@josxa/opencode-recall

@josxa/opencode-recall

Hybrid semantic + lexical recall over your local OpenCode session history, with cursor‑paginated ChatML transcript reads.

OpenCode keeps every session in a local SQLite database. This plugin makes that history first‑class context for the agent: find a relevant historical anchor with one ranked search, then page a bounded window of messages around it as a ChatML transcript — without loading an entire session into the model.

┌───────────────────┐    history_search    ┌──────────────────────┐
│  OpenCode agent   │ ───────────────────► │  Ranked anchors      │
│ (your assistant)  │                      │  (opaque cursors)    │
└───────────────────┘ ◄─────────────┐      └──────────┬───────────┘
         │      history_read        │                 │
         ▼                          │ ChatML window   │
┌───────────────────┐   pagination  │                 │
│  ChatML window    │ ◄─────────────┘                 │
│  with <nav/>      │                                 │
└───────────────────┘                                 │
                                                      ▼
                                     ┌──────────────────────────────┐
                                     │  opencode.db   (read‑only)   │
                                     │  + sidecar embedding index   │
                                     └──────────────────────────────┘

Why

The default way to "remember" prior work is to dump entire sessions into context. That's expensive, noisy, and most of it is irrelevant. Recall flips this around:

• Search returns anchors, not content. Each hit is a small JSON row with a cursor.
• Reads are bounded windows. You pick around, next, prev, head, tail, or full relative to a cursor, with a hard message limit.
• The transcript renderer is source‑faithful. Tool calls, patches, file attachments, and text are normalized into ChatML with clear truncation markers, so the model sees structure instead of a wall of JSON.
• Ranking is hybrid. Lexical SQL search runs against the OpenCode DB; semantic search runs against a sidecar embedding index built from the same rows. Results are deduped, term‑filtered, session‑diversified, and a bounded semantic‑rescue lane catches high‑confidence paraphrases.

Install

opencode plugin @josxa/opencode-recall -gf

This installs the package and wires it into your global OpenCode configuration automatically.

Manual installation

bun add -d @josxa/opencode-recall
# or
npm install --save-dev @josxa/opencode-recall

Then register the package in your OpenCode config (opencode.json or ~/.config/opencode/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["@josxa/opencode-recall"]
}

That exposes two tools to the agent:

Tool	Purpose
history_search	Return ranked anchors (cursor + metadata + snippet) for a query.
history_read	Read a bounded ChatML window around a cursor.

Embeddings (Ollama)

Semantic search uses Ollama running locally. The default model is all-minilm — small, fast, and good enough for recall.

# install: https://ollama.com/download
ollama serve
ollama pull all-minilm

The first history_search call builds the sidecar index incrementally; subsequent calls reuse and sync it. If the elapsed sync time crosses a threshold, search output is prefixed with a <sync indexed_rows="…" seconds="…" /> notice so the agent (and you) know an index update happened.

Lexical search keeps working without Ollama — but you lose paraphrase recall, which is half the value.

Configuration

All configuration is environment‑driven so the plugin works inside MCP, the CLI, and CI without extra wiring.

Variable	Default	Description
OPENCODE_DB_PATH	$HOME/.local/share/opencode/opencode.db	OpenCode SQLite database (read‑only).
OPENCODE_RECALL_DB_PATH	$HOME/.local/share/opencode/opencode-recall-index.db	Sidecar embedding index path.
OPENCODE_RECALL_OLLAMA_URL	http://127.0.0.1:11434	Ollama base URL.
OPENCODE_RECALL_EMBED_MODEL	all-minilm	Embedding model name. Try mxbai-embed-large for higher quality.

Use bun run eval:embeddings to compare installed embedding models against the local regression cases in docs/real-history-regressions.md.

Tool reference

history_search

Arg	Type	Notes
q	string	Required. Free‑text query.
n	number	Max hits (default 8, max 25).
dir	string	Exact OpenCode session directory filter.
after	ISO date	Only messages at or after this timestamp.
before	ISO date	Only messages at or before this timestamp. Defaults to now − 30 s to exclude the live conversation.

Returns a JSON array of compact hits:

[
  {
    "cursor": "msg_…",     // opaque; pass to history_read
    "sid":    "ses_…",     // OpenCode session id
    "dir":    "/Users/you/projects/foo",
    "title":  "Figma MCP server on Azure API Center",
    "time":   "2026-04-12T08:21:44.000Z",
    "role":   "assistant",
    "score":  0.7421,
    "text":   "…snippet capped at 280 chars…"
  }
]

history_read

Arg	Type	Notes
cursor	string	Required. A msg_…, a ses_…, or an encoded cursor from history_search.
mode	string	around (default), next, prev, head, tail, or full.
n	number	Message count. Default 12, max 50. For full, default 200, max 500.

Returns a ChatML‑like transcript window:

<hist sid="ses_…" dir="/…" mode="around" range="42-53" anchor="48" total="120" title="…">
<|im_start|>user name="msg_…" index="42" time="2026-04-12T08:21:44.000Z"
…text and tool_call blocks…
<|im_end|>
…more messages…
<nav cur="…" prev="…" head="…" next="…" tail="…" full="…" />
</hist>

Tool calls, patches, and file attachments render as structured tags with explicit truncated / original_chars markers when content is capped. The <nav/> element gives the agent cursors to continue paging without re‑searching.

How it works

• Source of truth. opencode.db is opened read‑only. The plugin never writes to it.
• Sidecar index. A separate SQLite database (opencode-recall-index.db) stores text chunks, content hashes, and Float32 embedding blobs. Synthetic session-title:<id> rows are indexed so proper‑noun title queries beat noisy snippet matches.
• Sync. Each history_search call performs an incremental sync with a 30‑minute overlap window and a lock so concurrent agents don't fight. Stale rows are pruned by comparing part ids.
• Ranking. Lexical + semantic candidates are merged, scored with title/text/directory term ratios + phrase boosts, filtered to require enough query‑term overlap, and diversified to at most two hits per session. A small semantic‑rescue lane admits paraphrase matches that miss lexical filters but have high embedding similarity.
• Reads. Windows are computed by row_number() over (session_id, time_created, id), then parts are normalized into text | tool | patch | file and capped (tool input 2 000 chars, output 6 000 chars).

Development

bun install
bun run ai:check       # biome + tsgo type-check
bun test               # deterministic ranking + cursor tests
bun run eval:real-history  # regression suite against your local opencode.db
bun run build          # emits dist/

Code quality is enforced by biome and tsgo --noEmit. See AGENTS.md for the style guide.

License

MIT © JosXa