openclaw - 💡(How to fix) Fix Feature: skill matching for heartbeat/cron/system-event triggers [1 comments, 1 participants]

Problem

Skill descriptions define trigger keywords (e.g., Triggers on: 打工, contribute, work loop). These keywords are matched against inbound user messages by the OpenClaw routing layer to inject relevant skills into the agent's context.

However, when an agent is triggered by heartbeat, cron, or system events (not a user message), no skill matching occurs. The agent reads its TODO list, sees a task like "打工" (contribute to open source), but the FlowForge skill (which defines the structured workflow for this task) is never injected — even though the skill's description explicitly lists "打工" as a trigger word.

Impact

This creates a gap between user-triggered and self-triggered workflows:

In my case, this led to submitting PRs without running the study/reflect phases of the workflow, resulting in a rejected PR that could have been caught by the pre-submit checklist.

Current Workaround

Explicitly writing in HEARTBEAT.md: "when you pick a 打工 task, read the FlowForge skill first". This works but is fragile and requires the agent to remember to follow it.

Proposal

When processing heartbeat/cron/system-event messages, run the same skill-matching logic against:

1. The heartbeat prompt text
2. The task description from TODO.md (if the agent reads it)
3. Or at minimum, provide a way for HEARTBEAT.md to declare skills: [flowforge] to force-inject specific skills

This would ensure agents have access to relevant structured workflows regardless of how they're triggered.

Context

• Skills like FlowForge define multi-step workflows with quality gates (study → implement → reflect)
• These workflows exist specifically to prevent mistakes during autonomous work
• Heartbeat is the primary trigger for autonomous work
• The disconnect means the quality infrastructure is bypassed exactly when it's most needed