The streaming iterator hook timed the whole provider stream and logged that as the
guardrail duration, so PrometheusLogger added LLM generation time to
litellm_overhead_with_guardrails_latency_metric. The hook now accumulates the time
spent inside _filter_single_text per chunk and logs that sum, keeping start_time and
end_time as the wall-clock window.
Resolves LIT-7589
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
A page size that does not divide SPEND_LOGS_PAGINATION_COUNT_CAP left the last
page starting inside the capped window and reading past it, so the rows
disagreed with the total reported next to them. The page limit now stops at the
end of that window, and has_more plus next_session_cursor still hand back a
cursor for walking further.
Claude-Session: https://claude.ai/code/session_01ESi9JwaXDww1vP3Qsrr4Mz
A page starting at or past SPEND_LOGS_PAGINATION_COUNT_CAP lies outside the
total the client is given, so it now returns no rows without running the page
query and the grouped top-N sort bound stays capped.
Rewrites the offset test to page a fake session store instead of asserting on
the generated SQL, and covers the last page inside the cap next to the first
page past it.
Claude-Session: https://claude.ai/code/session_01ESi9JwaXDww1vP3Qsrr4Mz
LiteLLM sends the Bedrock Mantle GPT rows through Bedrock's Responses endpoint, which refuses minimal on gpt-5.4 and gpt-5.5 like every other Bedrock GPT row. The earlier commit measured the raw chat endpoint, which accepts it, and dropped the flag by mistake. The ladder test now matches what the proxy path returns
Local runs measured 3.5 KB per request, so 10 KB keeps close to 3x headroom
while tightening from the earlier 12 KB.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Local runs measured p50 0.19s, p90 0.23s, p99 0.69s, so 2s/3s/5s left several
times that as slack. 1s/2s/3s keeps a comfortable margin while catching a
smaller regression than the looser ceilings would have.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Three local runs measured 1.33x-1.36x, so 2x is the tightest bound the data
supports and still catches a regression far smaller than 4x would. Noted in the
comment that this is the ceiling to loosen first if a weekly run trips it,
since core count shifts how much of baseline CPU is fixed per-request work.
Co-Authored-By: Claude Code <noreply@anthropic.com>
RSS moved 0.91x-1.40x across three identical local runs, so it stays loose at
2x rather than the arbitrary 1.5x carried over from the pre-padding-payload
calibration. CPU per request held steady at 1.33x-1.36x across the same runs,
so 4x replaces the looser 6x it inherited from stale numbers.
Co-Authored-By: Claude Code <noreply@anthropic.com>
A ratio against the healthy phase cannot bound either metric. Once the Redis
circuit breaker opens, a request skips Redis instead of waiting on its socket
timeout, so the chaos phase can measure cheaper than the baseline it is compared
against: local runs came in at 0.61x baseline p90 while a log-bytes ratio read
724x. Splitting Budget into RatioBudget and AbsoluteBudget lets RSS and CPU keep
the ratio they need, since both are machine-shaped, while latency and log volume
get the wall-clock ceiling a user actually cares about.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Live calls to Bedrock Mantle and Converse on 2026-09-11: the gpt-5.6 luna, sol, and terra rows and gpt-6-astra return 200 on reasoning_effort=max, gpt-6-astra returns 400 on none, and Mantle gpt-5.4 and gpt-5.5 return 200 on minimal. The commercial Bedrock rows now carry exactly those flags, and the schema test asserts the measured ladder per row instead of a blanket mirror of the direct OpenAI rows
* fix(datadog_llm_obs): keep tool call and result structure under redaction and emit tool output tokens
Under datadog_llm_observability_params.turn_off_message_logging the span kept only one role plus "redacted-by-litellm" per message, so Datadog showed Tool Call 0, Tool Result 0 and no tool output token data. The shared CustomLogger hook collapsed the messages before the callback ran, and the Datadog redaction then dropped tool_calls and tool_results.
The Datadog callback now opts out of the shared message collapse (redacts_messages_itself) and redacts its own normalized messages, keeping roles, tool names, ids and types while replacing content, arguments and results. Tool result tokens are counted with litellm.token_counter before redaction and shipped as the tool_output_tokens metric. Other callbacks keep the inherited behavior.
Resolves LIT-7545
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(datadog_llm_obs): drop explanatory docstrings from the redaction change
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(ui): regenerate schema.d.ts for the classifier descriptions changed in #40655
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): write failure spend log for guardrail-blocked /mcp-rest/tools/call
call_tool_rest_api only translated exceptions to HTTP responses, so a pre_mcp_call
guardrail block never reached failure_handler / async_failure_handler /
post_call_failure_hook and no LiteLLM_SpendLogs failure row was written. Extract
the failure logging from call_mcp_tool into _fire_mcp_tool_call_failure_logging
and run it in the REST route for anything raised between
common_processing_pre_call_logic and execute_mcp_tool
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): keep the original REST tool error when failure logging raises
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): log virtual mcp_tool_call failures and keep REST success latency scoped to tool execution
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Pad the locust payload to tens of KB so per-request bookkeeping cost scales with
body size instead of hiding behind a 40-byte prompt. Turn on
use_redis_transaction_buffer in the chaos config and JSON_LOGS in the workflow so
the spend buffer, pod lock, and JSON-encoded breaker tracebacks are all part of
the measured chaos cost. Add a log-bytes-per-request budget alongside latency,
RSS, and CPU, reading the proxy's log file size at each phase split; its ceiling
is uncalibrated since no chaos run has measured it yet.
Co-Authored-By: Claude Code <noreply@anthropic.com>
The Anthropic Messages route reaches the same Redis touchpoints and cost-tracking
callback through its own request path, so a failure-path regression there would not
surface from chat completions alone. Each simulated user now picks one endpoint round
robin and stays on it, and the per-endpoint split is asserted and reported so a run
that silently drove only one route fails instead of passing.
Co-Authored-By: Claude Code <noreply@anthropic.com>
* fix(proxy): keep call_type and request start time on failed-request spend logs
post_call_failure_hook pops litellm_logging_obj before the failure callbacks
run, so the spend row built from request_data had a blank call_type and used
datetime.now() as the start time. A guardrail-blocked MCP tool call therefore
showed up in the Logs page as an LLM row with no call type and a 0s duration.
Lift call_type and start_time off the logging object alongside the fields
already lifted, and have the DB failure hook prefer the lifted start time.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(proxy): inject the spend writer into _ProxyDBLogger instead of patching a module global
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(router): fall back from unhealthy auto-router tier
Co-Authored-By: Claude Code <noreply@anthropic.com>
(cherry picked from commit 00c7fd8376)
* fix(router): treat budget and tag exhaustion as a no-capacity verdict
The eligibility probe only read typed router errors as "nothing here can
serve this". Provider and deployment budget exhaustion, and tag routing
with no matching deployment, report it as a bare ValueError carrying a
RouterErrors marker, so the probe read a spent tier as live, skipped the
peer and default recovery, and failed the request.
---------
Co-authored-by: Tin Chi Lo <tin@berri.ai>
Co-authored-by: Claude Code <noreply@anthropic.com>
* fix(proxy): authorize every Responses API id, not only the ones the proxy issued
The ownership check on the Responses API only ran when the id arrived in the
proxy's own encrypted format. An id in any other shape skipped the check and
was forwarded upstream, so a key that did not own the response could retrieve,
cancel, delete, or chain off it.
Every addressed id now goes through one authorization step shared by retrieve,
cancel, delete, list-input-items, and create's previous_response_id. An id the
proxy did not issue is refused with 403 unless the deployment opts in with
general_settings.allow_unmanaged_response_ids, has responses id security
disabled, has no signing key configured, or the caller is a proxy admin.
* fix(proxy): re-authorize the retained responses id instead of trusting it
CLIENT PAUSE ALL for the length of the chaos phase instead of CLIENT PAUSE WRITE, so every Redis touchpoint on the request path times out rather than just the writes. The pause is sized to the phase because it freezes the control connection too; teardown's CLIENT UNPAUSE is a safety net for a phase that overran
Latency, RSS and CPU are now budgeted as chaos-over-baseline ratios (p50/p90/p99 for latency and RSS, CPU seconds per request once) through a small phase_budget module, replacing the machine-shaped absolutes. The Redis timeout rate is reported but no longer asserted
The final /metrics scrape waits for litellm_deployment_failure_responses_total to stop moving, since that counter is bumped from the async logging queue and lagged the load generator by thousands of increments. The model group carries a unique marker so a deployment left behind by an aborted run cannot absorb this run's retries
Co-Authored-By: Claude Code <noreply@anthropic.com>
Starlette scans the route table in registration order, so a request pays one
regex match per route registered ahead of its own. The proxy registers several
hundred routes and left the liveness probe near position 280 and the lazy
loaded /v1/messages at the very end. Move /health/liveliness, /health/liveness,
/v1/chat/completions, /chat/completions and /v1/messages to the front of the
route table after startup registration and again after a lazy router loads.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Both reasoning_effort mappers ended their if/elif chain in a bare ValueError.
exception_type() has no branch for ValueError, so it fell through to the shared
APIConnectionError fallback and the proxy answered a malformed client request
with a retryable HTTP 500 carrying no hint of the accepted values.
Raise UnsupportedParamsError (400) instead, listing the supported set, matching
what the Anthropic and Bedrock transforms already do and what this same file
already does at its five other param-validation sites.
This also covers 'xhigh' and 'max', which are members of litellm's own
REASONING_EFFORT literal but have no Gemini mapping, so callers bridging from
OpenAI-shaped code were hitting the 500 without typing anything wrong.
Fixes#40474
Claude-Session: https://claude.ai/code/session_01XT1qsbjLwnhiN5sQ2hNUxr
* feat(proxy): make the in-memory management cache capacity configurable
Add general_settings.user_api_key_cache_max_size (positive int, default 200) to resize the
in-memory tier of the shared user_api_key_cache at startup and on DB config reloads, expose it
in the Admin UI general settings, and cover it with behavioral tests. Prior art: #34726
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(caching): resize the in-memory tier from DualCache so any cache instance honours the cap
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(proxy): wrap the cache capacity field description to the 120 col limit
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
