chore: sync upstream 2026-08-21

.circleci/config.yml

@@ -1025,7 +1025,7 @@ jobs:
           name: Run tests
           command: |
             mkdir -p test-results
-            TEST_FILES=$(circleci tests glob "tests/agent_tests/**/test_*.py" | grep -v "^tests/agent_tests/local_only_agent_tests/")
+            TEST_FILES=$(circleci tests glob "tests/agent_tests/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
               --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \

.github/ci-coverage-allowlist.yml

@@ -4,6 +4,25 @@ description: >-
   by a job nor listed here, so every entry below is a decision on the record.
 
 test_paths:
+  - reason: >-
+      The caching suite in tests/local_testing, which runs nowhere. Every job that globs that
+      directory either deselects it (local_testing_part1 and part2 carry `-k "... and not caching
+      and not cache"`) or keeps only another keyword (langfuse, router, assistants), and no job
+      names these files the way redis_caching_unit_tests names test_dual_cache.py. Measured
+      2026-08-20 by collecting the directory under each job's own selector: 118 tests across
+      these eight files are selected by none of them. Listed so the gap is a decision rather
+      than an accident, and so the --slices guard has a baseline to ratchet down from. Revisit
+      when tests/local_testing is ported off CircleCI, where the keyless part of this suite
+      belongs in a real job
+    paths:
+      - tests/local_testing/test_cache_preset_key.py
+      - tests/local_testing/test_caching.py
+      - tests/local_testing/test_caching_handler.py
+      - tests/local_testing/test_disk_cache_unit_tests.py
+      - tests/local_testing/test_gcs_cache_unit_tests.py
+      - tests/local_testing/test_prompt_caching.py
+      - tests/local_testing/test_responses_stream_cache_keys.py
+      - tests/local_testing/test_unit_test_caching.py
   - reason: >-
       The end-to-end suite runs against a deployed proxy from its own in-cluster rig rather than
       from a pull request; it needs a live gateway and provider credentials no PR job holds
@@ -21,72 +40,24 @@ test_paths:
       - tests/documentation_tests/test_requests_lib_usage.py
       - tests/documentation_tests/test_standard_logging_payload.py
   - reason: >-
-      Sibling files here are executed by name from the code-quality workflow; this one is referenced
-      by no job
+      Named like a test but shaped like a benchmark: it fetches live image URLs, times aiohttp
+      against httpx, prints the ratio, and asserts nothing, so pytest cannot collect it (its
+      functions take arguments, not fixtures) and running it beside its siblings in the
+      code-quality workflow would add a network dependency for a number nothing reads. Exempt
+      as a script rather than as an unresolved gap; revisit by deleting it once the aiohttp
+      choice it informed is settled
     paths:
       - tests/code_coverage_tests/test_aio_http_image_conversion.py
   - reason: >-
-      A second mirror of the package tree living beside tests/test_litellm, which is the mirror the
-      repo convention names; only test_no_hardcoded_secrets.py is invoked, from the linting
-      workflow, and whether this directory should exist at all is unresolved
+      The last file of a second mirror that sat beside tests/test_litellm and ran nowhere. Its
+      other 33 files landed in the real mirror during August 2026, 30 as moves and 3 by merging
+      their bodies into the live file of the same name. This one cannot follow either route yet:
+      its live twin was rewritten from 1268 lines to 9434, and of the 19 tests here 5 have no
+      counterpart while 25 assertions fail against today's code, so what survives that rewrite
+      is a judgement about the endpoints, not a merge. Revisit by deciding which of the five
+      behaviours still hold
     paths:
-      - tests/litellm/a2a_protocol/providers/pydantic_ai_agents/test_pydantic_ai_agent_headers.py
-      - tests/litellm/a2a_protocol/providers/pydantic_ai_agents/test_pydantic_ai_agent_transformation.py
-      - tests/litellm/integrations/helicone/test_helicone_gemini.py
-      - tests/litellm/litellm_core_utils/test_json_schema_validation.py
-      - tests/litellm/llms/anthropic/test_anthropic_reasoning_effort.py
-      - tests/litellm/llms/anthropic/test_anthropic_schema_filter.py
-      - tests/litellm/llms/azure/test_azure_embedding.py
-      - tests/litellm/llms/bedrock/embed/test_embedding.py
-      - tests/litellm/llms/bedrock/test_nova_imported_models.py
-      - tests/litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py
-      - tests/litellm/llms/gradient_ai/chat/test_gradient_ai_chat_transformation.py
-      - tests/litellm/llms/oci/chat/test_oci_chat_transformation.py
-      - tests/litellm/llms/openai_like/test_abliteration_provider.py
-      - tests/litellm/llms/openai_like/test_assemblyai_provider.py
-      - tests/litellm/llms/openai_like/test_empiriolabs_provider.py
-      - tests/litellm/llms/vertex_ai/agent_engine/test_transformation.py
-      - tests/litellm/llms/vertex_ai/gemini/test_transformation.py
-      - tests/litellm/llms/vertex_ai/text_to_speech/test_transformation.py
       - tests/litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
-      - tests/litellm/proxy/agent_endpoints/test_agent_rbac.py
-      - tests/litellm/proxy/common_utils/test_rbac_utils.py
-      - tests/litellm/proxy/management_endpoints/test_common_utils.py
-      - tests/litellm/proxy/management_endpoints/test_cost_estimate_endpoint.py
-      - tests/litellm/proxy/test_claude_code_marketplace.py
-      - tests/litellm/proxy/test_init_litellm_callbacks.py
-      - tests/litellm/proxy/test_prisma_engine_watchdog.py
-      - tests/litellm/proxy/vector_store_endpoints/test_vector_store_rbac.py
-      - tests/litellm/test_bedrock_extended_beta_models.py
-      - tests/litellm/test_bedrock_nemotron_super.py
-      - tests/litellm/test_proxy_auth.py
-      - tests/litellm/test_router_retry_backoff_headers.py
-      - tests/litellm/test_sambanova_model_metadata.py
-      - tests/litellm/test_stream_chunk_builder_images.py
-  - reason: >-
-      Legacy proxy suite superseded by the proxy shards; no job invokes it and whether it still
-      describes supported behaviour is unresolved
-    paths:
-      - tests/old_proxy_tests/tests/test_anthropic_context_caching.py
-      - tests/old_proxy_tests/tests/test_anthropic_sdk.py
-      - tests/old_proxy_tests/tests/test_async.py
-      - tests/old_proxy_tests/tests/test_gemini_context_caching.py
-      - tests/old_proxy_tests/tests/test_langchain_embedding.py
-      - tests/old_proxy_tests/tests/test_langchain_request.py
-      - tests/old_proxy_tests/tests/test_llamaindex.py
-      - tests/old_proxy_tests/tests/test_mistral_sdk.py
-      - tests/old_proxy_tests/tests/test_openai_embedding.py
-      - tests/old_proxy_tests/tests/test_openai_exception_request.py
-      - tests/old_proxy_tests/tests/test_openai_request.py
-      - tests/old_proxy_tests/tests/test_openai_request_with_traceparent.py
-      - tests/old_proxy_tests/tests/test_openai_simple_embedding.py
-      - tests/old_proxy_tests/tests/test_openai_tts_request.py
-      - tests/old_proxy_tests/tests/test_pass_through_langfuse.py
-      - tests/old_proxy_tests/tests/test_q.py
-      - tests/old_proxy_tests/tests/test_simple_traceparent_openai.py
-      - tests/old_proxy_tests/tests/test_vertex_sdk_forward_headers.py
-      - tests/old_proxy_tests/tests/test_vtx_embedding.py
-      - tests/old_proxy_tests/tests/test_vtx_sdk_embedding.py
   - reason: >-
       No job invokes this suite and its files mix pure transformation tests with ones driving live
       vendor vector stores, so assigning them needs a per-file decision
@@ -116,6 +87,14 @@ test_paths:
       - tests/load_tests/test_otel_load_test.py
       - tests/load_tests/test_vertex_embeddings_load_test.py
       - tests/load_tests/test_vertex_load_tests.py
+  - reason: >-
+      A local-only agent rig: test_a2a_completion_bridge.py needs a LangGraph server on
+      localhost:2024 and test_a2a.py drives a live A2A endpoint, so neither can run in a
+      pull request job. Until 2026-08-20 the CircleCI agent job hid them behind a grep -v
+      that this census could not see; the glob now excludes them structurally and this entry
+      is the decision on the record. Revisit when the A2A bridge gets a recorded-wire fixture
+    paths:
+      - tests/agent_tests/local_only_agent_tests
   - reason: >-
       Third-party integration tests that skip themselves without OCI configuration or sandbox
       credentials, neither of which a pull request job holds
@@ -124,14 +103,14 @@ test_paths:
       - tests/integration/test_oci_integration.py
       - tests/integration/test_oci_proxy_integration.py
   - reason: >-
-      Two prompt-factory tests sitting at the top level of tests/ instead of under the
-      tests/test_litellm mirror the shards enumerate; they need moving rather than a shard entry
-    paths:
-      - tests/litellm_core_utils/test_anthropic_dedup_factory.py
-      - tests/litellm_core_utils/test_bedrock_converse_dedup_factory.py
-  - reason: >-
-      A unit test for the proxy-extras package that no job invokes, while the package's other tests
-      live under tests/proxy_migration_tests
+      A unit test for the proxy-extras package that no job invokes, while the package's other
+      tests live under tests/proxy_migration_tests. Measured 2026-08-20: 24 of its 28 tests pass
+      and the 4 in TestMigrationSQLIdempotency fail, because 13 migrations from 2026-03 onward use
+      bare CREATE TABLE, ADD COLUMN, CREATE INDEX and ADD CONSTRAINT rather than the guarded forms
+      this file requires. It also matches those keywords inside SQL comments, so two further
+      migrations are reported that are in fact fine. Wiring it up means deciding what to do about
+      the 13 first, and they cannot simply be edited: Prisma checksums an applied migration, so a
+      changed one breaks migrate deploy for existing installs
     paths:
       - tests/litellm-proxy-extras/test_litellm_proxy_extras_utils.py
 

.github/fork-patches.txt

@@ -150,3 +150,8 @@ ui/litellm-dashboard/src/components/Settings/AdminSettings/PluginSettings/schema
 litellm/llms/bedrock/chat/invoke_handler.py | self.accumulated_reasoning_content: str = "" | 2026-08-20 sync merge corruption: AWSEventStreamDecoder.__init__ dropped the fork-only accumulated_reasoning_content init while a sibling `self.accumulated_reasoning_content += reasoning_content` accumulation line (added earlier, untouched by upstream's nearby edits) survived, raising AttributeError on every Converse streaming reasoning delta (7 test failures: test_converse_streaming_usage_accounts_for_streamed_reasoning, the Nova 2 streaming suite). Also restored the call site's dropped `reasoning_content=(self.accumulated_reasoning_content or None)` argument to converse_config.transform_usage, which the merge silently replaced with only the new upstream `thinking_ran` argument. Same drop-only-the-declaration/drop-only-the-argument class as the anthropic legacy-thinking translation and litellm_pre_call_utils import entries above. REMOVAL CONDITION: remove once upstream threads its own equivalent reasoning-content accumulator through AWSEventStreamDecoder.
 litellm/llms/bedrock/chat/converse_transformation.py | reasoning_tokens: Final = min(estimated_reasoning_tokens, output_tokens) | 2026-08-20 sync merge reverted transform_usage's fork-only reasoning-token clamp: our tokenizer re-estimates Converse's un-reported reasoning token count from raw text, and an overestimate must not drive completion_tokens_details.text_tokens negative. Upstream's version (which also added the new `thinking_ran` parameter, preserved here) has no clamp at all. Verified present and passing on the prior day's merged commit (5049a8f7d) before being silently dropped; broke test_transform_usage_clamps_reasoning_estimate_to_output_tokens (assert 11 == 1, text_tokens=-10). REMOVAL CONDITION: remove once upstream ships an equivalent reasoning-token clamp itself.
 basedpyright-code-budget.json | "reportPrivateUsage": {\n    "limit": 1833 | reportPrivateUsage 1823->1833 during the 2026-08-20 upstream sync: `type_check_gate.py --update` ratcheted 1148 errors down across 48 other rules (including reportOperatorIssue, fully resolved) but cannot raise an over-budget rule, so this one was bumped by hand to the exact verified current total, same class as the ruff-strict-budget.json/type-discipline-budget.json resets above. REMOVAL CONDITION: lower back toward 1823 via a dedicated cleanup pass, or accept future syncs nudging it further if upstream's own counts keep growing.
+enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py | attribution = self._get_job_attribution(job) | 2026-08-21 sync merge corruption, round 3 (same recurring class as the 2026-08-09/2026-08-14/2026-08-15 entries above): upstream independently grew its own second job-claim layer this cycle (_claim_job_for_costing / _release_job_claim, called from inside _track_completed_batch_cost right before the spend-log write) alongside the fork's pre-existing layer (_claim_job / _release_job_claim / _finalize_job, called from check_batch_cost before _track_completed_batch_cost even runs). Both additions land in disjoint hunks, so the merge auto-applied both with no conflict, silently producing two claim layers stacked on the same row: the outer _claim_job flips batch_processed=True before the inner _claim_job_for_costing runs, so the inner claim's own compare-and-swap (where batch_processed=False) always finds 0 rows and _track_completed_batch_cost always returns None without ever calling async_success_handler -- batches stop being billed at all, and the row is left oscillating between claim and release every poll cycle. Also silently shadowed: Python keeps only the LAST definition of a duplicate method name, so the fork's fenced _release_job_claim(status="pricing") was shadowed by upstream's simpler unfenced one, meaning the release path stopped restoring status to "validating" too. Fixed by deleting upstream's newly-added _claim_job_for_costing method, its duplicate _release_job_claim definition, and the wrapping claim-check/try-release-on-exception it added around the async_success_handler call in _track_completed_batch_cost, restoring this file to byte-identical with the pre-sync commit for this whole region. REMOVAL CONDITION: none; regression fix. Separately, NOT fixed by this patch: upstream's new claim timing (claim right before the spend-log write, after the results fetch) exists specifically because claiming earlier -- which is what the fork's _claim_job still does, unchanged since before this sync -- flips batch_processed=True before the output file has been read, and the managed-files deletion guard treats batch_processed=True as "safe to delete," so a concurrent delete can remove the very output file an in-flight costing run is reading. This is a real, pre-existing gap in the fork's design (not introduced by this sync) that tests/proxy_unit_tests/test_check_batch_cost.py::TestMultiPodBatchCostClaim (upstream's brand-new test class this cycle) catches; closing it requires moving the fork's claim point later while preserving its status="pricing" fencing, _reclaim_abandoned_pricing_claims sweep, and spend-already-recorded dedup marker, none of which upstream's simpler single-flag claim provides, so it is a deliberate follow-up, not part of this merge-corruption fix. REMOVAL CONDITION: track upstream's job-claim/locking layer maturity as in the 2026-08-09 entry above; additionally, reconcile the claim-timing gap once a deliberate redesign of the fork's claim point is scoped and reviewed.
+litellm/proxy/prisma_migration.py | LITELLM_PRISMA_CLIENT_PREBAKED | Skips the standalone entrypoint's own runtime `prisma generate` when set. Genuine upstream breakage, confirmed present on a clean upstream checkout (Dockerfile, docker/Dockerfile.non_root, and prisma_migration.py are all byte-identical to upstream/litellm_internal_staging): upstream's own PR #37692 (merged into this fork by the 2026-08-21 sync) made the entrypoint return prisma generate's exit code by default instead of only logging it, turning a previously-silent failure into a hard one. prisma-python's generator.generate() unconditionally re-copies schema.prisma into the installed package and chmod's the copy even when the content already matches (prisma/generator/generator.py compares paths, not bytes), and shutil.copy's chmod step (copymode) requires owning the destination file -- something no arbitrary non-root runtime uid does for a file baked into the image at build time, no matter how the image's permissions are set up. The runtime images already bake the generated client from this same schema.prisma at build time, so re-running generate at container start is pure redundant work; skipping it removes the only place that redundant call is made (migrations/run.py's ProxyExtrasDBManager never calls it). REMOVAL CONDITION: remove once upstream either stops re-running `prisma generate` in this entrypoint for prebaked images, or prisma-python skips schema.prisma re-copy/chmod when the destination already matches.
+Dockerfile | LITELLM_PRISMA_CLIENT_PREBAKED=true | Sets the runtime-stage env var read by litellm/proxy/prisma_migration.py's LITELLM_PRISMA_CLIENT_PREBAKED skip (see that file's fork-patches.txt entry for the full root cause). REMOVAL CONDITION: same as litellm/proxy/prisma_migration.py's entry.
+docker/Dockerfile.non_root | LITELLM_PRISMA_CLIENT_PREBAKED=true | Sets the runtime-stage env var read by litellm/proxy/prisma_migration.py's LITELLM_PRISMA_CLIENT_PREBAKED skip (see that file's fork-patches.txt entry for the full root cause); this Dockerfile's existing chown/chgrp of the venv's prisma package dir for GID 0 is necessary but not sufficient, since it cannot satisfy generate()'s chmod-requires-ownership check for an arbitrary uid either. REMOVAL CONDITION: same as litellm/proxy/prisma_migration.py's entry.
+enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py | _claim_job called from _track_completed_batch_cost | 2026-08-22 resolution of the claim-timing gap left open by the 2026-08-21 entry above. The fork claimed a row before fetching its results, which flipped batch_processed=True while the output file was still being read. Two consequences, both real: the managed-files deletion guard (_get_batches_referencing_file filters batch_processed=False) stopped holding the output file, so a concurrent delete could remove the very file the in-flight costing was reading; and a worker killed mid-fetch left the batch marked processed, unbillable by any other pod until _reclaim_abandoned_pricing_claims swept it hours later. Moved the claim to where upstream's own claim sits, inside _track_completed_batch_cost immediately before logging_obj.async_success_handler, WITHOUT adopting upstream's simpler single-flag claim: the fork's status="pricing" fencing, _reclaim_abandoned_pricing_claims sweep, _mark_spend_recorded dedup marker and fenced _finalize_job are all retained, so the crash-re-billing and disable_spend_logs protections they provide are unchanged. Losing the compare-and-swap now costs only a duplicated results fetch (the winner bills exactly once) and is signalled to the caller by the CLAIM_LOST sentinel, distinct from the unroutable-row None. Callers that previously released a claim they held across the fetch no longer do so: a failed fetch never took a claim and leaves the row untouched outright, while a failed spend-log write releases inside _track_completed_batch_cost. tests/proxy_unit_tests/test_check_batch_cost.py::TestMultiPodBatchCostClaim (upstream's class) now passes; its journal and claim-call assertions were adapted to the fork's write shape (claim carries status="pricing"; the marker and fenced finalize are separate writes) while every behavioural assertion upstream makes is kept. REMOVAL CONDITION: track upstream's job-claim/locking layer maturity as in the 2026-08-09 entry above; the claim-timing divergence itself is now closed, so only the extra fencing/sweep/marker remain fork-only.