v2026.08.14 — deployed fork commit (57d8b6664733)

Fork release marking the LiteLLM commit currently deployed to production.

Commit	57d8b6664733 — head of litellm_internal_staging
Deployed image	us-central1-docker.pkg.dev/rayward-prod/litellm/litellm:57d8b6664733
Cloud Run service	rayward-litellm-gateway (us-central1, rayward-prod)
Paired with	rayward-internal/llm-gateway-infra@v2026.08.14 · rayward-external/bifrost@v2026.08.14

3,288 commits since v2026.07.05, overwhelmingly upstream BerriAI/litellm — merged here through 25 upstream syncs between 2026-07-05 and 2026-08-14, the last being #199.

What is ours

This fork is deliberately close to vanilla; the customization rule keeps changes at the Cloudflare edge or in gateway config wherever possible. Fork-owned work in this window was mostly keeping our recorded patches alive across upstream syncs:

• Fork patches restored after a -X theirs upstream sync merge dropped them — the recurring failure mode this fork guards against, and the reason a patch manifest plus sync CI exists at all.
• success_handler.py OpenAI-compatible-provider dispatch restored.
• Pricing-claimed rows excluded from the stale batch-cleanup sweep.
• Budget ceilings raised for legitimate upstream growth carried in a sync.

The pass-through cost-tracking patches (the owner-approved exception to keep-forks-vanilla, F1–F7/G1/G2) remain in force. Provider-scoped pass-through cannot attribute spend to a virtual key without them, and a capability whose patch goes missing is withdrawn rather than served uncosted — do not "restore vanilla" by reverting them.

Note for whoever reads this next

The tag marks a deployed commit, not a release of the fork as software. Rebuilding it is litellm-build-push.yml in the infra repo with source_sha=57d8b6664733.

v2026.07.05

Coordinated fleet release — 2026-07-05. Deployed from litellm_internal_staging @ e831ec3.

Changes since v2026.06.28 (6 commits)

• revert: back out canonical-model-names fork patches (#89, #90, #91) — keep gateway model names aligned with public IDs
• feat: derive geo bucket from edge headers (#88)
• fix: honor Anthropic cache TTL on the LiteLLM chat path

v2026.06.28 — Anthropic prompt cache control + upstream sync + CI pins

Changes since v2026.06.04.

Highlights

• Added Anthropic prompt cache control passthrough coverage.
• Updated workflow action pins across the fork.
• Continued upstream syncs through 2026-06-27 on litellm_internal_staging.
• Refreshed security/dependency handling for OSV, Dependabot, pypdf, vcrpy, faraday, vitest, and form-data.
• Tightened CodeQL, alert-bridge, daily branch, and sync automation behavior.

Notable PRs

• #86 feat: inject anthropic prompt cache control
• #85 ci: update workflow action pins
• #84 chore: sync upstream 2026-06-27
• #83 fix: sync upstream security and quality fixes
• #82 fix: normalize Anthropic legacy thinking in chat
• #77 fix(osv): refresh osv-scanner.toml and ignore aging false positives
• #70/#68 CI security advisory and OSV noise reduction

Compare: v2026.06.04...v2026.06.28

v2026.06.04 — Upstream sync (→2026-06-02) + sync-autopilot & daily-branch CI

What's Changed

• chore: sync upstream 2026-05-29 by @shudonglin in #25
• chore: sync upstream 2026-06-01 by @shudonglin in #27
• chore: sync upstream 2026-06-02 by @shudonglin in #29
• ci: sync PR autopilot + run daily-branch workflows on the fork by @shudonglin in #30
• ci: unblock sync-autopilot and daily-branch workflow permissions by @shudonglin in #31

Full Changelog: v2026.05.30...v2026.06.04

v2026.05.30 — Geo route header support + upstream sync + security hardening

Rayward fork baseline release before Cloudflare gateway routing unification work.

Highlights:

• Reads x-geo-route with legacy fallback and emits x-geo-route-resolved for route observability.
• Added fork-patches.txt and refreshed container/dependency security hardening.
• Synced upstream BerriAI/litellm litellm_internal_staging through the 2026-05-28 update window.

Target commit: 238f188
Branch: litellm_internal_staging