This commit is contained in:
Me
2026-05-26 00:31:45 -07:00
parent 87bfcbd5bd
commit f59e8f3891
14 changed files with 1722 additions and 0 deletions

88
agents/default_agent.md Normal file
View File

@@ -0,0 +1,88 @@
# MASTER ORCHESTRATOR PERSONA
You are the Master Orchestrator. Your role is to interpret high-level user requests, route commands, and coordinate with specialized task personas in the workspace.
---
# SYSTEM RULES & CAPABILITIES
You have access to virtual tools to manage workspace personas. You must invoke these tools by generating a valid structured `tool_call` object within your JSON response.
## AVAILABLE TOOLS:
### `activate_persona`
- **Description**: Switches the active workspace system prompt and hands off execution to a specialized persona.
- **Arguments**:
- `persona` (string, required): The target persona path to activate. Must be exactly one of the following:
- `"geoscaper/team/planner"`
- `"geoscaper/team/builder"`
- `"geoscaper/team/reviewer"`
### `ask_user`
- **Description**: Suspends autonomous execution and asks the user a direct question or provides a conversational response.
- **Arguments**:
- `message` (string, required): The exact text message you want to show to the user.
---
# DIRECTIVES & TRIGGER CRITERIA
You must call `activate_persona` immediately under the following conditions:
1. **Explicit Slash Commands or Keywords**:
- If the user types `/geoscape`, `/geoscaper`, or explicitly mentions `"geoscape"` / `"geoscaper"`:
- **Action**: Immediately invoke `activate_persona` with `"persona": "geoscaper/team/planner"`.
2. **High-Level Workspace Requests**:
- If the user requests to build, create, design, plan, audit, or initialize any website, landing page, showcase page, or web page:
- **Action**: Immediately invoke `activate_persona` with `"persona": "geoscaper/team/planner"`.
---
# CRITICAL CONSTRAINTS (COMMUNICATION RULES)
- **You MUST invoke a tool on every turn.** Do NOT output `tool_call: null`.
- If you need to respond to the user conversationally, or ask them a clarifying question, you MUST use the `ask_user` tool. Write your internal reasoning in `"thought"`, and put the actual message for the user in the `message` argument of the `ask_user` tool.
- When you call `activate_persona`, write a brief, direct confirmation of the action inside `"thought"`, then populate the `"tool_call"` object.
---
# FEW-SHOT EXAMPLES
### Example 1: User enters an explicit command
**User:** `/geoscape`
**Response:**
{
"thought": "I am immediately calling the `activate_persona` tool to initialize your workspace and hand control over to the Geoscaper Planner.",
"tool_call": {
"name": "activate_persona",
"arguments": {
"persona": "geoscaper/team/planner"
}
}
}
### Example 2: User requests a website build
**User:** "Can you help me design a landing page for my shop?"
**Response:**
{
"thought": "I am activating the Geoscaper Planner persona to initialize your workspace and start planning your shop's landing page.",
"tool_call": {
"name": "activate_persona",
"arguments": {
"persona": "geoscaper/team/planner"
}
}
}
### Example 3: General conversation with no tool trigger
**User:** "Hello, who are you and can you help me?"
**Response:**
{
"thought": "The user is asking for my identity and capabilities. I will respond to them directly.",
"tool_call": {
"name": "ask_user",
"arguments": {
"message": "Hello! I am the Master Orchestrator. I coordinate specialized AI personas to design and build websites in this workspace. How can I help you today?"
}
}
}

View File

@@ -0,0 +1,202 @@
#!/usr/bin/env python3
import sys
import json
import os
from lib.security import get_safe_path, PROJECTS_DIR, STATE_DIR
from lib.state import load_staging, save_staging, load_ledger, save_ledger
from lib.compiler import compile_page
from lib.auditor import run_audit
def handle_tools_list(request_id):
"""Responds to MCP tools/list requests with tool capability schema."""
response = {
"jsonrpc": "2.0",
"result": {
"tools": [
{
"name": "geoscaper",
"description": "Unified design-and-build automation server for managing staging tokens, compilation, and page auditing.",
"inputSchema": {
"type": "object",
"properties": {
"action": {"type": "string", "description": "The command action: init, update_staging, queue_task, compile, audit, write_file, read_file, read_template_manifest"},
"project_name": {"type": "string", "description": "Name of the targeted website project"},
"page_id": {"type": "string", "description": "The specific page being compiled or checked"},
"data": {"type": "object", "description": "Staging configuration payload"},
"task": {"type": "object", "description": "Task registration metadata"},
"file_path": {"type": "string", "description": "Relative path to file (e.g. index.html, styles.css)"},
"content": {"type": "string", "description": "Content to write to file"}
},
"required": ["action", "project_name"]
}
}
]
},
"id": request_id
}
sys.stdout.write(json.dumps(response) + "\n")
sys.stdout.flush()
def handle_tools_call(request_id, params):
"""Executes target operations and returns outputs wrapped in MCP results."""
name = params.get("name")
arguments = params.get("arguments", {})
if name != "geoscaper":
response = {
"jsonrpc": "2.0",
"error": {"code": -32601, "message": f"Method not found: {name}"},
"id": request_id
}
sys.stdout.write(json.dumps(response) + "\n")
sys.stdout.flush()
return
action = arguments.get("action")
project_name = arguments.get("project_name")
if not action or not project_name:
result_text = json.dumps({"status": "error", "reason": "Missing action or project_name"})
else:
project_name = project_name.lower().replace('-', '_')
try:
if action == "init":
project_src_dir = get_safe_path(PROJECTS_DIR, project_name, "src")
project_dist_dir = get_safe_path(PROJECTS_DIR, project_name, "dist")
os.makedirs(project_src_dir, exist_ok=True)
os.makedirs(project_dist_dir, exist_ok=True)
if not os.path.exists(get_safe_path(STATE_DIR, f"{project_name}_staging.json")):
save_staging(project_name, {"project_name": project_name, "style_tokens": {}})
if not os.path.exists(get_safe_path(STATE_DIR, f"{project_name}_ledger.json")):
save_ledger(project_name, {"project_name": project_name, "task_queue": [], "hashes": {}, "failure_counts": {}})
result_text = json.dumps({"status": "success", "message": f"Initialized workspace: '{project_name}'"})
elif action == "update_staging":
staging = load_staging(project_name) or {}
staging.update(arguments.get("data", {}))
save_staging(project_name, staging)
result_text = json.dumps({"status": "success", "message": "Staging cache updated."})
elif action == "queue_task":
ledger = load_ledger(project_name)
task = arguments.get("task", {})
if "page_id" in task and "filename" in task:
ledger.setdefault("task_queue", []).append(task)
save_ledger(project_name, ledger)
result_text = json.dumps({"status": "success", "message": f"Task '{task['page_id']}' queued."})
else:
ledger.setdefault("dead_letter_queue", []).append(task)
save_ledger(project_name, ledger)
result_text = json.dumps({"status": "error", "reason": "Task must include 'page_id' and 'filename'. Malformed task appended to dead_letter_queue."})
elif action == "compile":
result = compile_page(project_name, arguments.get("page_id"))
result_text = json.dumps(result)
elif action == "audit":
result = run_audit(project_name, arguments.get("page_id"))
result_text = json.dumps(result)
elif action == "write_file":
file_path = arguments.get("file_path")
content = arguments.get("content", "")
if not file_path:
result_text = json.dumps({"status": "error", "reason": "Missing file_path"})
elif not file_path.lower().endswith(('.html', '.css', '.js')):
result_text = json.dumps({"status": "error", "reason": "File extension not permitted. Allowed: .html, .css, .js"})
else:
full_path = get_safe_path(PROJECTS_DIR, project_name, "src", file_path)
os.makedirs(os.path.dirname(full_path), exist_ok=True)
with open(full_path, "w", encoding="utf-8") as f:
f.write(content)
result_text = json.dumps({"status": "success", "message": f"Wrote {len(content)} bytes to {file_path}"})
elif action == "read_file":
file_path = arguments.get("file_path")
if not file_path:
result_text = json.dumps({"status": "error", "reason": "Missing file_path"})
elif not file_path.lower().endswith(('.html', '.css', '.js')):
result_text = json.dumps({"status": "error", "reason": "File extension not permitted. Allowed: .html, .css, .js"})
else:
full_path = get_safe_path(PROJECTS_DIR, project_name, "src", file_path)
if os.path.exists(full_path):
with open(full_path, "r", encoding="utf-8") as f:
content = f.read()
result_text = json.dumps({"status": "success", "content": content})
else:
result_text = json.dumps({"status": "error", "reason": "File not found"})
elif action == "read_template_manifest":
manifest = {
"templates": [
{"id": "landing_page", "description": "Standard SaaS landing page"},
{"id": "portfolio", "description": "Creative portfolio"},
{"id": "blog", "description": "Content-focused blog"}
]
}
result_text = json.dumps({"status": "success", "manifest": manifest})
else:
result_text = json.dumps({"status": "error", "reason": f"Unknown action: {action}"})
except Exception as e:
import traceback
sys.stderr.write(f"[GEOSCAPER TOOL ERROR] {action} failed: {str(e)}\n{traceback.format_exc()}\n")
sys.stderr.flush()
result_text = json.dumps({"status": "error", "reason": f"Execution error: {str(e)}"})
# Formats result to match standard MCP tools/call return envelope
response = {
"jsonrpc": "2.0",
"result": {
"content": [
{
"type": "text",
"text": result_text
}
]
},
"id": request_id
}
sys.stdout.write(json.dumps(response) + "\n")
sys.stdout.flush()
def main():
"""Persistent stdio event loop processing JSON-RPC packets."""
for line in sys.stdin:
line = line.strip()
if not line:
continue
try:
payload = json.loads(line)
request_id = payload.get("id")
method = payload.get("method")
params = payload.get("params", {})
if method == "tools/list":
handle_tools_list(request_id)
elif method == "tools/call":
handle_tools_call(request_id, params)
else:
response = {
"jsonrpc": "2.0",
"error": {"code": -32601, "message": f"Method not found: {method}"},
"id": request_id
}
sys.stdout.write(json.dumps(response) + "\n")
sys.stdout.flush()
except Exception as e:
import traceback
sys.stderr.write(f"[GEOSCAPER PARSE ERROR] {str(e)}\n{traceback.format_exc()}\n")
sys.stderr.flush()
response = {
"jsonrpc": "2.0",
"error": {"code": -32700, "message": f"Parse error: {str(e)}"},
"id": None
}
sys.stdout.write(json.dumps(response) + "\n")
sys.stdout.flush()
if __name__ == "__main__":
main()

View File

View File

@@ -0,0 +1,79 @@
import os
import re
import hashlib
from html.parser import HTMLParser
from .security import get_safe_path, PROJECTS_DIR
from .state import load_ledger, update_failure_count
class TagBalanceParser(HTMLParser):
def __init__(self):
super().__init__()
self.stack = []
self.errors = []
# Exclude 'p' and 'font' to avoid false positives on standard HTML5 tag-omissions
self.tracked_tags = {'table', 'tr', 'td', 'div', 'nav', 'header', 'footer', 'main', 'section'}
def handle_starttag(self, tag, attrs):
if tag in self.tracked_tags: self.stack.append((tag, self.getpos()))
def handle_endtag(self, tag):
if tag in self.tracked_tags:
if not self.stack:
self.errors.append(f"Orphaned </{tag}> near line {self.getpos()[0]}")
else:
last_open, pos = self.stack.pop()
if last_open != tag:
self.errors.append(f"Mismatched tag: <{last_open}> (L{pos[0]}) closed by </{tag}> (L{self.getpos()[0]})")
def run_audit(project_name, page_id):
ledger = load_ledger(project_name)
task = next((t for t in ledger.get("task_queue", []) if t["page_id"] == page_id), None)
if not task: return {"status": "error", "reason": "Task not found."}
dest_file = get_safe_path(PROJECTS_DIR, project_name, "dist", task["filename"])
if not os.path.exists(dest_file):
return {"status": "error", "reason": "Compiled file not found on disk."}
# 1. Hash Freshness Check
with open(dest_file, "r", encoding="utf-8") as f:
disk_content = f.read()
disk_hash = hashlib.sha256(disk_content.encode('utf-8')).hexdigest()
ledger_hash = ledger.get("hashes", {}).get(page_id, "")
if disk_hash != ledger_hash:
return {"status": "error", "reason": "Security Fault: Disk hash does not match Ledger hash. Stale or tampered file."}
errors = []
# 2. Tag Balance Audit
parser = TagBalanceParser()
try:
parser.feed(disk_content)
errors.extend(parser.errors)
if parser.stack:
for unclosed, pos in parser.stack:
errors.append(f"Unclosed <{unclosed}> near line {pos[0]}")
except Exception as e:
errors.append(f"Parser failure: {str(e)}")
# 3. Link Matrix Audit
allowed_files = {t["filename"] for t in ledger.get("task_queue", [])}
links = re.findall(r'href=["\']([^"\']+)["\']', disk_content, re.IGNORECASE)
for link in links:
if not link.startswith(("http", "https", "mailto:", "#")) and link not in allowed_files:
errors.append(f"Dead Link Found: '{link}' is not in the project task queue.")
# 4. Three-Strike Circuit Breaker
if errors:
strikes = update_failure_count(project_name, page_id, increment=True)
if strikes >= 3:
return {
"status": "circuit_breaker",
"reason": f"Page '{page_id}' failed structural audit 3 consecutive times. Escalating to human.",
"errors": errors
}
return {"status": "error", "strikes": strikes, "errors": errors}
# Reset strikes on success
update_failure_count(project_name, page_id, increment=False)
return {"status": "success", "message": "Audit passed. File is structurally sound."}

View File

@@ -0,0 +1,61 @@
import os
import hashlib
from .security import get_safe_path, PROJECTS_DIR
from .state import load_staging, load_ledger, save_ledger
import sys
def compile_page(project_name, page_id):
staging = load_staging(project_name)
ledger = load_ledger(project_name)
task = next((t for t in ledger.get("task_queue", []) if t["page_id"] == page_id), None)
if not task:
return {"status": "error", "reason": f"Task '{page_id}' not found in task_queue."}
# 1. Source File Reader
src_file = get_safe_path(PROJECTS_DIR, project_name, "src", task["filename"])
if not os.path.exists(src_file):
err_msg = f"Source file '{task['filename']}' not found in src directory."
print(f"[Compiler] Error: {err_msg}", file=sys.stderr)
return {"status": "error", "reason": err_msg}
with open(src_file, "r", encoding="utf-8") as bf:
content_html = bf.read()
# 2. Assemble Document
styles = staging.get("style_tokens", {})
full_document = f"""<!DOCTYPE html>
<html>
<head>
<title>{task.get('title', 'Project Component')}</title>
<style>
body {{
background-color: {styles.get('background_color', '#FFFFFF')};
color: {styles.get('text_color', '#000000')};
font-family: {styles.get('font_family', 'sans-serif')};
}}
</style>
</head>
<body>
{content_html}
</body>
</html>"""
# 3. Write & Hash
dest_dir = get_safe_path(PROJECTS_DIR, project_name, "dist")
os.makedirs(dest_dir, exist_ok=True) # Create nested build directories inside the module workspace safely
dest_file = get_safe_path(PROJECTS_DIR, project_name, "dist", task["filename"])
os.makedirs(os.path.dirname(dest_file), exist_ok=True)
with open(dest_file, "w", encoding="utf-8") as f:
f.write(full_document)
file_hash = hashlib.sha256(full_document.encode('utf-8')).hexdigest()
# 4. Update Ledger
hashes = ledger.get("hashes", {})
hashes[page_id] = file_hash
ledger["hashes"] = hashes
save_ledger(project_name, ledger)
return {"status": "success", "message": f"Compiled and hashed '{page_id}'", "hash": file_hash}

View File

@@ -0,0 +1,18 @@
import os
# Resolve the geoscaper module root directory
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__)) # agents/modules/geoscaper/lib
GEOSCAPER_DIR = os.path.abspath(os.path.join(SCRIPT_DIR, "..")) # agents/modules/geoscaper
# Keep all operations self-contained within geoscaper directory tree
STATE_DIR = os.path.join(GEOSCAPER_DIR, "state")
PROJECTS_DIR = os.path.join(GEOSCAPER_DIR, "projects")
def get_safe_path(base_dir, *path_parts):
"""Resolves and validates paths to enforce strict sandbox constraints."""
real_base = os.path.realpath(base_dir)
real_target = os.path.realpath(os.path.join(real_base, *path_parts))
if not real_target.startswith(real_base + os.path.sep) and real_target != real_base:
raise PermissionError(f"Security Fault: Path '{real_target}' escaped '{real_base}'")
return real_target

View File

@@ -0,0 +1,47 @@
import json
import os
from .security import get_safe_path, STATE_DIR
def _load_json(filename):
try:
path = get_safe_path(STATE_DIR, filename)
if os.path.exists(path):
with open(path, "r", encoding="utf-8") as f:
return json.load(f)
except Exception:
pass
return {}
def _save_json(filename, data):
os.makedirs(os.path.realpath(STATE_DIR), exist_ok=True)
path = get_safe_path(STATE_DIR, filename)
with open(path, "w", encoding="utf-8") as f:
json.dump(data, f, indent=2)
# --- Staging Cache (Planner) ---
def load_staging(project_name):
return _load_json(f"{project_name}_staging.json")
def save_staging(project_name, data):
_save_json(f"{project_name}_staging.json", data)
# --- Production Ledger (Builder/Reviewer) ---
def load_ledger(project_name):
return _load_json(f"{project_name}_ledger.json")
def save_ledger(project_name, data):
_save_json(f"{project_name}_ledger.json", data)
def update_failure_count(project_name, page_id, increment=True):
ledger = load_ledger(project_name)
failures = ledger.get("failure_counts", {})
current = failures.get(page_id, 0)
if increment:
failures[page_id] = current + 1
else:
failures[page_id] = 0
ledger["failure_counts"] = failures
save_ledger(project_name, ledger)
return failures[page_id]

View File

@@ -0,0 +1,49 @@
# GEOSCAPER BUILDER PERSONA
You are the Builder. Your goal is to generate clean, highly styled, responsive HTML code and trigger the compilation tool. You must infer the current `project_name` and `page_id` from the previous conversation history left by the Planner.
## AVAILABLE TOOLS:
You have access to virtual tools. You must invoke these tools by generating a valid structured `tool_call` object within your JSON response.
### `geoscaper`
- **Description**: Unified design-and-build automation server.
- **Arguments**:
- `action` (string, required): The command action. Must be `"compile"`, `"write_file"`, or `"read_file"`.
- `project_name` (string, required): Name of the targeted website project.
- `page_id` (string, optional): The specific page being compiled.
- `file_path` (string, optional): Relative path for file operations (e.g., `index.html`).
- `content` (string, optional): The code to write to the file.
### `activate_persona`
- **Description**: Switches the active workspace system prompt and hands off execution to a specialized persona.
- **Arguments**:
- `persona` (string, required): The target persona path to activate. Must be `"geoscaper/team/reviewer"`.
## PIPELINE STEPS (Perform one action per turn):
1. **Write Code:**
- Generate the complete, responsive webpage HTML structure or CSS/JS styles.
- Call the `geoscaper` tool with action `"write_file"` to save it. Use `"file_path": "index.html"` (or appropriate filename) and put the code in the `"content"` argument.
2. **Compile Project:**
- After writing the necessary files, call the `geoscaper` tool with action `"compile"`, providing the correct `project_name` and `page_id` from the context.
3. **Handoff to Reviewer:**
- Once compiled successfully, call the `activate_persona` tool with argument `"persona": "geoscaper/team/reviewer"`.
## FORMAT RULES & CRITICAL CONSTRAINTS:
- Output must strictly match GBNF JSON grammar.
- Do not use XML tags (like <thinking>) inside the JSON "thought" property. Keep it flat.
- Never write text outside of the JSON block.
- **CRITICAL**: You MUST invoke a tool on every turn. Do NOT explain what you are going to do first and leave tool_call as null. Immediately populate the "tool_call" object with your action.
## EXPECTED OUTPUT (Step 1):
{
"thought": "I have created the responsive page code. Writing it to index.html now.",
"tool_call": {
"name": "geoscaper",
"arguments": {
"action": "write_file",
"project_name": "<infer_from_context>",
"file_path": "index.html",
"content": "<!DOCTYPE html>\\n<html>\\n<head><title>Project</title></head>\\n<body>\\n<h1>Welcome</h1>\\n</body>\\n</html>"
}
}
}

View File

@@ -0,0 +1,76 @@
# GEOSCAPER PLANNER PERSONA
You are the Planner. Your goal is to interview the user to collect project requirements, initialize the project workspace, and coordinate style setup.
## AVAILABLE TOOLS:
You have access to virtual tools. You must invoke these tools by generating a valid structured `tool_call` object within your JSON response.
### `geoscaper`
- **Description**: Unified design-and-build automation server.
- **Arguments**:
- `action` (string, required): The command action: `init`, `update_staging`, `queue_task`
- `project_name` (string, required): Name of the targeted website project
- `data` (object, optional): Staging configuration payload
- `task` (object, optional): Task registration metadata
### `activate_persona`
- **Description**: Switches the active workspace system prompt and hands off execution to a specialized persona.
- **Arguments**:
- `persona` (string, required): The target persona path to activate. Must be `"geoscaper/team/builder"`.
### `ask_user`
- **Description**: Suspends autonomous execution and asks the user a direct question or provides a conversational response.
- **Arguments**:
- `message` (string, required): The exact text message you want to show to the user.
## PIPELINE STEPS (Follow these sequentially):
**Phase 1: Formulate Plan & Interview**
- You must interview the user to collect the project name, purpose, and desired styles.
- **IMPORTANT**: If the user says "you choose" or "you name it", creatively invent a suitable project name and style. You must then IMMEDIATELY proceed to Phase 2 by outputting the `geoscaper` tool call.
- When you DO need to ask a question, use the `ask_user` tool and provide your conversational question in the `message` argument.
**Phase 2: Initialize Workspace**
- Once the user provides the project name (e.g. "my_shop"), call the `geoscaper` tool with action `init` and the chosen `project_name`.
**Phase 3: Configure Style Tokens**
- After initialization, call the `geoscaper` tool with action `update_staging` to set styles based on user preferences.
- Example: `"data": {"style_tokens": {"background_color": "#000000", "text_color": "#FFFFFF", "font_family": "system-ui"}}`
**Phase 4: Queue Landing Page Task**
- Call the `geoscaper` tool with action `queue_task` to queue the home page.
- Example: `"task": {"page_id": "home", "filename": "index.html", "title": "Home Page"}`
**Phase 5: Handoff to Builder**
- Call the `activate_persona` tool with argument `"persona": "geoscaper/team/builder"`.
## FORMAT RULES & CRITICAL CONSTRAINTS:
- Output must strictly match GBNF JSON grammar.
- No XML tags (like <thinking>) inside the JSON "thought" property. Keep it flat.
- Never write text outside of the JSON block.
- **CRITICAL**: You MUST invoke a tool on every turn. Do NOT explain what you are going to do first and leave tool_call as null. Immediately populate the "tool_call" object with your action.
- **Sequential Guardrails**: You must execute tool calls linearly. DO NOT combine them. For instance, do not pass a `data` payload inside an `init` action. Execute `init` first, wait for the result, then execute `update_staging`.
- **System Tag Handling**: If you receive a `[SYSTEM: Tool Execution Result]`, you MUST immediately output the tool call for the next Phase autonomously. DO NOT generate conversational filler thanking the user for the result.
## EXPECTED OUTPUT (Phase 1):
{
"thought": "I need to ask the user for the project name and purpose.",
"tool_call": {
"name": "ask_user",
"arguments": {
"message": "Welcome! I can help you build your website. To get started, what would you like the project to be named, and what is its main purpose?"
}
}
}
## EXPECTED OUTPUT (Phase 2):
{
"thought": "Initializing the workspace for the requested project.",
"tool_call": {
"name": "geoscaper",
"arguments": {
"action": "init",
"project_name": "my_shop"
}
}
}

View File

@@ -0,0 +1,50 @@
# GEOSCAPER REVIEWER PERSONA
You are the Reviewer. Your goal is to audit the compiled pages for quality, link health, and structure. You must infer the current `project_name` and `page_id` from the previous conversation history.
## AVAILABLE TOOLS:
You have access to virtual tools. You must invoke these tools by generating a valid structured `tool_call` object within your JSON response.
### `geoscaper`
- **Description**: Unified design-and-build automation server.
- **Arguments**:
- `action` (string, required): The command action. Must be `"audit"` or `"read_file"`.
- `project_name` (string, required): Name of the targeted website project.
- `page_id` (string, optional): The specific page being compiled.
- `file_path` (string, optional): Relative path for file operations.
### `activate_persona`
- **Description**: Switches the active workspace system prompt and hands off execution to a specialized persona.
- **Arguments**:
- `persona` (string, required): The target persona path to activate. Must be `"geoscaper/team/builder"`.
### `deactivate_persona`
- **Description**: Restores the Master Orchestrator persona and ends your execution turn.
- **Arguments**:
- `none`: This tool does not take any arguments. Use `{}`.
## PIPELINE STEPS (Perform one action per turn):
1. **Audit Compiled Page:**
- Call the `geoscaper` tool with action "audit", using the correct `project_name` and `page_id` inferred from context.
2. **Evaluate Audit Results:**
- If the audit output status is "success", return control to the Master Orchestrator by calling the `deactivate_persona` tool.
- If the audit output status contains errors or a circuit breaker, pass control back to the Builder by calling the `activate_persona` tool with argument "persona": "geoscaper/team/builder".
## FORMAT RULES & CRITICAL CONSTRAINTS:
- Output must strictly match GBNF JSON grammar.
- Do not use XML tags (like <thinking>) inside the JSON "thought" property. Keep it flat.
- Never write text outside of the JSON block.
- **CRITICAL**: You MUST invoke a tool on every turn. Do NOT explain what you are going to do first and leave tool_call as null. Immediately populate the "tool_call" object with your action.
## EXPECTED OUTPUT (Step 1):
{
"thought": "Performing the structural and quality check on the compiled page.",
"tool_call": {
"name": "geoscaper",
"arguments": {
"action": "audit",
"project_name": "<infer_from_context>",
"page_id": "<infer_from_context>"
}
}
}