← Docs hub

OpenCode / OpenClaw adapter

Reads .jsonl session transcripts written by the OpenCode / OpenClaw agents — both use an identical schema.

Runs on bare llmwiki sync when OpenCode / OpenClaw app-config session dirs exist on disk.

Session store

The adapter auto-detects the store across platforms:

• Linux: ~/.config/opencode/sessions/ and ~/.config/openclaw/sessions/
• macOS: ~/Library/Application Support/opencode/sessions/ and ~/Library/Application Support/openclaw/sessions/
• Windows: %APPDATA%\opencode\sessions\ and %APPDATA%\openclaw\sessions\

Both nested (<project>/<session>.jsonl) and flat (<project>-<session>.jsonl) layouts are handled.

What it reads

Each session is a JSONL stream of {role, content} records:

{"role": "user",      "content": "start a new feature"}
{"role": "assistant", "content": "…"}
{"role": "tool",      "content": "…"}

normalize_records() translates that schema into the Claude-style {type, message: {role, content}} that the shared renderer expects:

Enable it

python3 -m llmwiki sync --adapter opencode

To turn it off after enabling via config:

{ "adapters": { "opencode": { "enabled": false } } }

Automated (headless) sessions

No verified automation-launch markers in the store today — sessions are treated as not headless under filters.exclude_headless. See Multi-agent setup.

Output layout

Standard raw/sessions/<YYYY-MM-DDTHH-MM>-<project>-<slug>.md.

Code

• llmwiki/adapters/contrib/opencode.py
• Tests: tests/test_opencode_adapter.py (23 cases)
• Issue history: #43 (initial)

See also

• All adapters — comparison table of every agent adapter llmwiki supports out of the box.