[Infra] Merging RC Branch with Main

[yuneng-jiang]

Relevant issues

Fully Passing CI/CD: https://app.circleci.com/pipelines/github/BerriAI/litellm?branch=litellm_rc_branch

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🚄 Infrastructure

Changes

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 23, 2026 5:16pm

[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.
2 out of 3 committers have signed the CLA.

✅ ishaan-jaff
✅ yuneng-jiang
❌ cursoragent
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_rc_branch (6140927) with main (3dccdde)}

[greptile-apps]

Greptile Summary

This PR merges the RC branch into main as a v1.82.3 release, bundling two bug fixes, a test stabilisation pass, CI improvements, and a large set of pre-built frontend assets.

Key changes:

• litellm/cost_calculator.py – Fixes a regression where per-request custom pricing passed via Router returned 0.0 cost. The router registers each deployment under a hash ID in model_cost (without pricing fields); the fix now checks whether that entry actually carries pricing data before using it over the raw model name.
• litellm/proxy/proxy_cli.py – Fixes a TypeError crash in the Prometheus multiproc-dir auto-setup when callbacks/success_callback/failure_callback are configured as a plain string instead of a list.
• tests/test_litellm/proxy/test_proxy_cli.py – The test_startup_fails_when_db_setup_fails test now requires --enforce_prisma_migration_check to trigger the exit-on-failure path, reflecting an intentional (but backwards-incompatible) default-behavior change: the proxy now continues startup when DB migration fails unless the flag is explicitly set.
• tests/mcp_tests/test_mcp_server.py – Adds a missing mock for set_auth_context so the streamable-HTTP handler test does not crash when a None auth context is returned by the already-mocked auth extractor.
• CI / infra – Adds bsdmainutils to the security scan dependency list, adds an Anthropic pass-through env flag to the CI job, and bumps the version to 1.82.3.

Confidence Score: 4/5

• PR is safe to merge with one notable concern: the DB-migration-failure default behavior was flipped from exit-on-failure to continue-on-failure, which is a backwards-incompatible default change.
• Both code bug fixes are well-reasoned and accompanied by regression tests. The flaky-test stabilisation and mock additions are legitimate. The main risk is the silent behavior change to DB migration failure handling — users who relied on the proxy exiting when migration fails will now see it continue without opting in to --enforce_prisma_migration_check.
• tests/test_litellm/proxy/test_proxy_cli.py — the updated test now requires --enforce_prisma_migration_check to assert exit-on-failure, masking the default-behavior regression for existing users.

Important Files Changed

Filename	Overview
litellm/cost_calculator.py	Bug fix: when router_model_id is present in model_cost but has no pricing fields, fall back to the model name. Minor: only checks input_cost_per_token / input_cost_per_second, skipping output_cost_per_token.
litellm/proxy/proxy_cli.py	Bug fix: normalizes string callbacks/success_callback/failure_callback values to lists before concatenation, preventing TypeError when Prometheus multiproc dir auto-setup runs.
tests/test_litellm/proxy/test_proxy_cli.py	Test updated to add --enforce_prisma_migration_check flag, which reflects a backwards-incompatible default behavior change: the proxy no longer exits on DB setup failure without the explicit flag.
tests/test_litellm/proxy/test_prometheus_cleanup.py	New parametrized test verifies that string-typed callback values (not lists) are handled without a TypeError — good regression coverage for the proxy_cli.py fix.
tests/test_litellm/test_cost_calculator.py	New regression test for per-request custom pricing with Router correctly exercises the cost_calculator fix and verifies the router_model_id is not used when it has no pricing data.
tests/local_testing/test_router_utils.py	Adds @pytest.mark.flaky(retries=3, delay=1) and increases sleep to 2 s for a timing-sensitive test — reasonable stabilisation for a mock-based async test.
litellm/proxy/example_config_yaml/custom_auth_basic.py	Adds user_role=LitellmUserRoles.PROXY_ADMIN to the example custom auth handler, improving the example's completeness.
.circleci/config.yml	Adds LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES=true env var to the pass-through CI job, enabling the Anthropic messages route test.
ci_cd/security_scans.sh	Adds bsdmainutils to the apt-get install list, which provides the column utility needed by the security scanning script on Debian/Ubuntu CI runners.
pyproject.toml	Version bump from 1.82.2 to 1.82.3 in both [tool.poetry] and [tool.commitizen] sections.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Router.acompletion / completion] --> B{custom_pricing=True?}
    B -- No --> C[Use completion_response model / base_model / model]
    B -- Yes --> D{router_model_id in model_cost?}
    D -- No --> E[return_model = model]
    D -- Yes --> F{entry has input_cost_per_token\nor input_cost_per_second?}
    F -- Yes --> G[return_model = router_model_id]
    F -- No --> E
    G --> H[cost_calculator uses router_model_id pricing]
    E --> I[cost_calculator uses model-name pricing]

    subgraph proxy_cli fix
        J[litellm_settings callbacks value] --> K{is string?}
        K -- Yes --> L[wrap in list]
        K -- No --> M[use as-is]
        L --> N[check 'prometheus' in all_callbacks]
        M --> N
    end

Loading

_{Last reviewed commit: 6140927}

[greptile-apps]

Backwards-incompatible default behavior change

The test was updated to add --enforce_prisma_migration_check to the invocation, reflecting a change in proxy_cli.py where the proxy now continues running when PrismaManager.setup_database returns False (unless the new flag is explicitly set). This is a backwards-incompatible change to the default behavior per rule b48b7341.

Previously, a failed DB setup caused a sys.exit(1) unconditionally. Now the proxy will silently continue with a warning unless --enforce_prisma_migration_check / ENFORCE_PRISMA_MIGRATION_CHECK=true is set. Users who relied on the proxy exiting to signal a broken DB migration (e.g., in health checks or orchestration scripts) will no longer get that signal by default.

Consider whether the old default (exit on failure) should be preserved, with an opt-out flag like --ignore_prisma_migration_failure instead of an opt-in flag.

Rule Used: What: avoid backwards-incompatible changes without... (source)

[greptile-apps]

Missing check for output_cost_per_token

The pricing-existence check only gates on input_cost_per_token or input_cost_per_second, but a router entry could theoretically have only output_cost_per_token set (e.g., a model where input is free but output is charged). In that scenario, the entry would be treated as "no pricing" and fall back to the raw model name.

Consider extending the guard to cover additional pricing fields for completeness:

Suggested change

	if custom_pricing is True:
	if router_model_id is not None and router_model_id in litellm.model_cost:
	return_model = router_model_id
	entry = litellm.model_cost[router_model_id]
	if (
	entry.get("input_cost_per_token") is not None
	or entry.get("input_cost_per_second") is not None
	):
	return_model = router_model_id
	else:
	return_model = model
	else:
	return_model = model
	if (
	entry.get("input_cost_per_token") is not None
	or entry.get("input_cost_per_second") is not None
	or entry.get("output_cost_per_token") is not None
	):
	return_model = router_model_id
	else:
	return_model = model