larc-runtime

larc-runtime

Overview

This is a thin governance skill used when the OpenClaw Agent handles Lark/Feishu business operations.

• Thinking / dialogue: OpenClaw Agent
• Business governance: larc
• Atomic Feishu execution: official openclaw-lark plugin

The purpose of this skill is for the OpenClaw Agent to operate through LARC's queue / gate / lifecycle rather than carelessly hitting APIs directly.

Basic principles

1. First, larc ingress openclaw get the next action bundle with
2. Do not skip approval / preview on your own
3. Perform actual Feishu operations with the official openclaw-lark plugin as much as possible
4. Return state transitions to larc ingress done|fail|followup back to

Basic flow

1. Get the next governed action

bin/larc ingress openclaw --agent main --days 14

When specifying a particular queue item:

bin/larc ingress openclaw --queue-id <queue-id> --days 14

What this output includes:

• openclaw_command
• recommended_commands
• gate
• authority
• task_types

2. Pass it to OpenClaw as-is

Local embedded agent:

bin/larc ingress openclaw --queue-id <queue-id> --execute

Via Gateway:

bin/larc ingress openclaw --queue-id <queue-id> --gateway --execute

3. Read the context before execution

pending / preview item:

bin/larc ingress context --queue-id <queue-id> --days 14

delegated item:

bin/larc ingress handoff --queue-id <queue-id> --days 14

4. Execution and state update

Execution plan:

bin/larc ingress execute-stub --queue-id <queue-id>

Safe adapter execution:

bin/larc ingress execute-apply --queue-id <queue-id> --dry-run

Completion:

bin/larc ingress done --queue-id <queue-id> --note "Completed by OpenClaw"

Failure:

bin/larc ingress fail --queue-id <queue-id> --note "Reason"

When manual follow-up remains:

bin/larc ingress followup --queue-id <queue-id>

Decision rules

• gate: approval

  • Do not execute until approval is finished
  • larc ingress approve と larc ingress resume pass through

• gate: preview

  • Do not run the real execution right away; first read context / execute-stub

• authority: user

  • Operations that require real user permissions

• authority: bot

  • Operations that close on the bot/tenant side

• authority: user or bot

  • Compound task. Handle it by splitting according to the bundle's flow

Division of roles on the execution side

• LARC:

  • queue
  • permission
  • gate
  • delegation
  • memory retrieval
  • lifecycle

• official openclaw-lark plugin:

  • IM
  • Docs
  • Drive
  • Base
  • Wiki
  • Calendar
  • Task

Coordination with Addness goal management

The OpenClaw Agent links coding work with Addness goals. Details: ../addness/SKILL.md

At work start

# git ブランチから自動検出
addness-cli work start

# ゴールID直接指定
addness-cli work start --goal <GOAL_ID>

During work (progress recording)

addness-cli progress --message "設計完了、実装着手"

At work completion (required at the same time as LARC queue completion)

# larc ingress done と合わせて実行する
bin/larc ingress done --queue-id <queue-id> --note "Completed by OpenClaw"
addness-cli work done --message "<完了内容>" --pr <PR_URL>

Summary check

addness-cli summary --all

Constraints

• Writing to the banana org is strictly prohibited(read-only)
• When the goal ID is unknown: addness-cli -- goal list --assigned-to me --json

Things you must not do

• Directly execute an item that requires approval
• Finish only the work without updating the queue
• Complete Lark's source-of-truth data only locally
• Skip LARC and abuse APIs directly
• Create, update, or delete goals in the banana org