proactive-claw（主动爪）

🦞 Proactive Claw

Proactive Claw is a proactive execution engine that collaborates with you and learns from you.

It helps you be more productive by making smart suggestions at the right moment — and (by default) only applying what you approve.

Everything can run locally.

Safety in one glance ✅

✅ Asks before applying changes (default) ✅ Writes only what you approve (no surprise calendar edits) ✅ Local-first core bundle (no external integration helpers) ✅ Everything can run locally (small/local scoring model recommended) ✅ Local state folder is transparent and deletable

Why people install it (what you feel in week 1)

• You show up prepared more often.
• Your calendar becomes more realistic (buffers appear where they matter).
• Deep work blocks stop getting shredded by “just one quick meeting.”
• Follow-ups stop slipping.
• Proactive Claw learns your style quickly:

prep durations, preferred times, what you reject, what you always accept.

In short: it becomes your personal assistant.

Graph A — the simple loop (how it works at a glance)

+--------------------------+        +--------------------------+
| Calendar                 |        | Chat                     |
| Google Calendar/Nextcloud|        | (your control loop)      |
+------------+-------------+        +------------+-------------+
             |  events <-> suggestions           | prompts <-> decisions
             |                                   |
             +-------------------+---------------+
                                 |
                                 v
                  +--------------------------------------+
                  |            Proactive Claw             |
                  |  - notices what's coming              |
                  |  - prioritizes what matters           |
                  |  - suggests prep / buffers / followup |
                  |  - learns from your feedback          |
                  |  - applies only what you approve      |
                  +--------------------------------------+

Key idea: Proactive Claw is proactive, but you remain the decision-maker (by default).

Quick start (2 minutes)

1) Run readiness checks:

bash scripts/setup.sh --doctor

2) If dependencies are missing, print one install command:

bash scripts/setup.sh --print-install-cmd google
# or: bash scripts/setup.sh --print-install-cmd nextcloud

3) Bootstrap calendar connection:

bash scripts/setup.sh

4) Run a safe preview:

bash scripts/quickstart.sh

5) Approve a suggestion in chat, then run once:

python3 scripts/daemon.py

Modes (choose your vibe)

• Suggest (default): asks before applying changes (max_autonomy_level=confirm)
• Background (manual): run local daemon yourself (python3 scripts/daemon.py --loop)
• Autonomous (advanced): explicit opt-in only (not recommended until you’ve used it for a while)

Presets (easy setup)

You don’t need to tune dozens of settings. Pick a preset.

Presets are just a starting point. Proactive Claw adjusts to you with every interaction (approvals, edits, rejections), and quickly becomes personal.

A) Calm mode (minimal interruptions)

Choose this if you want only high-value suggestions.

• fewer prompts
• only important events get prep/buffers

B) Builder mode (deep work protected)

Choose this if you code/build and need long focus blocks.

• defends 2–3h deep work windows
• suggests moving meetings that fragment focus
• stronger buffers + recovery time

C) Meetings-heavy week

Choose this when your schedule is packed.

• adds buffers + reset breaks
• prevents back-to-back collapse
• proposes follow-ups so action doesn’t slip

(If you want, these presets can be expressed as a small copy/paste block in config.json.)

Chat scoring model (built-in; small/local recommended)

Proactive Claw includes a chat scoring model to rank signals (importance, urgency, disruption risk) so it knows:

• what matters,
• when to ask you,
• and when to stay quiet.

Recommendation: use a small model, ideally local (fast, cheap, private). Your planning model can be bigger; scoring should stay small and fast.

Scenarios (mini transcripts)

1) Presentation tomorrow → prep + buffer

You: Tomorrow I’m presenting the roadmap to the board. Claw (score = 0.92): High impact. You usually prep ~70 minutes for presentations. Claw: Want me to reserve 08:40–09:50 for prep + a 10-min buffer before the 10:00 meeting?

2) Deep work protection (developer mode)

Claw (score = 0.78): This meeting would split your deep work block (09:30–12:00). Claw: Prefer to move it to 13:00, or keep it and add a 20-min recovery buffer after?

3) Back-to-back day → buffers + reset

Claw (score = 0.81): Tomorrow is back-to-back from 10:00–14:00. Claw: Add 10-min buffers between meetings + a 25-min reset break around midday?

4) Follow-up that tends to slip

You: I promised I’ll send the summary. Claw (score = 0.66): Time-sensitive. Claw: Want a 20-min follow-up block right after the meeting (or at 17:10)?

5) It becomes personal fast (recurring pattern)

Claw (score = 0.71): I’ve noticed a pattern for “Client Review” meetings:

• you prefer 30-min prep the same day
• 10-min buffer before
• no meetings after 18:00

Claw: Want me to propose this pattern whenever a new “Client Review” appears (still asking you to approve)?

Graph B — technical architecture (backend = Google OR Nextcloud)

TECHNICAL ARCHITECTURE (backend = Google OR Nextcloud)

             +----------------------------------------------+
             |        Calendar backend (choose ONE)         |
             |  +-----------------------+  OR  +----------+ |
             |  | Google (OAuth/API)    |      | Nextcloud| |
             |  +-----------------------+      | (CalDAV) | |
             |                                 +----------+ |
             +-----------------------+----------------------+
                                     ^
        events / changes (propose+apply, Actions only)       |
                                     |
                                     v
+----------------------+      +------------------------------------------+      +----------------------+
|  OpenClaw Chat UI    | <--> |           Proactive Claw Core            | <--> |  Local state         |
| prompts <-> decisions|      | 1) Ingest: events + chat signals         |      | config/token/SQLite  |
| feedback loop        |      | 2) Score: urgency/importance (local rec) |      | logs (optional)      |
+----------------------+      | 3) Plan : prep + buffers + follow-ups    |      +----------------------+
                              | 4) Approve (default): ask/confirm in chat|
                              | 5) Apply : write approved changes        |
                              | 6) Learn : from your approvals/edits     |
                              +------------------------------------------+
                                     ^
                                     |
                           Optional (explicit opt-in)
                                     |
                                     v
                              +----------------------+
                              | Optional daemon      |
                              | periodic scan/suggest|
                              +----------------------+

Setup

0) Requirements

• python3 available on your machine
• Google backend: credentials.json (OAuth client)

OR

• Nextcloud backend: CalDAV URL + app password (or token)

1) Validate dependencies + bootstrap config

bash scripts/setup.sh --doctor
bash scripts/setup.sh --print-install-cmd google
# install dependencies manually
bash scripts/setup.sh

2) Choose your calendar backend

Option A — Google Calendar (OAuth)

You provide:

• credentials.json (OAuth client)

Proactive Claw creates:

• token.json after OAuth

Option B — Nextcloud (CalDAV)

You provide:

• CalDAV base URL
• app password

3) Configure behavior

Interactive wizard:

python3 scripts/config_wizard.py

Safe defaults:

python3 scripts/config_wizard.py --defaults

4) Core boundaries (this package)

• Includes: scripts/daemon.py for manual local runs.
• Excludes from published bundle: cross-skill connectors, team awareness, daemon installers, optional remote OAuth bootstrap, voice bridge, and llm rater.
• No background installer is shipped in this package.

FAQ

Will this mess up my calendar?

By default: no. Proactive Claw suggests changes in chat and applies only what you approve (max_autonomy_level = confirm).

Where does it store data?

Locally under: ~/.openclaw/workspace/skills/proactive-claw/

Does everything run locally?

Yes — it can. The scoring model is recommended to be small/local. The calendar backend still connects to your chosen provider because that’s where your calendar lives.

Can I use it without Google?

Yes — use Nextcloud (CalDAV).

Does core include daemon installer helpers?

No. Core includes scripts/daemon.py for manual runs. No launchd/systemd installer is published in this package.

What does “learning” mean here?

It learns your preferences from your approvals/edits over time: prep durations, preferred times, buffer sizes, meeting types that matter, and deep work rules.

What is the chat scoring model?

A lightweight model that assigns numeric scores (e.g., 0.66, 0.92) to decide what matters and when to prompt you. Recommended: small/local model.

Troubleshooting (by symptom)

• “Too many prompts” → use Calm mode; reject a few times; it will adapt
• “Not proactive enough” → run python3 scripts/daemon.py --loop for continuous local scanning
• “Prep blocks are wrong length” → edit them twice; it will converge
• “OAuth issues” → re-run setup; revoke token and re-auth if needed

Uninstall

1) Delete: ~/.openclaw/workspace/skills/proactive-claw/ 2) Revoke Google OAuth access if you used Google: myaccount.google.com/permissions

Glossary

• Prep block: time reserved before an event to prepare
• Buffer: short gap that prevents schedule collisions
• Deep work: uninterrupted focus block
• Daemon: background loop process (python3 scripts/daemon.py --loop)
• Score: 0–1 number indicating importance/urgency

To improve this 100×, here’s what would help:

1) Your config key list (or config.example.json) so I can turn presets into copy/paste blocks safely. 2) Whether you want scores displayed as 0.92 or 92/100 (both numeric; one is friendlier). 3) If you want an “Advanced” section for autonomous mode or prefer to keep it out of the main doc.