cyber-girlfriend（网络女友）

Cyber Girlfriend

Use this skill when the user wants an agent to feel like a real companion instead of a reactive assistant.

This skill is for designing and wiring:

• owner-only proactive messages
• relationship-aware tone and pacing
• configurable quiet hours, daily limits, and cooldowns
• lightweight relationship memory and preference learning
• optional topical share hooks such as X trending caches
• runtime integration without hardcoding private IDs, secrets, or model choices

Current OpenClaw Architecture

On OpenClaw, treat the system as two layers:

1. Live cron jobs own behavior delivery

- Each scheduled task has its own real cron payload. - morning, afternoon, evening, night, heartbeat, and any custom labels are all valid examples, not fixed limits. - Those payloads can contain the full task prompt, search steps, generation instructions, and delivery behavior. - On OpenClaw, prefer running companion cron jobs in a dedicated persistent custom session such as session:companion-owner with payload.kind: "agentTurn". - Use main + systemEvent only as a legacy compatibility path, because it can bleed into unrelated heartbeat or owner-chat context. - Do not duplicate full cron payloads in config.

1. scripts/companion_ping.py owns state and context

- pacing checks - quiet-hours / cooldown / daily-limit checks - lightweight relationship memory - style/content hint selection - hotspot cache consumption - recent-owner-message context extraction

Keep these responsibilities separate. Do not turn the state script back into a campaign/orchestration script.

What To Build

Produce these parts unless the user asks for a subset:

1. A local config file with private/runtime-specific values externalized
2. A companion behavior policy:

- quiet hours - daily limit - cooldown - owner-only routing

1. A lightweight state/context script
2. Lightweight state:

- last proactive time - last mode/style/content type - learned preference scores - relationship state inferred from owner replies

1. Optional share-source cache:

- X trending - RSS/blogs/channels - other user-approved sources

1. Live cron jobs that call the state script and then perform per-mode behavior

- Users may define any number of scheduled tasks, any task names, and any time slots. - The skill should guide them, not force a four-slot preset.

Hard Rules

• Never hardcode secrets.
• Keep proactive behavior owner-only unless the user explicitly wants broader scope.
• Default to restraint. A believable companion is low-frequency and context-aware, not spammy.
• Prefer strong tone control and natural delivery over long roleplay blocks.
• On OpenClaw, treat live cron as the source of truth for scheduled prompts.
• Keep runtime-specific values in local config or environment variables, not in published skill defaults.

First-Run Onboarding For OpenClaw

When a user wants to set this up from scratch on OpenClaw, do not jump straight into editing files and cron jobs blindly.

Use this order:

1. Confirm the proactive delivery target from the user's local config (delivery.channel / delivery.owner_target / delivery.account).
2. Ask whether they want the recommended starter profile or a custom schedule.
3. If they do not have a strong opinion yet, default to the recommended starter profile:

- morning - afternoon - evening - night - optional lightweight heartbeat

1. Create or update the local config first.
2. Then create the live cron jobs.
3. Then run one dry test for one mode.
4. Only after a successful user-visible delivery should the handler call --mark-sent.

For a detailed setup conversation, starter cron blueprint, copy-pasteable live cron templates, and profile decision flow, read:

• references/onboarding.md
• references/cron-blueprints.md
• references/live-cron-templates.md

Build Order

1. Define The Local Config Surface

Read references/configuration.md and create a local config from assets/cyber-girlfriend.config.example.json.

For new users, ask only the minimum configuration questions first:

• Which delivery channel should proactive messages use?
• What is the owner target/account on that channel?
• Do they want the recommended four-slot starter schedule or a custom schedule?
• What quiet hours, cooldown, and daily limit do they want?
• Do they want a lightweight heartbeat check-in in addition to scheduled modes?

If the user asks for “something like the current polished setup”, use the starter cron blueprints from references/cron-blueprints.md. If they need something directly copy-pasteable for OpenClaw cron creation, use references/live-cron-templates.md instead of improvising from scratch.

At minimum, externalize:

• owner routing info
• quiet hours
• pacing limits
• session/state/cache paths
• health-check commands
• optional share-source settings

The local config's delivery block is the canonical target for proactive outbound messages. Do not implicitly post into the current UI/session just because a cron or heartbeat wakeup happened there.

2. Define The Behavior Model

Use these mode buckets as the recommended starter profile unless the user wants something custom:

• morning
• afternoon
• evening
• night

Also support heartbeat and user-defined custom modes.

Keep style/content selection lightweight and script-driven. The script should output hints; the live cron prompt can decide how to express them.

Important framing for agents reading this skill:

• the four-slot profile is a starter recommendation, not a hardcoded product limit
• users may define any number of modes, any labels, and any cron times
• if the user asks for “something like yours”, the recommended starter profile is the correct default

3. Add Relationship Memory

Use lightweight heuristics, not opaque lore dumps.

Track:

• reply delay after proactive messages
• recent owner reply snippet
• preference counters such as service, clingy, curious, teasing, wrapup
• current relationship state such as secure, light_touch, present, slightly_needy, misses_him

4. Add Optional Share Sources

Read references/share-sources.md.

For X-based sharing on OpenClaw:

• prefer a local cache file over mandatory API coupling
• let cron jobs or another approved refresher update the cache
• keep Chrome path, URLs, cache path, and enablement in config
• let companion_ping.py consume cache rather than scrape live itself

5. Wire The Runtime

If the user is on OpenClaw, read references/openclaw-integration.md.

For new-user setup, prefer this delivery sequence:

• config first
• cron creation second
• one dry test third
• wording/style iteration last

Do not spend the user's time polishing prompt copy before the delivery path and cron plumbing actually work.

For the actual per-mode starter prompts, prefer the blueprint set in references/cron-blueprints.md and then customize only where the user has a clear preference. When the user or another agent needs concrete OpenClaw job objects, read references/live-cron-templates.md.

Use this pattern:

• live cron job wakes on schedule
• cron payload runs companion_ping.py <mode> --config ...
• if the script returns skip, the turn exits quietly
• if the script returns ok, the cron payload continues with the richer per-mode behavior
• proactive outbound delivery must follow the local config's delivery block rather than whichever session/UI received the wakeup
• only after the user-visible message is actually delivered successfully, run companion_ping.py <mode> --config ... --mark-sent

Any cron job may call the script with any mode label the user has configured.

Important pacing rule:

• heartbeat has its own cooldown bucket
• heartbeat must not consume the normal morning/afternoon/evening/night cooldown
• heartbeat is not a once-per-day mode; it should not be blocked by mode_already_sent_today

Keep the cron payloads rich and explicit. Keep the state script thin.

Operational Signal Contract

When the runtime script detects operational state, it should output a lightweight operational block for the calling agent/runtime to interpret.

Preferred shape:

• operational.gateway_healthy
• operational.cron_issues
• operational.signal

- level: none | medium | high - kind: e.g. none | cron_issue | gateway_unhealthy - blend: e.g. none | soft_service_note | service_report - should_mention: boolean

• operational.guidance

- mention_briefly: boolean - avoid_alarmist_tone: boolean

Default interpretation:

• level = none: ignore operational state in the final message
• level = medium: if mentioned at all, blend it in gently as a light service-flavored aside; do not let it take over the interaction
• level = high: it is acceptable to shift into a more explicit service/report posture

This contract belongs in the skill/runtime layer, not duplicated verbatim into every cron payload.

Output Expectations

When implementing this skill for a user:

• create or update the local config file
• if the user asked for setup from scratch, guide them through a clear onboarding flow instead of assuming they already know cron structure
• offer the recommended starter profile first unless they explicitly want a custom schedule
• keep the state script reusable and config-driven
• keep the scheduled prompts in live cron, not duplicated in config
• make sure the runtime output contract is documented when adding fields such as operational.signal and operational.guidance
• verify syntax and one end-to-end dry run when possible

Publishing Checklist

Before calling the skill publishable, confirm:

• no personal IDs or secrets remain in the publishable files
• no provider/model is mandatory unless the user explicitly wants that
• local config and runtime state are excluded from what gets published
• OpenClaw docs describe live cron as canonical if that is the chosen architecture
• documentation does not reference deleted scripts or retired flows