Skip to content

CHANGELOG.md

Latest commit

 

History

History
128 lines (112 loc) · 6.57 KB

CHANGELOG.md

File metadata and controls

128 lines (112 loc) · 6.57 KB

Changelog

All notable changes to amp-proxy are documented in this file. The format is loosely based on Keep a Changelog and this project adheres to Semantic Versioning.

Unreleased

0.1.1 - 2026-04-16

Added

  • amp-proxy init subcommand — interactive wizard that prompts for the custom provider URL and Bearer token, generates a random local API key, and writes a complete ready-to-run config.yaml. The full 9-entry Amp CLI model mapping table, force-model-mappings: true, gemini-route-mode: "translate", and localhost-only management are baked into the generated file so operators only have to answer two questions to get a working proxy. Pass -force to overwrite an existing config and -config <path> to write to a non-default path.
  • README bilingual quickstart update — both READMEs now lead with amp-proxy init as the recommended first-run flow, with the manual config.example.yaml path kept below for operators who prefer to hand-edit YAML.

0.1.0 - 2026-04-16

First tagged release. Binaries for macOS (Intel + Apple Silicon), Linux (x86_64 + arm64) and Windows (x86_64 + arm64) are attached to the GitHub release; grab the archive matching your platform and see the README for setup.

Added

  • customproxy package — model-keyed routing to any OpenAI-compatible endpoint. A request whose (possibly remapped) model field is claimed by a custom-providers entry in config.yaml is forwarded to that upstream with the provider's Bearer token.
  • Anthropic /v1/messages stream upgrade — non-streaming Amp CLI requests (used by the librarian subagent) are silently upgraded to SSE on the wire and collapsed back into a single JSON body downstream. Works around an upstream content-loss bug where the non-streaming response comes back with an empty content array despite non-zero usage.output_tokens.
  • Gemini ↔ OpenAI Responses translator — new gemini-route-mode: "translate" config switch rewrites Amp CLI's finder subagent's Google v1beta1 generateContent requests into OpenAI Responses API calls, forwards them through the custom provider, and translates the reply back into Gemini JSON before the client reads it. Covers both single-turn and multi-turn conversations, drops upstream-opaque thoughtSignature bytes, synthesises call_ids for functionCall ↔ functionResponse alignment, and lowercases Gemini's uppercase JSON-schema type values. Ships with unit and integration tests.
  • Model mappings — rewrite the model field pre-routing so that, for example, claude-opus-4-6 can be redirected to gpt-5.4(high) on a local endpoint. force-model-mappings: true makes mappings win over local OAuth providers when both match.
  • Ready-to-use Amp CLI routing defaults — config.example.yaml ships pre-filled with the full 9-entry mapping table covering the claude / gpt / gemini model families Amp CLI actually requests. Operators only need to point custom-providers at their gateway URL.
  • scripts/restart.ps1 — Windows helper that kills any stale amp-proxy.exe, relaunches the server, and redirects stdout/stderr to run.log.
  • scripts/test_gemini_translate.js — Node.js end-to-end smoke test that drives the Gemini translate path against a running amp-proxy instance and verifies the response shape.
  • Debug body capture middleware — opt-in via debug.capture-path-substring in config. Writes raw request/response bytes for matching URL paths to ./capture/*.log. Intended for local development only.
  • Bilingual README — Chinese is the primary README.md; English lives alongside at README_en.md, with language-toggle links at the top of both files.
  • GitHub Actions CI — runs go build, go vet, and go test ./internal/customproxy/... on every push and PR to main.
  • GoReleaser pipeline — tagged commits trigger a release workflow that produces cross-platform archives, checksums, and a GitHub release with auto-grouped changelog notes.

Fixed

  • Empty output array warning on /v1/responses — sseRewriter now injects the accumulated response.output_item.done items into response.completed when the upstream sends an empty array, keeping Amp CLI's Stainless SDK from silently discarding the reply.
  • Windows NTFS filename collision in bodyCapture — URL paths containing : (Google v1beta generateContent calls) previously wrote zero-byte files because : is an alternate-data-stream separator on NTFS. Replaced with _ so capture files stay portable.
  • restart.ps1 stderr loss — stdout and stderr are now both redirected into run.log; the run.log.err sibling file is kept for legacy callers but no longer receives content by default.
  • Amp CLI ContentLength mismatch after rewriteModelInRequest — custom provider forwarding now realigns both the c.Request.ContentLength field and the Content-Length header with the rewritten body length, preventing the net/http: ContentLength=X with Body length Y panic on the next hop.
  • pre-existing logAmpRouting noise — the RouteTypeAmpCredits path is now logged at Errorf with an explicit message telling the operator the model leaked past the routing table and is billable.

Changed

  • sdk/access/ moved to internal/access/ — amp-proxy has no external SDK consumers, so the sdk/ root directory was flattened into internal/. Import alias sdkaccess is preserved in the three call sites so future cherry-picks from upstream stay diff-friendly.
  • internal/amp/fallback_handlers.go — the custom-provider routing hook now skips Google v1beta native paths unless gemini-route-mode: translate is on; otherwise it falls through to the ampcode.com proxy to preserve protocol fidelity.

Notes

  • amp-proxy is a derivative of CLIProxyAPI. Every file that diverges from the upstream baseline is listed in NOTICE.md with a short rationale.
  • The extraction deliberately excludes upstream's OAuth login flows for ChatGPT / Claude Code / Gemini CLI. For OAuth-backed providers, run a separate bridge that terminates the OAuth flow and exposes a plain Bearer endpoint, then point a custom-providers entry at it.