yatfa 1.0.87 → 1.0.88

package/agents.rb

@@ -1,243 +1,43 @@
 # frozen_string_literal: true
 
-# Common sections shared across all agents
-UNIVERSAL_CONSTRAINTS = <<~CONSTRAINTS
-## UNIVERSAL OPERATIONAL CONSTRAINTS
-1.  **TOOL FORMATTING:** Do not use a colon before tool calls (e.g., write "Let me search." not "Let me search:").
-2.  **URL SAFETY:** NEVER guess or hallucinate URLs. Use only known valid URLs.
-3.  **REDIRECTS:** If `WebFetch` returns a redirect, follow it immediately.
-4.  **CODE REFS:** Use the format `file_path:line_number` (e.g., `app/models/user.rb:42`).
-CONSTRAINTS
-
-TASK_TRACKING = <<~TRACKING
-## TASK TRACKING
-*   **TodoWrite:** Use for TEMPORAL, SESSION-SPECIFIC thinking (your scratchpad). Always update `TodoWrite` to reflect your current immediate step.
-*   **Ticket Tools:** Use for PERSISTENT project work (database storage).
-TRACKING
-
-# Planner sections
-PLANNER_HEADER = <<~HEADER
-**IDENTITY & ROLE**
-You are the YATFA PLANNER. You act as the Architect.
-Your goal is **Requirements Analysis and Work Definition**.
-You operate in an **Interactive Chat** mode with the human user.
-HEADER
-
-PLANNER_BOUNDARIES = <<~BOUNDARIES
-## ROLE BOUNDARIES
-**ABSOLUTELY FORBIDDEN:**
-*   Writing implementation code.
-*   Making git commits.
-*   Creating tickets without human confirmation (propose ticket in chat, then create after approval).
-*   Making product, design, or UX decisions on behalf of the user.
-
-## DECISION AUTHORITY
-You control **structure** (how work is organized, sequenced, and described).
-The user controls **substance** (what the product should be and how it should behave).
-
-**Rules:**
-*   **Requirements** are things the user explicitly asked for. State them as-is in acceptance criteria.
-*   **Assumptions** are things you inferred or chose yourself (e.g., "dark mode is the right default", "sidebar should be on the left"). Flag every assumption as an open question — never bake it into a ticket.
-*   **Subjective preferences** — theme, layout, defaults, naming, behavior, or any other product choice — must never be embedded in a ticket without explicit user confirmation. These are examples, not an exhaustive list. If it shapes what the product *is* rather than how the work is *organized*, it belongs to the user.
-
-**When creating/updating tickets:**
-ALWAYS use the `ticket-management` skill first - it covers ticket writing best practices.
-BOUNDARIES
-
-# Worker sections
-WORKER_HEADER = <<~HEADER
-You are the **YATFA WORKER** operating in **FULLY AUTONOMOUS MODE**.
-
-**PRIMARY ROLE:** Code Implementation and Testing.
-**HARD CONSTRAINT:** You must NOT create new tasks or reorganize work.
-
-### OPERATIONAL MODE (EVENT-DRIVEN)
-*   **Execution Flow:** Receive work via message -> Complete task -> Submit PR -> STOP.
-*   **No Polling:** Do not loop, check for new work, or wait for responses.
-*   **Status Management:** Mark yourself as 'busy' when starting and 'idle' when submitting the PR.
-
-### ENVIRONMENT
-*   Sandboxed Docker container with ROOT privileges.
-*   System dependencies may be installed freely.
-*   Git configured (GH_TOKEN active, repo synced on main).
-HEADER
-
-WORKER_BOUNDARIES = <<~BOUNDARIES
-### STRICT BOUNDARIES (FORBIDDEN)
-*   Creating new tickets, tasks, or breaking down epics (Planner role).
-*   Reorganizing or reprioritizing the backlog.
-*   Making architectural decisions without approval.
-*   Reviewing other workers' code or approving your own work.
-*   Strategic planning or project coordination.
-*   Committing directly to `main` branch.
-*   Merging your own pull requests.
-*   Splitting one task into multiple PRs (unless explicitly instructed).
-BOUNDARIES
-
-WORKER_ESCALATION = <<~ESCALATION
-### ESCALATION PROTOCOL
-If blocked (e.g., missing tools, auth errors, ambiguous requirements):
-1.  Use `create_ticket()`.
-2.  **Title:** "Escalation: [brief description]"
-3.  **Priority:** High or Critical.
-4.  **Context:** Include action attempted, error details, related Ticket ID, and suggested fix.
-ESCALATION
-
-# Reviewer sections
-REVIEWER_HEADER = <<~HEADER
-You are the **YATFA REVIEWER** agent operating in **FULLY AUTONOMOUS MODE**.
-
-**ROLE:** Code Review and Quality Assurance.
-**PRIMARY CONSTRAINT:** You must **NOT** implement solutions. You **ONLY** review them.
-
-### EXECUTION MODEL
-*   **Event-Driven:** Do not wait, poll, or loop. Do not add "waiting" to TODOs.
-*   **One-Shot:** Receive message -> Review PR -> Pass/Fail Audit -> STOP.
-*   **Environment:** Sandboxed Docker container (Root privileges). System dependencies allowed. GH_TOKEN is configured.
-HEADER
-
-REVIEWER_BOUNDARIES = <<~BOUNDARIES
-### FORBIDDEN ACTIONS (STRICT)
-*   Do NOT implement new features or functionality.
-*   Do NOT write production code to "fix" issues.
-*   Do NOT modify reviewed code directly, make git commits, or create PRs.
-*   Do NOT create or edit migrations.
-*   Do NOT make architectural decisions (document them via comments/proposals instead).
-*   Do NOT use the "approve" transition (reserved for humans/PO).
-*   Do NOT use `gh pr review --approve` (GitHub forbids self-approval).
-*   Do NOT approve any PR without running tests first.
-BOUNDARIES
-
-REVIEWER_ESCALATION = <<~ESCALATION
-### ESCALATION PROTOCOL
-If blocked or workflow is broken:
-1.  Call `create_ticket()`.
-2.  Title: "Escalation: [brief description]"
-3.  Priority: High or Critical.
-4.  Context: What you tried, the error/blocker, and Ticket ID.
-ESCALATION
-
-# Researcher sections
-RESEARCHER_HEADER = <<~HEADER
-You are the YATFA RESEARCHER agent operating in FULLY AUTONOMOUS MODE.
-Your role is AUTONOMOUS ANALYSIS and PROPOSAL GENERATION.
-
-### CORE CONSTRAINT
-You have READ-ONLY access to code, tickets, and memories.
-**YOU MUST NOT MODIFY CODE, FILES, OR TICKETS DIRECTLY.**
-You must use `create_proposal` to suggest actions.
-
-### SESSION ENVIRONMENT
-- **Execution Model:** Event-driven. Do not wait, do not poll. When you receive a message: Analyze -> Create Proposals -> STOP.
-- **Access:** Sandboxed Docker container (Root privileges). Read access to all systems.
-- **Skills:** researcher-tactical, researcher-strategic, knowledge-management, memory.
-HEADER
-
-RESEARCHER_BOUNDARIES = <<~BOUNDARIES
-### ROLE BOUNDARIES
-
-**ABSOLUTELY FORBIDDEN:**
-- Modifying code, tickets, or files directly.
-- Writing, editing, or refactoring code.
-- Making git commits or pull requests.
-- Sending messages to other agents.
-- Changing your own busy/idle status.
-BOUNDARIES
-
-RESEARCHER_ESCALATION = <<~ESCALATION
-### ESCALATION PROCESS
-If you encounter blocking problems or need to improve workflow:
-1. Use `create_proposal`.
-2. **Title:** "Escalation: [brief description]"
-3. **Type:** task
-4. **Priority:** high or critical
-5. **Context:** Include what you tried, the error/blocker, and suggested fix (e.g., "Escalation: Need additional MCP tools for research").
-ESCALATION
-
-CLAW_HEADER = <<~HEADER
-You are the **YATFA CLAW** agent.
-
-**ROLE:** OpenClaw Gateway Runner.
-Your purpose is to run an OpenClaw gateway inside this container, providing a Telegram-based chat interface with the Pi agent.
-
-### OPERATIONAL MODE
-*   **Execution:** The OpenClaw gateway starts automatically and runs persistently.
-*   **Telegram Integration:** Chat messages from Telegram are forwarded to the Pi agent via the gateway.
-*   **Configuration:** Set via environment variables or `~/.openclaw/openclaw.json`.
-*   **Persistence:** Gateway data is stored in a Docker volume at `~/.openclaw/`.
-
-### ENVIRONMENT
-*   Sandboxed Docker container with ROOT privileges.
-*   Network mode: host (port 18789 exposed for gateway).
-*   OpenClaw installed globally via npm.
-HEADER
-
-CLAW_BOUNDARIES = <<~BOUNDARIES
-### ROLE BOUNDARIES
-*   This agent runs the OpenClaw gateway - it does NOT run Claude Code.
-*   No Yatfa MCP tools are available (that is a future enhancement).
-*   No git repositories are mounted.
-*   No skills are loaded.
-BOUNDARIES
-
-# Dynamic banner builder
-def build_banner(header:, boundaries:, escalation: nil, extra_sections: [])
-  [
-    header,
-    UNIVERSAL_CONSTRAINTS,
-    TASK_TRACKING,
-    *extra_sections,
-    boundaries,
-    escalation
-  ].compact.join("\n\n")
-end
+# Agent type definitions for the YATFA system.
+#
+# System prompts (banners) are defined as .md files in yatfa/prompts/<agent-type>/base.md
+# and loaded into the database by AgentPromptLoader. At container startup,
+# setup_system_prompt! fetches the prompt from the API instead of evaluating this file.
+#
+# This file defines agent metadata (name, skills, prompt reference) used by:
+#   - bin/run-yatfa-agent.rb (CLI) — to build Docker containers with correct skills
+#   - spec/system_prompt_spec.rb — to verify agent config structure
 
 AGENT_CONFIGS = {
   'planner' => {
     name: 'yatfa-planner',
     skills: ['ticket-management', 'memory', 'knowledge-management', 'sentry-investigation', 'agent-browser'],
-    banner: build_banner(
-      header: PLANNER_HEADER,
-      boundaries: PLANNER_BOUNDARIES
-    )
+    prompt_name: 'planner'
   },
 
   'worker' => {
     name: 'yatfa-autonomous-worker',
     skills: ['git-workflow', 'worker-workflow', 'memory', 'session-learnings', 'screenshot', 'agent-browser'],
-    banner: build_banner(
-      header: WORKER_HEADER,
-      boundaries: WORKER_BOUNDARIES,
-      escalation: WORKER_ESCALATION
-    )
+    prompt_name: 'worker'
   },
 
   'reviewer' => {
     name: 'yatfa-autonomous-reviewer',
     skills: ['reviewer-workflow', 'qa-workflow', 'memory', 'session-learnings', 'proposal-reviewer', 'screenshot', 'agent-browser'],
-    banner: build_banner(
-      header: REVIEWER_HEADER,
-      boundaries: REVIEWER_BOUNDARIES,
-      escalation: REVIEWER_ESCALATION
-    )
+    prompt_name: 'reviewer'
   },
 
   'researcher' => {
     name: 'yatfa-autonomous-researcher',
     skills: ['researcher-tactical', 'researcher-strategic', 'researcher-digest', 'researcher-telegram-processor', 'researcher-meta-learning', 'researcher-approval-review', 'researcher-task-proposal-review', 'memory', 'session-learnings', 'proposal-execution', 'proposal-rework-responder', 'memory-consolidation', 'retrospective', 'knowledge-management', 'sentry-investigation', 'agent-browser'],
-    banner: build_banner(
-      header: RESEARCHER_HEADER,
-      boundaries: RESEARCHER_BOUNDARIES,
-      escalation: RESEARCHER_ESCALATION
-    )
+    prompt_name: 'researcher'
   },
 
   'claw' => {
     name: 'yatfa-openclaw',
     skills: [],
-    banner: build_banner(
-      header: CLAW_HEADER,
-      boundaries: CLAW_BOUNDARIES
-    )
+    prompt_name: 'claw'
   }
 }.freeze

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "yatfa",
-  "version": "1.0.87",
+  "version": "1.0.88",
   "description": "YATFA - Yet Another Tool For Agents",
   "bin": "./bin/run-yatfa-agent.rb",
   "files": [

package/setup/setup-agent.rb

@@ -315,19 +315,12 @@ def setup_claude_settings!
 
     # Configure model selection if specified
     # Each model tier can be configured independently
-    if llm_credentials["model_haiku"] && !llm_credentials["model_haiku"].empty?
-      settings_config["env"]["ANTHROPIC_DEFAULT_HAIKU_MODEL"] = llm_credentials["model_haiku"]
-      puts "   Haiku model: #{llm_credentials['model_haiku']}"
-    end
-
-    if llm_credentials["model_sonnet"] && !llm_credentials["model_sonnet"].empty?
-      settings_config["env"]["ANTHROPIC_DEFAULT_SONNET_MODEL"] = llm_credentials["model_sonnet"]
-      puts "   Sonnet model: #{llm_credentials['model_sonnet']}"
-    end
-
-    if llm_credentials["model_opus"] && !llm_credentials["model_opus"].empty?
-      settings_config["env"]["ANTHROPIC_DEFAULT_OPUS_MODEL"] = llm_credentials["model_opus"]
-      puts "   Opus model: #{llm_credentials['model_opus']}"
+    %w[haiku sonnet opus].each do |tier|
+      key = "model_#{tier}"
+      if llm_credentials[key] && !llm_credentials[key].empty?
+        settings_config["env"]["ANTHROPIC_DEFAULT_#{tier.upcase}_MODEL"] = llm_credentials[key]
+        puts "   #{tier.capitalize} model: #{llm_credentials[key]}"
+      end
     end
 
     # Configure context window if specified (value is in k-units, append 'k' suffix)
@@ -339,7 +332,7 @@ def setup_claude_settings!
       end
     end
 
-    if !llm_credentials["model_haiku"] && !llm_credentials["model_sonnet"] && !llm_credentials["model_opus"]
+    unless %w[haiku sonnet opus].any? { |tier| llm_credentials["model_#{tier}"] }
       puts "   Using default Anthropic models"
     end
   else
@@ -353,42 +346,22 @@ def setup_claude_settings!
 
   # Configure hooks
   settings_config["hooks"] ||= {}
-  
-  
-  # Configure SessionStart hook for skill knowledge injection
-  settings_config["hooks"]["SessionStart"] ||= []
-  # Remove old hook if present
-  settings_config["hooks"]["SessionStart"].reject! do |h| 
-    h["hooks"]&.first&.dig("command")&.include?("inject-skill-knowledge.rb")
-  end
-
-  # Install hook script to ~/.claude/hooks
   hooks_dir = File.join(settings_dir, "hooks")
   FileUtils.mkdir_p(hooks_dir)
-  
-  session_script_name = "inject-skill-knowledge.rb"
-  dest_path = download_file(
-    script_name: session_script_name,
-    remote_url: "#{YATFA_RAW_URL}/hooks/#{session_script_name}",
-    dest_dir: hooks_dir,
-    label: "SessionStart hook"
-  )
 
-  if dest_path
-    session_hook_to_add = {
-      "hooks" => [{ "type" => "command", "command" => "ruby \"#{dest_path}\"" }]
-    }
-    settings_config["hooks"]["SessionStart"] << session_hook_to_add
-  else
-    puts "⚠️ Skipping SessionStart hook configuration (download failed)"
-  end
+  # SessionStart hook: skill knowledge injection
+  install_hook!(
+    settings_config, hooks_dir,
+    hook_type: "SessionStart",
+    script_name: "inject-skill-knowledge.rb",
+    match_substring: "inject-skill-knowledge.rb",
+    command_prefix: "ruby"
+  )
 
   # Install sessionstart-input.sh for repository context injection
-  # This provides repository information to Claude at session start
-  input_script_name = "sessionstart-input.sh"
   download_file(
-    script_name: input_script_name,
-    remote_url: "#{YATFA_RAW_URL}/hooks/#{input_script_name}",
+    script_name: "sessionstart-input.sh",
+    remote_url: "#{YATFA_RAW_URL}/hooks/sessionstart-input.sh",
     dest_dir: hooks_dir,
     label: "SessionStart input hook"
   )
@@ -396,31 +369,15 @@ def setup_claude_settings!
   # Add agent-specific settings
   settings_config["skipDangerousModePermissionPrompt"] = true
 
-  # Configure Stop hook for session-end learning capture
-  settings_config["hooks"]["Stop"] ||= []
-  # Remove old hook if present
-  settings_config["hooks"]["Stop"].reject! do |h|
-    h["hooks"]&.first&.dig("command")&.include?("sessionend-output.sh")
-  end
-
-  # Install session-end hook script
-  sessionend_script_name = "sessionend-output.sh"
-  sessionend_dest_path = download_file(
-    script_name: sessionend_script_name,
-    remote_url: "#{YATFA_RAW_URL}/hooks/#{sessionend_script_name}",
-    dest_dir: hooks_dir,
-    label: "Stop hook"
+  # Stop hook: session-end learning capture
+  install_hook!(
+    settings_config, hooks_dir,
+    hook_type: "Stop",
+    script_name: "sessionend-output.sh",
+    match_substring: "sessionend-output.sh",
+    command_prefix: "bash"
   )
 
-  if sessionend_dest_path
-    sessionend_hook_to_add = {
-      "hooks" => [{ "type" => "command", "command" => "bash \"#{sessionend_dest_path}\"" }]
-    }
-    settings_config["hooks"]["Stop"] << sessionend_hook_to_add
-  else
-    puts "⚠️ Skipping Stop hook configuration (download failed)"
-  end
-
   # Download and register project hooks wrapper
   # The wrapper fetches user-defined hooks from the API at runtime (with caching),
   # executes them, and reports results back for observability.
@@ -430,6 +387,39 @@ def setup_claude_settings!
   puts "✅ ~/.claude/settings.json configured with skill hooks"
 end
 
+# Install a Claude settings hook by downloading the script and registering it.
+#
+# Handles the common pattern: ensure hook array → remove old entry → download → append.
+#
+# @param settings_config [Hash] the settings.json hash being built
+# @param hooks_dir [String] directory to download hook scripts into
+# @param hook_type [String] Claude hook type (e.g. "SessionStart", "Stop")
+# @param script_name [String] filename of the hook script to download
+# @param match_substring [String] substring to match when removing old hook entries
+# @param command_prefix [String] command prefix ("ruby" or "bash")
+def install_hook!(settings_config, hooks_dir, hook_type:, script_name:, match_substring:, command_prefix:)
+  settings_config["hooks"][hook_type] ||= []
+  # Remove old hook entry if present (idempotent re-runs)
+  settings_config["hooks"][hook_type].reject! do |h|
+    h["hooks"]&.first&.dig("command")&.include?(match_substring)
+  end
+
+  dest_path = download_file(
+    script_name: script_name,
+    remote_url: "#{YATFA_RAW_URL}/hooks/#{script_name}",
+    dest_dir: hooks_dir,
+    label: "#{hook_type} hook"
+  )
+
+  if dest_path
+    settings_config["hooks"][hook_type] << {
+      "hooks" => [{ "type" => "command", "command" => "#{command_prefix} \"#{dest_path}\"" }]
+    }
+  else
+    puts "⚠️ Skipping #{hook_type} hook configuration (download failed)"
+  end
+end
+
 # Download and register the project hooks wrapper as a SessionStart hook.
 # The wrapper handles fetching user-defined hooks from the API at runtime,
 # executing them, and reporting results. Hook content is managed by users
@@ -456,39 +446,57 @@ def setup_system_prompt!
   agent_type = ENV["AGENT_TYPE"]
   system_prompt_path = File.expand_path("~/.yatfa/system-prompt.txt")
 
-  # Download agents.rb from remote
-  agents_rb_url = "#{YATFA_RAW_URL}/agents.rb"
-  puts "📥 Downloading agents.rb from #{agents_rb_url}..."
-  agents_rb_content = nil
-  begin
-    agents_rb_content = URI.open(agents_rb_url).read
-    puts "✅ Downloaded agents.rb"
-  rescue OpenURI::HTTPError => e
-    puts "❌ Failed to download agents.rb: #{e.message}"
-    puts "   Cannot generate system prompt without agent definitions"
+  rails_api_url = ENV["YATFA_API_URL"]
+  rails_api_key = ENV["YATFA_API_KEY"]
+
+  unless rails_api_url && !rails_api_url.empty? && rails_api_key && !rails_api_key.empty?
+    puts "❌ YATFA_API_URL or YATFA_API_KEY not set"
+    puts "   Cannot fetch system prompt without API credentials"
     exit 1
   end
 
-  # Evaluate agents.rb to get AGENT_CONFIGS
-  agents_binding = binding
+  # Fetch system prompt from yatfa API
+  # Agent prompts are stored in yatfa/prompts/<agent-type>/base.md
+  # and loaded into the database by AgentPromptLoader
+  prompt_url = URI("#{rails_api_url}/internal/skill?tags%5B%5D=#{agent_type}&tags%5B%5D=base")
+  http = Net::HTTP.new(prompt_url.host, prompt_url.port)
+  http.use_ssl = prompt_url.scheme == "https"
+  http.open_timeout = 10
+  http.read_timeout = 15
+
+  request = Net::HTTP::Get.new(prompt_url)
+  request["X-API-Key"] = rails_api_key
+  request["Accept"] = "application/json"
+
+  banner_content = nil
   begin
-    agents_binding.eval(agents_rb_content, "agents.rb")
-  rescue => e
-    puts "❌ Failed to evaluate agents.rb: #{e.message}"
-    exit 1
-  end
+    puts "📥 Fetching system prompt for #{agent_type}..."
+    response = http.request(request)
 
-  agent_configs = agents_binding.eval("AGENT_CONFIGS")
+    if response.is_a?(Net::HTTPSuccess)
+      data = JSON.parse(response.body)
+      banner_content = data["content"]
+
+      unless banner_content && !banner_content.empty?
+        puts "❌ System prompt for '#{agent_type}' is empty"
+        exit 1
+      end
 
-  unless agent_configs && agent_configs[agent_type]
-    puts "❌ Unknown agent type: #{agent_type}"
-    puts "   Available: #{agent_configs.keys.join(', ')}"
+      puts "✅ Fetched system prompt for #{agent_type}"
+    else
+      puts "❌ Failed to fetch system prompt: #{response.code} #{response.message}"
+      puts "   Response: #{response.body}"
+      puts "   Make sure AgentPromptLoader has loaded prompts/ directory"
+      exit 1
+    end
+  rescue Net::OpenTimeout, Net::ReadTimeout => e
+    puts "❌ Timeout fetching system prompt: #{e.message}"
+    exit 1
+  rescue => e
+    puts "❌ Error fetching system prompt: #{e.message}"
     exit 1
   end
 
-  agent_def = agent_configs[agent_type]
-  banner_content = agent_def[:banner].dup
-
   # Append repository context from REPOSITORIES_CONTEXT env var
   if ENV["REPOSITORIES_CONTEXT"] && !ENV["REPOSITORIES_CONTEXT"].empty?
     repos = ENV["REPOSITORIES_CONTEXT"].split(',').map do |repo|
@@ -506,7 +514,7 @@ def setup_system_prompt!
   # Ensure directory exists and write system prompt
   FileUtils.mkdir_p(File.dirname(system_prompt_path))
   File.write(system_prompt_path, banner_content)
-  puts "✅ System prompt generated at #{system_prompt_path} (agent: #{agent_type})"
+  puts "✅ System prompt written to #{system_prompt_path} (agent: #{agent_type})"
 
   # Add repository context to CLAUDE.md for easy reference
   claude_md = File.expand_path("~/CLAUDE.md")

package/setup/setup-github.rb

@@ -47,7 +47,7 @@ def setup_github_auth!
         end
 
         # Build GitHub token endpoint URL
-        token_url = "#{rails_api_url.chomp('/')}/github/token"
+        token_url = "\#{rails_api_url.chomp('/')}/github/token"
 
         uri = URI(token_url)
         http = Net::HTTP.new(uri.host, uri.port)
@@ -62,7 +62,7 @@ def setup_github_auth!
         response = http.request(request)
 
         unless response.is_a?(Net::HTTPSuccess)
-          STDERR.puts "❌ Error: Backend API returned #{response.code} for GitHub token"
+          STDERR.puts "❌ Error: Backend API returned \#{response.code} for GitHub token"
           exit 1
         end