playback reducer and its pacing. liveRows lists the traces the worker is reading, from job.reading, followed by completed reviews newest first, keyed by execution_id so a trace keeps its row when it finishes.
* feat(lens): show what the worker is reading and make View run obvious
Each trace in flight gets a highlighted row with the model and a live timer, and becomes its completed row in place. Completed rows show the real review time. View run is an outline button next to the progress line, and clicking anywhere on the strip opens it too.
* feat(lens): sum up a finished live run with time taken
doneLine reads like "Reviewed 30 traces in 31s with", measured from when reading started.
* feat(lens): slide one model rectangle over the traces being read
A single rounded rectangle carrying the provider logo and model wraps the real in-flight rows from job.reading. It translates and resizes over 250ms as traces finish in place. Before job.reading arrives it sits on a top slot showing the honest progress line, and when the run completes it fades out over 400ms. Rows have a fixed height and stable execution_id keys, so polls don't cause jumps or flicker.
* feat(lens): add in-flight runs to jobs and worker progress
* feat(lens): store in-flight runs from progress and clear them when a job ends
* refactor(lens): route progress, cancel and results through shared job transitions
* feat(lens): report each run as in flight when its review starts
* feat(lens): send in-flight runs with worker progress
* test(lens): cover in-flight runs across progress, old workers and terminal states
* test(lens): cover in-flight reporting under original run ids
* chore(ui): regenerate api types for lens in-flight runs
* feat(lens): model live reading lanes from in-flight runs and reviews
* feat(lens): show a now reading stage that types each trace's reasoning
* feat(lens): put the now reading stage above the trace list in View run
* fix(lens): resolve the analysis provider logo from the model catalog
* fix(lens): give demo jobs an empty in-flight list
* style(lens): format endpoint tests
* refactor(lens): name the run now handler in investigations view
* refactor(lens): name now reading conditions
* refactor(lens): name inline objects in the live run
* style(lens): format live run files
* fix(lens): keep worker settings inside the standalone worker package
* refactor(lens): keep update retry settings next to the repository
* fix(lens): start review history over when a run is reclaimed
* chore(lens): drop the unused review fixture
* refactor(lens): remove dead live helpers and use generated in-flight types
* fix(lens): keep polling a finished run until its last reviews arrive
* perf(lens): tick fast only while reasoning is typing
* fix(lens): isolate retried reviews and finding identities
* fix(lens): space the model name in run summary
* feat(lens): integrate confined workspace analysis with live reviews
* fix(lens): synchronize confined Python process monitoring
* Update review.md
* fix(lens): allow mixed context capacities and correct review assertions
* fix(lens): retrieve evidence on demand and isolate failed reviews
* fix(lens): isolate incomplete evidence reads from peer reviews
* test(lens): await trace status filter option
* test(lens): wait for reclaimed review state to settle
* fix(lens): recover from incomplete cross-session evidence
---------
Co-authored-by: Ishaan Jaff <ishaan@berri.ai>
* fix(sso): let CLI and Claude Code gateway sign-in through on DISABLE_ADMIN_UI nodes
* test(proxy): cover CLI SSO sign-in on a UI-disabled node
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(proxy): add models column to the end user table
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(proxy): enforce the end user models allowlist in model access checks
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(proxy): accept and return models on the customer endpoints
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(proxy): cover the customer models allowlist
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(proxy): resolve team aliases before the end user model check
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(lens): add per-trace review models to jobs and progress
* feat(lens): append worker reviews to the job, capped, and count every review
* feat(lens): report a review with reasoning for each screened trace
* chore(ui): regenerate api types for lens job reviews
* feat(lens): type job reviews and fill them in lens fixtures
* feat(lens): add live review playback model
* feat(lens): pick the analysis model and slow single-review pacing
* feat(lens): add sample reviews for previewing the live run
* feat(lens): add live run layout with queue, reading trace and conclusions
* feat(lens): show the live run on investigations and open it from run now
* feat(lens): stream large review backlogs at 150ms or less and list newest first
* fix(lens): show the live run only for real reviews and keep fixtures test-only
* refactor(lens): restyle the live run as the native progress panel
* fix(lens): retry contended investigation updates with jittered backoff
* feat(lens): add a reading ticker line and replay for finished runs
* feat(lens): collapse the live run to an ambient line with show work
* fix(ui): crop the cerebras logo viewBox to its mark so it reads at icon size
* feat(lens): format review span previews as readable messages
* feat(lens): derive strip status, honest issue counts and drawer focus from a job
* feat(lens): track active jobs before their first review
* feat(lens): add a live trace results drawer with readable spans
* feat(lens): put the live strip under the progress bar and drop the inline panel
* feat(lens): add an ambient live strip that opens the drawer
* fix(lens): wait out provider rate limits and retry model calls four times
* style(lens): format repository contention tests
* feat(lens): read review spans as a conversation timeline
Turns spans into the user's ask, tool calls with args and results, and the agent's reply, dropping system prompts. Also handles a preview cut that lands inside the Output header.
* fix(lens): list recorded agents in the run now dialog
The run now agent field used a native datalist, whose suggestions do not show inside the modal dialog, so the agent list looked empty even though /lens/agents returned names. Use the same Combobox as investigation setup.
* feat(lens): pace live playback so each trace stays readable
Every trace now stays up for at least 1.5s. A backlog is cleared by skipping to the newest few instead of flickering through them. Conclusions count traces per check and kind, and new helpers cover share bars, group filters and flashes.
* feat(lens): keep the live run ambient until View run is clicked
The drawer no longer opens on Run or when entering a running investigation. LiveRun takes reviews as a prop so it can move to a dedicated reviews endpoint.
* feat(lens): show the live run as a two-pane trace and conclusions view
Left pane: the trace being reviewed as a readable timeline, followed by Lens's reasoning and the verdict. Right pane: ranked conclusion groups with share bars, plus a trace list you can filter.
* fix(lens): run several investigations per worker and poll every two seconds
* feat(lens): add worker slot and poll interval settings
* feat(lens): add list summaries and an incremental review filter
* perf(lens): strip reviews and run attributes from the lens list and serve reviews separately
* test(lens): cover list summaries, review polling and review access
* feat(lens): explain why a queued investigation is waiting
Works out whether no worker is connected, the worker is busy (with its running investigations and an estimated start time), or it is just being picked up.
* feat(lens): show the queue reason and what the worker is doing in the live strip
The progress header and the strip replace "Queued for your worker" with the concrete reason. While waiting, the strip lists the busy worker's investigations; click one to open it.
* feat(lens): add a review page model carrying the total reviewed count
* fix(lens): page live reviews by index so out-of-order reviews are never skipped
* feat(lens): take an index cursor on the reviews endpoint
* test(lens): cover index cursors across out-of-order and rolled-over reviews
* chore(ui): regenerate api types for the lens reviews endpoint
* feat(lens): page job reviews by index cursor
Adds api.reviews for GET /lens/{id}/runs/{job}/reviews?after=N, with a demo implementation. appendPage adds pages in arrival order and keeps the latest 200. liveJob now keys off reviewed, since the list no longer carries reviews.
* feat(ui): add a lens reviews query that polls the index cursor while live
* fix(lens): feed the live run from the reviews endpoint and keep View run open
LiveRun now gets its reviews from useJobReviews instead of the list, which no longer carries them. View run stays clickable while a run is queued or running, and before the first trace the opened view says what the worker is doing.
* fix(lens): split live conclusions into issues and patterns
A check could show up twice with the same label, once as an issue and once as a pattern.
* fix(lens): group live conclusions by check with short labels
There is now one group per check_id: issue traces are the main count and pattern traces a secondary note, so there are no duplicate red and grey cards. A long instruction falls back to the humanized check id. Adds briefReasoning and traceRows for the simplified trace list, and drops helpers nothing uses.
* feat(lens): simplify View run to traces and conclusions
The left pane is the trace list. A soft highlighter carrying the provider and model slides to the trace being reviewed, and clicking a row shows just Lens's reasoning and verdicts. The right pane keeps one conclusion card per check.
* refactor(lens): drop client-side replay in favour of real in-flight rows
Removes the playback reducer and its pacing. liveRows lists the traces the worker is reading, from job.reading, followed by completed reviews newest first, keyed by execution_id so a trace keeps its row when it finishes.
* feat(lens): show what the worker is reading and make View run obvious
Each trace in flight gets a highlighted row with the model and a live timer, and becomes its completed row in place. Completed rows show the real review time. View run is an outline button next to the progress line, and clicking anywhere on the strip opens it too.
* feat(lens): sum up a finished live run with time taken
doneLine reads like "Reviewed 30 traces in 31s with", measured from when reading started.
* feat(lens): slide one model rectangle over the traces being read
A single rounded rectangle carrying the provider logo and model wraps the real in-flight rows from job.reading. It translates and resizes over 250ms as traces finish in place. Before job.reading arrives it sits on a top slot showing the honest progress line, and when the run completes it fades out over 400ms. Rows have a fixed height and stable execution_id keys, so polls don't cause jumps or flicker.
* feat(lens): add in-flight runs to jobs and worker progress
* feat(lens): store in-flight runs from progress and clear them when a job ends
* refactor(lens): route progress, cancel and results through shared job transitions
* feat(lens): report each run as in flight when its review starts
* feat(lens): send in-flight runs with worker progress
* test(lens): cover in-flight runs across progress, old workers and terminal states
* test(lens): cover in-flight reporting under original run ids
* chore(ui): regenerate api types for lens in-flight runs
* feat(lens): model live reading lanes from in-flight runs and reviews
* feat(lens): show a now reading stage that types each trace's reasoning
* feat(lens): put the now reading stage above the trace list in View run
* fix(lens): resolve the analysis provider logo from the model catalog
* fix(lens): give demo jobs an empty in-flight list
* style(lens): format endpoint tests
* refactor(lens): name the run now handler in investigations view
* refactor(lens): name now reading conditions
* refactor(lens): name inline objects in the live run
* style(lens): format live run files
* fix(lens): keep worker settings inside the standalone worker package
* refactor(lens): keep update retry settings next to the repository
* fix(lens): start review history over when a run is reclaimed
* chore(lens): drop the unused review fixture
* refactor(lens): remove dead live helpers and use generated in-flight types
* fix(lens): keep polling a finished run until its last reviews arrive
* perf(lens): tick fast only while reasoning is typing
* fix(lens): isolate retried reviews and finding identities
* fix(lens): space the model name in run summary
* Update review.md
* test(lens): await trace status filter option
---------
Co-authored-by: moe-berri <moe@berri.ai>
* fix(cost): bill per-second transcription models outside chat modes
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost): restore request-time billing for per-second transcription models
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost): bill audio length for per-second transcription models when known
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(openai/realtime): drop model from upstream URL for intent=transcription
* fix(openai/realtime): keep model on the upstream URL for non-OpenAI hosts
A custom api_base on the openai provider can be a gateway that routes on
the model query param, so the transcription model drop now applies only
to OpenAI's own hosts. xAI's URL is unchanged again
* test(integration): audit realtime transcription upstream URL on a custom api_base
Twenty-two integration cells cover transcription and conversation sessions on every realtime route, the OpenAI SDK sync and async clients, malformed and duplicated query params, unauthenticated and refused upgrades, an unknown model, idempotent spend logging, a twenty-session burst, an upstream outage mid burst, a worker kill, and a proxy restart, each asserting the exact query pairs the upstream received. The scripted upstream now records every websocket upgrade as an observation
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
access already means access groups in this codebase; authz names what the
module decides (who may act on a team). Pure rename: every importer now uses
authz, and access.py stays as a re-export because the published
litellm-enterprise 0.1.73 wheel imports is_team_admin from it.
* fix Bedrock json mode bug
* fix finish reason when max tokens hit
* fix hardcoded completion for streaming responses
* fix: echo Responses API request params onto bridged response
* refactor(responses): drop explanatory comments and build null finish_reason test without mutation
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): type the terminal event and Bedrock mock helpers
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): keep truncated WebSocket turns in history and drop invalid echoed params
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Mrinal Chanshetty <mrinal@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(types): prove runtime types at harness, search, rag and client boundaries
Replace Any with adapter-validated types in the litellm.agent() harness, the
search provider transformations, RAG ingestion and query, the vector store
pre-call hook and registry, the galileo and opik logging integrations and the
proxy client CLI. Each boundary gets unit tests for well-formed and malformed
payloads.
* chore(typing): prove types at more provider boundaries and restore search transformations
Second pass over a2a, embedding, rerank, image, audio and small provider
modules. Search transformations go back to their previous form because
validating their response bodies would change the proxy status for
malformed upstream bodies from 400 to 500.
* refactor(types): prove types at logging, files, rerank, image, audio and management boundaries
Replace Any with validated or annotation-only types in 37 more files: logging
integrations, token counters, provider files/rerank/image generation/audio
transcription transformations, pass-through logging handlers and management
endpoints. No proxy HTTP status or error type changes.
* refactor(types): prove types at repository, spend, files and router boundaries
Replace Any with repository table accessors, validated mappings and
annotation-only types in 33 more files: Prisma repositories, the enterprise
batch and responses cost checkers, budget reservation, files endpoints,
management endpoints, the policy registry, the adaptive and complexity routers
and the secret managers. No proxy HTTP status or error type changes.
* test(types): run the aiohttp transformation test in-process and cover repository row conversion
The aiohttp chat transformation test no longer starts a server. It feeds the
transformation a response whose json() returns the body under test.
The proxy unit shards now exercise stored model rows whose params are JSON
strings and the object permission create and update paths.
* ci: run unit selections from GHA test-path and drop the CircleCI unit jobs
* ci: keep existing shard token order and header note
* ci: throwaway, drop tests/unit/repositories from the unit shard to show assert-ci-coverage fails
* ci: revert throwaway assert-ci-coverage check
* ci: stop crediting --ignore paths as invoked in assert_ci_coverage
* ci: install the caching, extra_proxy and proxy-runtime extras in the GHA unit sync
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* test(tracing): pin HTTP request compatibility
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(tracing): generate existing HTTP request models from Rust schemas
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(tracing): bind trace query params to generated request models
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* ci(rust): raise the native wheel size gate to 48 MB
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* perf(tracing): read the trace list clock without a thread-pool dispatch
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(traces): share the trace page-size bounds between schema and reader
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(ui): type trace request queries against the generated OpenAPI schema
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* docs: encode the trace contract boundary in AGENTS.md
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(ui): format trace request aliases
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Yujong Lee <yujong@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(guardrails): add llm shield pii redaction and rehydration guardrail
LLM Shield is a self-hosted PII gateway. This adds it as a guardrail so a
proxy operator can redact personal data out of outbound requests and have
the original values restored in the model's reply.
The substitution is reversible, which is the difference from a masking
guardrail. Outbound text is replaced with placeholders held in a session
vault inside the operator's own LLM Shield deployment, and the reply is
restored before it reaches the caller, so the end user still sees real
values while the provider never received them.
Streaming responses are restored incrementally. LLM Shield holds back only
the trailing characters that could still turn out to be part of a
placeholder, so tokens are forwarded as they arrive rather than the whole
response being collected first. A placeholder split across two chunks is
never emitted in fragments.
The integration talks to LLM Shield over HTTP and adds no dependency.
Notes for reviewers:
- The guardrail sets use_native_lifecycle_hooks, since redaction and
restoration need the native pre-call, post-call and streaming hooks
rather than the unified path.
- Per-request state lives on the request dict, never on the guardrail
instance, because the proxy registers a single instance process-wide.
The streaming carry-over is a local of the generator for the same reason.
- Every failure blocks the request. A redaction guardrail that fails open
would send the exact data it exists to protect to the provider.
* feat(ui): list llm shield in the guardrail garden
Adds the card, preset and logo so operators can pick LLM Shield from the
guardrails page the same way as the other partner guardrails.
* docs(guardrails): add llm shield example config
Shows both modes on one entry. Listing only pre_call redacts the request
and then hands the placeholders back to the end user, so the test asserts
both hooks are enabled.
* feat(ui): use the llm shield brand mark for the guardrail logo
* fix(guardrails): restore llm shield values in anthropic replies
The /v1/messages reply is a plain dict with a content block list and no
choices, so it fell through the restore path and went back to the caller
still carrying placeholders. The request was redacted correctly, which is
what made this easy to miss.
Found by running all three endpoints against a live provider; the mocked
tests all passed because they only built the OpenAI shape. Adds tests for
the message shape and for leaving non-text blocks alone.
* docs(guardrails): correct the llm shield start command
* fix(guardrails): redact every request shape and restore every reply shape
Three gaps, all of which let an enabled guardrail hand data to the provider
or hand placeholders to the caller.
Requests only walked `messages`. The Responses API `input` and tool call
`arguments` went out untouched. Measured against a live provider: a request
sent through `/v1/responses` reached the model with the real address in it
while the guardrail reported as enabled. Request traversal now covers chat
content (string and multimodal), tool call arguments, and `input` as a bare
string or a list of items.
Fixing that exposed the matching gap on the way back: the Responses API reply
carries `output` items rather than `choices`, so it returned to the caller
still holding placeholders. It now gets its own walk, handling text blocks as
dicts or objects.
The dashboard preset seeded only pre_call, so a guardrail created from the UI
would redact the request and return the placeholders to the user. Presets can
now seed both modes; the form already normalised either shape.
Adds tests for each request shape, for both Responses API reply forms, and
replaces a test that had asserted the `input` bypass as correct behaviour.
* fix(guardrails): narrow the stream delta before writing to it
basedpyright could not prove the delta was non-None on the write path, and
reportOptionalMemberAccess has a zero budget. The guard is also clearer than
relying on the text check to imply it.
* fix(guardrails): mint the vault id instead of trusting the caller's
The vault id was taken from caller-supplied session metadata, and every
caller shares one LLM Shield key. Someone who knew or guessed another
caller's session id could send a placeholder, have the model echo it back,
and get that caller's plaintext restored into their own reply.
Vault ids are now minted per request behind a per-process prefix, so a
caller cannot name a vault this process uses. Redaction mints, restoration
reads back, and a reply whose id does not match is left holding its
placeholders rather than resolved against some other vault.
Also covers two more request fields that were reaching the provider intact:
the Responses API `instructions`, and the legacy `function_call.arguments`
alongside `tool_calls`.
The collectors move to module level, which drops the traversal back under
the complexity limit and lets the code carry its own explanation instead of
the comments that were restating it.
* fix(guardrails): drop Final from a loop-assigned local
basedpyright rejects a Final assigned inside a loop, and
reportGeneralTypeIssues sits one over its budget ceiling.
* fix(guardrails): redact completion prompts and responses tool items
Two more provider-bound request shapes were reaching the model intact while
the guardrail reported as enabled.
/v1/completions carries its text in a top-level `prompt`, which the
traversal never looked at. It is handled as a string and as the array form,
where each entry is rewritten in place.
Responses input items hold tool data outside `content`: a function_call item
in `arguments`, a function_call_output item in `output`. Both are now
collected alongside the item's content.
Adds a test per shape.
* fix(guardrails): redact the anthropic system prompt and string-array input
Two more provider-bound shapes, found by walking the request types rather
than waiting for them to be reported.
/v1/messages carries its system prompt at the top level, as a string or a
list of text blocks. It is one of the endpoints this guardrail claims to
cover, and a system prompt is a natural place to put a customer's details.
`input` as an array of bare strings, the embeddings and moderations shape,
was skipped because the loop only handled item dicts.
Verified against a live provider: a system prompt holding an address now
reaches the model as a stand-in and is restored in the reply.
* fix(guardrails): narrow prompt and input to a list before iterating
Guarding with a conditional iterable left the value un-narrowed, so passing
it on was an argument-type error and the element checks read as unreachable.
An early return narrows it properly and reads better.
* fix(guardrails): restore every streaming choice, not just the first
Streaming rehydration read and rewrote choices[0] only, so with n>1 every
later choice went back to the caller still holding its placeholders.
Each choice is its own token stream, so the sliding window is now tracked
per choice index rather than once per stream. A single shared window would
have been worse than the bug: it would splice the characters held back for
one choice onto the next one's delta.
The final flush walks every choice the same way, and the two helpers that
only ever looked at choices[0] are gone.
Adds a test that both choices come back restored, and one that each choice
gets its own window handed back rather than its neighbour's.
* refactor(guardrails): name the guardrail llm_shield_proxy throughout
The integration was called llm_shield in code, llm-shield in the example
config, and LLM Shield in the dashboard, while the product and its PyPI
package are both llm-shield-proxy. An operator who saw the guardrail in
LiteLLM could not tell what to install.
One identifier now: llm_shield_proxy for the enum value, module, directory,
class, config model, logo and environment variables, with LLM Shield Proxy
as the display name. That matches `pip install llm-shield-proxy`.
Renames only; no behaviour change.
* feat(guardrails): redact the participant name on a message
`name` on a user or assistant turn identifies a person and was going to the
provider intact. The proxy this integrates with already redacts it, so the
integration was the weaker of the two.
On a tool or function turn the same field carries the function's name, which
has to arrive unchanged or the call stops routing. That case is skipped, and
a test asserts the value is never even sent to the shield.
* fix(guardrails): flush every held choice, and cover tool results and suffix
Three review findings.
The trailing flush walked the last chunk's choices, so a choice that finished
earlier and stopped appearing lost whatever text was still held for it and its
answer was truncated. It is now driven by the windows themselves and emits one
chunk per choice, synthesising the choice when the terminal chunk omits it.
That was data loss, not just under-redaction.
An Anthropic tool_result carries its own content, as a string or as further
blocks, and only each part's `text` was being collected. Handled recursively;
image and audio parts still fall through untouched.
The legacy completions `suffix` is forwarded to providers that support it and
was never collected. Note the placement: it has to be gathered before the
string-prompt early return, which is what the new test pins.
* fix(guardrails): walk nested tool results iteratively, with a depth bound
CI flagged _collect_content as recursive. It was, and worse, it was unbounded:
a tool_result nests its own content, the nesting is caller controlled, and the
descent had nothing to stop it. That is a JSON bomb, not a style issue.
Now an explicit queue with a depth bound of 8. Real payloads nest one or two
deep. The queue is walked in document order because the shield maps its replies
back by position, so collection order is part of the contract.
* fix(guardrails): redact Responses PromptObject variables
A Responses request can send `prompt` as a PromptObject rather than a string.
Its `variables` are substituted into the stored prompt on the provider side, so
they are caller text, and the dict shape was falling through untouched.
`id` and `version` pick which stored prompt to run and are left unchanged.
* test(guardrails): assert the depth bound instead of only reaching the end
The depth test asserted nothing, so it passed whether or not the bound held,
and the test-quality gate counted it as a zero-assert test. It now sends a
shallow value alongside a 200-deep chain and asserts the shallow one is
collected while the value past the bound is not.
* fix(guardrails): keep system-prompt values out of the restored reply
Redaction put every span of a request into one vault, and the reply was restored
against that same vault. System prompts are written by the application and the
caller never sees them, so a caller who got the model to echo a placeholder back
had its plaintext restored into their own reply -- a way to read a system prompt
they were never shown.
Server-authored spans now go into a vault of their own: system and developer
turns, Anthropic's top-level `system`, and the Responses API `instructions`.
Its id is deliberately never stored, so nothing restores against it. The reply
is restored against the caller's vault alone, and an echoed placeholder from a
system prompt comes back as the placeholder.
Values the caller also wrote themselves are unaffected -- they are in the
caller's vault too, and still restore. The extra round trip happens only when a
request actually carries server-authored text.
* style(guardrails): satisfy ruff format and annotate the new tests
`ruff format` wanted the widened `_collect_responses_fields` signature on one
line, and the three tests added with the split-vault fix needed return
annotations to keep ANN201 level with the base.
* fix(guardrails): restore tool calls in the LLM Shield guardrail
The request walk redacted a tool call's `arguments` -- plus the legacy `function_call`,
Anthropic `tool_use.input` leaves and the Responses API's `function_call` /
`function_call_output` fields -- while the response walk restored only `message.content`.
A placeholder therefore reached the caller inside a tool call, and nothing raised.
This is the same change as the out-of-tree example adapter this file is copied from, kept
body-identical on purpose: the response side now collects every restorable span in one
positional rehydrate batch, streaming keeps a window per (choice index, tool-call index)
and flushes each into the chunk carrying the finish_reason, and `apply_guardrail` restores
`inputs["tool_calls"]` on the response side. The declared limit on restoring values inside
a JSON string is documented in the module.
* fix(guardrails): import copy, keep the vault id off the provider, drop recursion
Three defects Greptile and veria-ai found on the reopened PR, all real:
- `copy.deepcopy` was called in `apply_guardrail` with no `import copy`, a
guaranteed NameError on every response carrying tool calls. It landed on
2026-09-13, ten days after the review that rated this branch safe, and no test
reached it: every tool-call test covered the request side. Adds the import and
a regression test on the response side.
- The vault session id was stored in `metadata`, which is forwarded to the
provider on /v1/responses. A provider holding the placeholders and the session
id can call the shield's rehydrate endpoint and read back the plaintext this
guardrail exists to withhold. Moves it to `litellm_metadata`, which is not
forwarded, and reads it back from there only.
- `_collect_json_leaves` recursed over model-controlled JSON; the repo's
recursive_detector gate rejects that. Rewritten with an explicit stack, same
depth bound.
52 tests pass. ruff format, ruff-strict and check_type_discipline all clean, with
LIT counts identical to the merge base.
* fix(guardrails): build llm_shield_proxy stream deltas without new mutable literals
The lint job's LIT002 budget gate failed on this PR: the file added 11
mutable-collection constructions and the tree sits at its limit. Build the
index-only tool-call continuation in one helper, keep read-only inputs as
tuples, and annotate the lists the delta and texts fields require.
Adds tests for the two tool-call flush paths the refactor touches, which
had no coverage: held arguments landing in the finish_reason chunk next to
that chunk's own fragment, and the trailing flush of a stream that ends
without a finish_reason.
* fix(guardrails): drop Final from loop-body locals in llm_shield_proxy
basedpyright rejects Final on a name assigned inside a loop, and the eleven
such locals put reportGeneralTypeIssues over its budget (112/101). The LIT010
Final rule already exempts loop-body assignments, so the annotations go.
* feat(guardrails): restore llm_shield_proxy placeholders on native streams
Anthropic /v1/messages and /v1/responses streams have no `choices`, so the
streaming hook passed them through with placeholders still in them. Both
are now restored incrementally, with the same per-stream windows as chat:
- /v1/messages arrives as raw SSE. Frames are cut at event boundaries,
text_delta and input_json_delta are restored per block index, and held
text is emitted as one more delta ahead of content_block_stop. Signed
thinking deltas, frames from other endpoints and non-SSE raw streams
pass through unchanged.
- /v1/responses events are restored per item and part. Held text goes out
as a copy of the stream's last delta before its .done event, and the
events that repeat the reply (.done, content_part.done, output_item.done,
response.completed) are restored in full.
The request side now also redacts Anthropic tool_use inputs and Responses
reasoning summaries, and sends tool and function descriptions (including
parameter schema descriptions) and the user / safety_identifier fields to
the non-restorable vault, like system prompts. Tool results stay
restorable: the model reads them to answer, so restoring them returns what
the caller would have seen without the guardrail.
* fix(guardrails): redact llm_shield_proxy predicted outputs and output schemas
`prediction.content` is the caller's own draft of the reply, so it is
redacted into the caller vault and restored with the reply. The
descriptions in a structured-output schema (Chat
response_format.json_schema, Responses text.format) are application
authored like tool schemas, so they go to the non-restorable vault.
* fix(guardrails): fail closed on deep llm_shield_proxy requests, widen coverage
- Request walks no longer skip what lies past their depth bound. Content
nested past it, and tool inputs or schemas past the new JSON bound, now
block the request instead of reaching the provider unredacted. The old
depth test asserted the skip; it now asserts the block.
- Tool and output schemas are walked by their JSON Schema structure, and
give up `title`, `examples` and `default` as well as `description`.
`enum` and `const` still go out as sent.
- Responses events are matched by shape: any `*.delta` with a string delta
is a token stream, and any `*.done` restores every non-identifier text
field plus the `part` or `item` it repeats. This covers
reasoning_summary_part.done and MCP arguments, and future families.
Audio deltas are left alone.
- An SSE stream whose first chunk ends partway through a field name
(`b"eve"`) is no longer taken for a non-SSE stream.
* fix(guardrails): scan llm_shield_proxy schemas by default
The schema walk collected an allowlist of keywords, so any keyword it did
not list -- draft-07 `dependencies`, `$comment`, vendor `x-` extensions --
went to the provider in clear. Invert it: every string is collected except
under keywords whose value must go out verbatim (types, formats, patterns,
references, required lists, enum, const). Name -> subschema maps still
treat their keys as property names, so a property called `type` is
walked, not skipped.
* fix(guardrails): redact llm_shield_proxy schema enum and const values
`enum` and `const` were skipped by the schema walk, so a value holding PII
went to the provider in clear. They now go to the caller's vault rather
than the non-restorable one: the model emits the stand-in in its tool
arguments or structured output, and restoring the reply turns it back into
the value the schema allows, so the call still routes.
* fix(guardrails): redact llm_shield_proxy web search user locations
Web search forwards the user's approximate location, and its free-text
`city` and `region` fields can hold an address. Collect them into the
non-restorable vault, from Chat `web_search_options.user_location` and
from the `user_location` of Responses and Anthropic web-search tools.
* fix(guardrails): drop unused llm_shield_proxy suppressions
Upstream added LIT013 (a *-ok marker that suppresses nothing) and LIT014
(at most one for and one if per comprehension). Remove the 34 markers
that no longer suppress anything and flatten the finished streams with
itertools.chain.from_iterable.
* fix(guardrails): type the llm_shield_proxy request and reply walks
Narrowing with isinstance(x, dict) leaves keys and values unknown, so
every call that passed a narrowed value counted against the
reportUnknownArgumentType budget. Parse into dict[str, object] and
list[object] once, in _as_object and _as_array, type the carry keys and
accumulators, and bind writers with functools.partial instead of lambdas.
The shield's batch reply is now also checked to hold only strings.
* fix(guardrails): keep restored llm_shield_proxy replies out of the cache, widen coverage
Addresses the open veria-ai and Cursor Bugbot findings on #42645.
- Restore a copy of the reply and of each stream chunk, never LiteLLM's own object.
LiteLLM caches and logs that object, and placeholders are numbered per request, so
two callers' redacted requests can share a cache key: restoring in place cached one
caller's plaintext for the next. The deployment hook no longer restores either,
since LiteLLM caches what it returns; the proxy's post-call hook restores
model-level guardrails after the cache write.
- Restore /v1/completions replies, streamed and not, which carry `choice.text`.
- Redact Responses replay fields the reply side already restores: tool output sent
as input_text parts, custom_tool_call `input`, code_interpreter_call `code`.
- Redact typed Responses prompt variables (`{"type": "input_text", "text": ...}`).
- Put Responses system and developer input items in the non-restorable vault, like
their Chat counterparts.
- Expose LLMShieldProxyGuardrailConfigModel through get_config_model, so the
dashboard can collect the Shield URL and key.
* fix(guardrails): redact llm_shield_proxy plain-text document blocks
An Anthropic document block carries text inline, in a text source's `data` or a
content source's `content`, and that text reached the provider unredacted. Collect
both, plus the block's `title` and `context`; base64, URL and file sources pass
untouched.
* fix(guardrails): redact llm_shield_proxy extra_body overrides
LiteLLM merges extra_body over the transformed request just before sending, so text
placed there (input, messages, system, ...) replaced the redacted field on the wire.
Walk extra_body with the same collectors as the request, keeping the caller /
application split.
* test(guardrails): import InMemoryCache directly in the llm_shield_proxy cache test
litellm keeps a deprecated module-level `caching` bool, so `litellm.caching.caching`
resolves to that bool once an earlier test in the same worker has set it, and the test
failed with AttributeError depending on test order.
* fix(guardrails): restore llm_shield_proxy replies for model-level use outside the proxy
71e68fd stopped the deployment post-call hook from restoring, so the response cache
never holds restored plaintext. Inside the proxy that is right: the proxy's post-call
hook restores after the cache write. But with model-level `guardrails` on the SDK,
the deployment hooks are the only redact and restore steps, so callers got
placeholders back.
When the deployment pre-call hook is the one that redacts, it now records the
request's vault id and marks the request no-cache / no-store; the deployment
post-call hook restores only when that record matches. The cache key there is built
from the redacted request and a cache hit skips the post-call hook, so a cached reply
could neither be restored nor safely shared. Proxy requests carry no record and keep
restoring in the proxy's post-call hook, after the cache write.
* fix(guardrails): don't repeat usage in llm_shield_proxy end-of-stream flush chunks
With n>=2 and stream_options.include_usage, the end-of-stream flush copies the last
chunk the stream carried, which is the one holding usage, so each synthetic flush
chunk repeated it and a consumer summing usage chunks counted the request twice.
The copy now drops `usage`, matching a normal mid-stream chunk. Reported by
@yucheng-berri.
* fix(guardrails): keep restored llm_shield_proxy values out of telemetry, refuse SDK streams
- The post-call restore hook no longer goes through log_guardrail_information,
which recorded its whole return value, the restored reply, as guardrail_response.
That field is exported to traces even with message logging turned off.
- A model-level stream outside the proxy is refused once redacted. Nothing restores
an SDK stream, and its cache writer reads the request from before the deployment
hook, so it also got cached despite the no-store bypass.
- Drop a narrating comment, and keep example_config.yaml to config only; the
how-to lives in the docs PR.
* refactor(guardrails): split llm_shield_proxy into payload, request walk and stream modules
The module had grown past 1,700 lines. Shared payload types and helpers move to
payload.py, the request walk to request_walk.py and the stream restorers to
stream_restorers.py; llm_shield_proxy.py keeps the guardrail class. No behaviour
change.
* style(guardrails): drop routine comments from llm_shield_proxy
AGENTS.md keeps source comments to tool directives and genuinely complex logic;
the rationale stays in the docstrings.
