mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-13 23:11:40 +00:00
79 commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
tin-berri
|
7c621b3141
|
fix(auto-router): accept every reminder marker pair a harness emits (#36029)
* fix(auto-router): accept every reminder marker pair a harness emits reminder_markers held one (open, close) pair, so a harness that wraps injected context differently per agent type only got the slice of traffic using the configured envelope stripped. Every other agent type kept hitting the original bug: its reminder-only turn never stripped to empty, won "newest human ask", and the harness blob got classified in place of the real question, choosing the tier and therefore the spend. The field now takes a list of ReminderMarkerPair, following the KeywordTierRule pattern already in this file so each pair validates itself and errors point at reminder_markers.N.close rather than a bare index. Blocks from different pairs can nest, which the gap construction could not handle: resuming the kept text at an inner block's end walks back inside the enclosing block and leaks its remainder. Running the block ends through a maximum collapses nested and overlapping spans without a separate merge pass, and stays linear in block count, which a fold over a growing tuple of merged spans would not. A single pair's ends already increase, so the maximum is the identity and the default path is byte-identical: verified against the shipped function over 200k generated inputs, and every existing reminder test passes unchanged. The prior single-pair config shape is rejected loudly at startup and at /model/new rather than silently stripping nothing. * docs(auto-router): document reminder_markers in the complexity router README * chore(ui): regenerate dashboard API types for the reminder_markers shape --------- Co-authored-by: Abhimanyu Kapur <38531241+akapur99@users.noreply.github.com> |
||
|
tin-berri
|
55e666a05f
|
feat(complexity_router): report LLM classifier cost per request via routing_decision and x-litellm-classifier-cost header (#36015) | ||
|
Abhimanyu Kapur
|
c76882b51b
|
fix(auto-router): stop the embedding model's context window from failing long requests (#35956)
* fix(auto-router): stop the embedding model's context window from failing long requests The auto-router embeds the last user message to pick a model and sent it to the embedding model unbounded. Embedding models carry 512 to 8k token windows while the chat models they route to carry 200k+, so any prompt over the encoder's window failed at the routing step with a 400 the destination model would never have raised. Cut every doc to a character cap inside LiteLLMRouterEncoder, which is the one choke point the auto-router, complexity-router, semantic guard and MCP tool filter all share. Default 2000 chars, roughly 500 tokens, which fits even a 512-token self-hosted encoder, overridable per deployment with auto_router_max_input_chars and globally with DEFAULT_MAX_EMBEDDING_INPUT_CHARS. Truncation alone cannot cover provider-side batch and byte limits, so any failure of the route call now falls back to the auto-router's default model instead of propagating. That path also fixes two latent bugs: a no-match left the auto-router alias in place as the model name, which fails downstream with "Unmapped LLM provider" rather than reaching default_model, and an empty route list raised IndexError. Fixes #17869 Fixes #20277 * fix(auto-router): make the embedding input cap opt-in so guards still see whole prompts Defaulting the cap inside the shared encoder truncated every consumer, not just the auto-router. The semantic guard builds the same encoder, so its pre-call check would have classified only the first 2000 characters while the full message still reached the model, which a benign opener in front of an injection payload walks straight past. The MCP tool filter and complexity router were silently narrowed the same way. The encoder now defaults to sending docs whole and cuts only when a caller passes max_input_chars. The auto-router is the only caller that does, so guard, MCP filter and complexity-router behaviour is unchanged from before this branch. DEFAULT_MAX_EMBEDDING_INPUT_CHARS becomes DEFAULT_AUTO_ROUTER_MAX_INPUT_CHARS, since it is now specific to the auto-router, and drops its env override: the per-deployment auto_router_max_input_chars already covers it, and every env var in constants.py has to be documented, which is what broke the documentation and code-quality checks. Also drops the added comments and the redundant type: ignore that review flagged. * test(auto-router): cover the max_input_chars wiring from litellm_params Nothing asserted that auto_router_max_input_chars on the deployment reaches the AutoRouter that embeds prompts. Dropping the wiring left every test green while the cap silently reverted to the default, so an operator with a 512-token embedding model could not lower it and every long prompt would fall back to the default model instead of being routed. * test(auto-router): cover the populated route-choice list branch The route layer can hand back a list, and picking its first element is where the IndexError lived: the empty case was covered but the populated one was not, so the branch that reads route_choice[0].name could be deleted with every test still green. |
||
|
Abhimanyu Kapur
|
bea65b6dcc
|
fix(autorouter): match CJK keyword_tier_rules that regex word boundaries miss (#35984)
* fix(autorouter): match CJK keyword_tier_rules that regex word boundaries miss Single-word keywords were matched with a \b...\b regex. Every CJK character is a regex word character and CJK is written without spaces, so \b never fires between two of them and a rule like 发票 silently missed 我需要开发票, falling through to complexity scoring instead of the configured tier. Keywords containing CJK now match as plain substrings, the same way multi-word phrases already did. The gate reads the keyword rather than the prompt, so a keyword with no CJK in it keeps word boundary matching regardless of the script the prompt is written in. * fix(autorouter): cover Han extensions in planes 2 and 3, not just up to U+2FA1F The supplementary range stopped at U+2FA1F, so Extension G and H ideographs kept the word boundary path and stayed unmatchable. Both planes are dedicated to CJK ideographs, so covering them whole also handles later extensions without chasing each new block. |
||
|
Abhimanyu Kapur
|
b8df48cd7f
|
feat(auto-router): let operators replace the LLM classifier's system prompt (#35855)
* feat(auto-router): let operators replace the LLM classifier's system prompt The complexity router's LLM classifier has always sent one built-in rubric, so the router could only ever grade difficulty. Operators can now supply their own system prompt, which replaces the rubric outright and repurposes the same tier machinery for whatever taxonomy the prompt defines, data sensitivity being the obvious case. Replacement is total: neither the rubric nor its closing line is appended, since both describe grading difficulty over a "current message" and a prompt grading something else is entitled to contradict them. That closing paragraph is also the classifier's prompt-injection defense, so the config field and the dashboard editor both warn that a replacement omitting it lets a caller ask for a tier and get it. The heuristic fallback still scores complexity, which is meaningless for a repurposed taxonomy, so classifier_fallback now chooses between the heuristic scorer and routing straight to default_model. The default_model path bypasses tier pools, the adaptive bandit, and escalation, because no tier was decided and the point of that fallback is a known destination. It reports itself as default_model_fallback in the spend logs. The dashboard's prompt editor prefills from a new /auto_router/classifier/default_prompt endpoint rather than a copy of the rubric in the frontend, and stores no override when the draft matches the default, so later rubric improvements still reach every router that never customized it. Tier names stay SIMPLE/MEDIUM/COMPLEX/REASONING; a custom prompt redefines what they mean, not what they are called. * fix(complexity-router): don't let the default_model classifier fallback bypass routing plugins * fix(complexity-router): don't pin a session to the default model after a classifier failure * fix(complexity-router): omit the tier from a default-model-fallback routing decision The classifier never answered, so no tier was decided. The record reported the tier whose pool happens to hold default_model, which reads in the spend log and the UI as if the request was classified. Matches how default_fallback already records a route that no tier produced. * fix(proxy): allowlist /auto_router/ on the UI backend component The new GET /auto_router/classifier/default_prompt is a UI-consumed management route, so it belongs on the control plane. Without the prefix it was exposed by neither component and test_gateway_plus_backend_covers_full_app failed. * docs(ui): reword the classifier prompt disclaimer Frames the closing paragraph as a strong recommendation rather than a description of what gets dropped, names prompt injection explicitly, and notes the tier names stay fixed regardless of their display names. * fix(complexity-router): stop logging a fabricated tier on the plugin fallback path The classifier-failed fallback resolves a tier so the routing-plugin pipeline has a pool to filter, but nothing about the request produced that tier. The non-plugin short-circuit already dropped it from the logged decision; the plugin path still reported it, so a spend log claimed a classification the request never received. Record the pool as a plugin-filtered-pool signal instead. Also name the real problem when the resolved tier has no models at all: that raised "No candidate models left after routing-plugin filtering" and sent operators hunting for a policy plugin that never narrowed anything. |
||
|
Abhimanyu Kapur
|
cc1c7d6101
|
feat(complexity_router): let operators rename the four complexity tiers (#35893)
* feat(complexity_router): let operators rename the four complexity tiers Adds an optional tier_labels map to complexity_router_config so a deployment can put its own vocabulary on the four tiers, e.g. Cheap / Standard / Premium / Deep, instead of reading SIMPLE / MEDIUM / COMPLEX / REASONING in its dashboard, its spend logs, and the rubric the LLM classifier reasons with. Labels are display-only. Every config key stays canonical, so tiers, keyword_tier_rules[].tier, and tier_boundaries are written exactly as they are without labels, and partial maps are fine with unlisted tiers keeping their default name. A validator rejects blank labels, two tiers sharing a label, and a label that is another tier's canonical name, since any of those would make a log row or a rubric line ambiguous. That validator runs on the /model/new and /model/update write path already, so an ambiguous config gets a 400 rather than being stored for the router to refuse later. Under the default heuristic scorer the names are cosmetic: the scorer maps a weighted score to a rung and never reads a tier name, verified by running the eval corpus with and without a rename and getting identical tier and identical score on all 29 cases. Under classifier_type: llm the labels are the names in the rubric and the values the classifier must return, so the response format's enum is now built from the configured labels and a reply is resolved back to its tier against labels first, then canonical names, case-insensitively. An unresolvable reply degrades to the heuristic on the existing fallback path. A test pins the generated schema for an unrenamed deployment as equal to the shipped TierClassification schema, so the wire shape can't drift. Spend logs keep routing_decision.tier canonical so rows from before and after a rename stay comparable, and gain routing_decision.tier_label on the tiers that were renamed. * refactor(complexity_router): drop added comments and the Counter construction Review feedback: the repository guide bans new comments, so the explanatory comments and the appended docstring paragraphs this branch added come back out. One-line docstrings stay in complexity_router.py, matching that file's own convention. The duplicate-label check no longer builds a Counter, which the mutable-collection budget counts, and the error text drops its list() reprs for joined strings. The labels are stripped in tier_label() now rather than by rewriting the field in the validator, so the stored config keeps exactly what the operator wrote. schema.d.ts is regenerated: ComplexityRouterConfig is exposed in the OpenAPI spec, so tier_labels surfaces there. * fix(ui): carry tier_labels through the auto-router preset prefill buildPresetPrefill maps every payload key onto form state, but the tier_labels key added by this branch had no line, so a preset shipping labels would apply its tiers and silently drop its names. |
||
|
tin-berri
|
4fcaf7d736
|
feat(spend): derive a default auto-router savings baseline from the hardest tier (#35907)
* feat(spend): derive a default auto-router savings baseline from the hardest tier The savings driver shipped off by default: unless an operator names litellm_settings.autorouter_savings_baseline_model, every auto-routed request records $0.00 and the dashboard card never populates. Nobody discovers a knob whose feature they have never seen work, so the default has to come from somewhere the proxy already knows. The router's own tier ladder is that place. Without a router a deployment runs one model that can carry the hardest request it will see, so the derived baseline is the priciest model in the hardest configured tier, REASONING when present, otherwise the most severe tier the router actually defines. A cheap tier is a choice the router made, not a ceiling it was bounded by. An earlier draft of #35521 derived this per request and was deleted for it: ranking candidates against the request that ran meant reading the request, and every input shape it could take produced its own review finding. This derivation is ranked against one fixed reference request instead, a cache-heavy shape matching real auto-routed traffic, so it never reads the request at all. Candidates still resolve through the router's deployments, so Azure base_model and per-deployment pricing overrides rank correctly. The deciding router records the result on its routing_decision, because one model name can carry several tag-scoped routers with different tier ladders and only the deciding instance knows which of them routed the request. The spend writer's precedence is: configured baseline, then the recorded one, then off. When the setting is present the router skips deriving entirely rather than pricing candidates per decision only to be ignored. Resolution never raises; an unresolvable baseline zeroes the driver instead of failing a live request. Rows queued by a pod on the previous release carry no recorded baseline and fall back to the configured setting, exactly as today. The schema.d.ts regeneration also picks up the reminder_markers field that UI-19232 (#35874) added without regenerating, so one hunk there is inherited staleness rather than part of this change. * fix(spend): cache the derived baseline, price it by deployment, keep it out of the routing preview Three review findings on the derived baseline, addressed together because they all sit on the same value's path from derivation to consumer. Derivation walked and priced the hardest tier's whole pool inside a property read on every routing decision, unbounded by pool size. The router now caches the result per instance with a 30 second TTL, None results included, so the hot path is a clock compare and a deployment edit still lands within a window no operator watches closer than. Ranking used each deployment's effective pricing but recorded only the model name, so the spend writer priced the winning baseline at its public rate: a hardest tier whose deployment carries a negotiated rate produced materially wrong savings. The decision now also records savings_baseline_deployment_id and the writer resolves it through Router.get_deployment_model_info, exactly as the selected arm already does. The id is ignored whenever the configured setting overrides the recorded baseline, since the setting names a model, not a deployment. /auto_router/test_routing returns the routing decision verbatim to team admins while only authorizing the classifier and embedding models, so a derived baseline would resolve another team's model-group alias into its backend provider/model mapping and hand it to a caller never authorized for it. The preview's throwaway router is built with derive_savings_baseline=False; its decisions are never spend-tracked, so nothing is lost, and a source-pinning test keeps the flag on the endpoint. Also strips the explanatory comments this PR had added. * refactor(spend): pin the derived baseline per router instance instead of a TTL Creating or editing a router already rebuilds its ComplexityRouter instance, through unregister and re-add on upsert and through the registry reset on a full model_list load, so a value derived once per instance refreshes on exactly the flows that can change it. That makes the TTL a solution to a problem the rebuild lifecycle already solves, and it goes. Derivation stays deferred to first use rather than running in __init__: during a config load this router can be constructed before the deployments its tiers name, and a baseline pinned at that moment would be empty for the process lifetime. The one behavior the TTL had that the pin does not: editing a tier deployment without touching the router itself refreshed the baseline within a window. That edit path rebuilds only the edited deployment's own strategies, so the pin holds the old answer until the router is next saved or the config next loads. A stale deployment id degrades to public-rate pricing rather than failing, which is where every other unresolvable baseline already lands. |
||
|
Abhimanyu Kapur
|
31a86daa85
|
feat(auto-router): make reminder marker pair configurable (#35874)
* feat(auto-router): make reminder marker pair configurable Some harnesses inject internal context using their own marker pair instead of Claude Code's <system-reminder>/</system-reminder> convention, and some send it as a separate follow-up user message rather than inline with the ask. Both cases fall out of the same root cause: the router's marker-matching is hardcoded, so foreign markers never strip to empty and the reminder-only turn wins "newest human ask" selection instead of being skipped. Add an optional reminder_markers field to ComplexityRouterConfig so operators can override the (open, close) pair via proxy config, with the existing skip-when-empty selection logic handling both cases once the markers match. * test(auto-router): drop unsolicited comments from the reminder-markers regression test Per Greptile review on #35874: no comments unless explicitly requested. |
||
|
tin-berri
|
9e3a8df6c0
|
feat(spend): add net auto-router savings to the cost-optimization dashboard (#35521)
* feat(spend): add net auto-router savings to the cost-optimization dashboard The dashboard credited compression and prompt caching but said nothing about the optimization that picks the model, so the driver with the largest lever on a bill was the one an operator could not see. Savings are the counterfactual: without a router a deployment runs one model, and it has to be one that can carry the hardest request, so the baseline is the priciest model in the router's hardest configured tier. A cheap tier is a choice the router made, not a ceiling it was bounded by. `auto_router_savings_baseline_model` overrides it for operators who would genuinely have run something else. Both are provider-qualified before pricing, because a bare name can resolve to a different vendor's rates or to nothing at all, and a deployment is priced by its `base_model` where it has one, which is how Azure deployments are priced everywhere else. Both arms price the request's real usage through `generic_cost_per_token` rather than re-deriving per-token arithmetic, so tiered rates, ephemeral cache-write tiers and regional uplifts stay consistent with what was actually billed. `prompt_tokens` already includes the cache buckets, so charging them again at the input rate would price the same tokens twice. Cache state is what makes this hard. The baseline serves every turn, so whether it had the prompt cached is whether the conversation was already underway. On a continuing conversation it wrote the prompt earlier and would only read it now, so this request's write is what switching cost and counts against the saving. On a first turn nothing was cached for any model, the baseline would have written the same prompt, and both arms carry the write at their own rates. Charging the write to both cases understates a first turn to a few percent of its value, and because the write premium is fixed by prompt size while the saving grows with completion length, it can render a profitable route as a loss. That shape is read off the conversation rather than remembered: a second human ask means an earlier turn was served. No cache, no session id, and no dependence on a caller sending a session header. It cannot see a switch on a turn the router did not classify, and it reads a few-shot prompt's synthetic turns as prior conversation; both err toward charging the write, which under-claims. The baseline and the shape ride on the existing `routing_decision` record, which is already carried from the router to the spend log, already classified for redaction, and already written-or-cleared per attempt. A fallback that re-enters the hook therefore cannot leave either fact behind to be attributed to a deployment that never routed, and no new metadata key crosses the trust boundary. The result is signed. Whether a switch pays off is a race between the rate gap and the cache-write cost, and a narrow gap loses; flooring at zero would hide exactly the routing behaviour an operator needs to see. The donut plots only drivers that saved, while the card and range total keep the sign. Savings accrue into a new `autorouter_savings_spend` column on the six daily rollup tables, declared `NotRequired` because rows queued by a pod on the previous release carry no such key. It is summed by the rollup merge the cross-pod Redis drain also runs, and carried through the aggregation query, the per-row accumulation and the response model, so the dashboard reads a value the API actually sends. Tests enumerate the drivers from the response model itself and assert each is summed, accumulated, carried and totalled, so one added later cannot be half-wired. * fix(spend): let the baseline pay for a continuing turn's own growth `_baseline_usage` moved every cache-creation token into the baseline's read bucket whenever the conversation was underway. That is right for a switch, where the baseline never left the model it was on and really would only read, but wrong for a turn that stayed put: the prompt grew, and the tokens written are that growth. They are new to every model, so the baseline would have paid to write them too. Forgiving it that write made the counterfactual cheaper than it was and shrank the reported saving on ordinary steady-state traffic, by about 2% per turn. The selected arm was never involved; it has always been priced on the real usage. The error sat entirely on the baseline. The condition is that the request read more than it wrote, not that it read anything. A switch onto a model already holding a small prefix of this prompt still writes most of it, and that write is the switch's own cost; keying off a nonzero read would have handed such a request the full rate gap, turning +$0.0056 into +$0.1177. Comparing the two buckets separates a warm continuation, which reads far more than it writes, from a cold arrival, which does the reverse, and it leaves the existing invariant intact: a request reading 0 and one reading 1 both still land in the same place. * fix(spend): price each arm under the key litellm billed it, and see agent turns Two ways the savings number read the wrong thing, both from identifying a model by its name when the name is not what it costs. The counterfactual was ranked and priced on the public rate for the model a deployment names. A deployment may not be charged that rate: the router registers its configured prices under the deployment's own id and deliberately keeps them off the shared model-name key so deployments sharing a backend model do not pollute each other. So a hardest-tier deployment configured above its public rate lost the ranking to a cheaper candidate, and once chosen was priced at a rate nobody pays. Which key prices a deployment is now `_select_model_name_for_cost_calc`'s decision, the resolver the real request is billed through, rather than a second rule here that would have to re-learn that per-second and tiered overrides count, that a partial override still counts, and that a deployment configured at zero is priced at zero rather than treated as unpriced. The arm being subtracted had the same fault and a sharper edge. It priced the spend log's `model`, which on Azure is the deployment name, absent from the cost map, so the whole driver silently read zero for that traffic. It no longer re-derives anything: `model_map_information.model_map_key` is what litellm actually billed the request under, recorded at request time by that same resolver with `base_model` and custom pricing already applied. Separately, the conversation-shape discriminator counted human asks, and an agent loop can run twenty turns on one of them. Its tool traffic rides `tool_result` blocks on user turns that flatten to empty text, and `tool` roles that are never read, so a long agentic conversation looked like its own first turn and was handed the arithmetic that leaves the cache write on both arms. That is the one direction this must never fail in, because it inflates. An assistant turn is the direct evidence that something answered earlier, and it is blind to how the tool plumbing is spelled on either surface. * fix(spend): give the cost-key resolver both inputs the selected arm needs The served model was resolved through one input at a time, and each choice broke the half the other fixed. `model_map_key` is the served model already resolved through `base_model`, which is the only way an Azure deployment name reaches the cost map at all; without it the selected arm priced a name absent from the map, returned nothing, and the whole driver silently read zero for that traffic. But it is built without `router_model_id`, so it never carries a deployment's own price overrides, and a custom-priced deployment was compared at its public rate while the baseline used the real override. On a deployment configured well above its public rate that inverted the answer outright: a route that lost $21.88 reported saving $0.10. `_select_model_name_for_cost_calc` takes both, so it gets both. Which key prices a deployment stays its decision rather than a rule restated here. * fix(spend): same model is only the same cost when it is the same deployment The short-circuit compared resolved model identity, so two deployments of one model collapsed to "no switch" and reported zero. They are not the same cost: a deployment can carry a negotiated rate, and routing from the dear one to the list-price one is a real saving the dashboard reported as $0.00 against a true $21.93. Both arms now carry the key litellm prices them under, so the comparison is between deployments rather than between names. * refactor(spend): price from resolved rates, not from a name we keep re-resolving Four review rounds landed on one mechanism: which identifier prices a deployment. base_model, then the deployment id, then cache-only overrides. Each round added a clause to a resolution rule that should not exist, and a wrong primitive fails once per input shape, so each shape arrived as its own finding. `Router.get_deployment_model_info` already owns this. It merges a deployment's configured prices over the built-in map, folds in `base_model` defaults for deployments whose name is not a model, and falls back to the model name when nothing is overridden. Every shape hand-rolled here (cache-only, partial, per-second, Azure) was that function re-implemented badly. `generic_cost_per_token` now accepts already-resolved rates instead of demanding a name it looks up itself, which is what forced the name-bending in the first place. Both arms resolve through the owner and pass what they got: the counterfactual by the deployment the router would have used, the served request by the deployment that served it. The invented cost-key resolver is gone, and `Baseline` carries a deployment id rather than a key we chose on litellm's behalf. Net 64 insertions against 79 deletions. * test(spend): follow _most_expensive onto the router that prices its candidates Ranking moved through `Router.get_deployment_model_info`, since what a deployment costs is the router's answer to give; these four cases were still calling the old free-function signature. * fix(spend): rank baseline candidates by what a request costs, not by two rates "Most expensive" was decided by comparing output rate then input rate. That is a property of a rate, not of a request: a deployment dearer per output token can be cheaper per cached token, so the comparison ordered cache-heavy traffic backwards and recorded the wrong counterfactual. Candidates are now costed on one reference request through the same engine the savings themselves use, which leaves cache read and write rates, tiered tables and every other billing dimension to that engine rather than to another rule restated here. The reference request is cache-heavy because auto-routed traffic is. * fix(spend): pick the baseline against the request that ran, not a stand-in for one Ranking happened in the pre-routing hook, where the request has not executed yet, so candidates were costed against a hard-coded reference workload: 20k prompt, 19k of it cached, 1k out. Which candidate is dearest depends on that mix, so a pooled hardest tier holding a deployment with non-proportional configured rates could be ranked for a request nothing like the one served. The mix is known on the spend path, so the ranking belongs there. The routing decision now carries the tier's candidates rather than a winner already chosen, and the baseline is resolved against the usage that actually happened. The reference workload is gone; nothing here assumes a traffic shape any more. The router is passed in rather than imported from `proxy_server` inside the computation, so the savings stay a pure function of their arguments and the caller owns where the router comes from. That also makes the spend path testable without a running proxy, which the previous shape was not. * refactor(spend): measure savings against one configured model, not a derived one The counterfactual was derived per request: enumerate the hardest tier's deployments, resolve each one's effective pricing, price them all, take the dearest. That machinery produced a review finding per input shape it had not anticipated, and every answer it gave was one an operator could have stated in a line of config. So they state it. `litellm_settings.autorouter_savings_baseline_model` names the model the traffic would have run on without a router, for every auto-router on the proxy, and unset means the driver is off rather than a model nobody named being guessed at. `savings_baseline.py` and its tests are deleted outright, along with the tier enumeration, the candidate list on the routing decision, and the per-deployment override that shadowed it. Cache-state handling is untouched: the baseline is still priced on this request's own read and write split, so a switch still pays for re-warming the cache and a first turn still charges the write to both arms. 45 insertions against 482 deletions. * refactor(router): compute the conversation shape once and pass it down `_classify_and_route` re-derived it from the messages the hook had already resolved, so an ordinary routed request walked the turn list twice for one boolean. The hook computes it and hands it over, which is also where the affinity-hit path already got it from. Also moves `_get_llm_router` below the imports it sat among. * fix(router): drop the dead conversation_continuing parameter off the hook It was added to `async_pre_routing_hook` by mistake and immediately overwritten by the value the hook computes, so it never did anything. It also widened a signature every pre-routing strategy shares with the protocol in `types/router.py`, leaving this one router diverged from `AutoRouter` and the interface for no reason. Also records why an unreadable request counts as continuing: no messages is no evidence a turn was served, so it pays the cache write and under-claims rather than being handed a first turn's larger saving on nothing. * fix(spend): charge a baseline its input rate for cache buckets it cannot price A model with no cache_creation_input_token_cost, which is every OpenAI, Azure and Gemini entry, resolved that rate to 0.0 and carried the whole written prompt for free, so a first turn routed onto a cheaper model reported a loss. Same hole on cache reads. Those tokens are plain input on such a model, so they move into the text bucket. * refactor(spend): build the daily upsert payloads in one shot `common_data` and `update_data` were constructed and then appended to: `request_id` conditionally for tag rows, `endpoint` unconditionally a few lines later. A dict that grows after its literal cannot be reasoned about by reading the literal, which is the whole point of building it at once. The conditional key resolves to a spreadable value before either payload, so both are single expressions and the tag branch appears once instead of twice. Not wrapped in MappingProxyType, though it was suggested: these go straight to prisma, whose query builder branches on `isinstance(value, dict)` to tell a nested node from a scalar. A mappingproxy is a Mapping but not a dict, so it falls through to the serializer and raises `TypeError: Type <class 'mappingproxy'> not serializable` inside the batch upsert, where the surrounding except would log it and leave the rollups silently unwritten. * fix(spend): keep the one-shot upsert payloads under the type-discipline budget Building both payloads as single literals traded a mutation for two dict literals, and LIT002 counts construction rather than mutation, so the change the review asked for is the one the gate charges for. The empty branch is the avoidable half: it is the same value every time, so it moves to a module constant built once instead of a literal per transaction, and it is a read-only mapping so none of the call sites that spread it can fill it in later. |
||
|
tin-berri
|
8ad5d144a1
|
feat(complexity_router): default session affinity off and expose it in the UI (#35714)
* feat(ui): expose an Auto-Router session affinity toggle session_affinity on ComplexityRouterConfig defaults to True, and neither the create form nor the edit modal ever emitted the key, so every auto-router built in the UI silently pinned each session to its first turn's model for an hour with no way to see or change that. Adds an "Advanced: Session Affinity" switch to both surfaces, defaulted on to match the backend field. Both paths now write the key explicitly instead of falling through to the backend default, so a stored config states what the router actually does. A stored config with the key absent hydrates as on, since those routers are running with affinity enabled today; showing them as off would report the opposite of reality and persist it on the next save. * feat(complexity_router): default session affinity off and expose it in the UI session_affinity defaulted to True and the Auto-Router UI never emitted the key, so every router built there silently pinned each session to whatever model its first turn classified into for an hour, refreshed on every hit. There was no way to see that from the UI and no way to change it without hand-editing config.yaml. The default flips to False, so every turn is classified on its own merits and lands on the cheapest adequate tier. Pinning is now opt-in. The toggle added in the previous commit follows the field: it renders off, and both the create tab and the edit modal keep writing the key explicitly, so a stored config states what the router does instead of inheriting a default that can move under it. Behavior change for existing routers: those created before this have no session_affinity key stored, so they pick up the new default and start reclassifying every turn. That gives up the provider prompt cache the pin was preserving, and a multi-turn session can now change model between turns. Set session_affinity: true to keep the old behavior. |
||
|
tin-berri
|
11ad3ff939
|
refactor(complexity_router): drop the tier-rubric override, close the rubric on the window it was given (#35504)
Two changes to the classifier's system role, both narrowing it rather than adding to it classifier_tier_rubric let an operator replace the tier definitions. It shipped in #35471 alongside the assistant-turn context window, but the two answer different halves of the same report and only the context window was asked for. The override carried a composed prompt, an overridable and a non-overridable half, a blank-is-unset rule, a length-warning validator and a pair of dashboard controls. All of it goes The rubric then closes on one of two lines, chosen by classifier_context_window_size. At 0 no conversation is quoted, so the line is the original one, byte for byte: a deployment that sends no context is told to classify the current message and nothing else, which is what it could see all along. Above 0 the turns are quoted, and the original line told the model to disregard them, which is how a request whose difficulty was established in an earlier turn came back SIMPLE on the word "yes". There the line instead says to classify the current message using the quoted turns as context, and to rate what a short reply approves rather than the reply The choice keys on the window and not on classifier_context_include_assistant_turns. Whether the quoted turns are the user's alone or include the assistant's replies does not change what the model needs told, and whose turn is whose is already on the turns. Keying it on the assistant toggle would put the default deployment back on the original line, which is the configuration the report was raised against Folds in #35508, which built the window-dependent framing on top of the override this removes; that PR is closed in favour of this one |
||
|
tin-berri
|
9bed4955d4
|
feat(complexity_router): let the classifier see assistant turns and rate what a short reply approves (#35471)
The LLM classifier's context window carried user turns only, so a conversation whose difficulty was stated by the model rather than by the user was classified without it. Asked to find events, the assistant answers "here is the plan, it is complex, should I execute?", the user answers "yes", and the router rates the word "yes" and picks the cheapest tier Two independent causes, so two changes that are each provable on their own classifier_context_include_assistant_turns adds assistant turns to the window. It is off by default because turning it on shifts tier decisions, and therefore spend, for an already-deployed router, and because assistant text is net-new egress to the classifier deployment. With it on, classifier_context_window_size counts the last N turns across both roles, which is what makes the assistant's own statement of difficulty land in the window Assistant text reaches the classifier payload and nothing else. The window is read only by _build_classifier_user_payload, while keyword_tier_rules, escalation matching, the heuristic scorer and the semantic embedding all read the human ask through _iter_human_asks_newest_first. Those are substring and vector matchers, so an assistant echoing an escalation keyword back to a user would choose the model, and the spend, with nobody having asked. Rather than widen the shared iterator, _iter_context_turns_newest_first is separate and feeds the window alone, which makes the boundary structural instead of a rule to remember The rubric ended "Classify only the current message", and the classifier applied it literally: a request whose difficulty was established earlier came back SIMPLE because the message being rated was the word "yes". A context window the rubric then tells the model to disregard buys nothing, so the wording now asks it to rate the work the current message approves, judged in the conversation it continues, while still forbidding it to rate a quoted section as if that section were the request classifier_tier_rubric lets an operator replace the tier definitions. The trust-boundary paragraph is appended and cannot be replaced: it defends the operator against their own callers, so an operator writing tiers without that threat in mind would otherwise hand every keyholder the top tier by omission. Blank reads as unset so an empty form field falls back rather than sending a rubric with no tiers in it Turns are labelled by role only when assistant turns can appear, so the prompt of every deployment that never asked for this is unchanged byte for byte |
||
|
tin-berri
|
2dbcb9a999
|
feat(spend-logs): record when a spend log row is the auto-router's own classifier call (#35300)
The complexity router's classifier sub-call copies the parent request's metadata verbatim, so its spend log row carries the caller's key, team and user and is indistinguishable from traffic the caller actually sent. Nothing on the row says otherwise: call_type is "acompletion" either way, model_group is overwritten to the classifier's own model group so the row never looks auto-routed, and routing_decision is absent exactly as it is on an ordinary request. Record the fact the system already knows at call time. internal_call_origin is declared on SpendLogsMetadata, which is the allowlist _get_spend_logs_metadata projects onto, and stamped in _classifier_call_metadata; both classifier paths already route through that one function and it feeds the metadata and litellm_metadata buckets alike, so every request surface is covered at one site. The key is reserved rather than caller-supplied, so it joins routing_decision in the untrusted-metadata strip and a caller cannot label their own traffic as router overhead. The classifier call also inherited no session identity, so the router minted a fresh trace id and the row landed in a session of its own. Forwarding the parent's session puts it in the trace of the request that triggered it, which is where an operator looks for what the routing cost. |
||
|
tin-berri
|
c8bec20443
|
fix: give ComplexityRouter LLM classifier prior-turn context (LIT-4981) (#35185)
The ComplexityRouter's LLM classifier saw only the last user message, so on a multi-turn conversation it classified whatever happened to be last rather than what the human actually asked, and a near-constant classifier input pinned a whole session to one tier. The blindness turned out to be narrower than first diagnosed, and the fix is correspondingly smaller. Tool output was never the problem: on the Messages surface it rides a user turn as tool_result content blocks, which are not text parts, so flattening to `type == "text"` already dropped those turns; on chat completions it arrives on a `tool` role the extractor never read. Both surfaces were already handled before this change. What actually leaked through was the harness `<system-reminder>` block, which arrives as ordinary text, survives flattening, and became the current ask on any turn that carried one. So reminders are stripped rather than used to reject the turn, because a harness injects them alongside the live ask and not as a turn of their own; rejecting the turn would lose the ask, and keeping the block would feed the classifier the near-constant boilerplate that flattens tier selection in the first place. An earlier revision of this change also pattern-matched serialized tool_result payloads. That check only ever fired on a hand-serialized string neither request surface produces, it was where every review finding in this PR lived, and it is deleted here; the tests now pin the real shapes instead of the synthetic one they were built on. The classifier call is split into a system role carrying the rubric plus the caller's own system prompt, which stays byte-stable across a session so a provider can prompt-cache it, and a user role carrying the variable context: a bounded window of prior user turns, a conversation-depth signal, and the current ask. The caller's system prompt rides every turn, so task constraints are never dropped. The depth signal measures content-parts messages too, since counting only string content reported ~0 tokens for exactly the deep Messages-surface conversations that most need an expensive tier, and it is omitted entirely on the prompt-only path rather than asserting a false zero. Prior turns are excluded by matching the current ask rather than by dropping the newest turn positionally, because `aclassify` takes `prompt` and `messages` separately and a caller may classify something other than the newest turn. Truncated turns carry a marker so the classifier can tell a turn was clipped. Only the LLM classifier's input changes. The heuristic scorer, keyword overrides, escalation matching and semantic embedding still read the extracted current ask, which is why that extraction has to yield one clean human-authored string: those are substring and vector matchers, and an escalation keyword sitting inside a reminder blob would otherwise trip a tier jump on its own. Defaults keep single-turn classification equivalent to before. The prior-turn window is on by default so existing LLM-classifier deployments actually get the fix; the config field documents that those turns reach the classifier model, which may be a different provider than the routed completion model, and that the call already carries the current ask and the caller's system prompt in full. Scoped to the ComplexityRouter; the semantic AutoRouter is not touched. |
||
|
tin-berri
|
71dfab7177
|
feat(router): record why the auto-router picked a tier and show it in the logs (#35016)
Auto-routed requests were indistinguishable from ordinary ones once logged: the spend log recorded the requested model group and the resolved deployment, but nothing about which tier was chosen or what chose it. That information existed only inside verbose_router_logger f-strings, so answering "why did my prompt land on the cheap model" required log access and a running proxy. The complexity, quality, and adaptive pre-routing strategies now return a typed StandardLoggingRoutingDecision on their PreRoutingHookResponse, and Router.async_pre_routing_hook records it once for every attempt. Those three previously side-channelled their own state through three different metadata keys; the decision now travels on the hook contract itself, so the bucket is resolved in one place, through get_or_create_metadata_bucket, which already owns the question of which dict holds proxy-internal metadata and replaces a non-dict value instead of skipping the write. Recording happens on every attempt rather than only on a successful route: a fallback from an auto-router group to a plain group re-enters the hook with the same request kwargs, and a decision left behind there would attribute the first router's tier to the deployment that actually served the retry. The log details drawer renders the result as a Routing card between Request Details and Metrics; the card is absent on rows that carry no decision, so ordinary and pre-upgrade rows are unchanged. Three defects surfaced while making the recorded cause truthful, each of which would have persisted a wrong answer. The complexity router hardcoded cause=complexity_scorer even when the LLM classifier decided, and its silent fallback to the heuristic on classifier failure meant a row could claim an LLM verdict the LLM never gave; the cause now reports the path that actually ran. The keyword that triggered a tier rule was discarded before logging, as was the escalation keyword. The 2-reasoning-marker override returned REASONING with a score far below the REASONING boundary and no marker saying so, which reads as a scoring bug to anyone comparing the two; it now emits a reasoning-override signal, and the card labels those rows as an override instead of claiming the score met a boundary. The LLM path no longer reports a synthetic score of 1.0, and heuristic decisions carry a snapshot of the tier boundaries that mapped the score, so a historical row stays interpretable after the boundaries change. Signals name a matched term only when the caller's own message contains it. Scoring still reads the system prompt, but a term matched solely there is reported as a count, since signals reach a spend row the caller can read and naming one would disclose a term from a prompt it cannot see. routing_decision is stripped from caller-supplied metadata at ingress, so a client cannot forge its own provenance. |
||
|
Tin Chi Lo
|
3d5b8e5960 |
fix(complexity_router): propagate turn_off_message_logging to internal sub-calls
The classifier and semantic-embedding sub-calls now capture proxy_server_request, but neither forwarded the caller's turn_off_message_logging opt-out. A caller who disabled message logging still had their prompt stored in the clear in these internal sub-calls' spend-log rows, since should_redact_message_logging reads the flag per-call and this internal call never inherited it. |
||
|
tin
|
78207064d8 |
fix(complexity_router): log the classifier request on chat completions too
The classifier read its metadata only from litellm_metadata, which the proxy populates just for LITELLM_METADATA_ROUTES (/v1/messages, /v1/responses, ...); /v1/chat/completions puts it under metadata, so the classifier call arrived unattributed and _should_track_cost_callback dropped it, leaving no spend-log row at all for the captured request body to show up in. Also log response_format in the wire shape litellm actually sends (type_to_response_format_param) instead of the bare pydantic JSON schema Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> |
||
|
Tin Chi Lo
|
2a84c39762 | fix(complexity_router): capture the classifier request body in spend logs | ||
|
tin-berri
|
7ac5172686
|
Merge pull request #33290 from mihidumh/fix/latency-routing-timedelta
fix(router_strategy): serialize latency for non-chat responses in lowest-latency routing |
||
|
Mihidum Hettiyahandi
|
ac8ee71512 |
test(router_strategy): cover chat-path normalization branches in both handlers
Codecov flagged the ModelResponse-branch lines as uncovered: add async per-token normalization, plus zero-completion-token fallback tests for both handlers (the else branch storing plain float seconds). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mihidum Hettiyahandi
|
b5c363e016 |
fix(router_strategy): serialize latency for non-chat responses in lowest-latency routing
log_success_event/async_log_success_event only converted the end_time - start_time timedelta to float seconds inside the isinstance(response_obj, ModelResponse) branch, so every embedding / speech / image response appended a raw timedelta to the latency list and broke the Redis cache sync with 'Object of type timedelta is not JSON serializable' (no cross-replica latency sharing for those model groups + error-log spam). Normalize response_ms to float seconds up-front in both handlers. Completes the partial fix from #14040. Fixes #33169 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
devin-ai-integration[bot]
|
cc45d18e9c
|
feat(complexity-router): add return_raw_model_name toggle for response model field (#33875)
* feat(complexity-router): optionally return raw model name Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(proxy): restore asyncio import Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * chore(tests): preserve staging asyncio import Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(proxy): drop unused local asyncio import Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * feat(dashboard): add complexity router raw model toggle Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * refactor(complexity-router): move metadata key constant to constants.py Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * style(proxy-tests): preserve module spacing Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Krrish Dholakia <krrishdholakia@berri.ai> Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> |
||
|
devin-ai-integration[bot]
|
b0a0f11b09
|
feat(complexity-router): user-triggered escalation keywords (#33656)
* feat(complexity-router): user-triggered escalation keywords Add an escalation_keywords config option to the complexity router so a user can force a bump to the next-higher complexity tier by including a phrase in their message (a stronger model, but not one they get to choose). Defaults to ['LITELLM ESCALATE'] when unset, case-sensitive so it only fires on the deliberate shouted form; admins can override the list or set [] to disable. Escalation applies across every routing path: heuristic/LLM classification, literal and semantic keyword_tier_rules overrides, adaptive routing, and session affinity (where it bumps relative to the pinned model and persists the higher tier for the rest of the session). Capped at the highest configured tier and skips unconfigured intermediate tiers. Expose it in the Auto-Router v2 UI as an Escalation Keywords field wired into the complexity_router_config payload. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(complexity-router): validate escalation keywords and pin at tier ceiling Strip blank/whitespace escalation keywords so an empty phrase can't match every message and escalate all traffic. Keep the exact pinned model when a session escalates at the highest configured tier instead of randomly hopping to a peer in a multi-model pool. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Krrish Dholakia <krrishdholakia@berri.ai> Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> |
||
|
devin-ai-integration[bot]
|
637fc1f60e
|
fix(router): tag-aware pre-routing strategy selection for shared model_name (#33691)
* fix(router): tag-aware pre-routing strategy selection for shared model_name Complexity/auto/adaptive/quality router registries were keyed by model_name alone, so a second deployment sharing a model_name but carrying different tags was rejected and every request used the first config. This made tag-based routing to distinct provider configs behind one alias impossible, surfacing as 401 'Not allowed to access model due to tags configuration' for the second tag. Each registry now holds a list of tag-scoped strategies and async_pre_routing_hook selects the entry whose tags match the request before classification, falling back to a default-tagged then first-registered entry. A repeat of the same (model_name, tags) pair is still rejected. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(router): cover tag-scoped pre-routing strategy registry helpers Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * chore: re-trigger CI Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Krrish Dholakia <krrishdholakia@berri.ai> Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> |
||
|
Yassin Kortam
|
21ba9692c3
|
fix(router): apply team/key enable_tag_filtering to tag routing (#33436)
Team/key router_settings.enable_tag_filtering was stored and echoed by /team/info but never applied at request time: the per-request override whitelist in route_llm_request.py dropped it, tag filtering only read the router-level flag, and UpdateRouterConfig silently discarded the field on /key/generate and /config/update. Requests from teams with the toggle on were load balanced across all deployments instead of tag-matched ones. - add enable_tag_filtering to the router_settings_override whitelist and strip any client-supplied copy from the request body first, so only the key/team value reaches the router - run tag filtering when the request carries enable_tag_filtering=True; a request-level False cannot disable a router-level True, so per-request settings can only scope down, never escape the global policy - add the field to UpdateRouterConfig so key and config update paths stop dropping it, and to all_litellm_params so it never leaks into provider request bodies - allow it through Router.update_settings/get_settings so the global UI toggle persists across DB config reloads Resolves LIT-4390 |
||
|
Yassin Kortam
|
c6778b79c3
|
fix(router): honor per-request routing_strategy from key/team router_settings (#33429)
* fix(router): honor per-request routing_strategy from key/team router_settings Key and team router_settings.routing_strategy was stored and shown in the UI but never forwarded to the shared Router, so the global strategy always won. Forward it through router_settings_override and resolve it in _get_routing_context: a validated per-request strategy takes precedence over routing groups and the top-level strategy, with lazily built cached selectors for strategies that need one. Unknown or unsupported strategy values are ignored with a warning instead of failing the request, and routing_strategy is registered in all_litellm_params so it is stripped before the provider call. * fix(router): sweep override selectors on strategy re-init and cover coverage-gate helpers routing_strategy_init now unregisters cached per-request override selectors so a later update_settings strategy change cannot leave a zombie selector receiving callback events. Adds direct tests for the two new helpers so the router code coverage gate passes. * docs(team): document mcp_rpm_limit in update_team docstring The documentation CI job walks management_endpoints and requires every UpdateTeamRequest field to appear in the update_team docstring; mcp_rpm_limit was added to the model without a docstring line, failing the job on unrelated PRs depending on walk order. Regenerates schema.d.ts since the docstring feeds the OpenAPI spec. |
||
|
devin-ai-integration[bot]
|
bf3a058781
|
feat(complexity_router): enable session_affinity by default (#33500)
Pin a session's first-turn model for the rest of the session by default instead of reclassifying every turn. Keeps multi-turn sessions on a single model, preserving provider prompt caches and avoiding cross-model conversation-history errors (e.g. Anthropic rejecting a thinking block produced by a different model). Requests without a resolvable session_id are unaffected. Set session_affinity: false to opt out. Co-authored-by: Krrish Dholakia <krrishdholakia@berri.ai> Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> |
||
|
devin-ai-integration[bot]
|
fac43df9b9
|
fix(complexity_router): return empty dict from _classifier_call_metadata when metadata is absent (#33452)
* fix(complexity_router): return empty dict from _classifier_call_metadata when metadata is absent
The LLM classifier reads request_kwargs.get("litellm_metadata"), but the proxy stores request metadata under "metadata", so this returned None. _classifier_call_metadata then passed None straight through to the classifier acompletion call, which assumes a dict and blows up with 'NoneType' object has no attribute 'update'; the router swallowed it and silently fell back to heuristic scoring, so the configured LLM classifier never ran. Returning an empty dict keeps the classifier call well-formed.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): cover complexity-router LLM classifier routes over the proxy
Add a live e2e regression for the complexity auto-router: a lexically simple but hard prompt ("Is P equal to NP?") is routed by the LLM classifier to the higher-tier anthropic backend, read back from the spend log's model. Before the metadata fix the classifier silently crashed and the router fell back to heuristic SIMPLE scoring on the openai backend, so this test fails pre-fix and passes post-fix.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Krrish Dholakia <krrishdholakia@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
|
||
|
Krrish Dholakia
|
b96460608d
|
feat(router): resolve auto-router routing plugins from proxy YAML config (#33251)
* feat(router): resolve auto-router routing plugins from proxy YAML config Router(plugins=[...]) was Python-SDK constructor only, so proxy/YAML users had no way to configure it, and the merged pipeline narrowed candidates from the outer model alias rather than the auto-router's actual tier pool, making it a no-op for auto_router deployments. Add complexity_router_config.plugins (dotted-path strings resolved via get_instance_fn, the same convention litellm_settings.callbacks uses) and run the resolved plugins against ComplexityRouter's tier pool at every model-pick site, so a policy plugin narrows what get_model_for_tier actually returns instead of the outer alias list. adaptive=True with plugins set now raises at config validation instead of silently ignoring the plugins, since the bandit selector doesn't consume narrowed pools yet. Also fixes a latent bug in Router._generate_model_id: it json.dumps every litellm_params dict value to build a deployment hash id, which crashed once a live plugin object could land inside complexity_router_config. * fix(router): use stable class name, not object repr, in model-id json fallback json.dumps(v, default=str) on a litellm_params dict containing a live RoutingPlugin instance fell back to object.__repr__'s default <module.Class object at 0x...>, embedding the instance's memory address. _generate_model_id's hash (and therefore the deployment id) changed on every process restart/hot-reload for any deployment with complexity_router_config.plugins configured, defeating the function's own "consistently generate the same id" contract and orphaning anything keyed on that id across restarts (e.g. Redis-backed per-deployment state). Use the plugin's fully-qualified class name instead, which is stable across restarts. * test(router): cover _json_default_stable_id for router_code_coverage gate router_code_coverage.py's AST scanner requires every router.py function be called by name somewhere in tests/, and flagged the new _json_default_stable_id helper from the previous commit. * fix(router): close two routing-plugin policy-bypass gaps flagged by Veria AI Session-affinity pin shortcut: async_pre_routing_hook returned a session's first-turn pinned model on every later turn without ever re-running it through the plugin pipeline, so a policy plugin (e.g. a budget cap crossed mid-session) was only enforced on turn one. Now the pin shortcut is disabled whenever plugins are configured, so every turn re-runs _classify_and_route (and therefore the plugins). Plugin resolution validation: get_instance_fn accepts any dotted path and returns whatever object it finds there, so a misconfigured complexity_router_config.plugins entry passed proxy startup silently and only surfaced as a confusing AttributeError on the first request that reached the plugin pipeline. Extracted the resolution logic into resolve_complexity_router_plugins() and added an isinstance(..., RoutingPlugin) check that fails proxy startup immediately with a clear error instead. * fix(router): raise instead of falling back to default_model on empty plugin-narrowed tier default_model was never checked against the configured plugins, so it functioned as an unconditional escape hatch around whatever policy a plugin enforces -- a tenant/budget plugin narrowing a tier to zero candidates could still be bypassed by the fallback. Drop the fallback entirely for this path; a plugin narrowing to zero is a policy decision, not something to route around, matching the fail-closed behavior the Router-level plugin pipeline already uses for the same situation. Flagged by Veria AI on PR #33251. * style: ruff format complexity_router.py * style(proxy): use modern str | None instead of Optional[str] in resolve_complexity_router_plugins * fix(router): stop default_model short-circuit from skipping plugins on no-user-message path self.config.default_model or await self._pick_model_for_tier(...) -- Python's `or` short-circuits on a truthy default_model, so _pick_model_for_tier (and therefore the plugin pipeline) never ran at all for the no-user-message path whenever default_model was configured. A tenant/budget plugin's decision was silently bypassable this way even after the other two policy-bypass fixes, since this call site had a different shape from the other three pick sites. Removed the short-circuit; falls through to _pick_model_for_tier -> get_model_for_tier, which already checks the MEDIUM tier before default_model -- the same priority every other call site uses. Flagged by Veria AI on PR #33251. * fix(router): address Greptile findings on the plugin-bypass fixes Preserve default_model-first priority in the no-user-message path when no plugins are configured, instead of unconditionally flipping to the MEDIUM tier -- the plugin-bypass fix must not silently change model selection for the (much larger) population of users who don't use plugins at all. Gated on self.config.plugins, matching the pattern already used elsewhere in this PR, per CLAUDE.md's guidance against backwards-compat flags when a plain conditional does the job. Also close a gap in the plugin validation added earlier: @runtime_checkable only checks that `run` exists as an attribute, not that it's a coroutine function, so a synchronous `def run(self, context)` passed isinstance(resolved_plugin, RoutingPlugin) at startup and only failed at request time with a confusing TypeError. Added an inspect.iscoroutinefunction check. Both flagged by Greptile on PR #33251. |
||
|
Krrish Dholakia
|
397c84678b
|
feat(router): opt-in session affinity for complexity router (#33126)
* feat(router): opt-in session affinity for complexity router Complexity router reclassified every turn, which could flip the routed model group mid-session and break provider-side prompt caching. Add a session_affinity config flag: when a session_id is resolvable, pin the model chosen on the first turn and reuse it for the rest of the session, skipping reclassification. Pinned turns still stamp the adaptive bandit's chosen-model metadata so reward feedback keeps working when adaptive=True. * fix(router): refresh session-affinity TTL on hit, scope pin by API key Two issues from review: the TTL was only set on the first classification, so an active session outliving session_affinity_ttl_seconds silently lost its pin instead of refreshing as documented. And the cache key was scoped only by session_id, which is client-supplied and unauthenticated, so two different callers reusing the same session_id could poison each other's routing pin. Refresh the TTL on every cache hit, and namespace the cache key by the proxy-derived API key hash when available. |
||
|
Krrish Dholakia
|
39e0efa11d
|
fix(auto-router): correct Responses API tool_choice shape and propagate alias litellm_params (#32974)
* fix(anthropic-messages): send bare-string tool_choice to Responses API, propagate router-alias litellm_params
The Anthropic /v1/messages -> Responses API adapter always wrapped
tool_choice in an object ({"type": "auto"}, {"type": "required"}), but
the Responses API's tool_choice schema for these cases is a bare
string ("auto"/"required"/"none"). Sending the object shape to an
OpenAI-compatible backend (e.g. vLLM) fails Pydantic validation with a
400. The "none" case also fell through to "auto" instead of mapping to
"none".
Separately, litellm_params configured directly on a router-alias
deployment (auto_router/complexity_router, adaptive_router,
quality_router, or semantic auto_router) - e.g.
cache_control_injection_points, drop_params - were silently dropped
for every request through that alias. async_pre_routing_hook swaps
`model` from the alias name to the selected tier/route's model before
the deployment lookup runs, so the outbound call only ever merged in
the tier deployment's own litellm_params, never the alias's. Register
non-routing-config litellm_params from the alias deployment and apply
them to the request whenever a pre-routing hook substitutes the model.
* fix: satisfy ruff-strict-budget UP006 and router coverage checker
Use builtin dict[...] generics instead of typing.Dict for the two new
annotations introduced in the previous commit, since they pushed
UP006 over the codebase ceiling in ruff-strict-budget.json. Add a
direct unit test for _register_pre_routing_alias_overrides so the
text-based router_code_coverage.py checker sees it exercised by name.
* fix(router): replace alias-param denylist with a tight allowlist
_PRE_ROUTING_ALIAS_RESERVED_PARAMS excluded router-init-only keys from
the alias's litellm_params before forwarding the rest as request
kwargs, but GenericLiteLLMParams also holds deployment-management
fields (tpm, rpm, weight, tags, max_budget, budget_duration,
use_in_pass_through, litellm_credential_name, ...) on the same object.
Any of those left off the denylist would get silently forwarded as if
they were request kwargs.
Replace the denylist with a tight allowlist of exactly the two
request-shaping params this feature exists for - drop_params and
cache_control_injection_points - so unrelated management fields never
reach the outbound call regardless of what else GenericLiteLLMParams
grows to hold.
* fix(router): re-register adaptive-alias overrides on set_model_list reload
set_model_list() unconditionally clears pre_routing_alias_overrides on
every call (e.g. /config/reload), but _finalize_adaptive_router_if_configured()
skips rebuilding an AdaptiveRouter whose model_name already exists in
self.adaptive_routers - so _register_pre_routing_alias_overrides() never
ran again for an auto_router/adaptive_router alias after a reload,
silently dropping its drop_params/cache_control_injection_points.
Build the Deployment unconditionally and re-register its overrides even
on the skip-existing-router path; only the (expensive) AdaptiveRouter
construction itself stays skipped.
* style: ruff format after merging litellm_internal_staging
* fix(router): drop the alias-param allowlist, exclude only model
Per review discussion: instead of a router.py-local allowlist of exactly
which litellm_params an alias (auto_router/complexity_router,
adaptive_router, quality_router, semantic auto_router) can forward to
the request it routes, _register_pre_routing_alias_overrides now
forwards everything except `model` (the alias marker itself, e.g.
auto_router/complexity_router, never a real provider model).
Router-init-only fields (complexity_router_config,
complexity_router_default_model, auto_router_config,
auto_router_config_path, auto_router_default_model,
auto_router_embedding_model, adaptive_router_config,
adaptive_router_default_model, quality_router_config,
quality_router_default_model) now flow into request_kwargs unfiltered
too. That's safe because litellm.completion()/acompletion() already
strips anything in litellm.types.utils.all_litellm_params before
building the provider request - added these 10 keys there, alongside
the deployment-management fields (tpm, rpm, weight, ...) already listed.
Verified live: without that addition, complexity_router_config lands in
extra_body and ships raw to the provider; with it, it's stripped.
This moves the "which fields aren't real LLM params" list from a
router.py-local allowlist to the single existing global list every
completion() call already depends on, instead of maintaining two.
* refactor(router): look up alias litellm_params on demand instead of caching them
_register_pre_routing_alias_overrides cached each alias's litellm_params
into self.pre_routing_alias_overrides at deployment-init time, which
required keeping that cache in sync with set_model_list() reloads - the
exact bug the previous adaptive-router-reload fix was patching around
(AdaptiveRouter survives a reload, but the cache didn't always get
refreshed to match).
Delete the cache and the registration method entirely. async_pre_routing_hook
now looks up the alias's own litellm_params directly from self.model_list
via self.model_name_to_deployment_indices at request time, the same
model_list that's already correctly rebuilt on every set_model_list()
call. No second piece of state to invalidate, so the reload staleness
bug class isn't possible anymore, and it's less code than before.
|
||
|
Krrish Dholakia
|
26ab730bfa
|
feat(router): soft-floor adaptive mode for complexity router (#32947)
* feat(router): soft-floor adaptive mode for complexity router Let complexity_router_config.adaptive=true Thompson-sample across the union of tier pools with a tier-distance penalty, and wire the existing adaptive post-call bandit so mis-tiered requests can still recover. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): reattach adaptive hooks for hybrid complexity Finalize was wiping every AdaptiveRouterPostCallHook and only re-registering standalone auto_router/adaptive_router deployments, so complexity adaptive=true never received bandit updates. Co-authored-by: Cursor <cursoragent@cursor.com> * chore(router): drop unnecessary hybrid docstrings Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): attribute adaptive feedback Credit user reactions to the model that produced the previous response while keeping current-response signals on the serving model Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): tune hybrid cold defaults Use the cost-weighted policy that beat equal-pool complexity in the full bakeoff, and make the committed harness compare identical tier pools Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): preserve hybrid cold quality floor Sample only unobserved models in the classified tier until feedback exists, then apply adaptive scoring without mis-penalizing models shared across tiers Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): bound feedback context cache Cap retained session feedback so unique session IDs cannot exhaust router memory Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): preserve exhaustion signals Include tool-result exhaustion in adaptive feedback and clear strict lint regressions blocking CI Co-authored-by: Cursor <cursoragent@cursor.com> * refactor(router): remove stale owner cache Remove obsolete attribution state, tighten the embedded router type, and keep the test diff focused on adaptive behavior Co-authored-by: Cursor <cursoragent@cursor.com> * refactor(router): centralize hook cleanup Use the callback manager to discover and remove adaptive hooks across every registered callback list Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Krrish Dholakia
|
85f9bdd412
|
feat(router): add Router(plugins=[...]) routing-plugin pipeline (#32972)
* feat(router): add Router(plugins=[...]) routing-plugin pipeline Runs a sequence of user-supplied plugins before the routing decision is made. Each plugin reads/mutates a RoutingContext (messages, candidate models, metadata, signals); the narrowed candidate list is enforced when picking a deployment, raising rather than silently falling back if a plugin narrows to zero candidates. Prototype for the routing-plugin pipeline discussed in #32168. * fix(router): use ruff-modern typing, add raw/structured messages to RoutingContext - Use dict/list/X|None instead of Dict/List/Optional in new code, staying within the ruff strict-rule budget ratchet - Extract the guardrail-translation message normalization ComplexityRouter already had into a shared resolve_structured_messages() helper (litellm_core_utils/prompt_templates/factory.py), reused by ComplexityRouter and the new routing-plugin pipeline instead of duplicating it - RoutingContext now exposes both raw_messages (as received) and structured_messages (normalized across chat completions / Anthropic messages / Responses API), mirroring CustomGuardrail.apply_guardrail's pattern, per review feedback on #32972 - Add direct unit tests for _run_routing_plugins and _filter_by_routing_plugin_candidates (router_code_coverage gate requires every router.py function be called by name somewhere in tests/) * fix(test): rename to test_router_routing_plugins.py router_code_coverage.py's AST scanner only inspects test files whose filename contains the substring "router" -- test_routing_plugins.py doesn't match (routing != router), so it silently skipped this file and flagged _run_routing_plugins/_filter_by_routing_plugin_candidates as untested despite the direct unit tests added for them. * fix(router): fail closed when plugins are configured but the resolved routing path can't run them Router.completion() (and other sync entry points) resolves deployments via the synchronous get_available_deployment(), which never runs async_pre_routing_hook and therefore never runs the routing-plugin pipeline. async_get_available_deployment() itself falls back to that same synchronous method for routing strategies without an async-native selector (e.g. legacy "usage-based-routing" v1). Both paths would let a policy plugin (e.g. a deny-all rule) be silently bypassed. Raise instead of silently proceeding when self.routing_plugins is configured and the sync path is reached, since applying the pipeline to every selector path is a larger change out of scope for this PR. Per review: https://github.com/BerriAI/litellm/pull/32972/changes/BASE..bdfb583c2c6f8df10004fb249e11629d41ce71fa#r3565373303 |
||
|
Krrish Dholakia
|
f717e3b2f0
|
feat(router): random-pick multi-model complexity tiers (#32967)
* feat(router): random-pick multi-model complexity tiers Tier pools already make sense without adaptive; stop pinning lists to index 0 and shuffle within the classified tier instead. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(ci): format complexity router config Co-authored-by: Cursor <cursoragent@cursor.com> * fix(ci): use PEP 585 types for tier pools Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Abhimanyu Kapur
|
3385a32413 |
feat(complexity_router): log the cause of each routing decision
The complexity router's info log didn't say what drove a routing decision. Literal and semantic keyword matches logged an identical "keyword rule fired" line (no way to tell which mechanism fired), and the scorer's line carried no consistent marker tying it to the same question. Emit one greppable line per decision naming the cause: literal_keyword_match, semantic_keyword_match, or complexity_scorer. The hook already knows which ran (the config's semantic_keyword_matching flag distinguishes lexical from semantic; the override-vs-scorer branch distinguishes keyword match from scorer), so this is label-only: no behavior change, no new types, no added latency. Adds regression tests asserting each decision path logs its cause; they fail if a label is swapped or the cause= marker is dropped. |
||
|
Abhimanyu Kapur
|
952127647c |
fix(complexity_router): review hardening - blank keywords, router registry eviction, edit-modal controls
- config: KeywordTierRule now strips and drops blank/whitespace keywords (a stray "" makes _keyword_matches match every prompt, silently forcing that tier for all traffic); still requires at least one real keyword to remain - frontend build_complexity_router_config: trim keywords and drop rules left empty so an unfilled "Add keyword rule" row no longer ships a rule the backend rejects with a 400 in the heuristic (non-semantic) flow, where the client-side semantic guard doesn't run - proxy clear_cache / delete_model: the auto_router/ prefix also covers quality_router/ and adaptive_router/, so pop the model_name from all four router registries (no-op where absent) instead of only auto/complexity; otherwise a DB quality_router's stale entry made reload raise "already exists" and abort, and adaptive left a leak - frontend ComplexityRouterConfig: only render the Keyword Tier Overrides and Semantic keyword matching sections when their change handlers are provided, so the edit-auto- router modal (which omits them) no longer shows interactive-but-dead controls |
||
|
mateo-berri
|
9388d03089
|
fix(complexity_router): sanitize budget reservation inside forwarded user_api_key_auth | ||
|
Abhimanyu Kapur
|
4a4262aa0d |
fix(complexity_router): preserve user_api_key_auth in sub-call metadata
Removing user_api_key_auth entirely from classifier/embedding sub-call metadata (as _BUDGET_RESERVATION_METADATA_KEYS previously did) prevented _filter_deployments_by_model_access_groups from scoping those sub-calls to the caller's authorized access groups. An access-group-scoped caller could therefore reach embedding/classifier deployments outside their group. Only strip user_api_key_budget_reservation, which is the actual budget- reservation state that must not reach sub-calls. user_api_key_auth is now kept so access-group filtering works correctly for both the embedding path and the LLM classifier path. |
||
|
Abhimanyu Kapur
|
eb0c966255 |
fix(complexity_router): use max aggregation for semantic keyword route scoring
SemanticRouter defaults to mean aggregation across a route's utterances. Since each tier's route holds one utterance per configured keyword, a real semantic match on one keyword was averaged together with the tier's other, unrelated keywords and dragged below match_threshold — e.g. a MEDIUM tier with keywords [beep, boop, new york] never fired for a genuine "new york" paraphrase, because mean(sim_to_beep, sim_to_boop, sim_to_new_york) landed well under the threshold even though sim_to_new_york alone cleared it. Pass aggregation="max" so a tier matches if the query is close enough to any one of its keywords, not the average of all of them. Verified against live Voyage embeddings: raw cosine similarity for "new york" vs a paraphrase was 0.54 (above a 0.5 threshold), but the route scored 0.28 under mean aggregation and never matched; max aggregation fixes it. Adds a regression test with a tier holding one matching and two unrelated keywords, asserting the tier still fires; fails without aggregation="max". |
||
|
Abhimanyu Kapur
|
186d083bc7 |
fix(complexity_router): build semantic route index once under concurrent cold-start
Concurrent first requests each hit asyncio.to_thread to build the SemanticRouter index, firing duplicate embedding calls for the static route utterances. Guard the lazy build with a per-router asyncio.Lock (double-checked) so the index is constructed exactly once regardless of how many callers race in cold. Adds a regression test asserting ten simultaneous cold-start requests build the index the same number of times as a single request, and reworks the fake embedding router to count builds by how often a route utterance is embedded (robust to which embedding path the library uses) while still recording sync-call thread ids for the off-event-loop assertion. |
||
|
Abhimanyu Kapur
|
3c714ed7a6 |
feat(auto_router): keyword tier overrides and semantic keyword matching for the complexity router
Add deterministic keyword-to-tier overrides and optional embedding-based (semantic) keyword matching to the complexity router, and surface both in the Add Auto Router UI behind a Router Type selector: "Auto-Router v2 [Recommended]" (complexity tiers + keyword overrides + semantic matching, the default) and "Semantic Router [to be deprecated]" (the existing utterance-based router, unchanged). Keyword-to-tier overrides resolve to the highest tier matched rather than the first keyword matched, so match order no longer affects the routing decision. Backend: - config: KeywordTierRule model plus keyword_tier_rules, semantic_keyword_matching, embedding_model, and match_threshold on ComplexityRouterConfig, with a validator requiring an embedding model and rules when semantic matching is on - complexity_router: evaluate keyword rules before scoring; lexical matches escalate to the most-severe matched tier (order-independent), and semantic mode reuses LiteLLMRouterEncoder + SemanticRouter to match paraphrases by cosine similarity, falling back to the scorer when nothing matches - model management: clear complexity_routers on cache reload so config edits take effect Frontend: - Add Auto Router tab restores the Router Type radio (Auto-Router v2 recommended by default, Semantic Router still available) and sends keyword_tier_rules plus the semantic settings on the recommended path, instead of flattening keywords into custom_technical_keywords - client-side guard blocks submit when semantic matching is enabled without an embedding model or without any keyword tier rules, mirroring the backend validator - moved the "How Classification Works" explainer below Custom Technical Keywords and above Keyword Tier Overrides - remove the Test Connection action from the recommended flow, which can't build a valid pre-save payload for a router (leaves a TODO for a JSON preview / config test follow-up) Tests cover lexical escalation, semantic matching via the real library with injected embeddings, the semantic config guard, config validation, the reload-clear regression, and the frontend payload builder |
||
|
Krrish Dholakia
|
109193f26a
|
feat(router): add LLM-based classifier option to complexity router (#32169)
* feat(router): add LLM-based classifier option to complexity router Adds classifier_type: "heuristic" | "llm" to complexity_router_config. When set to "llm", the router calls a configured model (e.g. a small model like haiku) via structured output to pick the complexity tier, falling back to the existing regex/keyword scorer on any error, empty response, or unparseable output. * feat(ui): add classifier_type option to complexity router UI, fix edit flow Adds an "Advanced: Classification Method" section to ComplexityRouterConfig with a heuristic/LLM toggle, revealing a classifier model picker and timeout when LLM is selected. Also fixes the auto router edit modal, which never rendered the complexity router UI at all (it only handled the semantic router), and the "Edit Auto Router" button visibility check, which was gated on auto_router_config and never matched complexity router deployments. * fix(router): attribute classifier calls to caller, raise default timeout Forwards the original request's litellm_metadata into the classifier's acompletion call. Without it, the proxy's cost-tracking gate sees no user_api_key/team_id/user_id and silently drops spend logging and budget accounting for every classifier call, letting an authenticated user rack up unaccounted provider spend via repeated requests. Also raises the default classifier timeout from 400ms to 3000ms (400ms undershoots real LLM latency and would silently degrade to the heuristic scorer on most requests) and corrects the module/class docstrings, which still claimed zero external API calls after the llm classifier path was added. * fix(ci): resolve ruff strict-budget and frontend-lint failures - Use PEP 585 generics (dict/tuple/list) in the new aclassify/_classify_with_llm signatures instead of typing.Dict/Tuple/List, and suppress BLE001 on the intentionally broad except in aclassify's fallback path with a reason. - Fix prettier formatting in ComplexityRouterConfig.tsx. - Regenerate eslint-metrics.json (was stale after the classifier UI changes). * fix(ci): regenerate stale eslint-metrics.json * fix(router): strip parent budget reservation from classifier metadata The classifier's internal acompletion call previously forwarded the parent request's full litellm_metadata, including its budget reservation (user_api_key_budget_reservation / user_api_key_auth). That reservation belongs to the routed completion the classifier is deciding on, not to the classifier call itself, so it's now stripped while key/team attribution fields are still forwarded for spend logging. |
||
|
Mateo Wang
|
f628b41400
|
feat(complexity_router): add custom_technical_keywords config (#32262) | ||
|
Mateo Wang
|
fecaf5c9e5
|
feat(router): tag routing denylist support via ! prefix (#31728)
Adds `!` prefix negation to tag-based routing so callers can exclude deployments by exact tag value without enumerating every allowed alternative. `!provider:anthropic` removes all deployments tagged exactly `provider:anthropic` before routing, and positive and negation tags compose. Matching is exact literal membership (frozenset intersection), so there is no regex or ReDoS surface for client-supplied tags. Ban-only requests that carry only negation tags stay within the default pool, mirroring untagged-request semantics so callers can't use negation to escape it. Fallback chains keep working because get_deployments_for_tag runs on each routing hop Copy of #31680; implementation credit to @deepanshululla Co-authored-by: deepanshululla <15312873+deepanshululla@users.noreply.github.com> |
||
|
Sameer Kankute
|
133da06aa3
|
chore: litellm oss staging (#31185)
* fix(ui): widen Y-axis gutter on Usage charts so large token/request labels aren't clipped
The Total Tokens Over Time and Total Requests Over Time AreaCharts on the
Usage page used Tremor's default yAxisWidth (~56 px), which is too narrow
once totals pass the hundred-million mark — leading digits of labels like
"100.00M" / "4500.00M" got clipped against the chart edge. The requests
chart was worse: it formatted with toLocaleString(), so billion-scale
request counts produced "1,000,000,000" (13 chars) and overflowed
immediately.
Fix in two places so neither alone has to carry the whole margin:
- activity_metrics.tsx: add yAxisWidth={80} to both AreaCharts, and
switch the requests chart to the shared valueFormatter so it uses the
same compact k/M/B suffixes as the tokens chart.
- value_formatters.tsx: add a >= 1e9 branch to valueFormatter /
valueFormatterSpend that emits a "B" suffix (4.50B, $4.50B), keeping
every formatted label at most 7 chars.
Co-Authored-By: Claude Opus 4 (1M context) <noreply@anthropic.com>
* Update ui/litellm-dashboard/src/components/UsagePage/utils/value_formatters.tsx
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* docs(readme): add Deploy on AWS/GCP with Terraform section
Adds a quickstart for the two published Terraform modules on the public
registry (BerriAI/litellm/aws and BerriAI/litellm/google). Copy-paste
main.tf for each cloud, the one-time GCP Artifact Registry remote-repo
command, and pointers to the registry pages for the full input surface.
Sits inside the Get Started section, between the gateway/SDK table and
Run in Developer Mode -- where someone scanning the README for "how do I
deploy this" will land.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* docs(readme): add 1-click deploy buttons for AWS + GCP
GCP gets the real 1-click: Open in Cloud Shell badge that clones the repo
and walks through `terraform apply` via the existing DeployStack
tutorial (already shipped at terraform/litellm/gcp/examples/default/
TUTORIAL.md). User just picks a project.
AWS gets a soft 1-click: a Launch in AWS CloudShell badge that opens an
in-browser, already-authenticated shell. User runs four commands
(clone + cd + cp tfvars + terraform apply) once inside. There's no
native AWS deeplink that pre-clones a repo + runs a tutorial -- CFN
"Launch Stack" + CodeBuild would be needed for that, and that's a
separate piece of work.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* docs(readme): move AWS + GCP deploy buttons next to Render button
* docs(readme): unify deploy button sizes and badge styles
* docs(readme): bump deploy button height to 48 to match Render/Railway
* docs(readme): bump AWS/GCP badge height to compensate for SVG padding
* docs(readme): bump AWS/GCP badge height to 72
* docs(readme): bump AWS/GCP badge height to 84
* fix(readme): make deploy buttons same height (48px)
https://claude.ai/code/session_01MxQRMHSDXbqJh74rF86UBc
* docs(readme): flag GCP project ID substitution in image_registry
* docs(readme): equalize deploy button heights and fix Cloud Shell button font
GitHub rewrites an image's height attribute to "height: auto; max-height: Npx", which only caps and never stretches, so each image renders at its intrinsic height. The AWS/GCP shields badges are intrinsically 28px while the Render/Railway buttons are 40px, leaving the row uneven regardless of the height="48" we set. Replace the two shields badges with committed 40px PNGs so all four header buttons render at the same 40px.
Also swap the Cloud Shell button from open-btn.svg to open-btn.png. The SVG renders its label as live text with font-family "Roboto, Sans" and no generic fallback; since neither font exists in GitHub's render environment, the text fell back to a serif (Times New Roman). The PNG bakes in the correct typeface.
* docs(readme): collapse Railway deploy anchor to a single line
The Railway button wrapped its img across indented lines, so the anchor contained leading and trailing whitespace. GitHub underlines link content, rendering that whitespace as a small blue underline beside the button. Put the anchor on one line like the other three buttons so there is no inner whitespace to underline.
* Add Claude Fable 5 cost map entries as a data-only hotfix
Backports only the model map changes from #30064 so deployments on
released litellm versions pick up Fable 5 pricing, context window, and
the adaptive thinking flag through the hosted cost map fetch without
upgrading. Includes the supports_sampling_params flag on the 28
Fable 5 / Opus 4.7 / Opus 4.8 entries (ignored by released code, read
by the gating that ships with the next release) and the matching
one-line schema declaration so the map validation test passes.
https://claude.ai/code/session_01MZarYYT3aS7DxaNjoax6Gm
* fix: correct context window tokens for GPT-5 Pro and GPT-5.4 Mini/Nano
Three bugs in model_prices_and_context_window.json:
1. gpt-5-pro and gpt-5-pro-2025-10-06: max_input_tokens and max_tokens
were SWAPPED. GPT-5 Pro has a 400K context window (input) with 128K
max output, but the values were set as max_input=128000,
max_tokens=272000. This caused token limit errors when sending
prompts over 128K tokens to GPT-5 Pro.
2. gpt-5.4-mini and gpt-5.4-mini-2026-03-17: max_input_tokens was
272000, but GPT-5.4 Mini shares the same 1,050,000 token context
window as GPT-5.4. This was inconsistent with the azure/ variants
which already correctly had 1,050,000.
3. gpt-5.4-nano and gpt-5.4-nano-2026-03-17: same issue as Mini,
max_input_tokens was 272000 instead of 1,050,000.
Source: OpenAI model documentation and contextwindows.dev which
aggregates official context window sizes.
Fixes #30928 (partially — the issue incorrectly claims gpt-5/gpt-5-mini
should be 400K; their 272K values are correct per OpenAI docs)
* fix: also correct max_output_tokens for gpt-5-pro (272000→128000)
Per reviewer feedback, max_output_tokens was left at 272000 while
max_tokens was corrected to 128000, causing an internal inconsistency.
Both should be 128000 per OpenAI docs.
* fix(cost): price gpt-image generated output tokens as image tokens (#31147)
The OpenAI Images endpoints (/v1/images/generations, /v1/images/edits) return
usage with no output token breakdown — litellm's `ImageUsage` has no
`output_tokens_details` field — so generated-image OUTPUT tokens were priced at
the text rate (`output_cost_per_token`) instead of the image rate
(`output_cost_per_image_token`). For gpt-image-2 that is $10/1M vs $30/1M, a ~3x
undercount on the dominant cost component (image output is ~74% of spend). This
also affects azure gpt-image, which shares this calculator.
The OpenAI gpt-image cost calculator re-implemented usage handling instead of
reusing `calculate_image_response_cost_from_usage`, the shared helper that
azure_ai/gemini/vertex_ai already use. That helper classifies generated output
tokens as image tokens when the provider does not itemize output, and splits
text/image when it does.
Fix: route the ImageUsage path through `calculate_image_response_cost_from_usage`
(pre-transformed chat Usage objects are still costed directly). Adds a regression
test for the no-breakdown ImageUsage case (gpt-image-2).
* fix(bedrock): route application-inference-profile ARNs to converse (#18258) (#31098)
A bare application-inference-profile ARN passed as bedrock/arn:... fell
through to the invoke route, which cannot derive a provider from the
opaque profile id and raised 'Unknown provider=None'. The converse route
needs no provider, so detect these ARNs in get_bedrock_route and route
them to converse, matching the behavior of the already-documented
bedrock/converse/arn:... workaround.
Explicit invoke/ prefixes still win, and they remain a dead end for these
ARNs by design (no provider derivable). System-defined inference-profile
ARNs that embed a known model, and other opaque ARN types
(provisioned-model, imported-model, custom-model-deployment) that are
frequently invoke-only, are deliberately left on their current routes;
tests guard both boundaries.
* fix(moonshot): stop mutating caller messages on tool_choice='required' (#31060)
_add_tool_choice_required_message appended the "select a tool" prompt to
the caller's messages list in place, so transform_request corrupted the
caller's conversation history and appended a duplicate prompt on every
retry. Build and return a new list instead so the call stays idempotent.
Adds a regression test asserting the input messages list is unchanged
across repeated transform_request calls.
Co-authored-by: Wassbdr <wassim.badraoui07@gmail.com>
* fix(transcription): accept fractional usage.seconds in diarized_json responses (#30996)
gpt-4o-transcribe and compatible ASR backends return a diarized_json
response with usage={"type": "duration", "seconds": <float>}, e.g. 295.8.
TranscriptionUsageDurationObject typed seconds as int, so parsing the
response raised a pydantic ValidationError (int_from_float). That error
surfaces as an APIConnectionError which the router treats as retryable, so
it keeps re-calling the upstream (200 every time) until the upstream
rate-limits and returns 429 to the caller.
OpenAI specs this field as a float (see openai SDK UsageDuration.seconds),
so widen seconds to float. With the parse succeeding there is no exception
left to retry, which removes the loop.
Co-authored-by: Neimar Avila <19142978+neimaravila@users.noreply.github.com>
* fix(deepseek): drop non-function tools before chat completions call (#30910)
* fix(deepseek): drop non-function tools before chat completions call
DeepSeek's /chat/completions only accepts tools of type "function".
Requests bridged from /v1/responses can carry responses-API-native tool
types, for example a Codex CLI tool typed "namespace", which DeepSeek
rejects with "unknown variant 'namespace', expected 'function'" so the
whole request fails (issue #30722).
Filter unsupported tool types in the DeepSeek request transform so the
function tools still go through; when nothing callable remains, also drop
the now-dangling tool_choice and parallel_tool_calls
Fixes #30722
* test(deepseek): cover async tool filtering and document tool_choice assumption
Add an async_transform_request regression test so the sync and async tool
filtering paths cannot silently diverge, and document in _drop_unsupported_tools
that only non-function tools are dropped, so a function-named tool_choice always
references a surviving tool
* feat(catalog): add zai/glm-5.1, zai/glm-4.7-flash, openrouter/z-ai/glm-5.1 (#29840)
* feat(ui): surface team budget on key overview when key has no own budget (#30801)
* feat(ui): surface team budget on key overview when key has no own budget
* fix(ui): replace IIFE with derived variable and use find() for team budget display
* fix(anthropic): emit replayable streaming thinking blocks (#31022)
* feat(proxy): read cold-storage prompts back in the logs detail view (#30364)
* feat(proxy): read cold-storage prompts back in the logs detail view
When a deployment offloads prompts and responses to cold storage instead of
Postgres, the spend-log row holds only "{}" placeholders plus a
metadata.cold_storage_object_key pointer, so the UI logs detail drawer showed
nothing. The detail endpoint only read the placeholder columns and never
fetched the object back.
Resolve the payload per row based on actual content, not a config flag: if
Postgres has content, return it; otherwise read the exact stored object key and
fetch from the configured cold storage backend through ColdStorageHandler.
Reading the persisted key is a single GET. The key embeds a microsecond
timestamp that cannot be reconstructed from the millisecond-precision startTime
column, and listing the day's prefix to match on request_id would be too
expensive for this per-open path.
Also teach the detail drawer's pretty-view parser to accept a bare messages
array. The cold storage payload carries the prompt as a top-level messages list
with no proxy_server_request, so without this the output rendered while the
input stayed blank.
ColdStorageHandler gains an optional injected logger so the resolver can be unit
tested without monkeypatching. Postgres-stored prompts are unaffected: the fast
path returns the existing columns and the request-body object still renders the
same way.
* Update litellm/proxy/spend_tracking/spend_management_endpoints.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* test(proxy): cover ColdStorageHandler resolution paths and cold-storage fetch failure
Add unit tests for ColdStorageHandler (injected logger, graceful None when no
logger is configured, and resolution of a configured logger from the callback
registry) and a regression test asserting a cold storage backend exception
degrades to the Postgres values instead of surfacing a 500.
---------
Co-authored-by: Bytechoreographer <Bytechoreographer@users.noreply.github.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix(mavvrik): advance metricsMarker after upload; fix scheduler startup (#31068)
* fix(mavvrik): advance metricsMarker after upload + fix scheduler startup
Two bugs fixed:
1. deliver() never called PATCH /metrics/agent/ai/{connectionId} after a
successful GCS upload, so metricsMarker stayed at 0 and every daily run
re-exported the same dates in an infinite catch-up loop.
Fix: add _update_metrics_marker(date_epoch) called at the end of deliver()
after _upload_to_gcs() succeeds. A 4xx warns but does not raise (the GCS
file is already committed). A 410 raises consistent with the rest of the
destination.
2. init_mavvrik_focus_background_job runs at proxy startup before any LLM call
has triggered lazy instantiation of MavvrikFocusLogger, so it found no
logger instance and silently skipped registering the daily export job.
Fix: if no instance is found but "mavvrik" is in litellm.callbacks, call
_init_custom_logger_compatible_class to force instantiation before
the APScheduler job is registered.
* fix(mavvrik): catch up from earliest window when metricsMarker=0
When the connector is freshly registered, metricsMarker=0 parses to None.
The catch-up block was guarded by `if last_ingested and ...` which skipped
it entirely for None, so only yesterday was exported instead of the full
_MAX_CATCHUP_DAYS window.
Fix: treat None as being _MAX_CATCHUP_DAYS behind (start from earliest_catchup).
The existing > 7 day warning only fires for non-None markers that are old.
* fix(mavvrik): use now as end_time for yesterday's export window
LiteLLM_DailyUserSpend rows for a given date get their updated_at
bumped by the spend flush job throughout the next morning. The core
database query filters on updated_at, so capping end_time at midnight
(yesterday + 1 day) missed any spend rows flushed after midnight.
Fix: pass now (cron fire time) as end_time for the daily "yesterday"
window so all fully-settled rows are captured regardless of when the
flush job ran.
Verified: claude-3-5-sonnet BilledCost went from 0.0 to ~$2.40 per
row in the exported FOCUS CSV.
* fix(mavvrik): also use now as end_time for catch-up windows
* fix(mavvrik_focus): pass required args to _init_custom_logger_compatible_class
Calling it with only logging_integration raised TypeError at proxy startup
because internal_usage_cache and llm_router have no defaults. Also fix test
name to reflect the actual status code (5xx not 4xx) used in the mock.
* ci: retrigger CI run
* feat: pass through optional `instruction` field in the rerank API (vLLM/Qwen3-Reranker) (#30757)
* Add optional `instruction` passthrough to the rerank API
vLLM's /v1/rerank and /v1/score accept an optional top-level `instruction`
field (folded into the model's chat_template_kwargs and consumed by the
chat template — e.g. Qwen3-Reranker). LiteLLM's managed rerank route silently
dropped it: RerankRequest / OptionalRerankParams had no such field, so the
outgoing body was rebuilt without it.
Thread an opt-in `instruction: Optional[str]` through rerank()/arerank(),
get_optional_rerank_params, and the hosted_vllm transformation into the
request body, only when non-None. When callers omit it, model_dump(exclude_none)
drops the field and the outgoing request is byte-for-byte unchanged — fully
backward-compatible. (DeepInfra already forwards `instruction` via
non_default_params; this formalizes the field in the shared types.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Address review: thread `instruction` as a typed param + cover rerank_utils
Per PR review (greptile P2 + codecov):
- Make `instruction` a typed, named argument on the rerank provider interface
instead of recovering it from the opaque `non_default_params` blob. Adds
`instruction: Optional[str] = None` to `BaseRerankConfig.map_cohere_rerank_params`
and every provider override, and forwards it explicitly from
`get_optional_rerank_params`. hosted_vllm now reads the named param directly.
It is still also surfaced in `non_default_params` so providers that read it
there (e.g. DeepInfra) keep working now that `rerank()` consumes `instruction`
as a named param rather than leaving it in **kwargs.
- Add get_optional_rerank_params unit tests (present + absent) to cover the
previously-uncovered threading line flagged by codecov.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: scan rerank `instruction` through request guardrails
The rerank guardrail translation (CohereRerankHandler.process_input_messages)
only scanned `query`, so the newly added `instruction` field reached the
backend model unscanned. Since instruction-aware rerankers (hosted vLLM /
Qwen3-Reranker) fold `instruction` into the prompt, an authenticated caller
could place content there to bypass configured rerank request guardrails.
Generalize the handler to scan every user-controlled text field (`query` and
`instruction`) in one apply_guardrail call and write each sanitized value back
by index. Query-only requests are unchanged (single-element list at index 0);
non-string fields are left untouched. Adds tests covering instruction
scanning, PII masking write-back, and the non-string case.
Addresses the Veria AI security review on PR #30757.
* test: narrow Optional results before len() to satisfy basedpyright budget
The lint gate (basedpyright delta-vs-base budget) flagged one new
reportArgumentType: len(result.results) where results is
List[RerankResponseResult] | None. Assert results is not None first to
narrow the type before len()/indexing.
* fix: read rerank `instruction` from kwargs to satisfy basedpyright budget
The basedpyright delta-vs-base gate flagged one new reportArgumentType: the
Router forwards rerank calls via an untyped `**kwargs` unpack
(`litellm.arerank(**{**data, **kwargs})`), and declaring `instruction` as a
typed named param on the public `rerank`/`arerank` entrypoints made pyright
check that key against `str | None`, adding an error at router.py with no real
safety gain. Read `instruction` from kwargs in `rerank` instead.
It remains fully typed where it matters - threaded as a typed argument through
`get_optional_rerank_params` and each provider's `map_cohere_rerank_params`
(the original Greptile P2 ask). Whole-repo reportArgumentType is back to the
base count (net 0); rerank hosted_vllm + cohere guardrail suites pass; ruff clean.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(github_copilot): synthesize empty choices at the provider seam (#30929)
Newer Copilot Claude models (opus-4.7, opus-4.8) return responses with
choices=[], either carrying Anthropic-native content blocks or, for the
max_tokens=1 probe Claude Code sends, no content at all. github_copilot
is dispatched through the OpenAI SDK handler, which calls
convert_to_model_response_object directly and never invokes
GithubCopilotConfig.transform_response, so the empty-choices guard there
surfaced as a 500
Instead of synthesizing choices inside the shared
convert_to_model_response_object (which would silently turn empty choices
into a fabricated success for every provider), add a no-op
transform_parsed_response_dict hook on BaseConfig. GithubCopilotConfig
overrides it to synthesize choices from Anthropic-native content, reusing
its existing parsing, and the OpenAI SDK handler routes its parsed
response through the hook before generic conversion. The core utility
keeps treating empty choices as an error for all other providers
Fixes: https://github.com/BerriAI/litellm/issues/30927
Signed-off-by: David J. M. Karlsen <david@davidkarlsen.com>
* fix(router): stop fallback lookups from mutating the router fallbacks config (#30624)
* fix: correct amazon.titan-embed-text-v2 input price to $0.02/1M tokens (#29693)
* fix: correct amazon.titan-embed-text-v2 input price to $0.02/1M tokens
* test: scope local cost map env var with monkeypatch to avoid test pollution
* fix(sensitive_data_masker): fully mask secrets at or below the reveal threshold (#30764)
* fix(sensitive_data_masker): fully mask secrets at or below the reveal threshold
_mask_value did partial reveal by showing the first visible_prefix and last
visible_suffix characters, but for a value whose length was at or below
visible_prefix + visible_suffix (8 by default) it returned the value verbatim.
A value of exactly 8 chars fell through the length guard and computed
masked_length == 0, reconstructing the original string with no mask characters;
anything shorter hit the early return. Either way short credentials were emitted
in plaintext.
mask_dict routes real secrets through this path, so an 8-char-or-shorter redis
password, api key, or token could be written to logs and the UI unmasked. The
sibling helper mask_sensitive_keys already guards this case; _mask_value now does
the same by fully masking any value at or below the threshold.
* fix(sensitive_data_masker): add mask_short_values opt-out for truncation callers
Fully masking short values is the right default for secret masking, but
CooldownCache reuses the masker purely to truncate exception messages to the
first 50 characters, and it relies on short messages being returned readable.
Masking those blanked out short exception text and broke its tests.
Add a mask_short_values flag (default True, secure) and have CooldownCache pass
False so it keeps the truncation behavior, while every secret-masking caller
still gets short values fully masked.
* fix(mcp_debug): opt out of short-value masking to keep diagnostic token preview
MCPDebug uses the masker to preview auth tokens in debug headers and documents
that values of 10 chars or fewer are shown unchanged so token types stay
distinguishable. Pass mask_short_values=False so that diagnostic behavior is
preserved while secret maskers keep masking short values.
* fix(mcp_debug): mask short auth values in debug headers instead of echoing them
Earlier this masker opted out of short-value masking to keep a token preview, but
that echoes short authorization and token values verbatim in debug response
headers, which is the same leak this change is meant to close. Auth material
should never be emitted in full, so mask short values here too; the first/last
character preview still applies to longer tokens. Only CooldownCache keeps the
opt-out, since it truncates exception text rather than masking secrets.
* test(mcp_debug): assert masked short value preserves length
* refactor(fireworks_ai): remove deprecated audio transcriptions endpoint (#30917)
Fireworks AI deprecated audio inference on 2026-06-10
(https://docs.fireworks.ai/updates/changelog#audio-inference-and-image-generation-deprecation).
Live API testing confirms the endpoint is already non-functional: a valid
Fireworks API key receives HTTP 401 "Unauthorized" from
api.fireworks.ai/inference/v1/audio/transcriptions for every request,
regardless of payload. The audio-prod.api.fireworks.ai host referenced in
the test suite returns 401 for every path; the entire host is decommissioned.
Remove the dead FireworksAIAudioTranscriptionConfig class and every
reference to it across the codebase:
- Delete litellm/llms/fireworks_ai/audio_transcription/ directory (17-line
config class that inherited from OpenAIWhisperAudioTranscriptionConfig)
- Remove the Fireworks branch from
ProviderConfigManager.get_provider_audio_transcription_config() in
litellm/utils.py; update the stale comment in
get_optional_params_transcription that referenced fireworks ai
- Remove the FireworksAIAudioTranscriptionConfig entries from
LLM_CONFIG_NAMES and _LLM_CONFIGS_IMPORT_MAP in
litellm/_lazy_imports_registry.py
- Remove the TYPE_CHECKING re-export in litellm/__init__.py
- Remove the transcription branch in the fireworks_ai case of
get_supported_openai_params() in
litellm/litellm_core_utils/get_supported_openai_params.py
- Remove the whisper-v3 and whisper-v3-turbo entries from
model_prices_and_context_window.json and
litellm/model_prices_and_context_window_backup.json (both had
mode: audio_transcription and zero-cost pricing)
- Remove the TestFireworksAIAudioTranscription test class and its
imports from tests/llm_translation/test_fireworks_ai_translation.py
No other provider is affected. The openai_compatible_providers list,
FireworksAIMixin, and the OpenAI Whisper transcription handler all stay
because they are shared with other Fireworks endpoints and other
providers. The provider_endpoints_support.json registry already had
audio_transcriptions set to false for fireworks_ai.
* feat: add darkbloom provider (#30876)
* feat: add darkbloom provider
* fix: document darkbloom provider endpoints
* fix: address darkbloom review feedback
* fix: update darkbloom tool metadata
* fix: fail fast for non-Postgres database URLs (#30883)
* fix(proxy): fail fast on non-PostgreSQL DATABASE_URL instead of hanging on startup
LiteLLM's Prisma datasource is pinned to provider = 'postgresql', so a sqlite:// or mysql:// DATABASE_URL can never connect.
Today that surfaces as an opaque startup stall where the port never binds, and a separate 'DB not connected' 500 on /key/generate when no DATABASE_URL is set at all leaves operators guessing what to configure.
Validate the DATABASE_URL / DIRECT_URL scheme in run_server before any Prisma call and exit with an actionable message naming the unsupported scheme.
Also reword CommonProxyErrors.db_not_connected_error to tell the operator to set DATABASE_URL to a postgresql:// connection string.
Add regression tests covering postgres acceptance and sqlite/mysql/mssql rejection.
* fix: resolve CI failures and proxy DB URL typing issue
* fix(proxy): fail fast on non-PostgreSQL DATABASE_URLs with clear startup errors instead of hanging
* Validate DIRECT_URL alongside DATABASE_URL startup guards
* fix(bedrock): surface modeled HTTP status for mid-stream error events so 5xx is retryable (#24608) (#30946)
* fix(bedrock): surface modeled HTTP status for mid-stream error events (#24608)
* test(bedrock): mid-stream server errors trigger streaming fallback (#24608)
* style(bedrock): black-format stream-error helper (#24608)
* fix(mcp): re-land native tool preservation with typed annotations (#30645)
* fix(mcp): preserve native tools in semantic filter hook with typed annotations
* fix(mcp): tighten _is_mcp_tool Chat Completions shape check
* fix(sambanova): return embeddings supported params instead of dropping them (#30937)
* fix(router): send fallback metadata when streaming (#30914)
When a streaming request triggers a fallback, there was previously no way to
know it happened. This commit addresses this in a few ways:
1. The response now correctly populates the fallback headers
(`x-litellm-attempted-fallbacks`) so callers know a fallback happened.
2. The correct model ID is passed in the streaming chunks.
3. A streaming chunk with the fallback error can be optionally sent back
to the client (opt-in) by passing `include_fallback_errors: true` in
the request.
The format of the fallback errors while streaming is intentionally OpenAI
compatible to not break existing libraries that parse these events. It was
tested with Vercel's AI SDK (ai-sdk.dev). It is also opt-in, so it is not
delieved unexpectedly to callers by default.
* fix(mistral): drop output-only reasoning fields from input messages (#30884)
LiteLLM attaches reasoning_content and thinking_blocks to assistant
responses. Replaying those assistant turns verbatim forwarded the fields
back to Mistral, whose input schema forbids unknown keys, so the whole
request failed with a 422 extra_forbidden and reasoning models became
unusable across multiple turns.
Strip both fields from assistant messages before the request is built, in
a spot that runs ahead of the image/file branch so it applies on every
path. Fixes #30835
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(perplexity): bill search queries at the per-request price, not 1/1000 of it (#30652)
* fix(perplexity): bill search queries at the per-request price, not 1/1000
The fallback cost calculator divided search_context_cost_per_query by
1000, but that field stores the per-request price in USD: sonar is
{low: 0.005, medium: 0.008, high: 0.012}, matching Perplexity's published
$5/$8/$12 per 1,000 requests expressed per request. The gemini cost
calculator reads the same field per request with no division (its
docstring calls it "the per-request cost").
The division understated search cost by 1000x on every Perplexity call
that falls back to manual calculation (i.e. when the API does not return
a pre-computed usage.cost). Use the value directly.
Update the tests that had encoded the /1000 factor in their expectations,
and drop an unused import flagged by ruff in the touched test file.
* test(perplexity): update integration test search-cost expectations to per-request
The integration tests still encoded the old /1000 search-cost factor, so
they failed once the fallback calculator was corrected to bill
search_context_cost_per_query per request. Update the four expected-cost
computations (and the high-volume dollar-value comments) to match.
* test(perplexity): drop unused mock imports flagged by ruff
* fix: include model_access_groups when expanding all-team-models in get_team_models (#30622)
* fix(fireworks_ai): return None for transcription in get_supported_openai_params
Fireworks AI deprecated audio inference on 2026-06-10; the endpoint is
decommissioned. Without an explicit transcription branch, requests with
request_type='transcription' fell through to the else and returned
FireworksAIConfig chat-completion params. Return None instead to signal
the provider does not support transcription.
* fix(proxy): gate include_fallback_errors behind expose_fallback_errors_to_caller setting
Without an operator gate, any authenticated caller could set include_fallback_errors=True,
trigger a fallback, and read raw upstream exception messages from the
x-litellm-fallback-errors header and the litellm-fallback-metadata SSE event.
Strip include_fallback_errors from request data in common_processing_pre_call_logic
when expose_fallback_errors_to_caller is not set, so the router never builds the
error list. Also gate _should_include_fallback_errors on the same setting as a
secondary check for the streaming SSE injection path.
* test(proxy): opt in to expose_fallback_errors_to_caller in streaming SSE test
The operator gate added in
|
||
|
Sameer Kankute
|
0ffab87da4
|
fix(router): enforce deployment budgets for dynamically added models (#29273)
* fix(router): enforce deployment budgets for dynamically added models Register deployment max_budget/budget_duration when models are added via upsert_deployment (e.g. /model/new) so RouterBudgetLimiting matches startup model_list behavior. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): address CI lint and router coverage for budget sync helpers Remove unused RouterBudgetLimiting import and add router unit tests for deployment budget helper methods required by router_code_coverage. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(router): clear stale deployment budget on upsert without limits Unregister deployment budget config when max_budget/budget_duration are removed, including upsert replace paths. Hoist provider budget logger lookup outside the provider loop. Co-authored-by: Cursor <cursoragent@cursor.com> * Fix mypy * fix black --------- Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Yassin Kortam
|
516b741de1 | feat: selectively apply routing strategy according to model name | ||
|
harish-berri
|
25cf097da8 |
add test(tag-routing): prevent header regex bypass for strict plain tags.
Add tests to validate the condition improve the conditional readability by naming the plain-tag check explicitly. |
||
|
Krrish Dholakia
|
e50f945ef7 |
refactor(adaptive_router): move update_queue out of litellm.proxy
Some checks failed
Unit Tests: Proxy DB Operations / proxy-db (auth-checks, tests/proxy_unit_tests/test_auth_checks.py tests/proxy_unit_tests/test_user_api_key_auth.py, 20, 8) (push) Has been cancelled
Unit Tests: Proxy DB Operations / proxy-db (key-generation, tests/proxy_unit_tests/test_key_generate_prisma.py, 30, 0) (push) Has been cancelled
Unit Tests: Proxy DB Operations / proxy-db (proxy-utils, tests/proxy_unit_tests/test_proxy_utils.py, 20, 8) (push) Has been cancelled
Unit Tests: Proxy DB Operations / proxy-db (remaining, tests/proxy_unit_tests --ignore=tests/proxy_unit_tests/test_key_generate_prisma.py --ignore=tests/proxy_unit_tests/test_auth_checks.py --ignore=tests/proxy_unit_tests/test_user_api_key_auth.py --ignore=tests/proxy_unit_tests/test_p… (push) Has been cancelled
Unit Tests: Security / security (push) Has been cancelled
P1 review: adaptive_router.py had a top-level import of
AdaptiveRouterUpdateQueue from litellm.proxy.db, which broke the
SDK/proxy boundary that every other router strategy respects. No
other router_strategy module imports from litellm.proxy at module
level.
The queue only depends on litellm._logging — it never needed to
live under litellm.proxy. Moved:
litellm/proxy/db/db_transaction_queue/adaptive_router_update_queue.py
→ litellm/router_strategy/adaptive_router/update_queue.py
tests/test_litellm/proxy/db/db_transaction_queue/
test_adaptive_router_update_queue.py
→ tests/test_litellm/router_strategy/adaptive_router/test_update_queue.py
Also switched the queue's logger from verbose_proxy_logger to
verbose_router_logger to match the new module's ownership.
P2 review: drop unused constant STAGNATION_JACCARD_EXACT from
config.py — it was defined but never referenced.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Krrish Dholakia
|
f1da202d9e |
fix(adaptive_router): P1 flusher hot-reload + P2 hook accumulation + CI
P1: start the adaptive-router flusher loop unconditionally at proxy boot
instead of gating on 'adaptive_routers is non-empty'. Adaptive routers
added via /config/reload after boot now have their queues drained.
State is lazy-loaded per router on first flush tick (new _state_loaded
flag on AdaptiveRouter) so hot-reloaded routers still get their
persisted priors.
P2: _finalize_adaptive_router_if_configured now prunes stale
AdaptiveRouterPostCallHook callbacks from every litellm callback list
before registering new ones. Without this, every Router replacement
left the old hooks wired up in litellm.callbacks and double-fired
signal recording for every request. Uses
logging_callback_manager.remove_callbacks_by_type (same pattern as the
semantic tool filter).
CI fixes:
- black --check failure: reformatted litellm/router.py
- schema migration diff: aligned @@index with the explicit index name
('idx_adaptive_router_session_activity') from the original migration
by adding 'map:' to all three schema.prisma copies. No new migration
needed.
Tests: 1 new covering the prune-on-hot-reload path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|