Skip to content

CHANGELOG.md

Latest commit

 

History

History
123 lines (104 loc) · 5.84 KB

CHANGELOG.md

File metadata and controls

123 lines (104 loc) · 5.84 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.2.1] - 2026-09-22

Fixed

  • Thinking works with any model that exposes plaintext reasoning, not just one delta shape: recognized delta fields are now reasoning, reasoning_text, and thinking, and thinking streamed via message.part.updated snapshot deltas is forwarded too (with a dedupe guard so backends sending both shapes never emit it twice). Verified live against mimo-v2.6-flash-free (non-streaming reasoning_content, opt-in streaming deltas, clean content by default).
  • Backend per-model failures that arrive as HTTP 200 with info.error (e.g. 402 Insufficient account funds, found live on non-free models) no longer return an empty 200: non-streaming endpoints answer 500 backend_error with the upstream message, and streams end with an error frame instead of a clean-but-empty [DONE].
  • Long agentic runs no longer look hung: streams now send an SSE heartbeat comment every STREAM_HEARTBEAT_MS (default 15000, 0 disables), so client/proxy idle timeouts don't kill runs while the backend works tools with no text deltas. Verified live (heartbeat interleaved mid-run, clean [DONE]).
  • Streaming master switch: ENABLE_STREAMING=0 (or --no-streaming; --streaming re-enables) serves stream: true requests blocking as regular JSON instead of SSE, so strict clients keep working. Forcing streaming onto non-streaming requests is deliberately not offered.
  • Ops config gaps closed: HOST bind address (--host, default keeps 0.0.0.0 for Docker; use 127.0.0.1 for localhost-only no-auth use), CORS_ORIGIN (--cors, off by default; preflights bypass auth), VISION_MODEL (--vision_model, was hardcoded), INCLUDE_REASONING (--include_reasoning, default-on thinking for clients that can't send the flag), LOG_LEVEL (--log_level, error|warn|info|debug, request lines without bodies/keys), and a --version/-v flag.
  • Help for humans and agents: bare help subcommand (node server.js help) plus richer --help with endpoints, request-field hints, copy-paste examples, and exit codes.

Notes

  • Some models (observed: muse-spark-1.3-contributor-free) keep thinking encrypted server-side (reasoningEncryptedContent, empty text, zero deltas). For those only usage.completion_tokens_details.reasoning_tokens comes through — there is no plaintext for the adapter to forward.

[1.2.0] - 2026-09-22

Added

  • Thinking output: backend reasoning parts are exposed OpenAI-style as reasoning_content (DeepSeek/OpenRouter/vLLM convention)
    • Streaming chat completions forward thinking as delta.reasoning_content chunks when the client opts in (include_reasoning: true; also accepts include_thinking, stream_options.include_reasoning, reasoning, reasoning_effort, thinking / enable_thinking). Default still drops thinking so content stays clean.
    • Reasoning parts are tracked via message.part.updated snapshots, so thinking never leaks into content even when the backend streams it with field: "text" on a reasoning part.
    • Non-streaming chat completions include message.reasoning_content whenever the backend produced thinking (unless explicitly disabled, e.g. reasoning: { effort: "none" }).
    • usage.completion_tokens_details.reasoning_tokens when the backend reports reasoning tokens.
  • Thinking request aliases are accepted without error but never forwarded to the backend: opencode tunes effort via model variants (opencode.json), the adapter only gates reasoning output. Effort tuning via tools-style per-message params is not supported by POST /session/:id/message ({ messageID?, model?, agent?, noReply?, system?, tools?, parts }).

[1.1.1] - 2026-09-22

Fixed

  • Plain node server.js now reads .env (zero-dependency built-in loader), so PORT (and OPENCODE_URL, API_KEY, timeouts, retries) from .env are honored instead of silently falling back to port 80. Precedence: CLI flags > shell env > .env file > defaults.

[1.1.0] - 2026-09-21

Added

  • True token streaming: chat and text completions forward backend message.part.delta events live via prompt_async + the /event bus (replaces buffered char-by-char replay); stream_options.include_usage adds a usage block to the final chunk; client disconnect aborts the backend run
  • Conversation memory: opt-in session_id request field pins turns to one backend session (no history replay); responses echo it back; unknown ids return 404 session_not_found
  • Resilience: retries with backoff on network failures and 5xx GETs (OPENCODE_RETRIES), configurable OPENCODE_TIMEOUT / OPENCODE_MESSAGE_TIMEOUT
  • OpenAI-shaped errors everywhere on /v1/* (401 invalid_api_key, 400 invalid_json, 404 session_not_found, 502 backend_unreachable)
  • npm binary: global install exposes the opencode-api-node command

[1.0.0] - 2026-09-20

Added

  • OpenAI-compatible proxy for opencode serve (server.js, Express, zero runtime deps besides express)
  • Endpoints: GET /, GET /health, GET /v1/models, POST /v1/chat/completions, POST /v1/completions
  • SSE character-by-character streaming for chat and text completions
  • Multi-turn context via noReply history replay, system prompt mapping, provider/model splitting, vision-model forcing with image_url inlining
  • Optional bearer auth (API_KEY), backend selection (OPENCODE_URL), port selection (PORT / --port)
  • --session [<id>] CLI action to dump backend session data as JSON
  • Docker image + compose setup with healthcheck
  • Test suite with mock backend (npm test)
  • MIT license