2.0.0-next.1 — unified embeddings interface (opencode V2)

Embedding access is now unified on a single OpenAI-compatible interface.

Migration (breaking — no backward-compatible aliases)

Old	New
embeddingBackend: "python"	embeddingBackend: "builtin"
embeddingBackend: "remote"	embeddingBackend: "endpoint"
remoteApiUrl	embeddingBaseUrl
remoteApiKey	embeddingApiKey

• Default model: nomic-embed-text-v1 → embeddinggemma-2 (multilingual). The index records its
  embedding-space identity and auto-rebuilds on the next session — both are 768-dim, so the change
  is detected by identity, not dimension.
• builtin is self-contained: the bundled python/embed_server.py serves OpenAI-compatible
  POST /v1/embeddings and auto-downloads the model into ~/.opencode-md-memory/models/<model>/onnx
  on first use (MDM_EMBED_MODEL_DIR to override, MDM_EMBED_AUTO_FETCH=0 to disable).
• endpoint: point at any OpenAI-compatible embeddings service — llama.cpp / Ollama / TEI /
  LM Studio / OpenAI — through the same interface.

Changes

• unify embedding backends on one OpenAI-compatible client (builtin / endpoint)
• python/embed_server.py: serves /v1/embeddings; atomic .part download; /health reports the model
• index rebuilds on any model change (including a dimension-preserving swap)
• universal loader: accepts any ONNX text model (sentence_embedding or mean-pool; token_type_ids
  and image/video/audio feature inputs handled when present)

Compatibility

• opencode V2 (@opencode/plugin >=2.0.0 <3) — this line. Pre-release.
• For opencode V1, install the latest tag (1.3.0).

1.3.0 — unified embeddings interface (opencode V1)

Embedding access is now unified on a single OpenAI-compatible interface.

Migration (breaking — no backward-compatible aliases)

Old	New
embeddingBackend: "python"	embeddingBackend: "builtin"
embeddingBackend: "remote"	embeddingBackend: "endpoint"
remoteApiUrl	embeddingBaseUrl
remoteApiKey	embeddingApiKey

• Default model: nomic-embed-text-v1 → embeddinggemma-2 (multilingual). The index records its
  embedding-space identity and auto-rebuilds on the next session — both are 768-dim, so the change
  is detected by identity, not dimension.
• builtin is self-contained: the bundled python/embed_server.py serves OpenAI-compatible
  POST /v1/embeddings and auto-downloads the model into ~/.opencode-md-memory/models/<model>/onnx
  on first use (MDM_EMBED_MODEL_DIR to override, MDM_EMBED_AUTO_FETCH=0 to disable).
• endpoint: point at any OpenAI-compatible embeddings service — llama.cpp / Ollama / TEI /
  LM Studio / OpenAI — through the same interface.

Changes

• unify embedding backends on one OpenAI-compatible client (builtin / endpoint)
• python/embed_server.py: serves /v1/embeddings; atomic .part download; /health reports the model
• index rebuilds on any model change (including a dimension-preserving swap)
• universal loader: accepts any ONNX text model (sentence_embedding or mean-pool; token_type_ids
  and image/video/audio feature inputs handled when present)

Compatibility

• opencode V1 (@opencode-ai/plugin >=1.4.0) — this line.
• For opencode V2, install opencode-md-memory@next (2.0.0-next.1).

v1.2.0

v1.2.0

Highlights

• 语义标题取自正文 H1：向量的语义标题不再绑定文件名 slug，而是取正文首个 # H1 标题（无 H1 时回退文件名）。内容换题只需改 H1，下次注入自动重嵌；纯正文编辑不改 H1 不触发重嵌，保持编辑零成本。

Bug fix

• 嵌入服务自愈：修复嵌入服务死亡（idle-timeout 自卸载 / 进程重启孤儿断管）后整场会话每条消息超时失败、注入完全失效的问题。
  • 请求失败 → 丢弃缓存 → 重建后端（python respawn / remote 重建客户端）→ 重试一次
  • loopback 地址（127.0.0.1）加入 NO_PROXY，死端口即刻失败（~8ms）而非经代理等待 ~3s
  • 正常日志（model loaded / [embed] 计时）改走 info 级，不再被 opencode TUI 当错误渲染
  • embed server stdout 打印容错 BrokenPipeError，孤儿断管服务不再导致 /embed 500

Chore

• 索引 title 提取改为 stat 键控缓存，未变更文件零读取（减少每次注入的全量读盘）

Notes

• 配置项无变更，可直接升级（opencode plugin opencode-md-memory -f -g；若未生效需 rm -rf ~/.cache/opencode/packages/opencode-md-memory* 后重装）

v1.1.0

Highlights

• Binary vector index — vectors now live in a flat vectors.bin (Float32Array) alongside a slim index.json (id + title only). Smaller index, faster loads.
• L1 index cache — parsed index is cached in-memory and invalidated by mtime, avoiding re-parsing on every injection.
• Similarity score in references — injected <memory_reference> entries now show their cosine score.

Config unification

semanticModel and remoteModel are removed. Use a single option for both backends:

{
  "plugin": [
    [
      "opencode-md-memory",
      {
        "embeddingBackend": "remote",
        "embeddingModel": "text-embedding-3-small"
      }
    ]
  ]
}

• embeddingModel is passed to the OpenAI-compatible /embeddings API (remote backend).
• For the python backend it resolves the model directory at ~/.opencode-md-memory/models/{model}/onnx/, and is forwarded to the embed server via the new --model-name flag.

Bug fix

• Dimension-mismatch safety — switching embedding models (a different output dimension) now rebuilds the index from scratch instead of mixing old/new vectors, which previously could corrupt vectors.bin or surface NaN scores. A defensive check in writeIndex fails loudly on any mixed-dimension write.

Breaking change

• semanticModel / remoteModel config keys removed → use embeddingModel.

Full Changelog: https://github.com/xykbear/opencode-md-memory/commits/v1.1.0