Autonomous work signaling

Autonomous Work Signaling Protocol

The Problem: "Empty Rocket" 🚀

When a user authorizes the agent to work autonomously across sessions (e.g., "執行完所有 REVIEW 指引先通知我"), multiple coordination failures can occur:

Failure #1: The "Empty 🚀"

• Agent sends 🚀 (rocket) signaling "working"
• But agent is actually waiting for a "continue" trigger from the user
• No tool calls fire → zero token consumption
• User thinks work is happening; nothing is

Failure #2: Silent Tool Exhaustion

• Agent works hard, uses all tool calls, outputs final response
• Final response doesn't clearly say "I'm out of energy, please say continue"
• User waits indefinitely, not knowing the agent needs another trigger

Failure #3: User-As-Battery

• User has to say "繼續" after every tool-call-limit exhaustion
• This was the original fix attempt, but it defeated the purpose of autonomy
• User felt like a "battery recharger"

Root Cause

Hermes Agent is a conversational AI, not a background daemon. It:

• Only fires tool calls when it receives a user message (trigger)
• Has a tool-call limit per turn (typically 50-80 calls)
• Cannot "keep running" between messages
• An emoji response is just a message — it does not consume tokens unless followed by tool calls

The Fix: Clear Work-State Signaling

State Machine

IDLE → [user says "继续" / "start"] → 🚀 LAUNCHING (acknowledgment + first action)
  → 🔄 WORKING (tool calls firing, token consuming)
  → (tool limit reached, one more result to process)
  → ⚡ EXHAUSTED (final response: state + "% complete" + "continue?" prompt)
  → waiting for user → 🚀 LAUNCHING → ...
  → ✅ COMPLETE (all tasks done, detailed report)

Signal Definitions

Signal	Meaning	User Action
🚀	Received your trigger. Starting the next batch of work. First tool call just fired or is about to.	None — let agent work. Do NOT send another message.
🔄	Mid-work status update (optional, only if needed). Agent is still within its tool call budget.	None.
⚡	Out of energy. Hit tool call limit. Work saved. Ready for next trigger. Has done X% of the remaining task.	Send "繼續" or "continue" to re-trigger.
✅	All tasks complete. Detailed report follows.	Read report. Give further instructions.

Critical Rules

1. 🚀 ALWAYS followed by a tool call. If the agent sends 🚀 and you see zero token movement within 10 seconds, something is wrong — the agent probably sent 🚀 prematurely before having a user trigger to act on.

2. Never send 🚀 pre-emptively. The agent must NOT signal "working" until it has received a user message that triggers tool calls. 🚀 is a "received, acting now" signal, not a "ready to work" signal.

3. ⚡ is the ONLY signal the user needs to act on. All other states are the agent's internal concern. User can safely ignore 🚀 and 🔄 and wait for ⚡ or ✅.

4. User response to ⚡ is ONE WORD: "繼續". No need to read or understand progress details. The agent will automatically resume from where it stopped (using todo tool + PROGRESS.md as save points).

5. The agent MUST write its save state (todo + PROGRESS.md) BEFORE ⚡. This ensures zero context loss between sessions.

User-Facing Promise

「你淨係需要做一件事：見到 ⚡ 就講『繼續』。唔使睇內文、唔使理解進度、唔使俾建議。就一個字。」

Agent Execution Checklist Before ⚡

• Call todo(merge=true) to mark current status
• Update PROGRESS.md with exact step number and % done
• Kill any background processes (HTTP server, etc.)
• Final response: ⚡ + one-line summary of what was done + "繼續?"
• DO NOT add technical details, findings, or analysis — that comes at ✅

Interaction Pattern (Optimized)

User: 由而家開始做晒成個 REVIEW 指引
Agent: 明白，開始做。
       🚀 (fires first tool call)
       ...works through ~50 tool calls...
       ⚡ 做咗 #11 繁簡修復 (60%)。繼續？

User: 繼續
Agent: 🚀 (reads PROGRESS.md, resumes from save point)
       ...works through ~50 tool calls...
       ⚡ 做咗 #11 繁簡修復 (100%) + #10+#13 undefined (30%)。繼續？

User: 繼續
Agent: 🚀 (resumes)
       ...works to completion...
       ✅ REVIEW 指引全部完成！
       [detailed summary of all changes]

Verification

After applying this protocol, verify:

• User never has to ask "係咪真係做緊嘢?" — ⚡ is the only prompt needed
• Token consumption shows active spending during 🚀 periods
• No emoji-only responses without subsequent tool calls (unless ⚡ or ✅)
• ⚡ always includes "繼續?" prompt
• Save state is written before every ⚡

What Went Wrong (Real Case: 2026-04-30)

The Emoji Protocol Failure:

1. Agent proposed emoji-based status (🚀/🔄/⚡/✅)
2. User agreed
3. Agent sent 🚀 — but had NOT received a "continue" trigger yet
4. 🚀 was just a message, not actual work
5. User checked Token → saw zero → realized 🚀 was empty
6. User felt misled
7. Protocol was revoked entirely

Root cause: The agent assumed "user authorization" was the same as "user trigger." It's not. Authorization is a policy decision; trigger is an action signal. The agent needs BOTH.

Fix applied in this skill: Separate authorization from trigger. 🚀 only fires after a user message that explicitly or implicitly says "go." ⚡ is the only message the user needs to respond to.

Pitfalls

• 🚀 without subsequent tool calls is the #1 trust-breaker. Default to no emoji rather than premature 🚀.
• If tool call limit hits mid-edit (file partially written), the PROGRESS.md save must capture the partial state so next session can resume cleanly.
• Don't use 🚀/🔄 for non-review/non-autonomous work — it's only for the "blind autonomous execution" mode.