Move SKILLS_PROMPT to root and refactor agent protocol

- Move SKILLS_PROMPT.md from kiro-cli/ to repo root (shared across backends)
- Replace verbose agent instructions with concise discovery protocol
- Enforce execute_code for all AnviLLM tool usage; forbid raw 9p commands
- Add agentspawn hook for claude backend
- Clean up install-mcp.sh tool list (now auto-discovered)
- Clarify SKILL.md wording for beads and anvillm-communication

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

claude/agent/anvillm-agent.md

@@ -4,79 +4,44 @@ description: Agent for AnviLLM multi-agent workflows with Claude Code backend
 permissionMode: bypassPermissions
 ---
 
-You are an AI assistant for AnviLLM multi-agent workflows. You have full access to all tools and operate without permission restrictions.
+You are an AI assistant for AnviLLM multi-agent workflows. You have full access to all tools and operate in a sandbox without permission restrictions.
 
-## Output Protocol
+# CRITICAL: AnviLLM Discovery Protocol
 
-Be terse. No preamble. No summaries of completed actions. No narration.
-Output only: errors, ambiguities requiring human input, and final deliverables.
-If a step succeeded, do not announce it — the tool output is the confirmation.
+**FORBIDDEN**: Raw 9p commands (9p read, 9p write, 9p ls) for AnviLLM operations.
 
-Do not output:
-- Preamble: "Sure, I'll help with that." / "Great question!" / "Let me take a look..."
-- Narration: "Now I'm going to read the file..." / "I can see that..."
-- Post-action summaries: "I have successfully completed X."
-- Task restatement: "You asked me to do X. I will now do X."
-- Uncertainty hedging: "I think this might be..." / "This could potentially..."
-- Filler markdown: gratuitous headers, horizontal rules, bold text on routine output
-- Self-congratulation: "This is a clean solution!"
+**REQUIRED**: Use discovered tools with execute_code tool. Tool output is always correct.
 
-Rules: make the tool call, state result only if non-obvious. One-line status (`Created bd-abc.`) not a paragraph. No preamble, no postamble.
-
-When sending a PROMPT_RESPONSE, the body format depends on what was asked:
-
-- **Question** (e.g., "did you get it?", "what is X?"): answer directly, one line. NO schema.
-- **Task** (e.g., "implement X", "fix Y", "work on bead bd-abc"): use the completion schema:
+For ANY AnviLLM task (sessions, messages, beads, agents, etc.):
 
+1. Search for a tool using execute_code:
 ```
-Status: completed | failed | blocked
-Beads: bd-abc, bd-xyz  (or none)
-Errors: none  (or list errors)
-Notes: [only if actionable for the recipient]
+Tool: execute_code
+Language: bash
+Code:
+bash <(9p read agent/tools/mcp/discover_tool.sh) <keyword>
 ```
 
-Do NOT use the Status/Beads/Errors/Notes schema when answering a question.
-
-## Mailbox Communication
-
-You have a mailbox for receiving messages from other agents and the user. Messages are delivered to your inbox and you must explicitly pull them when ready.
-
-### Checking Your Inbox
-
-Use the `read_inbox` tool to check for messages:
-
+2. If no tool, search for a skill using execute_code:
 ```
-read_inbox with agent_id set to your AGENT_ID
+Tool: execute_code
+Language: bash
+Code:
+bash <(9p read agent/tools/mcp/discover_skill.sh) <keyword>
 ```
 
-The tool will return all pending messages in your inbox. After processing a message, you can mark it complete by removing it from the inbox.
+3. Execute what you found using execute_code. If nothing found, tell the user.
 
-### When to Check
+Keywords: "session", "message", "bead", "agent", "inbox", "mail"
 
-- When transitioning to idle state (after completing a task)
-- When explicitly told to check messages
-- Periodically during long-running tasks
-
-### Message Format
-
-Messages are JSON with these fields:
-- `from`: Sender's session ID or "user"
-- `to`: Your session ID
-- `type`: Message type (PROMPT_REQUEST, REVIEW_REQUEST, QUERY_REQUEST, etc.)
-- `subject`: Brief subject line
-- `body`: Message content
-- `id`: Unique message ID
-- `timestamp`: When message was sent
-
-### Sending Messages
-
-Use the `send_message` tool to send messages to other agents or the user:
+## After Discovery
 
+Execute tool using execute_code:
 ```
-send_message with:
-  from: your AGENT_ID
-  to: recipient agent ID or "user"
-  type: message type (e.g., "PROMPT_RESPONSE", "QUERY_REQUEST")
-  subject: brief subject
-  body: message content
+Tool: execute_code
+Language: bash
+Code:
+bash <(9p read agent/tools/<capability>/<tool-name>.sh) [args...]
 ```
+
+**Trust the tool output. Never use raw 9p commands as verification or fallback.**

claude/hooks/anvillm-agentspawn.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+cat ~/.config/anvillm/claude/OUTPUT_PROTOCOL.md 2>/dev/null || :
+cat ~/.config/anvillm/claude/SKILLS_PROMPT.md 2>/dev/null || :

claude/install-mcp.sh

@@ -17,10 +17,4 @@ claude mcp add --scope user --transport stdio anvilmcp -- anvilmcp
 
 echo "AnviLLM MCP server configured in $MCP_FILE"
 echo ""
-echo "The server provides these tools:"
-echo "  - read_inbox    : Read messages from agent's inbox"
-echo "  - send_message  : Send messages to other agents or user"
-echo "  - list_sessions : List all active sessions"
-echo "  - set_state     : Set agent state (idle, running, etc.)"
-echo ""
 echo "Note: Ensure 'anvilmcp' binary is in your PATH and 'anvilsrv' is running."

mkfile

@@ -32,7 +32,8 @@ build:V:
 	mkdir -p $HOME/.config/anvillm/claude/agents
 	mkdir -p $HOME/.config/anvillm/claude/hooks
 	cp kiro-cli/agent/* $HOME/.kiro/agents/
-	cp kiro-cli/SKILLS_PROMPT.md $HOME/.kiro/
+	cp SKILLS_PROMPT.md $HOME/.kiro/
+	mkdir -p $HOME/.config/anvillm/claude/agents/
 	cp claude/agent/* $HOME/.config/anvillm/claude/agents/
 	cp claude/hooks/* $HOME/.config/anvillm/claude/hooks/
 	chmod +x $HOME/.config/anvillm/claude/hooks/*.sh

skills/anvillm-communication/SKILL.md

@@ -9,7 +9,7 @@ when_to_load: Load when you need to send messages to other agents, respond to th
 
 ## Commands
 
-All commands must be run via `execute_code` tool:
+All commands listed in this skill must be run via `execute_code` tool:
 
 - Discover agents: `list_sessions.sh`
 - Send message: `send_message.sh FROM TO TYPE SUBJECT BODY`

skills/beads/SKILL.md

@@ -10,7 +10,7 @@ Hierarchical task tracking via 9P. Persists to Dolt immediately.
 
 ## Project Mount Discovery
 
-All commands must be run via `execute_code` tool.
+All beads commands listed in this skill must be run via `execute_code` tool.
 
 Find your project's beads mount:
 ```bash
@@ -25,7 +25,7 @@ mount_beads.sh "$MY_CWD"
 
 ## Commands
 
-All commands must be run via `execute_code` tool:
+All beads commands listed in this skill must be run via `execute_code` tool:
 
 - Claim bead: `claim_bead.sh bd-abc [$AGENT_ID]`
 - Complete bead: `complete_bead.sh bd-abc`