fix(router): clamp least_busy request counter at zero (salvage of #25393, credit @mango766)

[Bartok9]

Salvage

Rebases and lands the core fix from #25393 by @mango766 (also adjacent to #25325 by @rudra717). Originals went quiet / CONFLICTING against current tips.

Problem

least_busy increments the per-deployment counter pre-call and decrements on success/failure. Under races the count can go negative, after which that deployment always wins the min and starves the rest.

Fix

Clamp every decrement with max(request_count_value - 1, 0) on sync/async success and failure handlers.

Tests

tests/test_litellm/router_strategy/test_least_busy_clamp.py

Credit

Original approach and issue framing: @mango766 (#25393). Related clamp lwork: @rudra717 (#25325).

AI-assisted rebuild; human-reviewed. Does not include the tie-break jitter portion of #25393 — clamp only for a clean micro-merge.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/router_strategy/least_busy.py	0.00%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[CLAassistant]

All committers have signed the CLA.

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing Bartok9:salvage/least-busy-clamp-credit-mango766 (debbca5) with main (6a919ae)¹}

Footnotes

1. No successful run was found on litellm_oss_daily_2026_07_20 (5d4c4d0) during the generation of this report, so main (6a919ae) was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[Bartok9]

Sera doctor_ci: triage of 3 red checks:

• osv-scan — pre-existing upstream lockfile vulns (per culture); not ours, not chasing.
• lint / LiteLLM Linting — Can't find action.yml...cache-prisma-binaries. Upstream workflow references .github/actions/cache-prisma-binaries which checkout did not resolve (upstream CI infra bug, not our diff).
• misc / Run tests — 2 failures in tests/test_litellm/interactions/test_openapi_compliance.py: test_status_enum_values (daily branch OpenAPI spec now includes queued but test expects old set) and test_content_schema_uses_discriminator (daily spec now has discriminator). Both are daily-branch drift, addressed by fix(interactions): add queued to Interaction status enum (OSS daily) #34318 when it lands. Our PR touches least_busy router only — zero overlap with these tests.

CI last ran 2026-08-10 on stale daily tip. No code change needed here; will re-evaluate after #34318 lands. Agent-Owner: sera

[Bartok9]

doctor_ci 2026-08-20: triage of 2 red checks on this draft.

• misc / Run tests — 3 failures in test_openapi_compliance.py (discriminator, Turn, queued). These are base-drift: upstream main already has the fixes (fix(interactions): add queued to the Interaction status enum #34135 + fix(tests): assert Content variants are identified by type, not by the discriminator keyword #35161), but the daily branch litellm_oss_daily_2026_07_20 this PR targets has not yet picked them up. Not caused by this PR's changes.
• osv-scan — pre-existing upstream lockfile vulns (gitpython/pyasn1/brace-expansion/js-yaml/sharp). Per campaign culture, not contributor-fixable.

This PR's own code and test_least_busy_clamp.py are unaffected. No code change needed from our side.

[Bartok9]

Doctor watch (2026-08-21): misc / Run tests failed, but our change is isolated to least_busy.py (4-line clamp) + a trivial max() test that does not import litellm. The base branch litellm_oss_daily_2026_07_20 is the latest daily but is stale; this looks like flake / pre-existing base-drift on the daily branch, not a real conflict in our diff. osv-scan is pre-existing upstream lockfile vuln. No code change needed — will re-check on next watch.