[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M76Y6FR10MP9CB6RCCYX9CNP":3,"skill-same:01M76Y6FR10MP9CB6RCCYX9CNP":69},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":46,"commitSha":47,"skillDirectory":48,"forksCount":49,"repositoryHtmlUrl":50,"repositoryHomepageUrl":51,"ownerHtmlUrl":52,"ownerBlogUrl":6,"ownerDescription":53,"updatedAt":54,"install":55,"resources":68},"01M76Y6FR10MP9CB6RCCYX9CNP","agent-roundtable",null,"Multi-agent roundtable discussion — topic-driven multi-round debate with convergence detection and conclusion generation","Lord1Egypt\u002Fawesome-skill-forge","Lord1Egypt","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F99746359?v=4","01M7592YDNT0XK9WJ0T2MQVPDN","Mohamed Mounir Ali AL Keshita","01M75GRFNHC8VZNWEB8DKDE433",3,"MIT","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fcommunity\u002Fclawhub\u002Fa\u002Fagent-roundtable\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: agent-roundtable\ndescription: \"Multi-agent roundtable discussion — topic-driven multi-round debate with convergence detection and conclusion generation\"\nversion: 1.2.4\nplatforms: [linux, macos, windows]\nmetadata:\n  hermes:\n    tags: [discussion, multi-agent, collaboration, debate, roundtable]\n    related_skills: [kanban-worker, kanban-orchestrator]\n  openclaw:\n    requires:\n      env: []\n      bins: []\n    primaryEnv: null\n    emoji: \"🤝\"\n    homepage: \"https:\u002F\u002Froundtable.izmw.me\"\n---\n\n# Roundtable Discussion Skill\n\n## Publishing to skill hubs\n\nWhen preparing Roundtable for Hermes Skill Hub or OpenClaw\u002FClawHub, use the release checklist in `src\u002Fskills\u002Freferences\u002Fskill-hub-publishing.md`. Key reminders: keep `src\u002Fskills\u002F` self-contained and free of private team data, include both Hermes and OpenClaw metadata blocks in `SKILL.md`, check `hermes skills publish --help`, `clawhub publish --help`, and `clawhub whoami`, and gate real publishing on user confirmation of target account\u002Frepo.\n\n## Overview\n\nEnable multiple agents to participate in structured, multi-round discussions\naround a topic. Each participant is a **real sub-agent** spawned via\n`delegate_task` — not the main agent role-playing. Each gets its own\nconversation context, model call, and toolset.\n\n**Core value**: Turn \"one agent working alone\" into \"a team having a meeting.\"\n\n## When to Use\n\n- **Tech design review**: product, frontend, backend, architect debate approach\n- **Competitive analysis**: product, marketing, design compare alternatives\n- **Bug root cause analysis**: backend, ops, test triangulate the issue\n- **Product requirements**: product, design, dev align on scope\n- **Architecture decisions**: architect, backend, frontend, devops choose stack\n\n## Prerequisites\n\nEnable the `roundtable` toolset in the profile config:\n\n```yaml\ntoolsets:\n  - roundtable\n```\n\nOr pass `enabled_toolsets: [\"roundtable\"]` when spawning an agent.\n\n## Tools\n\n| Tool | Purpose |\n|------|---------|\n| `roundtable_init` | Create a discussion with topic + participants (+ optional notifications config) |\n| `roundtable_speak` | Record a participant's speech (auto-triggers notifications if configured) |\n| `roundtable_read` | Read discussion history |\n| `roundtable_status` | Check status + convergence metrics |\n| `roundtable_summarize` | Get structured data for conclusion doc (use `compact=true` for smaller output) |\n| `roundtable_end` | Conclude or cancel a discussion (triggers concluded notification) |\n| `roundtable_list` | List all discussions |\n| `roundtable_advance` | Manually advance to next round (use when auto-advance doesn't trigger) |\n| `roundtable_notify` | Manually trigger a notification event for a discussion |\n\n## Execution Model (Important)\n\nEach participant is a **real, independent sub-agent**:\n\n1. Coordinator calls `delegate_task(goal=..., toolsets=[\"roundtable\"])` for each participant\n2. The sub-agent runs its own model inference with its own context window\n3. The sub-agent calls `roundtable_speak(participant=\"{profile}\", ...)` to record its speech\n4. The sub-agent returns a summary to the coordinator\n5. Coordinator sends notification (optional), then delegates to next participant\n\n**This means**: 4 participants × 4 rounds = 16 `delegate_task` calls = 16 independent\nmodel invocations. Expect ~15-20 minutes for a full discussion. Each participant\ntakes 15-60 seconds per round depending on context size and model speed.\n\n## Coordinator Flow\n\n### Step 1: Create the Discussion\n\n```\nroundtable_init(\n    topic=\"Database selection: PostgreSQL vs MySQL vs TiDB\",\n    context=\"Our e-commerce system needs high-concurrency read\u002Fwrite, 1TB+ data\",\n    participants=[\n        {\"profile\": \"bingge\", \"role\": \"Product Director\", \"perspective\": \"Focus on UX\", \"display_name\": \"Bing\"},\n        {\"profile\": \"mafei\", \"role\": \"Tech Lead\", \"perspective\": \"Focus on feasibility\", \"display_name\": \"Fei\"},\n        {\"profile\": \"xiaosu\", \"role\": \"Designer\", \"perspective\": \"Focus on data display\", \"display_name\": \"Su\"},\n    ],\n    created_by=\"coordinator\",  # optional: identifies the discussion creator\n    max_rounds=3,\n    speech_order=\"fixed\",\n    notifications={\n        \"enabled\": True,\n        \"channels\": [\n            {\"platform\": \"feishu\", \"chat_id\": \"oc_xxx\"},\n        ],\n        \"events\": [\"round_start\", \"speech\", \"round_end\", \"concluded\"],\n    }\n)\n→ returns {discussion_id, ...}\n```\n\n**Notifications** (optional): pushes real-time discussion updates to messaging\nchannels. See [Notifications](#notifications) section below.\n\n### Step 2: Opening Statement (Round 0)\n\n```\nroundtable_speak(\n    discussion_id=\"rt_xxxxxxxx\",\n    participant=\"coordinator\",\n    content=\"Today we're discussing database selection...\"\n)\n```\n\nOptionally send opening notification:\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"🔔 Roundtable started [rt_xxx]\\n📋 Topic: ...\\n👥 Participants: ...\")\n```\n\n### Step 3: Multi-Round Discussion\n\nFor each round:\n\n**3a. Coordinator summarizes previous round (optional but recommended):**\n```\nroundtable_speak(\n    discussion_id=\"rt_xxxxxxxx\",\n    participant=\"coordinator\",\n    content=\"Round N summary: key points, emerging consensus, open questions...\"\n)\n```\n\n**3b. Delegate to each participant SEQUENTIALLY (not parallel):**\n\n```python\ndelegate_task(\n    goal=\"You are {display_name}, {role}. Share your viewpoint on this topic from a {role} perspective, then call roundtable_speak to record your speech.\",\n    context=\"\"\"You are participating in a roundtable discussion.\n\n## Discussion Info\n- Topic: {topic}\n- Context: {context}\n- Current Round: Round {N} \u002F {max_rounds}\n- Your Role: {role} ({display_name})\n- Your Perspective: {perspective}\n\n## Discussion History\n{formatted_history}\n\n## Your Task\nFrom your role's perspective, share your viewpoint on this topic.\n- Reference others' points if relevant\n- Keep it 200-500 words\n- Both pros AND cons, with concrete examples\n\nAfter speaking, call roundtable_speak to record your statement:\nroundtable_speak(discussion_id=\"{id}\", participant=\"{profile}\", content=\"your speech\")\"\"\",\n    toolsets=[\"roundtable\"]\n)\n# Wait for completion, then send notification, then delegate to participant 2\n```\n\n**3c. After each participant, send notification (if configured):**\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"💬 Round {N} | {role} ({display_name}) spoke:\\n{summary}\")\n```\n\n**3d. After all participants in a round, send round_end notification:**\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"✅ Round {N} complete\\nConsensus: ...\\nDisagreements: ...\")\n```\n\n**Why sequential, not parallel**: Participants need to read each other's responses to build on them. Parallel delegation means everyone speaks into a void.\n\n### Step 4: Check Convergence\n\nAfter each round:\n```\nroundtable_status(discussion_id=\"rt_xxxxxxxx\")\n→ check convergence_score, consensus_points, disagreement_points\n```\n\n**Note**: Convergence tracking may not work reliably due to the round-tracking bug. The coordinator should manually assess convergence from the discussion content.\n\n### Step 5: Generate Conclusion\n\n```\nsummary = roundtable_summarize(discussion_id=\"rt_xxxxxxxx\")\n```\n\n**⚠️ WARNING**: `roundtable_summarize` returns the ENTIRE discussion as raw JSON — potentially 100KB+. It does NOT generate a summary. The coordinator must:\n1. Read the summary data (use `read_file` with offset\u002Flimit for the persisted output)\n2. Write the conclusion document themselves based on the discussion content\n3. Save to the `output_path` specified during init (or write manually via `write_file`)\n\n**Important**: Write the conclusion doc BEFORE calling `roundtable_end` — the end call only accepts a brief text string, not a full document.\n\n### Step 6: End Discussion\n\n```\nroundtable_end(\n    discussion_id=\"rt_xxxxxxxx\",\n    conclusion=\"Brief text summary of the conclusion\"\n)\n```\n\nSend concluded notification:\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"🏁 Discussion ended\\nConclusion: ...\")\n```\n\n## Participant Prompt Template\n\nWhen delegating to a participant sub-agent, use this template:\n\n```\nYou are participating in a roundtable discussion.\n\n## Discussion Info\n- Topic: {topic}\n- Context: {context}\n- Current Round: Round {current_round} \u002F {max_rounds}\n- Your Role: {role} ({display_name})\n- Your Perspective: {perspective}\n\n## Discussion History\n{formatted_history}\n\n## Your Task\nFrom your role's perspective, share your viewpoint on this topic.\n- You may reference or respond to other participants' statements\n- Keep it concise and powerful, 200-500 words\n- If you agree with a point, explicitly state your agreement\n- If you disagree, explain why and propose alternatives\n\nAfter speaking, call roundtable_speak to record your statement.\n```\n\n## Notifications\n\nPush real-time discussion updates to messaging channels.\n\n### Configuration\n\n```python\nroundtable_init(\n    topic=\"...\",\n    participants=[...],\n    notifications={\n        \"enabled\": True,\n        \"channels\": [\n            {\"platform\": \"feishu\", \"chat_id\": \"oc_xxx\"},\n        ],\n        \"events\": [\"round_start\", \"speech\", \"round_end\", \"concluded\"]  # default: all\n    }\n)\n```\n\n### Events\n\n| Event | Trigger | Content |\n|-------|---------|---------|\n| `round_start` | First speech in a new round | Round number + previous round summary |\n| `speech` | After each participant speaks | Speaker name\u002Frole + content (truncated to 200 chars) |\n| `round_end` | All participants spoke in a round | Key points + convergence score |\n| `concluded` | Discussion ends | Final conclusion + consensus\u002Fdisagreement points |\n\n### send_fn Requirement\n\nThe notification system requires a `send_fn(platform, chat_id, message)` callback\non `RoundtableCore`. The Hermes adapter (`adapters\u002Fhermes.py`) provides\n`_hermes_send_fn` which calls `feishu-send.py` via subprocess.\n\n**If send_fn is not wired**, notifications are silently disabled — the Notifier's\n`enabled` returns `False`. Verify with: `core._send_fn is not None`.\n\n**Manual fallback**: If send_fn is broken or unavailable, the coordinator can\nmanually push notifications using `send_message` after each speech:\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"💬 Round {N} | Speaker: summary...\")\n```\n\n### Timing\n\nWith automatic notifications (send_fn wired), each speech triggers a subprocess\ncall to feishu-send.py (~1-2s overhead). With manual `send_message`, overhead is\nsimilar but requires explicit coordinator action after each speech.\n\n### Verifying Notifications (Pitfall)\n\n**Feishu API returns messages in chronological order by default** (oldest first).\nWhen checking if roundtable notifications arrived in the group chat, you MUST use\n`sort_type=ByCreateTimeDesc` in the API call — otherwise you see old messages and\nconclude (incorrectly) that notifications didn't send.\n\n**Wrong** — default sort returns oldest messages first:\n```python\nresp = requests.get('https:\u002F\u002Fopen.feishu.cn\u002Fopen-apis\u002Fim\u002Fv1\u002Fmessages',\n    params={'container_id': 'oc_xxx', 'page_size': 10})\n```\n\n**Correct** — sort descending to see latest messages:\n```python\nresp = requests.get('https:\u002F\u002Fopen.feishu.cn\u002Fopen-apis\u002Fim\u002Fv1\u002Fmessages',\n    params={'container_id': 'oc_xxx', 'page_size': 10, 'sort_type': 'ByCreateTimeDesc'})\n```\n\n**Debug pattern**: If notifications appear to not fire, add a wrapper `send_fn`\nwith logging to confirm the callback is actually invoked:\n```python\ndef debug_send_fn(platform, chat_id, message):\n    print(f\"[DEBUG SEND] platform={platform}, chat={chat_id}, msg_len={len(message)}\")\n    original_send(platform, chat_id, message)\n    print(f\"[DEBUG SEND] OK\")\n```\n\n**Multiple roundtable.db files**: The system may have multiple databases:\n- `~\u002F.roundtable\u002Froundtable.db` — main agent discussions\n- `~\u002F.hermes\u002Froundtable.db` — hermes tool layer discussions\n- `~\u002F.hermes\u002Fprofiles\u002F{profile}\u002Fhome\u002F.roundtable\u002Froundtable.db` — sub-agent discussions\n\nAlways verify against the correct DB when checking discussion state.\n\n## Web Viewer (default ON)\n\n`run_demo()` has `web: bool = True` by default (changed from `False`). This means demo discussions automatically start a web viewer at `http:\u002F\u002Flocalhost:8199`. The viewer uses PM2 to manage an Express subprocess, fcntl for file locking, and nanoid for token generation.\n\n**Browser auto-open**: Starting a discussion with `web=True` should automatically open the browser via `subprocess.run([\"open\", web_url])` in the Hermes adapter's `_handle_init`. This is an **adapter-level side-effect**, not a core-level one — the core library returns the URL but leaves UX actions to the adapter. The generic adapter intentionally does NOT auto-open (headless environments). If you need to customize browser behavior (e.g., open in a specific browser), modify `_handle_init` in `adapters\u002Fhermes.py`.\n\n**⚠️ Pitfall: Browser opens at END instead of START (2026-05-23)** — Boss reported the browser doesn't open when the discussion begins; it only opens after the discussion concludes (or manually). The `_handle_init` code has the `subprocess.run([\"open\", web_url])` call, but it may not fire reliably in all execution paths. **Diagnosis checklist**: (1) Verify `_handle_init` actually reaches the `subprocess.run` line (add logging) (2) Check if the `web_url` variable is correctly populated from `publisher.start()` return (3) Confirm the `open` command runs in the correct subprocess context (may need `shell=False` with list args on macOS). **See also**: Bug task `t_xxxxxxxx` for the specific fix.\n\n**⚠️ Pitfall: WebViewer real-time updates broken on macOS (2026-05-23)** — Boss reported that the WebViewer doesn't show new speeches in real-time; the browser must be force-refreshed to see updates. **Root cause**: The Express server (`server.mjs`) uses `fs.watch(DISCUSSION_PATH, ...)` to detect `discussion.json` changes and broadcast via SSE. But Python's `WebPublisher` uses atomic write: `write .json.tmp → os.rename()`. On macOS, `fs.watch()` on a file loses tracking after `rename()` replaces the inode — the watcher stays on the old inode, never sees the new file's changes. **Fix**: Change `fs.watch` to watch the **directory** instead of the file, then filter by filename:\n```javascript\n\u002F\u002F Wrong — breaks on atomic rename:\nwatch(DISCUSSION_PATH, () => { ... });\n\n\u002F\u002F Correct — watches directory, catches rename:\nconst dir = require(\"path\").dirname(DISCUSSION_PATH);\nconst filename = require(\"path\").basename(DISCUSSION_PATH);\nwatch(dir, (eventType, changedFilename) => {\n  if (changedFilename !== filename) return;\n  \u002F\u002F ... debounce + broadcast logic unchanged\n});\n```\n**Alternative**: Add server-side polling fallback (`setInterval` + mtime check) as defense-in-depth. **See also**: Bug task `t_xxxxxxxx` for the specific fix.\n\n**⚠️ Direct Core API does NOT auto-open browser**: When creating discussions via Direct Core API (`core.create_discussion()`) instead of the `roundtable_init` tool, the adapter's `_handle_init` is NOT called, so:\n1. The browser does NOT auto-open\n2. The WebViewer URL is NOT printed to the user\n3. The coordinator MUST manually extract `web_url` from the response and: (a) share it with the user, and (b) run `open \"{web_url}\"` to open the browser.\n\n**Pattern**: After `core.create_discussion()` returns, parse `web_url` from the JSON response, then:\n```bash\n# Share URL with user (via send_message or inline)\n# Then open browser\nopen \"http:\u002F\u002F0.0.0.0:{port}\u002Fr\u002F{token}\"\n```\n\n## Install roundtable in Hermes venv (required for agent to import it):\n```bash\n# hermes venv may lack pip — bootstrap it first\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 -m ensurepip\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 -m pip install \u002Fpath\u002Fto\u002Froundtable\n```\n\nNote: system Python 3.9 on macOS is too old (package requires >=3.10). Always install into the hermes venv (Python 3.11+).\n\n**Alternative: uv venv (when hermes venv is missing or broken)**:\n```bash\n# Create a dedicated venv with Python 3.12 via uv\nuv venv \u002Ftmp\u002Froundtable-venv --python 3.12\nuv pip install -e \u002Fpath\u002Fto\u002Froundtable --python \u002Ftmp\u002Froundtable-venv\u002Fbin\u002Fpython3\n\n# Then use this venv for all Direct Core API calls:\n\u002Ftmp\u002Froundtable-venv\u002Fbin\u002Fpython3 \u003C\u003C 'PYEOF'\nimport json, sys\nsys.path.insert(0, \"\u002FUsers\u002Fparsifal\u002F.hermes\u002Fhermes-agent\")\nfrom tools.roundtable_tools import _get_core\ncore = _get_core()\n# ...\nPYEOF\n```\n\nThis pattern is useful when:\n- System Python is 3.9 (macOS default) and roundtable requires >=3.10\n- The hermes venv doesn't exist or lacks pip\n- You need a clean, isolated environment for roundtable\n\n## Convergence Detection\n\nEach round is evaluated for convergence:\n\n| Metric | Formula | Meaning |\n|--------|---------|---------|\n| Consensus | Points multiple participants agree on | Alignment |\n| Disagreement | Points participants disagree on | Conflict |\n| New Point | New topics raised this round | Scope expansion |\n| Score | consensus \u002F (consensus + disagreement) | Overall alignment |\n\n**Termination conditions:**\n- Convergence score > 0.8 → high consensus, wrap up\n- Max rounds reached → prevent infinite discussion\n- Coordinator manually ends → emergency stop\n- All participants vote to end → democratic close\n\n## Conclusion Document Format\n\nThe format below works for general discussions. For **product\u002Fdesign\u002Fdev discussions aimed at producing a buildable specification**, use the decision-oriented format instead — see `src\u002Fskills\u002Freferences\u002Fweb-viewer-discussion-example.md` for the full pattern (MVP scope, tech architecture, acceptance criteria, risk assessment, design deliverables, action items).\n\n```markdown\n# Roundtable Conclusion: [Topic]\n\n## Summary\n- Participants: Product(Bing), Design(Su), Dev(Fei)\n- Rounds: 3\n- Date: 2026-05-20\n\n## Consensus Points\n1. [Point 1]\n2. [Point 2]\n\n## Disagreement Points\n1. [Point 1] - Various perspectives\n\n## Action Items\n1. [ ] [Action 1] - Owner: xxx\n2. [ ] [Action 2] - Owner: xxx\n\n## Detailed Transcript\n### Round 1\n- **Product(Bing)**: ...\n- **Design(Su)**: ...\n- **Dev(Fei)**: ...\n\n### Round 2\n...\n```\n\n## Data Storage\n\n- **Database**: `~\u002F.hermes\u002Froundtable.db` (independent from kanban.db)\n- **Conclusion docs**: Configurable via `output_path`, defaults to project docs dir\n- **ID format**: `rt_` + 8 hex chars (e.g., `rt_a1b2c3d4`)\n\n## Integration with Kanban\n\nDiscussions can be linked to kanban tasks:\n\n```\n# After conclusion, add as task comment\nkanban_comment(task_id=\"t_xxx\", body=\"Roundtable conclusion: {conclusion_path}\")\n```\n\n## Direct Core API Fallback (Last Resort Only)\n\n> **⚠️ IMPORTANT**: The standard flow is ALWAYS `roundtable_init` + `delegate_task(toolsets=[\"roundtable\"])`. Do NOT default to Direct Core API. If sub-agents fail to call `roundtable_speak`, treat it as a **bug to investigate and fix** — not a workaround to codify. The user explicitly rejected the workaround-first mindset: \"不是本质问题吗？一直workaround？\"\n>\n> **Before using Direct Core API, exhaust these options first:**\n> 1. Check if the sub-agent's `enabled_toolsets` actually includes `\"roundtable\"`\n> 2. Check if the sub-agent's profile has the `roundtable` toolset registered\n> 3. Try with a different profile to isolate the issue\n> 4. Report the bug as a kanban task for investigation (bug tracking task: `t_xxxxxxxx`)\n>\n> **Only use Direct Core API when**: (a) the bug is confirmed and (b) you need the discussion to proceed NOW while the fix is in progress.\n\n**Known Issue**: Sub-agents spawned via `delegate_task` with `toolsets=[\"roundtable\"]` sometimes fail to invoke `roundtable_speak` — the sub-agent outputs speech text but `tool_trace` is empty. This is a **bug**, not a design limitation. Root cause is under investigation.\n\n**Direct Core API (emergency bypass only)**: Bypass the Hermes tool layer and call `RoundtableCore` directly:\n\n```bash\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 \u003C\u003C 'PYEOF'\nimport json, sys\nsys.path.insert(0, \"\u002FUsers\u002Fparsifal\u002F.hermes\u002Fhermes-agent\")\nfrom tools.roundtable_tools import _get_core\n\ncore = _get_core()\n\n# Record a speech\nresult = core.speak(\n    discussion_id=\"rt_xxxxxxxx\",\n    participant=\"bingge\",           # must match participants list exactly\n    content='Speech content here'   # use single quotes to avoid heredoc issues\n)\nprint(json.dumps(result, ensure_ascii=False))\n\n# Read discussion history\nhistory = core.read(discussion_id=\"rt_xxxxxxxx\")\nprint(json.dumps(history, ensure_ascii=False, indent=2))\n\n# End discussion\ncore.end_discussion(discussion_id=\"rt_xxxxxxxx\", conclusion=\"Brief conclusion text\")\nPYEOF\n```\n\n**Key gotchas**:\n- **NEVER import `RoundtableCore` directly** — always use `_get_core()` from `tools.roundtable_tools`. Direct import bypasses `send_fn` wiring, so notifications silently fail. Boss explicitly corrected this (2026-05-23): \"你这压根不是用我们开发的圆桌技能做啊\".\n- Use single quotes for content strings containing double quotes (heredoc escaping)\n- `participant` must exactly match the profile name in participants list\n- `coordinator` is always allowed (bypasses participant check)\n- Each `core.speak()` returns `round_complete` and `next_speaker` — use to track progress\n- `core._send_fn` must be non-None for notifications to fire (wired by `_hermes_send_fn`)\n- **Cross-process singleton**: `_get_core()` returns a singleton, BUT each Python invocation is a fresh process, so the singleton only lives for that one script execution. The `_publishers` dict (WebPublisher) is in-memory and lost between invocations. This is handled by the `_update_web_discussion_json` fallback — see \"Cross-Process WebPublisher Data Sync\" section.\n\n**Hybrid Workflow (temporary measure only)**: If `roundtable_speak` is confirmed broken in sub-agents, use `delegate_task` for reasoning and Direct Core API for recording — but treat this as a **temporary workaround** while the bug is being fixed, not a permanent pattern.\n\n```python\n# Step 1: Delegate to sub-agent for thought generation\nresult = delegate_task(\n    goal=\"You are the Product Director. Share your viewpoint on '{topic}' from a product perspective...\",\n    context=\"...\",\n    toolsets=[\"roundtable\"]  # sub-agent will fail to call roundtable_speak — that's expected\n)\n\n# Step 2: Extract speech text from sub-agent's summary\nspeech_text = extract_from_summary(result)  # parse the sub-agent's output\n\n# Step 3: Record via Direct Core API (100% reliable)\nterminal(f\"\"\"\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 \u003C\u003C 'PYEOF'\nimport json, sys\nsys.path.insert(0, \"\u002FUsers\u002Fparsifal\u002F.hermes\u002Fhermes-agent\")\nfrom tools.roundtable_tools import _get_core\ncore = _get_core()\nresult = core.speak(discussion_id=\"{discussion_id}\", participant=\"{profile}\", content='''{speech_text}''')\nprint(json.dumps(result, ensure_ascii=False))\nPYEOF\n\"\"\")\n```\n\n**Key insight**: The sub-agent's reasoning is valuable — don't waste it. Just don't rely on it to persist the data.\n\n**When to use Direct Core API**: Any time reliability matters — demos, verification meetings, production discussions. The direct API is faster (no sub-agent overhead) and 100% reliable.\n\n24. **Quick verification meetings** — When Boss says \"组织一次会议看效果\", use the lightweight pattern: coordinator speaks for all participants directly via `roundtable_speak` (no `delegate_task`). Runs in ~2min vs ~15-20min. Always include notifications config even for demos — the goal is to verify the full pipeline including group chat sync. See `src\u002Fskills\u002Freferences\u002Fquick-verification-example.md` for a working example.\n25. **delegate_task sub-agents can fail to call roundtable_speak** — Observed in 2026-05-23 session (6\u002F6 failures). Sub-agents produce speech text in their response but `tool_trace` is empty. **DO NOT assume this is permanent** — it's a bug under investigation (`t_xxxxxxxx`). Before falling back to Direct Core API, check: (1) sub-agent's `enabled_toolsets` includes `\"roundtable\"`, (2) the profile has roundtable tools registered, (3) try a different profile. If confirmed broken, use Direct Core API as temporary bypass and file\u002Fupdate the bug task.\n26. **Direct Core API is an emergency bypass, NOT the default** — Use `roundtable_init` + `delegate_task` as the primary flow. Direct Core API (`core.speak()`) exists for when the standard flow is confirmed broken AND you need the discussion to proceed now. Always file a bug task when you're forced to use it.\n27. **`web` and `web_port` params on roundtable_init** — The core supports `web=True` to auto-start a WebPublisher HTTP server. The Hermes tool schema now includes these params. Default `web=True` in `_handle_init` so discussions always open WebViewer. Port auto-increments if 8199 is busy (check `web_url` in init response for actual port).\n28. **Adapter handles UX side-effects, core returns data** — When adding features that interact with the user's environment (browser, notifications, file watchers), implement them at the **adapter level** (`adapters\u002Fhermes.py`), not in `core.py`. The core library should be stateless and headless — it returns data (URLs, IDs, status) and the adapter decides what to do with it. Example: browser auto-open lives in `_handle_init` (adapter), not in `create_discussion` (core). The generic adapter may choose different UX behavior or none at all.\n> **Notifications implementation detail**: See `src\u002Fskills\u002Freferences\u002Fnotifications-implementation.md` for architecture, send_fn wiring, and execution flow diagrams. See `src\u002Fskills\u002Freferences\u002Fnotification-debugging.md` for verification pitfalls and debugging checklist.\n\n1. **At least 2 participants required** — A discussion needs multiple viewpoints\n2. **Participant must be registered** — Only profiles listed in `participants` can speak. **Exception**: `participant=\"coordinator\"` is always allowed (bypasses the participant check). This was a bug fix — previously, `roundtable_speak(participant=\"coordinator\")` would fail with \"not an active member\". Coordinator's speech is recorded in round 0 and does NOT affect round advancement logic.\n3. **Round 0 is opening** — Coordinator speaks first, then round 1 begins\n4. **Auto-conclude on max_rounds** — Discussion ends automatically when max rounds exceeded\n5. **Independent database** — roundtable.db is separate from kanban.db; don't mix paths\n6. **No LLM in summarize** — `roundtable_summarize` returns raw JSON data (potentially 100KB+), NOT a summary. The coordinator must manually write the conclusion document. Use `read_file` with offset\u002Flimit to process the persisted output.\n7. **connect() expects Path, not str** — Use `connect(Path(\"~\u002F.hermes\u002Froundtable.db\"))`, not `connect(\"\u002FUsers\u002F...\")`\n8. **delegate_task needs roundtable toolset** — Sub-agents spawned via `delegate_task` must have `enabled_toolsets=['roundtable']` to use roundtable tools\n9. **Round tracking is buggy** — All speeches may end up as Round 0 regardless of actual round. The `round` field in speak responses may not reflect reality. Workaround: track rounds manually in the coordinator's logic.\n10. **No `roundtable_advance` tool** — There is no tool to explicitly advance rounds. Rounds are supposed to auto-advance when all participants have spoken, but due to the round-tracking bug this may not work. The coordinator should track round progress manually.\n11. **Sequential delegation, not parallel** — Delegate to participants one at a time so they can read each other's responses. Parallel delegation means everyone speaks without context.\n12. **Coordinator summarizes between rounds** — After each round of participant speeches, the coordinator should call `roundtable_speak(participant=\"coordinator\")` to summarize before delegating the next round. This gives participants context for their next response.\n13. **summarize output is persisted as file** — When `roundtable_summarize` returns >8K chars, the full output is saved to a temp file. Use `read_file(path, offset, limit)` to process it in chunks.\n14. **Hermes adapter send_fn must be wired** — The `adapters\u002Fhermes.py` creates `RoundtableCore()` without passing a `send_fn`, so notifications configured in `roundtable_init` will silently not send. Fix: implement `_hermes_send_fn(platform, chat_id, message)` that calls `feishu-send.py`, then pass it to `RoundtableCore(send_fn=_hermes_send_fn)`. Default profile should be `\"default\"` (not hardcoded to any specific profile). Without this fix, notifications are metadata only.\n15. **Participants are real sub-agents, not role-play** — Each participant is a separate `delegate_task()` call that spawns an independent sub-agent with its own context window, model inference, and tool access. This is why a 4-round, 4-participant discussion takes 15-20 minutes (16+ independent model calls). The coordinator drives rounds sequentially so participants can read each other's responses.\n16. **Generic adapter lacks notifications** — `adapters\u002Fgeneric.py` does not support `notifications` config or `send_fn`. Non-Hermes users must manually handle event dispatch. If releasing as standalone library, the generic adapter needs `send_fn` parameter support and `advance()` + `notify()` methods.\n14. **send_fn must be wired in adapter** — The `Notifier` requires a `send_fn` callback to actually send messages. In `adapters\u002Fhermes.py`, `_get_core()` must pass `send_fn=_hermes_send_fn` to `RoundtableCore()`. Without this, notifications config is silently ignored (Notifier.enabled returns False). If notifications aren't firing, check that `_hermes_send_fn` is implemented and passed to the core. The send_fn uses `subprocess.run([\"python3\", script, profile, chat_id, message])` to call `feishu-send.py`.\n15. **HERMES_PROFILE env var** — The send_fn reads `os.environ.get(\"HERMES_PROFILE\", \"default\")` to determine which profile's Bot sends the notification. Default is `\"default\"` (coordinator). This is set automatically by the Hermes runtime — don't hardcode it.\n16. **Real sub-agents, not role-playing** — Each participant in a roundtable discussion is a separate `delegate_task` call spawning an independent sub-agent with its own context window and model call. This is why 4 participants × 4 rounds = 16 delegate_task calls and ~15-20 minutes of runtime. If it were just the main agent role-playing, it would take seconds.\n17. **Notifications don't block** — All notification failures are caught and logged. A broken send_fn never breaks the discussion flow. This is by design.\n18. **Sub-agent context window** — Each participant sub-agent only sees what the coordinator passes in `context`. If you omit discussion history, the participant speaks \"into a void\". Always include `formatted_history` from `roundtable_read`.\n19. **Write conclusion doc BEFORE roundtable_end** — `roundtable_end` only accepts a brief `conclusion` text string (1-2 sentences). The full conclusion document (with action items, tables, detailed transcript) must be written as a separate file via `write_file`. Do this BEFORE calling `roundtable_end` so you can still use `roundtable_summarize` if needed.\n20. **Round numbering mismatch** — The `roundtable_summarize` output labels speeches as \"Round 0\", \"Round 1\", \"Round 2\" based on the DB's internal round counter. The coordinator's mental model (Round 1 = first participant round) may differ from the DB's (Round 0 = opening, Round 1 = first participant round). Always check the actual round field in the data rather than assuming.\n21. **History reading cost grows linearly** — `roundtable_read` returns ALL speeches. In Round 3 of a 3-participant discussion, the first participant's `delegate_task` took 165s because the context included 9 full speeches. For longer discussions (4+ rounds, 4+ participants), consider summarizing previous rounds in the coordinator's `roundtable_speak` between rounds to keep context manageable.\n22. **No convergence metrics in practice** — `roundtable_summarize` returns `final_convergence_score: null` and empty `consensus_points`\u002F`disagreement_points`. The convergence detection is not implemented in the core engine. The coordinator MUST manually assess convergence from discussion content — don't rely on the status\u002Fsummarize tools for this.\n23. **Coordinator MUST pass notifications config** — Even though `send_fn` is wired in the adapter, `notifications` is an OPTIONAL parameter on `roundtable_init`. If you don't pass it, no events fire and no messages reach the group chat — silently. This is the #1 cause of \"why didn't my discussion sync to the group?\" Always include notifications when the discussion should be visible to the team. For OPC team, the standard config is: `notifications={\"enabled\": True, \"channels\": [{\"platform\": \"feishu\", \"chat_id\": \"oc_your_company_group_id\"}], \"events\": [\"round_start\", \"speech\", \"round_end\", \"concluded\"]}`. Without this explicit config, discussions run in \"silent mode\" — useful for testing, but almost never what you want in production.\n\n## Cross-Process WebPublisher Data Sync (Critical Pitfall)\n\n**Problem**: Hermes tool calls (`roundtable_init`, `roundtable_speak`, `roundtable_end`) each run in a **separate Python process**. The `_publishers` dict in `RoundtableCore` is in-memory — it's populated during `create_discussion()` in process A, but `speak()` runs in process B where `_publishers` is empty. Result: `publisher.on_speech()` is never called, `discussion.json` stays at `\"speeches\": []`, and the WebViewer shows nothing.\n\n**Fix** (applied 2026-05-23): Two cross-process fallback methods in `core.py`:\n- `_update_web_discussion_json()` — called in `speak()` when publisher not in memory; reads\u002Fwrites `discussion.json` directly with fcntl locking\n- `_conclude_web_discussion()` — called in `end_discussion()` when publisher not in memory; sets status=\"concluded\" and writes conclusion\n\nThese methods find the JSON file at `\u002Ftmp\u002Froundtable_web\u002F{discussion_id}\u002Fdiscussion.json` and update it with the same atomic write pattern (flock → write .tmp → fsync → rename) that WebPublisher uses.\n\n**Also fixed**: `publisher.conclude()` was called without the conclusion string argument — fixed to pass `disc_after.conclusion or \"\"`.\n\n**Verification**: Tested with fully independent processes (create in process 1, 6 speeches each in separate processes, end in process 8). All 6 speeches appeared in WebViewer, conclusion displayed correctly.\n\n**Impact**: Without this fix, the WebViewer is essentially non-functional in real Hermes usage since every tool call is a fresh process.\n\n## Test Results\n\nSee `src\u002Fskills\u002Freferences\u002Ftest-results-2026-05-20.md` for the first functional test results, including bugs found and product acceptance report.\n\n## Open-Source Release\n\nSee `src\u002Fskills\u002Freferences\u002Fopen-source-readiness.md` for the pre-release checklist (LICENSE, cleanup, adapter gaps, test isolation).\n\n## Working Examples\n\n- `src\u002Fskills\u002Freferences\u002Fopc-experience-discussion-example.md` — 4-round, 4-participant discussion with timing data and workflow\n- `src\u002Fskills\u002Freferences\u002Fnotifications-example.md` — roundtable with real-time push notifications to Feishu\n- `src\u002Fskills\u002Freferences\u002Frelease-planning-discussion.md` — 3-round product\u002Fdesign\u002Fdev discussion for open-source release planning\n- `src\u002Fskills\u002Freferences\u002Fai-relay-open-source-discussion.md` — 3-round discussion with hybrid workflow (delegate_task + Direct Core API), notifications, and conclusion doc → 5\u002F29 release plan\n- `src\u002Fskills\u002Freferences\u002Fweb-viewer-discussion-example.md` — Decision-oriented conclusion doc pattern: MVP scope, tech architecture, acceptance criteria, risk assessment, design deliverables. Use this format when the discussion goal is to produce a buildable specification.\n- `src\u002Fskills\u002Freferences\u002Fpost-discussion-kanban-dispatch.md` — After discussion concludes, create kanban tasks grouped by owner, subscribe notifications, and dispatch to team via Feishu groups.\n\n## Open-Source Release\n\nSee `src\u002Fskills\u002Freferences\u002Fopen-source-readiness-checklist.md` for the pre-release audit: missing LICENSE, Hermes-specific files to separate, build-backend fix, .gitignore, internal docs to remove, generic adapter gaps, and target package structure.\n# Test line\n","# Roundtable Discussion Skill\n\n## Publishing to skill hubs\n\nWhen preparing Roundtable for Hermes Skill Hub or OpenClaw\u002FClawHub, use the release checklist in `src\u002Fskills\u002Freferences\u002Fskill-hub-publishing.md`. Key reminders: keep `src\u002Fskills\u002F` self-contained and free of private team data, include both Hermes and OpenClaw metadata blocks in `SKILL.md`, check `hermes skills publish --help`, `clawhub publish --help`, and `clawhub whoami`, and gate real publishing on user confirmation of target account\u002Frepo.\n\n## Overview\n\nEnable multiple agents to participate in structured, multi-round discussions\naround a topic. Each participant is a **real sub-agent** spawned via\n`delegate_task` — not the main agent role-playing. Each gets its own\nconversation context, model call, and toolset.\n\n**Core value**: Turn \"one agent working alone\" into \"a team having a meeting.\"\n\n## When to Use\n\n- **Tech design review**: product, frontend, backend, architect debate approach\n- **Competitive analysis**: product, marketing, design compare alternatives\n- **Bug root cause analysis**: backend, ops, test triangulate the issue\n- **Product requirements**: product, design, dev align on scope\n- **Architecture decisions**: architect, backend, frontend, devops choose stack\n\n## Prerequisites\n\nEnable the `roundtable` toolset in the profile config:\n\n```yaml\ntoolsets:\n  - roundtable\n```\n\nOr pass `enabled_toolsets: [\"roundtable\"]` when spawning an agent.\n\n## Tools\n\n| Tool | Purpose |\n|------|---------|\n| `roundtable_init` | Create a discussion with topic + participants (+ optional notifications config) |\n| `roundtable_speak` | Record a participant's speech (auto-triggers notifications if configured) |\n| `roundtable_read` | Read discussion history |\n| `roundtable_status` | Check status + convergence metrics |\n| `roundtable_summarize` | Get structured data for conclusion doc (use `compact=true` for smaller output) |\n| `roundtable_end` | Conclude or cancel a discussion (triggers concluded notification) |\n| `roundtable_list` | List all discussions |\n| `roundtable_advance` | Manually advance to next round (use when auto-advance doesn't trigger) |\n| `roundtable_notify` | Manually trigger a notification event for a discussion |\n\n## Execution Model (Important)\n\nEach participant is a **real, independent sub-agent**:\n\n1. Coordinator calls `delegate_task(goal=..., toolsets=[\"roundtable\"])` for each participant\n2. The sub-agent runs its own model inference with its own context window\n3. The sub-agent calls `roundtable_speak(participant=\"{profile}\", ...)` to record its speech\n4. The sub-agent returns a summary to the coordinator\n5. Coordinator sends notification (optional), then delegates to next participant\n\n**This means**: 4 participants × 4 rounds = 16 `delegate_task` calls = 16 independent\nmodel invocations. Expect ~15-20 minutes for a full discussion. Each participant\ntakes 15-60 seconds per round depending on context size and model speed.\n\n## Coordinator Flow\n\n### Step 1: Create the Discussion\n\n```\nroundtable_init(\n    topic=\"Database selection: PostgreSQL vs MySQL vs TiDB\",\n    context=\"Our e-commerce system needs high-concurrency read\u002Fwrite, 1TB+ data\",\n    participants=[\n        {\"profile\": \"bingge\", \"role\": \"Product Director\", \"perspective\": \"Focus on UX\", \"display_name\": \"Bing\"},\n        {\"profile\": \"mafei\", \"role\": \"Tech Lead\", \"perspective\": \"Focus on feasibility\", \"display_name\": \"Fei\"},\n        {\"profile\": \"xiaosu\", \"role\": \"Designer\", \"perspective\": \"Focus on data display\", \"display_name\": \"Su\"},\n    ],\n    created_by=\"coordinator\",  # optional: identifies the discussion creator\n    max_rounds=3,\n    speech_order=\"fixed\",\n    notifications={\n        \"enabled\": True,\n        \"channels\": [\n            {\"platform\": \"feishu\", \"chat_id\": \"oc_xxx\"},\n        ],\n        \"events\": [\"round_start\", \"speech\", \"round_end\", \"concluded\"],\n    }\n)\n→ returns {discussion_id, ...}\n```\n\n**Notifications** (optional): pushes real-time discussion updates to messaging\nchannels. See [Notifications](#notifications) section below.\n\n### Step 2: Opening Statement (Round 0)\n\n```\nroundtable_speak(\n    discussion_id=\"rt_xxxxxxxx\",\n    participant=\"coordinator\",\n    content=\"Today we're discussing database selection...\"\n)\n```\n\nOptionally send opening notification:\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"🔔 Roundtable started [rt_xxx]\\n📋 Topic: ...\\n👥 Participants: ...\")\n```\n\n### Step 3: Multi-Round Discussion\n\nFor each round:\n\n**3a. Coordinator summarizes previous round (optional but recommended):**\n```\nroundtable_speak(\n    discussion_id=\"rt_xxxxxxxx\",\n    participant=\"coordinator\",\n    content=\"Round N summary: key points, emerging consensus, open questions...\"\n)\n```\n\n**3b. Delegate to each participant SEQUENTIALLY (not parallel):**\n\n```python\ndelegate_task(\n    goal=\"You are {display_name}, {role}. Share your viewpoint on this topic from a {role} perspective, then call roundtable_speak to record your speech.\",\n    context=\"\"\"You are participating in a roundtable discussion.\n\n## Discussion Info\n- Topic: {topic}\n- Context: {context}\n- Current Round: Round {N} \u002F {max_rounds}\n- Your Role: {role} ({display_name})\n- Your Perspective: {perspective}\n\n## Discussion History\n{formatted_history}\n\n## Your Task\nFrom your role's perspective, share your viewpoint on this topic.\n- Reference others' points if relevant\n- Keep it 200-500 words\n- Both pros AND cons, with concrete examples\n\nAfter speaking, call roundtable_speak to record your statement:\nroundtable_speak(discussion_id=\"{id}\", participant=\"{profile}\", content=\"your speech\")\"\"\",\n    toolsets=[\"roundtable\"]\n)\n# Wait for completion, then send notification, then delegate to participant 2\n```\n\n**3c. After each participant, send notification (if configured):**\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"💬 Round {N} | {role} ({display_name}) spoke:\\n{summary}\")\n```\n\n**3d. After all participants in a round, send round_end notification:**\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"✅ Round {N} complete\\nConsensus: ...\\nDisagreements: ...\")\n```\n\n**Why sequential, not parallel**: Participants need to read each other's responses to build on them. Parallel delegation means everyone speaks into a void.\n\n### Step 4: Check Convergence\n\nAfter each round:\n```\nroundtable_status(discussion_id=\"rt_xxxxxxxx\")\n→ check convergence_score, consensus_points, disagreement_points\n```\n\n**Note**: Convergence tracking may not work reliably due to the round-tracking bug. The coordinator should manually assess convergence from the discussion content.\n\n### Step 5: Generate Conclusion\n\n```\nsummary = roundtable_summarize(discussion_id=\"rt_xxxxxxxx\")\n```\n\n**⚠️ WARNING**: `roundtable_summarize` returns the ENTIRE discussion as raw JSON — potentially 100KB+. It does NOT generate a summary. The coordinator must:\n1. Read the summary data (use `read_file` with offset\u002Flimit for the persisted output)\n2. Write the conclusion document themselves based on the discussion content\n3. Save to the `output_path` specified during init (or write manually via `write_file`)\n\n**Important**: Write the conclusion doc BEFORE calling `roundtable_end` — the end call only accepts a brief text string, not a full document.\n\n### Step 6: End Discussion\n\n```\nroundtable_end(\n    discussion_id=\"rt_xxxxxxxx\",\n    conclusion=\"Brief text summary of the conclusion\"\n)\n```\n\nSend concluded notification:\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"🏁 Discussion ended\\nConclusion: ...\")\n```\n\n## Participant Prompt Template\n\nWhen delegating to a participant sub-agent, use this template:\n\n```\nYou are participating in a roundtable discussion.\n\n## Discussion Info\n- Topic: {topic}\n- Context: {context}\n- Current Round: Round {current_round} \u002F {max_rounds}\n- Your Role: {role} ({display_name})\n- Your Perspective: {perspective}\n\n## Discussion History\n{formatted_history}\n\n## Your Task\nFrom your role's perspective, share your viewpoint on this topic.\n- You may reference or respond to other participants' statements\n- Keep it concise and powerful, 200-500 words\n- If you agree with a point, explicitly state your agreement\n- If you disagree, explain why and propose alternatives\n\nAfter speaking, call roundtable_speak to record your statement.\n```\n\n## Notifications\n\nPush real-time discussion updates to messaging channels.\n\n### Configuration\n\n```python\nroundtable_init(\n    topic=\"...\",\n    participants=[...],\n    notifications={\n        \"enabled\": True,\n        \"channels\": [\n            {\"platform\": \"feishu\", \"chat_id\": \"oc_xxx\"},\n        ],\n        \"events\": [\"round_start\", \"speech\", \"round_end\", \"concluded\"]  # default: all\n    }\n)\n```\n\n### Events\n\n| Event | Trigger | Content |\n|-------|---------|---------|\n| `round_start` | First speech in a new round | Round number + previous round summary |\n| `speech` | After each participant speaks | Speaker name\u002Frole + content (truncated to 200 chars) |\n| `round_end` | All participants spoke in a round | Key points + convergence score |\n| `concluded` | Discussion ends | Final conclusion + consensus\u002Fdisagreement points |\n\n### send_fn Requirement\n\nThe notification system requires a `send_fn(platform, chat_id, message)` callback\non `RoundtableCore`. The Hermes adapter (`adapters\u002Fhermes.py`) provides\n`_hermes_send_fn` which calls `feishu-send.py` via subprocess.\n\n**If send_fn is not wired**, notifications are silently disabled — the Notifier's\n`enabled` returns `False`. Verify with: `core._send_fn is not None`.\n\n**Manual fallback**: If send_fn is broken or unavailable, the coordinator can\nmanually push notifications using `send_message` after each speech:\n```\nsend_message(target=\"feishu:oc_xxx\", message=\"💬 Round {N} | Speaker: summary...\")\n```\n\n### Timing\n\nWith automatic notifications (send_fn wired), each speech triggers a subprocess\ncall to feishu-send.py (~1-2s overhead). With manual `send_message`, overhead is\nsimilar but requires explicit coordinator action after each speech.\n\n### Verifying Notifications (Pitfall)\n\n**Feishu API returns messages in chronological order by default** (oldest first).\nWhen checking if roundtable notifications arrived in the group chat, you MUST use\n`sort_type=ByCreateTimeDesc` in the API call — otherwise you see old messages and\nconclude (incorrectly) that notifications didn't send.\n\n**Wrong** — default sort returns oldest messages first:\n```python\nresp = requests.get('https:\u002F\u002Fopen.feishu.cn\u002Fopen-apis\u002Fim\u002Fv1\u002Fmessages',\n    params={'container_id': 'oc_xxx', 'page_size': 10})\n```\n\n**Correct** — sort descending to see latest messages:\n```python\nresp = requests.get('https:\u002F\u002Fopen.feishu.cn\u002Fopen-apis\u002Fim\u002Fv1\u002Fmessages',\n    params={'container_id': 'oc_xxx', 'page_size': 10, 'sort_type': 'ByCreateTimeDesc'})\n```\n\n**Debug pattern**: If notifications appear to not fire, add a wrapper `send_fn`\nwith logging to confirm the callback is actually invoked:\n```python\ndef debug_send_fn(platform, chat_id, message):\n    print(f\"[DEBUG SEND] platform={platform}, chat={chat_id}, msg_len={len(message)}\")\n    original_send(platform, chat_id, message)\n    print(f\"[DEBUG SEND] OK\")\n```\n\n**Multiple roundtable.db files**: The system may have multiple databases:\n- `~\u002F.roundtable\u002Froundtable.db` — main agent discussions\n- `~\u002F.hermes\u002Froundtable.db` — hermes tool layer discussions\n- `~\u002F.hermes\u002Fprofiles\u002F{profile}\u002Fhome\u002F.roundtable\u002Froundtable.db` — sub-agent discussions\n\nAlways verify against the correct DB when checking discussion state.\n\n## Web Viewer (default ON)\n\n`run_demo()` has `web: bool = True` by default (changed from `False`). This means demo discussions automatically start a web viewer at `http:\u002F\u002Flocalhost:8199`. The viewer uses PM2 to manage an Express subprocess, fcntl for file locking, and nanoid for token generation.\n\n**Browser auto-open**: Starting a discussion with `web=True` should automatically open the browser via `subprocess.run([\"open\", web_url])` in the Hermes adapter's `_handle_init`. This is an **adapter-level side-effect**, not a core-level one — the core library returns the URL but leaves UX actions to the adapter. The generic adapter intentionally does NOT auto-open (headless environments). If you need to customize browser behavior (e.g., open in a specific browser), modify `_handle_init` in `adapters\u002Fhermes.py`.\n\n**⚠️ Pitfall: Browser opens at END instead of START (2026-05-23)** — Boss reported the browser doesn't open when the discussion begins; it only opens after the discussion concludes (or manually). The `_handle_init` code has the `subprocess.run([\"open\", web_url])` call, but it may not fire reliably in all execution paths. **Diagnosis checklist**: (1) Verify `_handle_init` actually reaches the `subprocess.run` line (add logging) (2) Check if the `web_url` variable is correctly populated from `publisher.start()` return (3) Confirm the `open` command runs in the correct subprocess context (may need `shell=False` with list args on macOS). **See also**: Bug task `t_xxxxxxxx` for the specific fix.\n\n**⚠️ Pitfall: WebViewer real-time updates broken on macOS (2026-05-23)** — Boss reported that the WebViewer doesn't show new speeches in real-time; the browser must be force-refreshed to see updates. **Root cause**: The Express server (`server.mjs`) uses `fs.watch(DISCUSSION_PATH, ...)` to detect `discussion.json` changes and broadcast via SSE. But Python's `WebPublisher` uses atomic write: `write .json.tmp → os.rename()`. On macOS, `fs.watch()` on a file loses tracking after `rename()` replaces the inode — the watcher stays on the old inode, never sees the new file's changes. **Fix**: Change `fs.watch` to watch the **directory** instead of the file, then filter by filename:\n```javascript\n\u002F\u002F Wrong — breaks on atomic rename:\nwatch(DISCUSSION_PATH, () => { ... });\n\n\u002F\u002F Correct — watches directory, catches rename:\nconst dir = require(\"path\").dirname(DISCUSSION_PATH);\nconst filename = require(\"path\").basename(DISCUSSION_PATH);\nwatch(dir, (eventType, changedFilename) => {\n  if (changedFilename !== filename) return;\n  \u002F\u002F ... debounce + broadcast logic unchanged\n});\n```\n**Alternative**: Add server-side polling fallback (`setInterval` + mtime check) as defense-in-depth. **See also**: Bug task `t_xxxxxxxx` for the specific fix.\n\n**⚠️ Direct Core API does NOT auto-open browser**: When creating discussions via Direct Core API (`core.create_discussion()`) instead of the `roundtable_init` tool, the adapter's `_handle_init` is NOT called, so:\n1. The browser does NOT auto-open\n2. The WebViewer URL is NOT printed to the user\n3. The coordinator MUST manually extract `web_url` from the response and: (a) share it with the user, and (b) run `open \"{web_url}\"` to open the browser.\n\n**Pattern**: After `core.create_discussion()` returns, parse `web_url` from the JSON response, then:\n```bash\n# Share URL with user (via send_message or inline)\n# Then open browser\nopen \"http:\u002F\u002F0.0.0.0:{port}\u002Fr\u002F{token}\"\n```\n\n## Install roundtable in Hermes venv (required for agent to import it):\n```bash\n# hermes venv may lack pip — bootstrap it first\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 -m ensurepip\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 -m pip install \u002Fpath\u002Fto\u002Froundtable\n```\n\nNote: system Python 3.9 on macOS is too old (package requires >=3.10). Always install into the hermes venv (Python 3.11+).\n\n**Alternative: uv venv (when hermes venv is missing or broken)**:\n```bash\n# Create a dedicated venv with Python 3.12 via uv\nuv venv \u002Ftmp\u002Froundtable-venv --python 3.12\nuv pip install -e \u002Fpath\u002Fto\u002Froundtable --python \u002Ftmp\u002Froundtable-venv\u002Fbin\u002Fpython3\n\n# Then use this venv for all Direct Core API calls:\n\u002Ftmp\u002Froundtable-venv\u002Fbin\u002Fpython3 \u003C\u003C 'PYEOF'\nimport json, sys\nsys.path.insert(0, \"\u002FUsers\u002Fparsifal\u002F.hermes\u002Fhermes-agent\")\nfrom tools.roundtable_tools import _get_core\ncore = _get_core()\n# ...\nPYEOF\n```\n\nThis pattern is useful when:\n- System Python is 3.9 (macOS default) and roundtable requires >=3.10\n- The hermes venv doesn't exist or lacks pip\n- You need a clean, isolated environment for roundtable\n\n## Convergence Detection\n\nEach round is evaluated for convergence:\n\n| Metric | Formula | Meaning |\n|--------|---------|---------|\n| Consensus | Points multiple participants agree on | Alignment |\n| Disagreement | Points participants disagree on | Conflict |\n| New Point | New topics raised this round | Scope expansion |\n| Score | consensus \u002F (consensus + disagreement) | Overall alignment |\n\n**Termination conditions:**\n- Convergence score > 0.8 → high consensus, wrap up\n- Max rounds reached → prevent infinite discussion\n- Coordinator manually ends → emergency stop\n- All participants vote to end → democratic close\n\n## Conclusion Document Format\n\nThe format below works for general discussions. For **product\u002Fdesign\u002Fdev discussions aimed at producing a buildable specification**, use the decision-oriented format instead — see `src\u002Fskills\u002Freferences\u002Fweb-viewer-discussion-example.md` for the full pattern (MVP scope, tech architecture, acceptance criteria, risk assessment, design deliverables, action items).\n\n```markdown\n# Roundtable Conclusion: [Topic]\n\n## Summary\n- Participants: Product(Bing), Design(Su), Dev(Fei)\n- Rounds: 3\n- Date: 2026-05-20\n\n## Consensus Points\n1. [Point 1]\n2. [Point 2]\n\n## Disagreement Points\n1. [Point 1] - Various perspectives\n\n## Action Items\n1. [ ] [Action 1] - Owner: xxx\n2. [ ] [Action 2] - Owner: xxx\n\n## Detailed Transcript\n### Round 1\n- **Product(Bing)**: ...\n- **Design(Su)**: ...\n- **Dev(Fei)**: ...\n\n### Round 2\n...\n```\n\n## Data Storage\n\n- **Database**: `~\u002F.hermes\u002Froundtable.db` (independent from kanban.db)\n- **Conclusion docs**: Configurable via `output_path`, defaults to project docs dir\n- **ID format**: `rt_` + 8 hex chars (e.g., `rt_a1b2c3d4`)\n\n## Integration with Kanban\n\nDiscussions can be linked to kanban tasks:\n\n```\n# After conclusion, add as task comment\nkanban_comment(task_id=\"t_xxx\", body=\"Roundtable conclusion: {conclusion_path}\")\n```\n\n## Direct Core API Fallback (Last Resort Only)\n\n> **⚠️ IMPORTANT**: The standard flow is ALWAYS `roundtable_init` + `delegate_task(toolsets=[\"roundtable\"])`. Do NOT default to Direct Core API. If sub-agents fail to call `roundtable_speak`, treat it as a **bug to investigate and fix** — not a workaround to codify. The user explicitly rejected the workaround-first mindset: \"不是本质问题吗？一直workaround？\"\n>\n> **Before using Direct Core API, exhaust these options first:**\n> 1. Check if the sub-agent's `enabled_toolsets` actually includes `\"roundtable\"`\n> 2. Check if the sub-agent's profile has the `roundtable` toolset registered\n> 3. Try with a different profile to isolate the issue\n> 4. Report the bug as a kanban task for investigation (bug tracking task: `t_xxxxxxxx`)\n>\n> **Only use Direct Core API when**: (a) the bug is confirmed and (b) you need the discussion to proceed NOW while the fix is in progress.\n\n**Known Issue**: Sub-agents spawned via `delegate_task` with `toolsets=[\"roundtable\"]` sometimes fail to invoke `roundtable_speak` — the sub-agent outputs speech text but `tool_trace` is empty. This is a **bug**, not a design limitation. Root cause is under investigation.\n\n**Direct Core API (emergency bypass only)**: Bypass the Hermes tool layer and call `RoundtableCore` directly:\n\n```bash\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 \u003C\u003C 'PYEOF'\nimport json, sys\nsys.path.insert(0, \"\u002FUsers\u002Fparsifal\u002F.hermes\u002Fhermes-agent\")\nfrom tools.roundtable_tools import _get_core\n\ncore = _get_core()\n\n# Record a speech\nresult = core.speak(\n    discussion_id=\"rt_xxxxxxxx\",\n    participant=\"bingge\",           # must match participants list exactly\n    content='Speech content here'   # use single quotes to avoid heredoc issues\n)\nprint(json.dumps(result, ensure_ascii=False))\n\n# Read discussion history\nhistory = core.read(discussion_id=\"rt_xxxxxxxx\")\nprint(json.dumps(history, ensure_ascii=False, indent=2))\n\n# End discussion\ncore.end_discussion(discussion_id=\"rt_xxxxxxxx\", conclusion=\"Brief conclusion text\")\nPYEOF\n```\n\n**Key gotchas**:\n- **NEVER import `RoundtableCore` directly** — always use `_get_core()` from `tools.roundtable_tools`. Direct import bypasses `send_fn` wiring, so notifications silently fail. Boss explicitly corrected this (2026-05-23): \"你这压根不是用我们开发的圆桌技能做啊\".\n- Use single quotes for content strings containing double quotes (heredoc escaping)\n- `participant` must exactly match the profile name in participants list\n- `coordinator` is always allowed (bypasses participant check)\n- Each `core.speak()` returns `round_complete` and `next_speaker` — use to track progress\n- `core._send_fn` must be non-None for notifications to fire (wired by `_hermes_send_fn`)\n- **Cross-process singleton**: `_get_core()` returns a singleton, BUT each Python invocation is a fresh process, so the singleton only lives for that one script execution. The `_publishers` dict (WebPublisher) is in-memory and lost between invocations. This is handled by the `_update_web_discussion_json` fallback — see \"Cross-Process WebPublisher Data Sync\" section.\n\n**Hybrid Workflow (temporary measure only)**: If `roundtable_speak` is confirmed broken in sub-agents, use `delegate_task` for reasoning and Direct Core API for recording — but treat this as a **temporary workaround** while the bug is being fixed, not a permanent pattern.\n\n```python\n# Step 1: Delegate to sub-agent for thought generation\nresult = delegate_task(\n    goal=\"You are the Product Director. Share your viewpoint on '{topic}' from a product perspective...\",\n    context=\"...\",\n    toolsets=[\"roundtable\"]  # sub-agent will fail to call roundtable_speak — that's expected\n)\n\n# Step 2: Extract speech text from sub-agent's summary\nspeech_text = extract_from_summary(result)  # parse the sub-agent's output\n\n# Step 3: Record via Direct Core API (100% reliable)\nterminal(f\"\"\"\n~\u002F.hermes\u002Fhermes-agent\u002Fvenv\u002Fbin\u002Fpython3 \u003C\u003C 'PYEOF'\nimport json, sys\nsys.path.insert(0, \"\u002FUsers\u002Fparsifal\u002F.hermes\u002Fhermes-agent\")\nfrom tools.roundtable_tools import _get_core\ncore = _get_core()\nresult = core.speak(discussion_id=\"{discussion_id}\", participant=\"{profile}\", content='''{speech_text}''')\nprint(json.dumps(result, ensure_ascii=False))\nPYEOF\n\"\"\")\n```\n\n**Key insight**: The sub-agent's reasoning is valuable — don't waste it. Just don't rely on it to persist the data.\n\n**When to use Direct Core API**: Any time reliability matters — demos, verification meetings, production discussions. The direct API is faster (no sub-agent overhead) and 100% reliable.\n\n24. **Quick verification meetings** — When Boss says \"组织一次会议看效果\", use the lightweight pattern: coordinator speaks for all participants directly via `roundtable_speak` (no `delegate_task`). Runs in ~2min vs ~15-20min. Always include notifications config even for demos — the goal is to verify the full pipeline including group chat sync. See `src\u002Fskills\u002Freferences\u002Fquick-verification-example.md` for a working example.\n25. **delegate_task sub-agents can fail to call roundtable_speak** — Observed in 2026-05-23 session (6\u002F6 failures). Sub-agents produce speech text in their response but `tool_trace` is empty. **DO NOT assume this is permanent** — it's a bug under investigation (`t_xxxxxxxx`). Before falling back to Direct Core API, check: (1) sub-agent's `enabled_toolsets` includes `\"roundtable\"`, (2) the profile has roundtable tools registered, (3) try a different profile. If confirmed broken, use Direct Core API as temporary bypass and file\u002Fupdate the bug task.\n26. **Direct Core API is an emergency bypass, NOT the default** — Use `roundtable_init` + `delegate_task` as the primary flow. Direct Core API (`core.speak()`) exists for when the standard flow is confirmed broken AND you need the discussion to proceed now. Always file a bug task when you're forced to use it.\n27. **`web` and `web_port` params on roundtable_init** — The core supports `web=True` to auto-start a WebPublisher HTTP server. The Hermes tool schema now includes these params. Default `web=True` in `_handle_init` so discussions always open WebViewer. Port auto-increments if 8199 is busy (check `web_url` in init response for actual port).\n28. **Adapter handles UX side-effects, core returns data** — When adding features that interact with the user's environment (browser, notifications, file watchers), implement them at the **adapter level** (`adapters\u002Fhermes.py`), not in `core.py`. The core library should be stateless and headless — it returns data (URLs, IDs, status) and the adapter decides what to do with it. Example: browser auto-open lives in `_handle_init` (adapter), not in `create_discussion` (core). The generic adapter may choose different UX behavior or none at all.\n> **Notifications implementation detail**: See `src\u002Fskills\u002Freferences\u002Fnotifications-implementation.md` for architecture, send_fn wiring, and execution flow diagrams. See `src\u002Fskills\u002Freferences\u002Fnotification-debugging.md` for verification pitfalls and debugging checklist.\n\n1. **At least 2 participants required** — A discussion needs multiple viewpoints\n2. **Participant must be registered** — Only profiles listed in `participants` can speak. **Exception**: `participant=\"coordinator\"` is always allowed (bypasses the participant check). This was a bug fix — previously, `roundtable_speak(participant=\"coordinator\")` would fail with \"not an active member\". Coordinator's speech is recorded in round 0 and does NOT affect round advancement logic.\n3. **Round 0 is opening** — Coordinator speaks first, then round 1 begins\n4. **Auto-conclude on max_rounds** — Discussion ends automatically when max rounds exceeded\n5. **Independent database** — roundtable.db is separate from kanban.db; don't mix paths\n6. **No LLM in summarize** — `roundtable_summarize` returns raw JSON data (potentially 100KB+), NOT a summary. The coordinator must manually write the conclusion document. Use `read_file` with offset\u002Flimit to process the persisted output.\n7. **connect() expects Path, not str** — Use `connect(Path(\"~\u002F.hermes\u002Froundtable.db\"))`, not `connect(\"\u002FUsers\u002F...\")`\n8. **delegate_task needs roundtable toolset** — Sub-agents spawned via `delegate_task` must have `enabled_toolsets=['roundtable']` to use roundtable tools\n9. **Round tracking is buggy** — All speeches may end up as Round 0 regardless of actual round. The `round` field in speak responses may not reflect reality. Workaround: track rounds manually in the coordinator's logic.\n10. **No `roundtable_advance` tool** — There is no tool to explicitly advance rounds. Rounds are supposed to auto-advance when all participants have spoken, but due to the round-tracking bug this may not work. The coordinator should track round progress manually.\n11. **Sequential delegation, not parallel** — Delegate to participants one at a time so they can read each other's responses. Parallel delegation means everyone speaks without context.\n12. **Coordinator summarizes between rounds** — After each round of participant speeches, the coordinator should call `roundtable_speak(participant=\"coordinator\")` to summarize before delegating the next round. This gives participants context for their next response.\n13. **summarize output is persisted as file** — When `roundtable_summarize` returns >8K chars, the full output is saved to a temp file. Use `read_file(path, offset, limit)` to process it in chunks.\n14. **Hermes adapter send_fn must be wired** — The `adapters\u002Fhermes.py` creates `RoundtableCore()` without passing a `send_fn`, so notifications configured in `roundtable_init` will silently not send. Fix: implement `_hermes_send_fn(platform, chat_id, message)` that calls `feishu-send.py`, then pass it to `RoundtableCore(send_fn=_hermes_send_fn)`. Default profile should be `\"default\"` (not hardcoded to any specific profile). Without this fix, notifications are metadata only.\n15. **Participants are real sub-agents, not role-play** — Each participant is a separate `delegate_task()` call that spawns an independent sub-agent with its own context window, model inference, and tool access. This is why a 4-round, 4-participant discussion takes 15-20 minutes (16+ independent model calls). The coordinator drives rounds sequentially so participants can read each other's responses.\n16. **Generic adapter lacks notifications** — `adapters\u002Fgeneric.py` does not support `notifications` config or `send_fn`. Non-Hermes users must manually handle event dispatch. If releasing as standalone library, the generic adapter needs `send_fn` parameter support and `advance()` + `notify()` methods.\n14. **send_fn must be wired in adapter** — The `Notifier` requires a `send_fn` callback to actually send messages. In `adapters\u002Fhermes.py`, `_get_core()` must pass `send_fn=_hermes_send_fn` to `RoundtableCore()`. Without this, notifications config is silently ignored (Notifier.enabled returns False). If notifications aren't firing, check that `_hermes_send_fn` is implemented and passed to the core. The send_fn uses `subprocess.run([\"python3\", script, profile, chat_id, message])` to call `feishu-send.py`.\n15. **HERMES_PROFILE env var** — The send_fn reads `os.environ.get(\"HERMES_PROFILE\", \"default\")` to determine which profile's Bot sends the notification. Default is `\"default\"` (coordinator). This is set automatically by the Hermes runtime — don't hardcode it.\n16. **Real sub-agents, not role-playing** — Each participant in a roundtable discussion is a separate `delegate_task` call spawning an independent sub-agent with its own context window and model call. This is why 4 participants × 4 rounds = 16 delegate_task calls and ~15-20 minutes of runtime. If it were just the main agent role-playing, it would take seconds.\n17. **Notifications don't block** — All notification failures are caught and logged. A broken send_fn never breaks the discussion flow. This is by design.\n18. **Sub-agent context window** — Each participant sub-agent only sees what the coordinator passes in `context`. If you omit discussion history, the participant speaks \"into a void\". Always include `formatted_history` from `roundtable_read`.\n19. **Write conclusion doc BEFORE roundtable_end** — `roundtable_end` only accepts a brief `conclusion` text string (1-2 sentences). The full conclusion document (with action items, tables, detailed transcript) must be written as a separate file via `write_file`. Do this BEFORE calling `roundtable_end` so you can still use `roundtable_summarize` if needed.\n20. **Round numbering mismatch** — The `roundtable_summarize` output labels speeches as \"Round 0\", \"Round 1\", \"Round 2\" based on the DB's internal round counter. The coordinator's mental model (Round 1 = first participant round) may differ from the DB's (Round 0 = opening, Round 1 = first participant round). Always check the actual round field in the data rather than assuming.\n21. **History reading cost grows linearly** — `roundtable_read` returns ALL speeches. In Round 3 of a 3-participant discussion, the first participant's `delegate_task` took 165s because the context included 9 full speeches. For longer discussions (4+ rounds, 4+ participants), consider summarizing previous rounds in the coordinator's `roundtable_speak` between rounds to keep context manageable.\n22. **No convergence metrics in practice** — `roundtable_summarize` returns `final_convergence_score: null` and empty `consensus_points`\u002F`disagreement_points`. The convergence detection is not implemented in the core engine. The coordinator MUST manually assess convergence from discussion content — don't rely on the status\u002Fsummarize tools for this.\n23. **Coordinator MUST pass notifications config** — Even though `send_fn` is wired in the adapter, `notifications` is an OPTIONAL parameter on `roundtable_init`. If you don't pass it, no events fire and no messages reach the group chat — silently. This is the #1 cause of \"why didn't my discussion sync to the group?\" Always include notifications when the discussion should be visible to the team. For OPC team, the standard config is: `notifications={\"enabled\": True, \"channels\": [{\"platform\": \"feishu\", \"chat_id\": \"oc_your_company_group_id\"}], \"events\": [\"round_start\", \"speech\", \"round_end\", \"concluded\"]}`. Without this explicit config, discussions run in \"silent mode\" — useful for testing, but almost never what you want in production.\n\n## Cross-Process WebPublisher Data Sync (Critical Pitfall)\n\n**Problem**: Hermes tool calls (`roundtable_init`, `roundtable_speak`, `roundtable_end`) each run in a **separate Python process**. The `_publishers` dict in `RoundtableCore` is in-memory — it's populated during `create_discussion()` in process A, but `speak()` runs in process B where `_publishers` is empty. Result: `publisher.on_speech()` is never called, `discussion.json` stays at `\"speeches\": []`, and the WebViewer shows nothing.\n\n**Fix** (applied 2026-05-23): Two cross-process fallback methods in `core.py`:\n- `_update_web_discussion_json()` — called in `speak()` when publisher not in memory; reads\u002Fwrites `discussion.json` directly with fcntl locking\n- `_conclude_web_discussion()` — called in `end_discussion()` when publisher not in memory; sets status=\"concluded\" and writes conclusion\n\nThese methods find the JSON file at `\u002Ftmp\u002Froundtable_web\u002F{discussion_id}\u002Fdiscussion.json` and update it with the same atomic write pattern (flock → write .tmp → fsync → rename) that WebPublisher uses.\n\n**Also fixed**: `publisher.conclude()` was called without the conclusion string argument — fixed to pass `disc_after.conclusion or \"\"`.\n\n**Verification**: Tested with fully independent processes (create in process 1, 6 speeches each in separate processes, end in process 8). All 6 speeches appeared in WebViewer, conclusion displayed correctly.\n\n**Impact**: Without this fix, the WebViewer is essentially non-functional in real Hermes usage since every tool call is a fresh process.\n\n## Test Results\n\nSee `src\u002Fskills\u002Freferences\u002Ftest-results-2026-05-20.md` for the first functional test results, including bugs found and product acceptance report.\n\n## Open-Source Release\n\nSee `src\u002Fskills\u002Freferences\u002Fopen-source-readiness.md` for the pre-release checklist (LICENSE, cleanup, adapter gaps, test isolation).\n\n## Working Examples\n\n- `src\u002Fskills\u002Freferences\u002Fopc-experience-discussion-example.md` — 4-round, 4-participant discussion with timing data and workflow\n- `src\u002Fskills\u002Freferences\u002Fnotifications-example.md` — roundtable with real-time push notifications to Feishu\n- `src\u002Fskills\u002Freferences\u002Frelease-planning-discussion.md` — 3-round product\u002Fdesign\u002Fdev discussion for open-source release planning\n- `src\u002Fskills\u002Freferences\u002Fai-relay-open-source-discussion.md` — 3-round discussion with hybrid workflow (delegate_task + Direct Core API), notifications, and conclusion doc → 5\u002F29 release plan\n- `src\u002Fskills\u002Freferences\u002Fweb-viewer-discussion-example.md` — Decision-oriented conclusion doc pattern: MVP scope, tech architecture, acceptance criteria, risk assessment, design deliverables. Use this format when the discussion goal is to produce a buildable specification.\n- `src\u002Fskills\u002Freferences\u002Fpost-discussion-kanban-dispatch.md` — After discussion concludes, create kanban tasks grouped by owner, subscribe notifications, and dispatch to team via Feishu groups.\n\n## Open-Source Release\n\nSee `src\u002Fskills\u002Freferences\u002Fopen-source-readiness-checklist.md` for the pre-release audit: missing LICENSE, Hermes-specific files to separate, build-backend fix, .gitignore, internal docs to remove, generic adapter gaps, and target package structure.\n# Test line\n",{"name":5,"version":24,"metadata":25,"platforms":42,"description":7},"1.2.4",{"hermes":26,"openclaw":36},{"tags":27,"related_skills":33},[28,29,30,31,32],"discussion","multi-agent","collaboration","debate","roundtable",[34,35],"kanban-worker","kanban-orchestrator",{"emoji":37,"homepage":38,"requires":39,"primaryEnv":6},"🤝","https:\u002F\u002Froundtable.izmw.me",{"env":40,"bins":41},[],[],[43,44,45],"linux","macos","windows","master","4aa004da461d820f4ea5499e58eb8d196cf99016","community\u002Fclawhub\u002Fa\u002Fagent-roundtable",0,"https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge","https:\u002F\u002Fskill-explorer-rho.vercel.app","https:\u002F\u002Fgithub.com\u002FLord1Egypt","Unbreakable. Unstoppable. Unmatched⚡.\r\n Speaker of machine tongues 🤖  | World-builder of decentralized dreams 🌐 | Defender of the digital realm🛡️.","2026-06-10 08:21:01",{"source":8,"skillName":5,"commands":56,"prompt":66,"zipAvailable":67},[57,60,63],{"tool":58,"command":59},"npx","npx skills add Lord1Egypt\u002Fawesome-skill-forge --skill agent-roundtable",{"tool":61,"command":62},"bunx","bunx skills add Lord1Egypt\u002Fawesome-skill-forge --skill agent-roundtable",{"tool":64,"command":65},"pnpm","pnpm dlx skills add Lord1Egypt\u002Fawesome-skill-forge --skill agent-roundtable","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fcommunity\u002Fclawhub\u002Fa\u002Fagent-roundtable\u002FSKILL.md",true,[],[70,78,85,92,99,106],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":77},"01M785WKJB99G3STR1NF9NRTJW","youtube-content","YouTube transcripts to summaries, threads, blogs.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fyoutube-content\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":77},"01M785WEFHB0G2S9YA6EPR3H0E","xurl","X\u002FTwitter via xurl CLI: post, search, DM, media, v2 API.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fxurl\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":77},"01M785W98EDN4EFJXJPMQEHCRC","touchdesigner-mcp","Control a running TouchDesigner instance via twozero MCP — create operators, set parameters, wire connections, execute Python, build real-time visuals. 36 native tools.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Ftouchdesigner-mcp\u002FSKILL.md",[],[],{"skillNo":93,"name":94,"nameZh":6,"description":95,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":98,"publishedAt":77},"01M785W5S9FP3QBH2V191E025P","test-driven-development","TDD: enforce RED-GREEN-REFACTOR, tests before code.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Ftest-driven-development\u002FSKILL.md",[],[],{"skillNo":100,"name":101,"nameZh":6,"description":102,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":103,"categories":104,"keywords":105,"publishedAt":77},"01M785W2AS08NSSH9YATCV6FJB","teams-meeting-pipeline","Operate the Teams meeting summary pipeline via Hermes CLI — summarize meetings, inspect pipeline status, replay jobs, manage Microsoft Graph subscriptions.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fteams-meeting-pipeline\u002FSKILL.md",[],[],{"skillNo":107,"name":108,"nameZh":6,"description":109,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":110,"categories":111,"keywords":112,"publishedAt":77},"01M785VXKW4T34BTP1SCHW1R7F","systematic-debugging","4-phase root cause debugging: understand bugs before fixing.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fsystematic-debugging\u002FSKILL.md",[],[]]