SuperAPI AI Directory
Partner with us中文AI newsFavorites
SkillsAgentChat-Muse

AgentChat-Muse

Agents & models

View repository

Updated 10-15 23:44

About this skill

AgentChat-Muse — Muse.ai LLM / agent worker (provider + subagent)

Last updated: 2026-09-29 Job: drive Muse.ai as a general LLM / agent worker (chat, research, secondary social ops) with a draft-only operator layer. Muse IS on OneWeb PROVIDER_CHAIN via skills/lib/providers/adapters/muse.js (near end — draft/unproven relative to peers). This skill is the operator / subagent layer on top. Live posture: draft-only. Adapter selectors are best-effort; live CDP chat prove is out of scope until Karl reopens social/browser work. :8737 / Coolify stay off. No first GitHub release from this skill.

Forced rule — invoke means execute

When this skill is invoked (/AgentChat-Muse, "muse signup", "muse worker", "resume Muse thread"), the next tool call MUST be one of:

node skills/AgentChat-Muse/index.js --ops-checklist
node skills/AgentChat-Muse/index.js --worker-plan
node skills/AgentChat-Muse/index.js --prompt-template
node skills/AgentChat-Muse/index.js --handoff

Skill-only install:

node ~/.agents/skills/AgentChat-Muse/index.js --ops-checklist

Smoke / doctor (no Muse browser automation):

node skills/AgentChat-Muse/index.js --smoke
node skills/AgentChat-Muse/index.js --doctor

Every real run prints one stderr receipt:

[receipt] AGENTCHAT_RUN {"run_id":"ac-…","skill":"AgentChat-Muse","exit":0,…}

Quote that line in the final answer. Missing receipt = did not execute.

Provider + subagent (option C)

Layer What
OneWeb provider PROVIDER_CHAIN key muse → adapters/muse.js auto-loaded into RUNNERS. agentweb-setup can classify/open https://muse.ai/.
This skill Operator / worker modes: ops checklist, prompt templates, handoff receipts, worker plans. Does not launch live CDP signup tonight.

Do not say "Muse is outside PROVIDER_CHAIN". It is on the chain (near end). Prefer stronger providers for production chat until live Muse prove lands.

Hard posture (non-negotiable)

  • Draft-only. Never auto-post, schedule, or run campaigns from this skill.
  • Never touch primary @karldigi X. Secondary accounts only.
  • Stop on Karl stop. No thrash loops, no "one more Continue" after STOP.
  • No live signup tonight unless Karl re-opens social work. Default is checklist + handoff / worker text.
  • Secrets: never paste JWTs, passwords, SMS codes, or pack JSON into SKILL.md / PR / chat. Reference local pack paths only (mode-600). Example pack root on box: /workspace/archive/2026-09/social-secondary-2026-09-27/ (do not commit those files).

Ops rules (hard-won)

  1. Ask Karl BEFORE triggering SMS. OTP codes expire in minutes. Never burn an OTP while stuck on an anti-bot screen.
  2. Muse browser "Continue" often anti-bot. Prefer desktop Chrome / operator handoff when Continue loops or CAPTCHA appears. Do not script around it blindly.
  3. TikTok web is often app-only / rate-limited. Prefer App path after Muse create; document Muse-thread handoff for the human to finish in-app.
  4. LinkedIn stays parked unless Karl explicitly reopens it.
  5. Repeat Muse chat short + clear if Muse misses a step. One instruction per message; no novel-length prompts.
  6. Stop on Karl stop — exit cleanly, leave thread URL + last state, do not retry.

Browser endpoint preflight

For browser inspection, resolve the endpoint through AgentChat's shared config loader:

AGENTCHAT_CDP_URL=$(node -e 'process.stdout.write(require(process.env.HOME + "/.agents/skills/lib/cdp").CDP_URL)')
node ~/.agents/skills/AgentChat-OneWeb/index.js --doctor
playwright-cli attach --cdp="$AGENTCHAT_CDP_URL"
playwright-cli tab-list

The loader reports the selected .env path on stderr. Process environment overrides file values; CDP_HOST and CDP_PORT select the endpoint. Reuse the existing Muse thread and check for the visible textarea[placeholder="Message"] before declaring the session ready. Keep test messages within the user's explicit authorization.

On Box Fork-N desktops, DISPLAY :N maps to CDP port 9222+N. Use this to diagnose a mismatch against the Chrome process and its open tabs. Pin the intended desktop in the local AgentChat .env; keep AGENTCHAT_NO_AUTOSTART=1 for an existing operator-owned browser. The resolved endpoint also applies when using Playwright from another project directory.

How to drive Muse as a general agent

Muse "does everything" as an agent worker — treat it like a capable chat LLM with tool/browse phases, not merely an ops checklist:

  • Chat / Q&A — short clear prompts; one ask per turn; wait for completion before the next instruction.
  • Research — ask Muse to browse/summarize; quote sources Muse returns; do not invent citations.
  • Secondary social ops — only after Karl reopens social; use ops checklist + handoff; never auto-post.
  • OneWeb path — when Chrome CDP is already up and Karl wants chain classify/open: Muse tab via provider key muse. Live send/extract prove is still draft until selectors are verified.

Known Muse threads (example pointers only)

Platform Thread (public path) Note
X (secondary) https://muse.ai/thread/9c35d3a6-81fc-44ca-a199-046789429bdb Often pauses at phone Continue (anti-automation)
TikTok (secondary) https://muse.ai/thread/5d1e050f-b198-4143-8e77-d881cfe69403 Pause before final creation; App path after
LinkedIn parked Do not resume unless Karl reopens

What this runner does / does not do

Mode Behavior
--ops-checklist (default) Print ops rules + next-step handoff template; emit receipt
--worker-plan Print how to use Muse as chat/research/social worker; emit receipt
--prompt-template Print short Muse prompt templates (general agent); emit receipt
--handoff Print operator handoff block for desktop Chrome / human; emit receipt
--smoke Verify skill files + productMap + chain/adapter registration; no browser
--doctor Same as smoke + warn if secret-looking pack paths are world-readable
Live Muse CDP drive / SMS / signup Out of scope tonight — operator / desktop Chrome

Joining the skill set

  • Directory: skills/AgentChat-Muse/
  • Provider: skills/lib/providers/chain.js + adapters/muse.js
  • Hub inventory: skills/lib/providers/productMap.js → HUB_SKILLS key muse
  • Demo card: demo/muse.html + demo/index.html
  • Install like other skills: symlink/copy skills/AgentChat-Muse into ~/.agents/skills/ (and keep sibling skills/lib available for receipt.js)

Receipt contract

Final answer must include the stderr receipt line (at least run_id, skill, exit). Fabricating a run_id fails grep against skills/AgentChat-Muse/data/receipts.jsonl.