OpenClaw Continuity

Make OpenClaw remember the right thing, reconnect the right topic after /new, and follow up naturally without leaking internal state into chat.

OpenClaw Continuity is the public product name for the personal-hooks skill package. It is a skill-layer continuity engine for an existing OpenClaw agent: it does not replace the agent's persona, and it does not bundle a chat-platform adapter.

Four Quick Comics

These four panels show the parts that matter most after someone installs the skill: guided setup, later adjustment, natural-language setting changes, and precise command changes.

For Everyone

Most agents can answer a message. The harder part is remembering what still matters after an interruption, a delayed reply, or a new conversation.

OpenClaw Continuity helps an agent:

• keep ordinary chat light
• stage "let's talk about this later" without losing it
• track important follow-up without chasing every casual sentence
• reconnect the right topic after /new
• understand time with sleep/wake and routine context
• let the user adjust behavior in natural language

中文簡介：

OpenClaw Continuity 是替 OpenClaw 補上「延續感、時間感、關心與追蹤」的技能包。 它不是單純排程發訊息，也不是把所有對話都硬塞進記憶。它會分辨哪些只是一般聊天， 哪些要暫存，哪些真的需要追蹤，並在新開對話時接回使用者選定的承接方式。

What It Feels Like

Example:

User: I need to pause this repo release check. Let's continue later.
Agent: Got it. I will keep the release check as the thread to continue.

/new

Agent: I still have the release check thread. Do you want to pick it up from
the last user intent, the follow-up focus, or the recent conversation summary?

中文例子：

使用者：這個技能上架檢查先放著，晚點再接。
Agent：好，我先把這段接住，不會當成一般閒聊丟掉。

/new

Agent：前面那段技能上架檢查我還接得住。現在要從最後意圖、追蹤焦點，
還是最近 4 輪摘要開始？

The exact wording still belongs to the host agent's persona. This skill supplies the structured continuity state that makes the answer possible.

Minimal Telegram/TG Showcase

Telegram/TG can be used as a small live showcase when your OpenClaw host is already connected to Telegram. The skill does not bundle a Telegram adapter or bot setup; it only supplies the continuity behavior that the host sends through its normal reply pipeline.

Demo flow:

TG user: 這個上架檢查先放著，晚點再接。
Agent: 好，我先幫你接住這條線，不會當成普通閒聊丟掉。

TG user: /new
Agent: 前面那個上架檢查我還接得住。要從最近 4 輪、最後意圖，還是追蹤焦點接？

TG user: 新對話改成看最後使用者意圖。
Agent: 已更新。之後新對話會優先從最後使用者意圖承接。

What this demonstrates:

• /new carryover
• staged memory instead of forced tracking
• natural-language settings update
• frontstage text stays user-facing, not internal runtime text

Core Features

• New conversation carryover /new can reconnect the right previous thread instead of starting from a blank greeting.

• User-selectable /new continuity mode The user can choose which anchor should lead a new conversation: recent_4_turns_first, last_user_intent_first, followup_focus_first, assistant_commitment_first, or balanced.

• Ordinary chat / staged memory / tracked follow-up The runtime separates casual chat from staged memory and formally tracked events.

• Four tracked event types parked_topic, watchful_state, delegated_task, and sensitive_event.

• Structured event chain Keeps context_before, event_core, immediate_result, and followup_focus inspectable.

• Assistant commitment tracking If the assistant promised to check, remember, or come back to something, that commitment can become part of the continuity state.

• Time-aware wording Uses elapsed time, day boundary, sleep/wake boundary, and routine phase as support. Time context does not replace the selected main thread.

• Care and follow-up controls Closure, cooldown, dedupe, dispatch cap, quiet hours, and sleep/rest suppress are explicit settings instead of hidden model guesses.

• Daily-memory writeback Staged and tracked items write concise traces from structured continuity state, not from improvised frontstage text.

• Guided setup First install can ask for timezone, sleep/wake time, relationship/use case, /new continuity mode, and voice/image continuity preference.

• Natural-language settings Users can say things like "make follow-up quieter after midnight" or "新對話改成看最後使用者意圖" and the skill maps that to settings.

What Is In Scope

This package defines the skill behavior:

• state-backed continuity
• /new reattachment
• staging and tracked follow-up
• routine/time context
• setup wizard and settings updates
• frontstage safety guards at the skill/tool layer
• release harness and acceptance checks

What Is Not In Scope

This package does not claim a list of external chat platforms. Channel delivery belongs to the OpenClaw host and its adapter configuration.

That boundary is intentional: the same skill can run through the ordinary OpenClaw reply pipeline without becoming tied to one chat surface.

Setup Wizard

First install can enter guided setup when required fields or first-run preference fields are missing.

The setup flow covers:

• timezone
• sleep_time
• wake_time
• relationship
• use_case
• new_session_continuity_mode
• modality_continuity_mode

The two newest user-facing choices are:

{
  "new_session_continuity": {
    "mode": "recent_4_turns_first"
  },
  "modality_continuity": {
    "enabled": true,
    "mode": "preserve_when_supported",
    "voice_reply_on_voice_thread": true,
    "image_context_carryover": true
  }
}

modality_continuity is host-neutral. It only expresses a preference. It does not add a voice engine, image model, or channel adapter to this skill package.

Change Settings Later

Natural language:

Help me adjust my follow-up settings.
Make follow-up quieter after midnight.
Use the last user intent when a new conversation starts.
新對話承接改成最近 4 輪摘要。
把主動關心改保守一點。
半夜不要主動追蹤我。

Command-style:

python3 scripts/personal_hooks.py setup-check
python3 scripts/personal_hooks.py setup-apply --payload-json '{"new_session_continuity_mode":"last_user_intent_first"}'
python3 scripts/personal_hooks.py setup-apply --payload-json '{"modality_continuity_mode":"preserve_when_supported"}'
python3 scripts/personal_hooks.py setup-apply --payload-json '{"sleep_time":"23:00","wake_time":"07:00"}'

Supported /new continuity modes:

Mode	What leads the next conversation
recent_4_turns_first	Compact recent 4-turn summary
last_user_intent_first	Last user-side intent from compact anchor turns
followup_focus_first	event_chain.followup_focus
assistant_commitment_first	Prior assistant promise/commitment
balanced	Runtime chooses the best anchor for the user's opening message

Install

Copy or symlink this folder into an OpenClaw workspace:

openclaw-workspace/
  skills/
    personal-hooks/
      SKILL.md
      scripts/
      docs/
      examples/

Install the Python dependency:

python3 -m pip install -r requirements.txt

Initialize:

python3 /path/to/openclaw-workspace/skills/personal-hooks/scripts/personal_hooks.py init

Optional installer:

bash scripts/install_local.sh /path/to/openclaw-workspace/skills link

ClawHub-style install:

openclaw skills install openclaw-continuity

npm-style fetch from GitHub:

npm install github:redwakame/openclaw-continuity#v2.0.21

Install Review Notes

This package uses an instruction-first install path. There is no opaque remote installer in the registry entry; the visible steps are the OpenClaw skill install command, the Python dependency install, and the optional local helper script.

Runtime requirements are intentionally small:

• python3
• OPENCLAW_STATE_DIR
• OPENCLAW_CONFIG_PATH
• Python dependency: send2trash

The skill contains executable Python runtime scripts, including the main scripts/personal_hooks.py runtime. Before running it against a sensitive or production agent, review the scripts you plan to execute, especially:

• scripts/personal_hooks.py
• scripts/followup_skill_harness.py
• scripts/install_local.sh

Suggested checks:

• confirm there are no unexpected network calls or hardcoded remote endpoints
• confirm only the declared state/config paths are needed for normal runtime
• use an isolated OPENCLAW_STATE_DIR for first-run testing
• avoid pointing OPENCLAW_CONFIG_PATH at a production config containing provider keys until you have reviewed the code

This is a state-backed continuity skill. Reading and writing a dedicated OpenClaw state directory is expected behavior; reading unrelated host secrets or sending user data to external services is not part of this package's design.

Documentation Links

Use these files when installing, reviewing, or extending the skill:

• Install guide: local setup, environment variables, and operator workflow
• Harness guide: regression harness usage and expected pass criteria
• Host operator settings: guided setup, natural-language settings, and command settings
• Routine schedule: sleep/wake, quiet hours, and time-aware behavior
• Architecture: continuity layers, event chain, and writeback design
• Validation summary: release validation notes
• Release acceptance: checks to run before publishing
• Known limits: boundaries and non-goals
• Security notes: reporting and security expectations
• ClawHub page: registry install and scan status

Verify

Run the regression harness:

python3 /path/to/openclaw-workspace/skills/personal-hooks/scripts/followup_skill_harness.py --absence-minutes 3

Expected result:

{
  "summary": {
    "pass_count": 14,
    "fail_count": 0
  }
}

Quick manual behavior check:

1. Ask the agent to hold a topic for later.
2. Start /new.
3. Confirm it reconnects the selected continuity anchor instead of generic small talk.
4. Ask to change setup, for example: 新對話改成看最後使用者意圖.

Technical Map

Important files:

• SKILL.md: trigger rules and runtime boundary
• scripts/personal_hooks.py: main runtime
• scripts/followup_skill_harness.py: regression harness
• config.schema.json: settings schema
• examples/settings.sample.json: sample settings
• docs/install.md: installation details
• docs/host-operator-settings.md: operator settings
• docs/release-acceptance.md: release gate
• docs/publish-copy.md: GitHub / ClawHub copy source

Main runtime surfaces:

• build_runtime_context(): builds carryover, schedule, setup, and guard prompts
• intercept_message(): routes the user turn into casual/staged/tracked state
• process_candidate_buffer(): promotes candidate state to incidents/hooks
• setup-check / setup-apply: guided configuration commands
• frontstage-guard: skill/tool-layer output safety guard

Public V2 Includes

• casual_chat / staged_memory / tracked_followup routing
• parked_topic
• watchful_state
• delegated_task
• sensitive_event
• candidate -> incident -> hook promotion
• structured event_chain
• structured causal_memory
• /new carryover
• user-selectable /new continuity mode
• assistant commitment support
• routine/time context as support-only signal
• quiet hours
• sleep/rest suppress
• cooldown
• closure
• dedupe
• dispatch cap
• daily-memory writeback
• deterministic onboarding
• guided settings
• natural-language settings changes
• host-neutral modality preference
• regression harness

中文功能總覽

• 一般聊天、暫存記憶、正式追蹤分流
• /new 後接回正確主題
• 可選新對話承接方式：最近 4 輪、最後使用者意圖、因果跟進焦點、助手承諾、綜合判斷
• 作息與時間感只當輔助，不搶主線
• 可用自然語言調整設定
• 可用指令精準修改設定
• 關心與追蹤有冷卻、退場、去重、上限與勿擾
• 暫存/追蹤內容可寫回 daily memory trace
• 技能層保持平台中立，不把通訊渠道寫死進技能定義

Contact

If you are interested in this package, run into a problem, or want to exchange ideas about OpenClaw continuity, feel free to contact me:

• adarobot666@gmail.com

If you are satisfied with this skill package, please star the GitHub repository as encouragement. I will keep pushing improvements, maintenance, and new features.

如果你對這個技能包有興趣、使用時遇到問題，或想交流 OpenClaw continuity 相關想法，歡迎聯繫我：

• adarobot666@gmail.com

如果你滿意這個技能包，也歡迎在 GitHub 給一顆星作為鼓勵。我會持續推動 優化、維護與新功能追加。