@homeflare/distilled-litellm

This package is not a HomeFlare SDK. It is an unmodified copy of @distilled.cloud/litellm, generated by the distilled pipeline from the LiteLLM Proxy's own app.openapi() document at tag v1.103.0, published here under its own name only because @distilled.cloud/litellm does not exist on npm yet — the kit builds it in a local clone of alchemy-run/distilled but, per house policy, never pushes to that upstream. See ../alchemy/docs/distilled-interim.md for the full route every distilled-sourced vendor package follows, and why this one exists at all.

⛔ Do not hand-edit anything under src/. The whole point of the route above is that this directory is copied, not written — a fix belongs in the distilled clone's packages/litellm/, regenerated, then copied back here.

⚠️ License. Apache-2.0, not the kit's usual MIT — this content is a redistribution of alchemy-run/distilled's own Apache-2.0-licensed output, and relicensing a copy would misstate what it is. See LICENSE.

What's in it

91 service modules, one per LiteLLM Proxy OpenAPI tag — model_management, key_management, team_management, internal_user_management, budget_management, misc (the untagged /config/pass_through_endpoint* family and other UI/debug/login/MCP routes), and every other admin/inference tag LiteLLM's own API reference groups by (964 operations total, measured from the generated models; v1.100.0 was 917 in 105 modules). Every operation gets an Effect-typed, catchTag-able error for every status the spec declares, plus the undeclared 400/403/404/409 responses this package's own source-reading found on the primary CRUD routes of the five admin tags above (patches/{budget,key,team,internal_user,model}_management/*.json — each patch cites the exact source file:line it was grounded in, at v1.100.0; re-measured against v1.103.0: every patched status is still raised, only the lines moved; see ../../litellm/docs/errors.md and ../../litellm/docs/patch-reverification.md in the distilled clone for the full account). POST /key/delete also types its 404 and 403 as resource-specific tags, KeyNotFound (message No keys found) and KeyDeleteForbidden (message not authorized to delete this key), each matched on the status and that phrase of LiteLLM's own message (patches/key_management/delete_key_fn_key_delete_post.json). The credential by-name read and delete type their 404 as CredentialNotFound (message Credential not found), and the credential PATCH carries credential_name in its body as a second member (credential_name_body, wire-named credential_name) because the vendor's UpdateCredentialItem requires it there and a Smithy member has one binding (patches/credential_management/{get_credential_by_name_credentials_by_name__credential_name__get,delete_credential_credentials__credential_name__delete,update_credential_credentials__credential_name__patch}.json). GenerateKeyResponse.key and NewUserResponse.key — the literal virtual-key value /key/generate, /key/regenerate and /user/new return — decode to Redacted.Redacted<string> rather than a plain string.

Changed at v1.103.0. LiteLLM moved the routes of fifteen per-provider * Pass-through tags (anthropic_pass_through, azure_pass_through, bedrock_pass_through, vertex_ai_pass_through, …) under one llm_passthrough tag (107 operations), so those fifteen modules are gone and Services.llmPassthrough is new; open_ai_pass_through shrank from 10 to 5 operations. The kit's own consumers use only budget_management and misc, which kept every operation they call.

Typed data-plane bodies. POST /rerank, /v1/rerank, /v2/rerank and /mcp-rest/tools/call read the raw request body, so LiteLLM's OpenAPI document declares no requestBody for them and the request used to be S.Struct({}). They now carry typed bodies (RerankRequest: model, query, documents required; CallMcpToolRestRequest: name required, server_id, arguments), added by distilled patches grounded in the v1.103.0 source — never by editing generated files. Responses stay body: unknown, and 72 other operations still have no declared body: see ../../litellm/docs/data-plane-bodies.md in the distilled clone. scripts/smoke.ts asserts the typed fields reach the wire as the JSON body the handler reads.

import * as Litellm from '@distilled.cloud/litellm'; // aliased onto this package — see docs/distilled-interim.md
import * as Effect from 'effect/Effect';

const program = Litellm.Services.keyManagement
  .generateKeyFnKeyGeneratePost({ key_alias: 'my-service' })
  .pipe(
    Effect.provide(Litellm.LitellmProtocol),
    Effect.provide(Litellm.credentials({ apiKey: '...', baseUrl: 'https://litellm.example.com' })),
  );

Kit code never imports @homeflare/distilled-litellm directly — always @distilled.cloud/litellm, aliased in the consuming package's package.json ("@distilled.cloud/litellm": "npm:@homeflare/distilled-litellm@<version>"), so the eventual cutover to the real published package is a one-line alias swap with no import changes anywhere in the kit.

Updating it

From the distilled clone's homeflare/litellm worktree, with the spec dumped the way codegen/manifest.json's litellm-openapi entry records (copy it to packages/litellm/specs/.local/specs/):

DISTILLED_SPECS_LOCAL=1 pnpm generate litellm      # convert + generate + format
pnpm exec tsc -b packages/litellm --noCheck false  # scoped; never the whole graph
pnpm specs:check

Then mirror src/ here byte for byte — rsync -a --delete, not cp -R: the generator never removes a module whose tag disappeared (v1.103.0 dropped fifteen), and scripts/prune-services.ts in the clone only cleans the clone. Check with diff -rq, add a changeset (a minor while the generated surface changes shape) and let the kit's normal release flow bump the version and publish it — never edit version or CHANGELOG.md by hand (docs/releasing.md).

License

Apache-2.0 (see LICENSE) — not the kit's usual MIT. src/ is an unmodified copy of @distilled.cloud/litellm's own output from alchemy-run/distilled, which is itself Apache-2.0; this package's own files (package.json, tsconfig.json, scripts/smoke.ts, this README) are offered under the same terms.