chore(release): backport 1.84.8 patches + #30220 deps to stable/1.86.x and cut 1.86.6

[yuneng-berri]

Relevant issues

Backports the patch set that shipped in 1.84.8 onto the 1.86.x stable line, plus two
extras (#30220 dependency bumps and #29493 budget-reservation), and cuts 1.86.6. Every
fix was confirmed absent on 1.86.x before picking; 1.86.x branched from staging on
2026-05-16, so it predated almost the entire set. Nothing here is already on the line.

What is included

Cherry-picks in staging merge order (each carries a cherry picked from footer to its
staging commit):

• fix(router): use forwarded model_id for native Azure container IDs #27921 fix(router): use forwarded model_id for native Azure container IDs
• fix(proxy): expose Prisma idle/connect timeout + extra DB URL params #28395 fix(proxy): expose Prisma idle/connect timeout + extra DB URL params
• feat(proxy): add disable_budget_reservation general setting (#27639) #29493 feat(proxy): add disable_budget_reservation general setting
• fix(proxy): recover from cached-plan errors by reconnecting the Prisma client #29983 fix(proxy): recover from cached-plan errors by reconnecting the Prisma client
• feat(proxy): add option to disable server-side prepared statements for DB lookups #29984 feat(proxy): add option to disable server-side prepared statements for DB lookups
• fix(proxy): return 5xx on DB infra errors during auth; reserve 401 for genuine auth failures #29986 fix(proxy): return 5xx on DB infra errors during auth; reserve 401 for genuine auth failures
• fix(anthropic_passthrough): resolve costing model from message_start chunk, litellm_params and model_group instead of 'unknown' #30160 fix(passthrough): resolve costing model when body model is unknown
• feat: litellm oss 110626 #30202 (targeted subset) fix(passthrough): skip [DONE] sentinels and non-JSON SSE frames in Anthropic streaming logging
• fix(proxy): grace-period key rotation 401s; return deprecated-key lookup result directly #30327 fix(proxy): return deprecated-key lookup result directly in get_data combined view
• chore(deps): bump vitest, brace-expansion, pypdf and tornado #30220 dependency bumps (pypdf, tornado, aiohttp constraint, vitest, brace-expansion), applied as a deps backport
• bump: version 1.86.5 -> 1.86.6, then a uv.lock refresh

Adaptation notes

The source diff of every code pick is byte-identical to its staging commit; the adaptations
below are confined to tests, build artifacts, and the dependency manifests.

• fix(router): use forwarded model_id for native Azure container IDs #27921: the bundled _experimental/out/...js UI build artifact was dropped from the pick;
  build artifacts are regenerated and never hand-carried across lines. Its container tests
  came along and pass.
• feat(proxy): add disable_budget_reservation general setting (#27639) #29493: its only staging-reachable form is inside aggregator 32c88ca; the granular
  squash 1032dd7 is no longer on a ref. Picked 1032dd7 (diff content-verified equal to the
  aggregator's hunks) with the footer citing 32c88ca. Its tests came along and pass.
• fix(proxy): recover from cached-plan errors by reconnecting the Prisma client #29983, fix(proxy): grace-period key rotation 401s; return deprecated-key lookup result directly #30327: the regression test file test_prisma_client_get_data.py does not exist
  on 1.86.x (the test tree was reorganized after 1.86.x branched), so these ship source-only.
• feat(proxy): add option to disable server-side prepared statements for DB lookups #29984: the generated schema.d.ts does not exist on 1.86.x and was dropped; the
  proxy_cli tests came along and pass.
• fix(proxy): return 5xx on DB infra errors during auth; reserve 401 for genuine auth failures #29986: the test_auth_exception_handler.py additions patch seed_request_identity, an
  OTEL symbol that exists on staging but not on 1.86.x, so those were dropped (the source
  fix does not reference that symbol). The db/test_exception_handler.py regression tests
  (195 lines) came along and pass.
• fix(anthropic_passthrough): resolve costing model from message_start chunk, litellm_params and model_group instead of 'unknown' #30160: its handler test additions conflicted with the line's test file and were dropped
  (source-only).
• feat: litellm oss 110626 #30202 subset: only the anthropic_passthrough_logging_handler hardening hunks and their
  four tests; the rest of that staging batch is intentionally excluded. Tests came along and pass.
• chore(deps): bump vitest, brace-expansion, pypdf and tornado #30220: 1.86.x exact-pins the proxy-runtime extras, so pypdf moves ==6.10.2 -> ==6.13.1
  (staging used >=6.12.0,<7.0); the [tool.uv] constraints (tornado>=6.5.6, aiohttp>=3.13.5,<3.14)
  match staging. uv.lock and the dashboard package-lock.json were regenerated on this line
  rather than cherry-picked: pypdf 6.13.1, tornado 6.5.7, aiohttp held at 3.13.5, vitest 3.2.6,
  brace-expansion 5.0.6.

Known noise on this line

• tests/.../test_azure_container_transformation.py::...uses_azure_env_var fails locally
  only: litellm's find_dotenv walks up to the repo .env, which carries a real
  AZURE_OPENAI_API_KEY that the Azure key-resolution chain prefers over the test's
  monkeypatched value. The test is byte-identical to staging fix(router): use forwarded model_id for native Azure container IDs #27921 (already green on
  stable/1.84.x and in CI). It passes in an environment without that key.

Type

🐛 Bug Fix

Changes

See the cherry-pick list above. Source content for every pick is byte-identical to staging;
adaptations are test/artifact/dependency-manifest only.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
3 out of 4 committers have signed the CLA.

✅ Ar-maan05
✅ Sameerlite
✅ yuneng-berri
❌ yassin-berriai
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[yuneng-berri]

@greptileai

[codecov]

Codecov Report

❌ Patch coverage is 27.08333% with 105 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
..._handlers/anthropic_passthrough_logging_handler.py	10.00%	36 Missing ⚠️
litellm/proxy/proxy_cli.py	3.70%	26 Missing ⚠️
litellm/proxy/container_endpoints/ownership.py	14.81%	23 Missing ⚠️
litellm/proxy/db/exception_handler.py	75.86%	7 Missing ⚠️
litellm/proxy/utils.py	0.00%	7 Missing ⚠️
litellm/llms/custom_httpx/container_handler.py	0.00%	2 Missing ⚠️
litellm/proxy/auth/user_api_key_auth.py	33.33%	2 Missing ⚠️
litellm/router.py	0.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

Structured backport cutting 1.86.6 from the stable/1.86.x line. Source hunks are byte-identical to their staging counterparts; adaptations are confined to tests, build artifacts, and dependency manifests.

• Auth/DB hardening: auth errors during DB outages now surface 503 instead of 401 (auth_exception_handler.py, exception_handler.py); the deprecated-key recursive lookup in get_data now returns the result directly instead of overwriting response and crashing on dict subscript of a LiteLLM_VerificationTokenView.
• Azure containers routing: native upstream container IDs (e.g. cntr_<hex>) now recover model_id from the ownership row captured at create time; _normalize_api_base/_extract_api_version strip endpoint-path suffixes and prefer the api-version embedded in api_base; all httpx container calls use params or None so an empty dict no longer strips ?api-version from the URL.
• New general_settings knobs: disable_budget_reservation, database_connect_timeout, database_socket_timeout, database_disable_prepared_statements, database_extra_connection_params exposed in _types.py and wired through proxy_cli.py; Prisma cached-plan recovery now reconnects the client instead of mutating the SQL.

Confidence Score: 4/5

The auth path changes are conservative additions (503 vs 401 on DB outage, direct return of the deprecated-key lookup result) and all have corresponding regression tests. The container routing refactor and httpx param fixes are well-isolated. The disable_budget_reservation flag is gated, clearly documented, and emits a per-request warning.

Each cherry-pick is byte-identical to its staging version, tests cover the targeted regressions comprehensively, and no backwards-incompatible changes were introduced without a flag. Two minor edge cases (broad OSError classification in the DB exception classifier, and silent query-param loss in _normalize_api_base) are unlikely to bite in practice but worth tracking.

litellm/proxy/db/exception_handler.py (broad OSError catch) and litellm/llms/azure/containers/transformation.py (_normalize_api_base dropping non-api-version query params).

Important Files Changed

Filename	Overview
litellm/proxy/utils.py	Fixes deprecated-key lookup returning a LiteLLM_VerificationTokenView (not a dict) then crashing on dict subscript; fix stores the recursive result in deprecated_response and returns it directly, skipping the dict-normalisation block.
litellm/proxy/db/exception_handler.py	Adds is_database_service_unavailable_error and is_prisma_engine_internal_error to distinguish infrastructure outages (→503) from genuine auth failures (→401); includes traceback-walk heuristic for the prisma engine AttributeError masquerade.
litellm/proxy/auth/auth_exception_handler.py	Raises 503 instead of 401 when the DB is unreachable during auth; PrismaDBExceptionHandler is already imported, control-flow ordering is correct.
litellm/proxy/auth/user_api_key_auth.py	Threads general_settings into _reserve_budget_after_common_checks; disable_budget_reservation flag skips optimistic reservation with a warning log on every affected request.
litellm/proxy/proxy_cli.py	Extracts _build_db_connection_url_params to centralise Prisma URL-param construction; adds connect_timeout, socket_timeout, disable_prepared_statements, and extra_params passthrough.
litellm/proxy/container_endpoints/ownership.py	Makes get_container_forwarding_params async; adds _CONTAINER_STORED_ID_CACHE and _get_stored_container_id to recover model_id for native upstream Azure container IDs from the ownership row captured at create time.
litellm/router.py	In _init_containers_api_endpoints, reads a forwarded model_id from kwargs as a fallback when the container_id carries no LiteLLM routing payload; security boundary test confirms the fallback is gated on container_id presence.
litellm/llms/azure/containers/transformation.py	Adds _normalize_api_base to strip endpoint-path suffixes and _extract_api_version to prefer the api-version embedded in api_base over the deployment's older chat api_version when building the containers URL.
litellm/llms/custom_httpx/container_handler.py	Replaces params=query_params with params=query_params or None throughout all httpx calls; fixes 404s caused by httpx stripping the ?api-version query-string when an empty dict is passed.
litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py	Hardens streaming logging: skips [DONE] sentinels and non-JSON SSE frames; adds _resolve_costing_model and _extract_model_from_anthropic_chunks to recover the model name when body model is "unknown".
pyproject.toml	Bumps version 1.86.5→1.86.6, pypdf 6.10.2→6.13.1, adds tornado>=6.5.6 and aiohttp>=3.13.5,<3.14 constraint-dependencies.

_{Reviews (1): Last reviewed commit: "chore: refresh uv.lock for 1.86.6" | Re-trigger Greptile}

[greptile-apps]

Broad OSError catch may misclassify non-DB errors as infrastructure outages

OSError is the base class for FileNotFoundError, PermissionError, IsADirectoryError, and many others. If a certificate file is missing, a TLS library raises ssl.SSLError (an OSError subclass), or a config file can't be opened during the auth flow, this check would return True and the caller would surface a 503 ("database temporarily unreachable") rather than letting the error propagate naturally. In practice the auth DB path is unlikely to trigger those, but the comment "OSError covers ConnectionError…" could give future contributors confidence to add more code near this block. Consider tightening to the concrete connection subclasses: ConnectionError, TimeoutError, socket.timeout, ssl.SSLError.

[greptile-apps]

_normalize_api_base silently drops non-api-version query params

When the path suffix is stripped, the reconstructed URL uses empty strings for the query and fragment components of urlunparse, discarding the entire original query string. Any param other than api-version in api_base — for example a subscription-key or custom auth query param some Azure deployments add — would be silently lost. _extract_api_version saves the version before normalisation, but nothing preserves the rest. If those extra params are ever needed downstream, they'll need to be forwarded explicitly.

[greptile-apps]

Greptile Summary

Backports 10 cherry-picked patches from staging onto stable/1.86.x and cuts 1.86.6. Every source hunk is byte-identical to its staging commit; adaptations are confined to test drops (where staging-only symbols are absent on this line), build artifacts, and dependency manifests.

• Auth correctness (fix(proxy): return 5xx on DB infra errors during auth; reserve 401 for genuine auth failures #29986): DB infrastructure failures (socket errors, prisma engine crash on malformed payload, cached-plan escapes) now surface as 503 instead of 401, preventing valid keys from being rejected during outages; the deprecated-key lookup in get_data was crashing when the result Pydantic model was subscripted as a dict — now it returns immediately (fix(proxy): grace-period key rotation 401s; return deprecated-key lookup result directly #30327).
• Azure container routing (fix(router): use forwarded model_id for native Azure container IDs #27921): get_container_forwarding_params is now async and recovers the deployment model_id for native upstream (hex) container IDs from the stored unified_object_id; AzureContainerConfig strips endpoint-path suffixes from api_base and prefers the api-version embedded in the URL; params={} → params=None prevents httpx from silently stripping the ?api-version=… query string.
• New opt-in DB knobs (fix(proxy): expose Prisma idle/connect timeout + extra DB URL params #28395, feat(proxy): add option to disable server-side prepared statements for DB lookups #29984) and budget-reservation escape hatch (feat(proxy): add disable_budget_reservation general setting (#27639) #29493): connect_timeout, socket_timeout, disable_prepared_statements, and extra_connection_params are wired from general_settings into the Prisma DATABASE_URL; disable_budget_reservation lets operators trade hard per-request enforcement for relief from phantom BudgetExceededError under high concurrency.

Confidence Score: 4/5

Safe to merge; all source changes are bug fixes with targeted test coverage and no backwards-incompatible defaults.

The backport is cleanly structured: every changed hunk has a corresponding test (or a documented reason why the staging test was dropped), new settings default to off/None so existing deployments are unaffected, and the two most impactful fixes (deprecated-key crash and 503-vs-401 during DB outages) are well-covered by the new mock test suites. The one minor style nit in router.py keeps the score off the top.

litellm/proxy/container_endpoints/ownership.py — the new _get_stored_container_id path makes a direct litellm_managedobjecttable.find_first call on cache-miss; worth confirming the table has an index on (model_object_id, file_purpose) in production.

Important Files Changed

Filename	Overview
litellm/proxy/utils.py	Fixes two bugs: (1) deprecated-key lookup now assigns to deprecated_response and returns immediately, avoiding a crash when the Pydantic LiteLLM_VerificationTokenView was subscripted as a dict; (2) the cached-plan fallback no longer mutates the SQL and instead triggers attempt_db_reconnect before retrying.
litellm/proxy/db/exception_handler.py	Adds is_prisma_engine_internal_error and is_database_service_unavailable_error to classify DB infrastructure failures so auth returns 503 instead of 401 during genuine DB outages.
litellm/proxy/auth/auth_exception_handler.py	Adds a 503 branch before the default 401 raise, using is_database_service_unavailable_error; control flow ordering is correct (placed after ProxyException re-raise).
litellm/proxy/auth/user_api_key_auth.py	Threads general_settings into _reserve_budget_after_common_checks; adds early-return guard when disable_budget_reservation is True with a per-request WARNING log.
litellm/proxy/proxy_cli.py	Extracts DB URL param construction into _build_db_connection_url_params; wires four new general_settings fields into Prisma DATABASE_URL/DIRECT_URL; handles string-to-bool coercion for YAML-sourced values.
litellm/proxy/container_endpoints/ownership.py	Makes get_container_forwarding_params async; adds _CONTAINER_STORED_ID_CACHE and _get_stored_container_id to recover the deployment model_id for native upstream Azure container IDs; includes negative sentinel and 60 s TTL cache to avoid per-request DB hits.
litellm/llms/azure/containers/transformation.py	Adds _normalize_api_base and _extract_api_version; fixes container URL construction for models whose api_base was stored as the responses endpoint URL.
litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py	Adds [DONE] sentinel skipping, json.JSONDecodeError tolerance, model extraction from streaming chunks, and _resolve_costing_model fallback for Anthropic-compatible providers.
litellm/router.py	Falls back to the proxy-forwarded model_id kwarg when container ID decoding yields no model_id; the fallback condition calls .strip() twice.
pyproject.toml	Version bump 1.86.5 → 1.86.6; pypdf pinned at 6.13.1; tornado and aiohttp constraint-dependencies added.

_{Reviews (2): Last reviewed commit: "chore: refresh uv.lock for 1.86.6" | Re-trigger Greptile}

[greptile-apps]

.strip() is called twice on _forwarded_model_id — once in the truthiness guard and again to produce the value. Since .strip() is idempotent the double-call is harmless, but computing it once and reusing the result makes the intent clearer.

Suggested change

	model_id = decoded.get("model_id") or (
	_forwarded_model_id.strip()
	if isinstance(_forwarded_model_id, str) and _forwarded_model_id.strip()
	else None
	)
	_stripped_model_id = (
	_forwarded_model_id.strip()
	if isinstance(_forwarded_model_id, str)
	else ""
	)
	model_id = decoded.get("model_id") or (_stripped_model_id or None)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!