openbiliclaw-adapter

Files (1)

• SKILL.md

SKILL.md (first lines)

# OpenBiliClaw Agent Bridge Skill

Use this skill when you are inside the OpenBiliClaw workspace and need current state or want to push feedback back into the learning loop. The bridge is host-neutral: OpenClaw, Hermes and WorkBuddy use the same JSON contract.

## Deployment Choice

Choose deployment by target machine capability:

1. Docker available: prefer Docker
2. No Docker: use local Python deployment

## Bootstrap

### Docker-first

Run:

```bash
docker compose up -d --build
docker exec -it openbiliclaw-backend openbiliclaw init
```

Keep the repository checkout available so the host can discover this workspace skill.

### Local fallback

If Docker is unavailable, bootstrap locally:

```bash
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
cp config.example.toml config.toml
```

Then initialize OpenBiliClaw once:

```bash
openbiliclaw init
```

If `config.toml` is still missing API Key or B 站 Cookie and the terminal is interactive, `openbiliclaw init` will guide the operator through setup. After init, verify the adapter bridge:

```bash
uv run python -m openbiliclaw.integrations.openclaw.cli doctor
```

For a longer setup guide, read `docs/openclaw-quickstart.md` and `docs/agent-integration.md`.

## Command Bridge

Always call the adapter through the JSON CLI bridge:

```bash
uv run python -m openbiliclaw.integrations.openclaw.cli <command> [flags]
```

Supported commands:

- `capabilities` — negotiate `agent-bridge/v2` and the complete capability list before caching tools
- `sync-account`
- `get-profile`
- `recommend --limit 5 [--source-platform <platform>] [--exclude-item-id <id>] [--request-id <stable-id>] [--stream-ref <stream>] [--realtime]`
- `reshuffle` / `append` — replace or append precomputed recommendation pages
- `restore` / `present` — restore a Director screen, then acknowledge only cards actually shown by the host
- `get-delight` / `respond-delight` — view, like, dislike, dismiss or chat about a surprise
- `activity-feed` / `platform-availability`
- `next-probe` — get the next speculative-interest hypothesis to ask the user about
- `respond-interest-probe --domain "..." --response confirm|reject|defer|chat [--message "..."]`
- `next-avoidance-probe` / `respond-avoidance-probe --domain "..." --response confirm|reject|defer|chat`
- `chat --message "..." [--session openclaw]` — durable Socratic dialogue turn
- `chat-history [--session openclaw]`
- `profile-edit-state` / `edit-profile` — read or update deterministic profile overlays
- `save-local`, `list-saved`, `remove-saved` — local-first saved lists
- `sync-saved --allow-state-changing` — explicitly authorized native-save synchronization
- `runtime-status`
- `submit-feedback --recommendation-id 7 --feedback-type like --request-id feedback-7-like-1 --note "很对胃口"`
- `listen` — long-running WebSocket stream for real-time push events (see below)

The complete source of truth is `openbiliclaw_get_capabilities` / `emit-skill-descriptors`; do not hard-code an older subset.

### Director screen protocol

When Director mode is enabled, create one stable request ID for each `recommend` / `append` / `reshuffle` action and reuse it with the same parameters on retries. Use a stable stream for one reading session. `append` advances the screen; `reshuffle` retains its current direction. Use `restore` to recover without generating or consuming another screen. Mode configuration is TOML-based; restart the complete backend after changes.

Preserve the returned `batch_ref`, `item_key`, and exact `recommendation_id`. Tool delivery alone is not exposure. After the host confirms which cards it actually displayed, call `present --batch-ref <batch> --ack-id <stable-ack> --item-key <shown-key>` (repeat `--item-key` for the displayed subset). Retry with the same ACK identity. If the host cannot confirm visibility, leave presentation unacknowledged; explicit user click/feedback provides its own evidence. Associate feedback with the exact recommendation and optional `--batch-ref`, never the newest row for the same content.

## Proactive Push (WebSocket)

Instead of polling `get-delight` / `next-probe`, OpenClaw can receive real-time push notifications via WebSocket:

```bash
uv run python -m openbiliclaw.integrations.openclaw.cli listen
```

This connects to the runtime stream and outputs one JSON line per event:

```json
{"ok": true, "data": {"status": "connected", "ws_url": "ws://127.0.0.1:8420/api/runtime-stream", "event_types": ["avoidance.chat", "avoidance.confirmed", "avoidance.deferred", "avoidance.probe", "avoidance.rejected", "delight.candidate", "delight.chat", "delight.disliked", "delight.liked", "delight.refreshed", "interest.chat", "interest.confirmed", "interest.deferred", "interest.probe", "interest.rejected"]}}
{"ok": true, "data": {"type": "delight.candidate", "bvid": "BV1xxx", "title": "...", "delight_reason": "...", "delight_score": 0.92, "delight_hook": "深层共鸣"}}
{"ok": true, "data": {"type": "interest.probe", "domain": "建筑美学", "reason": "...", "question": "我从你最近的轨迹里嗅到你可能对【建筑美学】感兴趣——... 这个方向你自己认不认？"}}
{"ok": true, "data": {"type": "avoidance.probe", "domain": "浅层热点复读", "reason": "...", "question": "我猜【浅层热点复读】可能是你想避开的方向——... 这个判断准吗？"}}
```

Default event types include `delight.candidate`, `interest.probe`, `avoidance.probe` and their confirmed/rejected/deferred result events. The command auto-reconnects on disconnection. Press Ctrl-C to stop.

Options:
- `--ws-url <url>` — override the WebSocket endpoint
- `--events <types>` — comma-separated event types to forward; omit it to use the current default manifest

## Socratic Dialogue & Interest Probing

The host can proactively ask the user to clarify or confirm interests and avoidances, then send the answer back into the learning loop.

### Get the next interest hypothesis

```bash
uv run python -m openbiliclaw.integrations.openclaw.cli next-probe
```

From whiteguo233/OpenBiliClaw, MIT.

More from whiteguo233/OpenBiliClaw