AGENT SKILL

live-debrief

Voice-driven After Action Review (AAR) system for the Rabbit R1 — guides a team through four structured debrief phases, logs timestamped observations per speaker, and exports a complete Markdown and JSON report to your local machine.

License 许可
MIT
最近更新
2026-09-04
Stars
★ 0
Forks
0

01 前言 Front Matter

name
live-debrief
author
rabbit-r1-skillz
version
1.0.0
description
Voice-driven After Action Review (AAR) system for the Rabbit R1 — guides a team through four structured debrief phases, logs timestamped observations per speaker, and exports a complete Markdown and JSON report to your local machine.

02 资源文件 3

03 SKILL.md 正文

LiveDebrief — OpenClaw Skill

LiveDebrief turns your Rabbit R1 into a military-style After Action Review facilitator. Using push-to-talk recording and on-device LLM categorization, it guides a team through four structured debrief phases (Intended, Actual, Delta, Lessons), logs every observation with speaker attribution and timestamp, supports a parallel timeline mode for event sequencing, and exports a complete Markdown document plus structured JSON report to your local machine when the AAR is complete.

The skill receives a single command from the R1 at the end of each AAR: the full rendered Markdown export and the structured JSON data. Your skill handler saves both to disk and responds within the 20-second receive window so the R1 can confirm the file was written.


Commands

Command Sent by R1 Description
export-aar Automatically on AAR completion Save the completed AAR as Markdown and JSON files to your configured output directory

There is one ocSend call in the Creation. It fires when the user finishes all four phases and the complete screen is shown.


Prerequisites

  • Node.js 18+ on your local machine
  • OpenClaw daemon installed and running (openclaw serve)
  • R1 and local machine on the same local network (OpenClaw requirement)
  • Write access to the output directory (default: ~/debriefs)

No external API keys are required. The skill writes files locally only.


Skill Installation

1. Create the skill directory

mkdir -p ~/.openclaw/skills/live-debrief

2. Install dependencies

cd ~/.openclaw/skills/live-debrief
npm init -y
npm install fs-extra

3. Create the configuration file

Create ~/.openclaw/skills/live-debrief/.env:

# Directory where AAR files are saved
DEBRIEF_OUTPUT_DIR=~/debriefs

# Optional: subdirectory organization ("flat" or "by-date")
DEBRIEF_ORG=by-date

4. Create the skill file

Create ~/.openclaw/skills/live-debrief/skill.js:

"use strict";

const path    = require("path");
const fse     = require("fs-extra");
require("dotenv").config({ path: path.join(__dirname, ".env") });

const OUTPUT_DIR = (process.env.DEBRIEF_OUTPUT_DIR || "~/debriefs")
  .replace(/^~/, require("os").homedir());

const ORG = process.env.DEBRIEF_ORG || "by-date";

// ── Entry point called by OpenClaw ────────────────────────────────────────────
module.exports = async function liveDebrief(command, params) {
  if (command !== "export-aar") {
    return { error: `Unknown command: ${command}` };
  }

  const { aarId, name, markdown, json, ts } = params;

  if (!aarId || !markdown || !json) {
    return { ok: false, error: "Missing required fields (aarId, markdown, json)" };
  }

  // ── Determine output directory ─────────────────────────────────────────────
  let outDir = OUTPUT_DIR;
  if (ORG === "by-date") {
    const date = new Date(ts || Date.now()).toISOString().split("T")[0];
    outDir = path.join(OUTPUT_DIR, date);
  }
  await fse.ensureDir(outDir);

  // ── Build safe filename from AAR name ─────────────────────────────────────
  const safeName = (name || "untitled")
    .toLowerCase()
    .replace(/[^a-z0-9]+/g, "-")
    .replace(/^-|-$/g, "")
    .slice(0, 60);

  const mdPath   = path.join(outDir, `${safeName}-${aarId}.md`);
  const jsonPath = path.join(outDir, `${safeName}-${aarId}.json`);

  // ── Write files ────────────────────────────────────────────────────────────
  await fse.writeFile(mdPath,   markdown,               "utf8");
  await fse.writeFile(jsonPath, JSON.stringify(json, null, 2), "utf8");

  console.log(`[live-debrief] AAR saved:\n  ${mdPath}\n  ${jsonPath}`);

  return {
    ok:       true,
    mdPath,
    jsonPath,
    savedAt:  new Date().toISOString(),
  };
};

5. Register the skill

Add to ~/.openclaw/openclaw.config.json:

{
  "skills": [
    {
      "name": "live-debrief",
      "path": "~/.openclaw/skills/live-debrief/skill.js",
      "description": "LiveDebrief AAR export — saves Markdown and JSON to local disk"
    }
  ]
}

6. Start the OpenClaw daemon

openclaw serve

Confirm live-debrief appears in the registered skills list before launching the R1 Creation.


How It Works — Step by Step

Boot

  1. The R1 Creation loads and calls loadAARIndex() from creationStorage.plain, reading the list of saved AARs by their index key (ld:aar-index).
  2. The boot screen shows saved AAR count. PTT starts a new AAR; long-press PTT opens the AAR select list.

Name the AAR

  1. User presses PTT and speaks the operation, event, or exercise name (STT via CreationVoiceHandler).
  2. The sttEnded message returns the transcript. The R1 displays it and prompts the user to long-press PTT to confirm.
  3. A new AAR object is created in memory and persisted to creationStorage.plain under ld:aar:<id>.

Add Participants

  1. The Participants screen lets users add speaker names or callsigns one at a time via PTT. Each name is stripped of introductory phrases ("I am", "this is") and deduplicated.
  2. Scrolling up/down selects the active speaker whose name will be attributed to the next logged observation.
  3. Long-press PTT advances to Phase 1 when participant setup is complete. A default "Team" participant is added automatically if none were spoken.

Four-Phase Debrief

The debrief progresses through four phases in order:

Phase Question Purpose
1 — INTENDED What was the plan? What did we expect? Capture stated objectives, resources, timeline, responsibilities
2 — ACTUAL What actually happened? Reconstruct the real sequence of events
3 — DELTA Why did actual differ from intended? Identify root causes, missed signals, planning failures
4 — LESSONS What do we sustain? What do we improve? Extract actionable takeaways for next time

For each phase:

  1. User presses PTT to open the recording overlay. The R1 STT captures the spoken observation.
  2. On PTT release, the transcript is sent to the R1's built-in LLM (R1.sendToLLM) with a structured prompt requesting:
    • text — a cleaned, concise 1–2 sentence version of the observation (max 20 words)
    • category — one of sustain, improve, event, root, or general
    • facilitatorPrompt — a Socratic follow-up question if the LLM judges one would deepen the analysis
  3. The observation is stored under the current phase with speaker attribution, timestamp, raw transcript, and LLM-structured text.
  4. If the LLM returns a facilitator prompt that has not already been shown, the R1 surfaces it on the Facilitator Prompt screen after 3 seconds. PTT records an answer; long-press PTT dismisses.
  5. Short long-press PTT (released before 1.2 s) triggers the phase advance confirm gate. The user must long-press PTT again to confirm.
  6. On confirmation, the R1 reads back the phase summary (up to 3 observations spoken aloud), then auto-advances to the next phase.

Timeline Mode

At any point during Phase 2 (ACTUAL) or later phases, a long-press PTT held for more than 1.2 seconds enters Timeline Mode. In this mode:

  1. A running elapsed timer (T+MM:SS) is displayed, anchored to when the timeline was first entered.
  2. Each PTT press records a timestamped event: the spoken text is stored with its relative elapsed time in both the global timeline array and the current phase's observations (category: event).
  3. Long-press PTT exits timeline mode and returns to the phase screen.

Export

  1. After all four phases are complete and the AAR Complete screen is shown, the R1 calls:
await R1.ocSend("live-debrief", "export-aar", {
  aarId:    aar.id,       // unique AAR identifier
  name:     aar.name,     // human-readable name (e.g. "Op Nightfall")
  markdown: exportDoc.markdown,
  json:     exportDoc.json,
  ts:       Date.now(),
});
await R1.ocReceive({ timeoutMs: 20000 });
  1. The OpenClaw daemon routes the command to your skill handler.
  2. The skill writes two files to the configured output directory and returns { ok: true, mdPath, jsonPath, savedAt }.
  3. The R1 receives the response within 20 seconds, marks the AAR as exported, and announces completion.
  4. If the skill does not respond within 20 seconds, or if OpenClaw is unreachable, the R1 displays "Export failed — saved locally." The AAR data remains intact in creationStorage.plain and can be re-exported.

export-aar Parameter Shape

{
  aarId:    string,   // unique AAR id (base-36 timestamp + random suffix)
  name:     string,   // human-readable AAR name
  markdown: string,   // rendered Markdown export document (see AAR Format below)
  json:     object,   // structured JSON export (see Response Shape below)
  ts:       number,   // Date.now() at export time
}

Skill Response Shape

Your skill handler should return:

{
  "ok":      true,
  "mdPath":  "/Users/you/debriefs/2026-06-14/op-nightfall-abc123.md",
  "jsonPath": "/Users/you/debriefs/2026-06-14/op-nightfall-abc123.json",
  "savedAt": "2026-06-14T21:34:55.000Z"
}

On error:

{
  "ok":   false,
  "error": "EACCES: permission denied, open '/debriefs/...'"
}

The R1 does not parse these fields — it only checks that a response arrived within the timeout window. The fields are logged by OpenClaw for debugging.


AAR Format

Markdown export structure

# After Action Review — <name>
**Date:** YYYY-MM-DD
**Participants:** Alice, Bob, Charlie

---

## Phase 1: INTENDED
*What was the plan? What did we expect?*

### General
- **Alice** (14:03): Objective was to secure the northern perimeter by 0600.

### Event
- **Bob** (14:05): [T+00:02] Timeline anchor set at mission launch.

## Phase 2: ACTUAL
...

## Phase 3: DELTA
...

## Phase 4: LESSONS
...

## Timeline of Events

- `T+00:00` **Alice:** Mission launch, all elements in position.
- `T+00:14` **Bob:** Northern element reported contact.

## Executive Summary

**Sustains (3)**
- Communication discipline maintained throughout.
- ...

**Improvements (2)**
- React time to contact was 4 minutes, target was 90 seconds.
- ...

---
*Generated by LiveDebrief on R1 · 6/14/2026, 2:34:55 PM*

JSON export structure

{
  "id": "lq8k2x4ab1c",
  "name": "Op Nightfall",
  "date": "2026-06-14",
  "participants": [
    { "name": "Alice" },
    { "name": "Bob" }
  ],
  "phases": [
    {
      "id": "intended",
      "observations": [
        {
          "id": "lq8k2x5de2f",
          "speaker": "Alice",
          "raw": "We planned to secure the northern perimeter by zero six hundred",
          "text": "Planned to secure northern perimeter by 0600.",
          "category": "general",
          "ts": 1718401380000,
          "timelineTs": null
        }
      ]
    },
    { "id": "actual",   "observations": [] },
    { "id": "delta",    "observations": [] },
    { "id": "lessons",  "observations": [] }
  ],
  "timeline": [
    {
      "id": "lq8k2x5de3g",
      "ts": 1718401440000,
      "relMs": 60000,
      "relLabel": "T+01:00",
      "speaker": "Bob",
      "text": "Northern element reported contact"
    }
  ],
  "stats": {
    "totalObservations": 12,
    "sustains": 3,
    "improves": 5,
    "timelineEvents": 4
  },
  "exportedAt": "2026-06-14T21:34:55.000Z"
}

Observation categories

Category Color Meaning
sustain Green What went well and should be repeated
improve Amber What should be done differently
event Blue A factual timeline event (what happened)
root Red A root cause or contributing factor to a gap
general Gray Planning context, intent, or background information

Configuration

Variable Default Description
DEBRIEF_OUTPUT_DIR ~/debriefs Directory where AAR files are written
DEBRIEF_ORG by-date File organization: by-date (subdirectory per date) or flat (all in output dir)

With by-date organization, files are grouped:

~/debriefs/
  2026-06-14/
    op-nightfall-lq8k2x4ab1c.md
    op-nightfall-lq8k2x4ab1c.json
  2026-06-15/
    morning-brief-lq9m3y5cd2d.md
    morning-brief-lq9m3y5cd2d.json

LED Behavior

State LED
Boot, AAR select, participant setup Blue
Active phase — logging observations Green
LLM structuring observation Amber
Facilitator prompt displayed Amber
Recording (STT active) Blue
AAR complete Teal

Controls

Action Screen Result
PTT Boot Go to AAR select
PTT AAR select Open selected AAR
Long-press PTT AAR select Start new AAR
PTT Name AAR Start/stop recording AAR name
Long-press PTT Name AAR Confirm name and proceed
PTT Participants Set active speaker / open add-participant
Long-press PTT Participants Done — enter Phase 1
PTT Phase Start/stop recording observation
Long-press PTT (< 1.2 s) Phase Request phase advance (confirm gate)
Long-press PTT (≥ 1.2 s) Phase Enter timeline mode
PTT Timeline Record timestamped event
Long-press PTT Timeline Exit timeline mode
PTT Facilitator prompt Record answer to prompt
Long-press PTT Facilitator prompt Dismiss prompt
PTT Confirm gate Cancel
Long-press PTT Confirm gate Confirm action
Scroll up/down AAR select, participants, phase feed, timeline Navigate / scroll list
PTT Complete screen Start new AAR
Long-press PTT Complete screen Return to AAR select

Security Note

Debrief content may be sensitive. Stored locally only. Never transmitted off your machine.

  • All AAR data is written to creationStorage.plain (per-Creation isolated storage, not bypassable from other JS contexts).
  • The OpenClaw export writes files only to the directory you configure in .env. No network calls are made by this skill.
  • Do not configure DEBRIEF_OUTPUT_DIR to a path synced by cloud storage (Dropbox, OneDrive, iCloud Drive) if the debrief content is operationally sensitive.
  • Treat the OpenClaw listen port as internal. Confirm your firewall blocks it from external interfaces before use in a sensitive environment.

Troubleshooting

R1 shows "Export failed — saved locally." / red LED

  • Confirm openclaw serve is running on your local machine.
  • Verify both devices are on the same network and the OpenClaw port is reachable.
  • Check openclaw serve terminal output for handler errors from skill.js.
  • The AAR is still fully stored in device storage. Re-exporting is not currently exposed as a standalone action — the data persists and the Creation will attempt the export again if you reload the completed AAR from the AAR select screen.

Files are not appearing in the expected directory

  • Check that DEBRIEF_OUTPUT_DIR in .env resolves to a writable path. Run node -e "console.log(require('os').homedir())" to verify tilde expansion.
  • If DEBRIEF_ORG=by-date, the files appear in a subdirectory named with today's date (YYYY-MM-DD), not directly in the output root.

The skill responds but the file is empty or malformed

  • The markdown and json fields are built entirely on the R1 side. If observations are missing, confirm the LLM structuring step completed before the phase was advanced — the "Structuring entry…" screen must resolve before the observation is stored.

ocReceive times out (20 s)

  • The OpenClaw daemon or skill handler took longer than 20 seconds. This usually indicates a slow or blocked disk write (antivirus scan, network-mounted path). Move DEBRIEF_OUTPUT_DIR to a local SSD path.

Observation shows raw transcript instead of structured text

  • The R1's built-in LLM returned a non-JSON response or the JSON parse failed. The Creation falls back to the first 80 characters of the raw transcript. This is expected behavior when the LLM is under load. The raw transcript is always preserved in the raw field of the JSON export.

Facilitator prompts are not appearing

  • Prompts are suppressed if the LLM did not return a facilitatorPrompt field, if the prompt string is shorter than 10 characters, or if that exact prompt was already shown during the session. Each unique prompt is shown at most once per AAR session.

04 同仓库更多 Skills

ajohnsonnow
ajohnsonnow/r1-warren

whisper-quiz

WhisperQuiz — voice-answered trivia game for Rabbit R1. Generates questions via the R1's built-in LLM, captures spoken answers via STT, and evaluates correctness with flexible LLM-based matching.

★ 0
ajohnsonnow
ajohnsonnow/r1-warren

walk-cast

Walking podcast production skill. Receives raw audio metadata and transcript from the R1, removes long pauses, applies noise reduction and normalization, generates AI chapter titles and episode metadata via LLM, and uploads the final episode to your configured podcast RSS host (Spotify for Podcasters / Anchor, Podbean, Buzzsprout, RSS.com, or any FTP-accessible feed).

★ 0
ajohnsonnow
ajohnsonnow/r1-warren

voice-novel

Local machine skill for VoiceNovel — handles story export to disk.

★ 0
ajohnsonnow
ajohnsonnow/r1-warren

voice-guide

Environment description and navigation assistant for blind and low-vision users. Analyzes camera frames via a vision LLM, reads visible text, identifies scene categories (people/vehicles/food/hazards/exits), performs 360° scans, retrieves GPS position, and sends emergency SMS alerts — all via audio feedback with zero screen dependency.

★ 0
ajohnsonnow
ajohnsonnow/r1-warren

voice-draft

Voice-driven document drafting for the Rabbit R1. Dictate stream-of-consciousness speech; the R1 formats it into Email, Blog Post, Report, Proposal, Meeting Notes, Creative Writing, Code Comments, Cover Letter, or Tweet Thread — then saves the result to a configured folder on your local machine via OpenClaw.

★ 0
ajohnsonnow
ajohnsonnow/r1-warren

voice-commit

Hands-free git and CI workflow voice controller for the Rabbit R1. Speak git/CI commands and hear results — commit, push, status, PR management, test runs, and deploy, all voice-driven via OpenClaw.

★ 0