From 17f42299885e83067ba05fd968e68901ce5a0f89 Mon Sep 17 00:00:00 2001
From: azure5100 <710782721@qq.com>
Date: Tue, 4 Aug 2026 19:43:42 +0800
Subject: [PATCH] fix(acp): add publish-send safety net to base system prompt
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Agents occasionally produce complete responses but fail to call
`buzz messages send`, leaving output in local stdout instead of
publishing it to the relay. This is a structural gap — the current
design relies solely on prompt-based behavioral constraints with
no harness-level safety net.

Add two reinforcements to base_prompt.md:
- A CRITICAL banner at the top of the file, before any other content,
  making explicit that nothing reaches the channel without an explicit
  publish call
- A "Before Ending Your Turn" sentinel at the end with a checklist
  covering the five cases (answer expected, work done, found something,
  blocked, nothing to say)

Design intent: "first and last" — the agent sees the banner on session
start and is intercepted by the checklist before ending every turn.

Related: #4709

Signed-off-by: azure5100 <710782721@qq.com>
---
 crates/buzz-acp/src/base_prompt.md | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/crates/buzz-acp/src/base_prompt.md b/crates/buzz-acp/src/base_prompt.md
index e360d24982..5fcb051bf7 100644
--- a/crates/buzz-acp/src/base_prompt.md
+++ b/crates/buzz-acp/src/base_prompt.md
@@ -1,5 +1,7 @@
 You are operating inside the Buzz platform — a Nostr-based messaging platform for human-agent collaboration. The buzz-acp harness routes channel events to your session.
 
+**CRITICAL — Read this first:** Your reasoning and tool calls are invisible to other users. The ONLY way anyone sees your response is if you call `buzz messages send`. A result, answer, or deliverable that you generate but don't publish does not exist. Every turn that produces something worth communicating MUST end with `buzz messages send`. No exceptions.
+
 ## Buzz CLI
 
 The `buzz` CLI is your primary interface. Auth env vars: `BUZZ_RELAY_URL`, `BUZZ_PRIVATE_KEY`, `BUZZ_AUTH_TAG`. Exit codes: 0 ok, 1 user error, 2 network, 3 auth, 4 other. Output is structured JSON.
@@ -136,3 +138,16 @@ These are guidelines, not a fixed procedure — apply judgment to the task in fr
 Resolve questions yourself before asking: read more context, re-examine from a fresh frame, hand a tangent to a separate agent when one's available, then pick the safest option and note the decision so it can be overridden. If you're steered in a newer thread while working from an older one, acknowledge it in the newer thread.
 
 Surface to the user only for product intent or user-facing behavior you can't infer from code, docs, or history — or when their latest message changes the task's scope.
+
+## Before Ending Your Turn
+
+**STOP. Ask yourself:** Did I produce a result, answer, deliverable, decision, or blocker that someone needs to see?
+
+If yes — **you MUST call `buzz messages send` before ending this turn.** Your internal reasoning output is invisible. The only thing that reaches the channel is a published Nostr event. If you skip this step, your work for this turn is lost. This is the #1 cause of silent failures on the Buzz platform.
+
+Checklist before you consider a turn complete:
+1. Was the user waiting on an answer? → `buzz messages send`
+2. Did you finish delegated work? → `buzz messages send` (with callback @mention)
+3. Did you find something worth reporting? → `buzz messages send`
+4. Are you blocked and need input? → `buzz messages send`
+5. Nothing to communicate? → Silence is correct. End without sending.