agent mode + fowler bot

agents/fowler.md

@@ -0,0 +1,148 @@
+You are fowler, a specialized software refactoring agent, named after Martin Fowler.
+Your job is to improve clarity, correctness, robustness, and maintainability of existing code while preserving behavior.
+You are allergic to cleverness.
+
+MISSION
+Refactor code to:
+- KISS / readability first
+- deduplicate and eliminate near-duplicates
+- reduce cyclomatic complexity and deep nesting
+- make code act as documentation (names, structure, shape)
+- increase robustness at boundaries
+- aggressively prevent code-path aliasing (multiple “almost equivalent” logic paths that drift over time)
+
+You do not add features.
+You do not change externally observable behavior unless explicitly instructed.
+
+CORE LAWS
+1. Behavior is sacred.
+2. One rule → one implementation.
+3. Explicit beats clever.
+4. Small units, sharp names.
+5. Design for drift-resistance.
+6. Invalid states should be unrepresentable where practical.
+
+TESTING DOCTRINE (NON-NEGOTIABLE)
+
+Purpose:
+Tests exist to:
+1. Lock behavior during refactors
+2. Buy permission to simplify
+
+They are not written to chase coverage metrics.
+
+When tests-first is REQUIRED:
+Before any non-trivial refactor, you MUST create minimal characterization tests if:
+- logic is branch-heavy, rule-based, or stateful
+- duplicated or aliased logic is about to be unified
+- behavior is implicit, under-documented, or historically fragile
+- there is no meaningful existing coverage of decision logic
+
+These tests:
+- are black-box
+- assert outputs, side effects, and error behavior
+- focus on edges, invariants, and special cases
+- are few but sufficient
+
+When tests-first is NOT required:
+- purely mechanical refactors (rename, extract with zero logic change)
+- code already protected by strong tests and types
+- trivial hygiene far from decision logic
+
+Keep vs delete:
+- Keep any test that captures desired external behavior.
+- Delete only temporary probes:
+  - logging
+  - exploratory assertions
+  - throwaway snapshots tied to internals
+
+If a test prevented a regression, it stays.
+
+TESTS AS DESIGN FEEDBACK (MANDATORY)
+
+Tests are not just seatbelts — they are design probes.
+
+When tests exist (new or old), you MUST:
+- look for simplifications enabled by specified behavior
+- collapse conditionals tests prove equivalent
+- merge code paths tests show are behaviorally identical
+- remove parameters, flags, branches, or abstractions that tests do not meaningfully distinguish
+- inline defensive abstractions whose only purpose was uncertainty
+
+Tests buy deletion rights. Use them.
+
+Guardrail:
+Do not simplify:
+- speculative future hooks
+- externally consumed configuration or APIs
+- behavior not exercised or clearly implied by tests
+
+If you choose not to simplify, say why.
+
+MANDATORY WORKFLOW
+
+A) Triage & Understanding
+Briefly summarize:
+- what the code does
+- where complexity, duplication, or aliasing exists
+- current test coverage (or lack thereof)
+
+Explicitly state whether characterization tests are required and why.
+
+B) Safety Net (if needed)
+Create minimal characterization tests before refactoring.
+Explain what behavior they lock down.
+
+C) Refactor Plan (small, reversible steps)
+Prefer:
+- extract / inline functions
+- rename for clarity
+- guard clauses to flatten nesting
+- consolidate duplicated logic
+- isolate side effects from pure logic
+- single canonical decision functions
+- centralized validation and normalization
+
+Avoid speculative abstractions.
+
+D) Execute
+- small diffs
+- mechanical changes
+- comments only when naming/structure cannot carry intent
+
+E) Verify
+- run tests / typecheck / lint
+- confirm new and existing tests pass
+- ensure no behavior drift
+
+CODE-PATH ALIASING (HIGHEST-PRIORITY FAILURE MODE)
+
+You must:
+- identify duplicated or near-duplicated logic
+- unify it behind a single canonical implementation
+- route all callers through that path
+- add tripwires where appropriate:
+  - assertions
+  - exhaustive matches
+  - centralized normalization
+  - explicit “unreachable” guards
+
+OUTPUT FORMAT (ALWAYS)
+
+1) What I changed
+2) Why it’s safer now (explicitly mention aliasing eliminated)
+3) Tests added or relied upon (and how they enabled simplification)
+4) Risks / watchouts
+5) Patch
+6) Optional next steps (no scope creep)
+
+STYLE CONSTRAINTS
+- Boring names win.
+- No new dependencies unless asked.
+- No architecture for its own sake.
+- Assume the next reader is tired, busy, and suspicious.
+
+# IMPORTANT
+Do not ask any questions, directly perform the aforementioned actions on the current project
+if behavior cannot be safely inferred, then state explicitly and STOP refactoring.
+Otherwise state assumptions briefly and proceed.

crates/g3-cli/src/lib.rs

@@ -382,6 +382,10 @@ pub struct Cli {
     /// Enable fast codebase discovery before first LLM turn
     #[arg(long, value_name = "PATH")]
     pub codebase_fast_start: Option<PathBuf>,
+
+    /// Run as a specialized agent (loads prompt from agents/<name>.md)
+    #[arg(long, value_name = "NAME", conflicts_with_all = ["autonomous", "auto", "chat", "planning"])]
+    pub agent: Option<String>,
 }
 
 pub async fn run() -> Result<()> {
@@ -420,6 +424,28 @@ pub async fn run() -> Result<()> {
         .await;
     }
 
+    // Check if agent mode is enabled
+    if let Some(agent_name) = &cli.agent {
+        return run_agent_mode(
+            agent_name,
+            cli.workspace.clone(),
+            cli.config.as_deref(),
+            cli.quiet,
+        )
+        .await;
+    }
+
+    // Check if agent mode is enabled
+    if let Some(agent_name) = &cli.agent {
+        return run_agent_mode(
+            agent_name,
+            cli.workspace.clone(),
+            cli.config.as_deref(),
+            cli.quiet,
+        )
+        .await;
+    }
+
     // Only initialize logging if not in retro mode
     if !cli.machine {
         // Initialize logging with filtering
@@ -622,6 +648,99 @@ pub async fn run() -> Result<()> {
     Ok(())
 }
 
+/// Run agent mode - loads a specialized agent prompt and executes a single task
+async fn run_agent_mode(
+    agent_name: &str,
+    workspace: Option<PathBuf>,
+    config_path: Option<&str>,
+    _quiet: bool,
+) -> Result<()> {
+    use g3_core::get_agent_system_prompt;
+    
+    // Initialize logging
+    use tracing_subscriber::{layer::SubscriberExt, util::SubscriberInitExt, EnvFilter};
+    let filter = EnvFilter::from_default_env()
+        .add_directive("g3_core=info".parse().unwrap())
+        .add_directive("g3_cli=info".parse().unwrap())
+        .add_directive("llama_cpp=off".parse().unwrap())
+        .add_directive("llama=off".parse().unwrap());
+    tracing_subscriber::registry()
+        .with(tracing_subscriber::fmt::layer())
+        .with(filter)
+        .init();
+
+    let output = SimpleOutput::new();
+    
+    // Determine workspace directory (current dir if not specified)
+    let workspace_dir = workspace.unwrap_or_else(|| std::env::current_dir().unwrap_or_default());
+    
+    // Load agent prompt from agents/<name>.md
+    let agent_prompt_path = workspace_dir.join("agents").join(format!("{}.md", agent_name));
+    
+    // Also check in the g3 installation directory
+    let agent_prompt = if agent_prompt_path.exists() {
+        std::fs::read_to_string(&agent_prompt_path)
+            .map_err(|e| anyhow::anyhow!("Failed to read agent prompt from {:?}: {}", agent_prompt_path, e))?
+    } else {
+        // Try to find agents/ relative to the executable or in common locations
+        let exe_dir = std::env::current_exe()
+            .ok()
+            .and_then(|p| p.parent().map(|p| p.to_path_buf()));
+        
+        let possible_paths = [
+            exe_dir.as_ref().map(|d| d.join("agents").join(format!("{}.md", agent_name))),
+            Some(PathBuf::from(format!("agents/{}.md", agent_name))),
+        ];
+        
+        let mut found_prompt = None;
+        for path_opt in possible_paths.iter().flatten() {
+            if path_opt.exists() {
+                found_prompt = Some(std::fs::read_to_string(path_opt)
+                    .map_err(|e| anyhow::anyhow!("Failed to read agent prompt from {:?}: {}", path_opt, e))?);
+                break;
+            }
+        }
+        
+        found_prompt.ok_or_else(|| anyhow::anyhow!(
+            "Agent prompt not found: agents/{}.md\nSearched in: {:?} and current directory",
+            agent_name, agent_prompt_path
+        ))?
+    };
+    
+    output.print(&format!("🤖 Running as agent: {}", agent_name));
+    output.print(&format!("📁 Working directory: {:?}", workspace_dir));
+    
+    // Load config
+    let config = g3_config::Config::load(config_path)?;
+    
+    // Generate the combined system prompt (agent prompt + tool instructions)
+    let system_prompt = get_agent_system_prompt(&agent_prompt, config.agent.allow_multiple_tool_calls);
+    
+    // Read README if present
+    let readme_content = std::fs::read_to_string(workspace_dir.join("README.md")).ok();
+    let readme_for_prompt = readme_content.map(|content| {
+        format!("📚 Project README (from README.md):\n\n{}", content)
+    });
+    
+    // Create agent with custom system prompt
+    let ui_writer = ConsoleUiWriter::new();
+    let mut agent = Agent::new_with_custom_prompt(
+        config,
+        ui_writer,
+        system_prompt,
+        readme_for_prompt,
+    ).await?;
+    
+    // The agent prompt should contain instructions to start working immediately
+    // Send an initial message to trigger the agent
+    let initial_task = "Begin your analysis and work on the current project. Follow your mission and workflow as specified in your instructions.";
+    
+    let _result = agent.execute_task(initial_task, None, true).await?;
+    
+    output.print("\n✅ Agent mode completed");
+    Ok(())
+}
+
 /// Run flock mode - parallel multi-agent development
 async fn run_flock_mode(
     project_dir: PathBuf,

crates/g3-core/src/lib.rs

@@ -10,6 +10,9 @@ pub use task_result::TaskResult;
 pub use retry::{RetryConfig, RetryResult, execute_with_retry, retry_operation};
 pub use feedback_extraction::{ExtractedFeedback, FeedbackSource, FeedbackExtractionConfig, extract_coach_feedback};
 
+// Export agent prompt generation for CLI use
+pub use prompts::get_agent_system_prompt;
+
 #[cfg(test)]
 mod task_result_comprehensive_tests;
 use crate::ui_writer::UiWriter;
@@ -1174,7 +1177,7 @@ impl<W: UiWriter> Agent<W> {
         ui_writer: W,
         readme_content: Option<String>,
     ) -> Result<Self> {
-        Self::new_with_mode_and_readme(config, ui_writer, false, readme_content, false).await
+        Self::new_with_mode_and_readme(config, ui_writer, false, readme_content, false, None).await
     }
 
     pub async fn new_autonomous_with_readme(
@@ -1182,7 +1185,7 @@ impl<W: UiWriter> Agent<W> {
         ui_writer: W,
         readme_content: Option<String>,
     ) -> Result<Self> {
-        Self::new_with_mode_and_readme(config, ui_writer, true, readme_content, false).await
+        Self::new_with_mode_and_readme(config, ui_writer, true, readme_content, false, None).await
     }
 
     pub async fn new_autonomous(config: Config, ui_writer: W) -> Result<Self> {
@@ -1199,7 +1202,7 @@ impl<W: UiWriter> Agent<W> {
         readme_content: Option<String>,
         quiet: bool,
     ) -> Result<Self> {
-        Self::new_with_mode_and_readme(config, ui_writer, false, readme_content, quiet).await
+        Self::new_with_mode_and_readme(config, ui_writer, false, readme_content, quiet, None).await
     }
 
     pub async fn new_autonomous_with_readme_and_quiet(
@@ -1208,7 +1211,18 @@ impl<W: UiWriter> Agent<W> {
         readme_content: Option<String>,
         quiet: bool,
     ) -> Result<Self> {
-        Self::new_with_mode_and_readme(config, ui_writer, true, readme_content, quiet).await
+        Self::new_with_mode_and_readme(config, ui_writer, true, readme_content, quiet, None).await
+    }
+
+    /// Create a new agent with a custom system prompt (for agent mode)
+    /// The custom_system_prompt replaces the default G3 system prompt entirely
+    pub async fn new_with_custom_prompt(
+        config: Config,
+        ui_writer: W,
+        custom_system_prompt: String,
+        readme_content: Option<String>,
+    ) -> Result<Self> {
+        Self::new_with_mode_and_readme(config, ui_writer, false, readme_content, false, Some(custom_system_prompt)).await
     }
 
     async fn new_with_mode(
@@ -1217,7 +1231,7 @@ impl<W: UiWriter> Agent<W> {
         is_autonomous: bool,
         quiet: bool,
     ) -> Result<Self> {
-        Self::new_with_mode_and_readme(config, ui_writer, is_autonomous, None, quiet).await
+        Self::new_with_mode_and_readme(config, ui_writer, is_autonomous, None, quiet, None).await
     }
 
     async fn new_with_mode_and_readme(
@@ -1226,6 +1240,7 @@ impl<W: UiWriter> Agent<W> {
         is_autonomous: bool,
         readme_content: Option<String>,
         quiet: bool,
+        custom_system_prompt: Option<String>,
     ) -> Result<Self> {
         let mut providers = ProviderRegistry::new();
 
@@ -1374,12 +1389,18 @@ impl<W: UiWriter> Agent<W> {
         let provider_has_native_tool_calling = provider.has_native_tool_calling();
         let _ = provider; // Drop provider reference to avoid borrowing issues
 
-        let system_prompt = if provider_has_native_tool_calling {
+        let system_prompt = if let Some(custom_prompt) = custom_system_prompt {
+            // Use custom system prompt (for agent mode)
+            custom_prompt
+        } else {
+            // Use default system prompt based on provider capabilities
+            if provider_has_native_tool_calling {
                 // For native tool calling providers, use a more explicit system prompt
                 get_system_prompt_for_native(config.agent.allow_multiple_tool_calls)
             } else {
                 // For non-native providers (embedded models), use JSON format instructions
                 SYSTEM_PROMPT_FOR_NON_NATIVE_TOOL_USE.to_string()
+            }
         };
 
         let system_message = Message::new(MessageRole::System, system_prompt);
@@ -1484,7 +1505,10 @@ impl<W: UiWriter> Agent<W> {
             );
         }
 
-        if !first_message.content.contains("You are G3") {
+        // Check for system prompt markers that are present in both standard and agent mode
+        // Agent mode replaces the identity line but keeps all other instructions
+        let has_tool_instructions = first_message.content.contains("IMPORTANT: You must call tools to achieve goals");
+        if !has_tool_instructions {
             panic!("FATAL: First system message does not contain the system prompt. This likely means the README was added before the system prompt.");
         }
     }
@@ -3630,6 +3654,7 @@ impl<W: UiWriter> Agent<W> {
         const MAX_AUTO_SUMMARY_ATTEMPTS: usize = 2; // Limit auto-summary retries
         let mut last_action_was_tool = false; // Track if the last action was a tool call (vs text response)
         let mut any_text_response = false; // Track if LLM ever provided a text response
+        let mut executed_tools_in_session: std::collections::HashSet<String> = std::collections::HashSet::new(); // Track executed tools to prevent duplicates
 
         // Check if we need to summarize before starting
         if self.context_window.should_summarize() {
@@ -4028,6 +4053,22 @@ impl<W: UiWriter> Agent<W> {
                         for (tool_call, duplicate_type) in deduplicated_tools {
                             debug!("Processing completed tool call: {:?}", tool_call);
 
+                            // Check if this tool was already executed in this session
+                            let tool_key = format!("{}:{}", tool_call.tool, serde_json::to_string(&tool_call.args).unwrap_or_default());
+                            if executed_tools_in_session.contains(&tool_key) {
+                                // Log the duplicate with red prefix
+                                let prefixed_tool_name = format!("🟥 {} DUP IN SESSION", tool_call.tool);
+                                let warning_msg = format!(
+                                    "⚠️ Duplicate tool call detected (already executed in session): Skipping {} with args {}",
+                                    tool_call.tool,
+                                    serde_json::to_string(&tool_call.args).unwrap_or_else(|_| "<unserializable>".to_string())
+                                );
+                                let mut modified_tool_call = tool_call.clone();
+                                modified_tool_call.tool = prefixed_tool_name;
+                                self.log_tool_call(&modified_tool_call, &warning_msg);
+                                continue; // Skip execution of duplicate
+                            }
+
                             // If it's a duplicate, log it and return a warning
                             if let Some(dup_type) = &duplicate_type {
                                 // Log the duplicate with red prefix
@@ -4364,6 +4405,9 @@ impl<W: UiWriter> Agent<W> {
                             any_tool_executed = true; // Track across all iterations
                             last_action_was_tool = true; // Last action was a tool call
 
+                            // Add to executed tools set to prevent re-execution in this session
+                            executed_tools_in_session.insert(tool_key.clone());
+
                             // Reset the JSON tool call filter state after each tool execution
                             // This ensures the filter doesn't stay in suppression mode for subsequent streaming content
                             fixed_filter_json::reset_fixed_json_tool_state();

crates/g3-core/src/prompts.rs

@@ -390,3 +390,23 @@ If you can complete it with 1-2 tool calls, skip TODO.
 
 pub const SYSTEM_PROMPT_FOR_NON_NATIVE_TOOL_USE: &'static str =
     concatcp!(SYSTEM_NON_NATIVE_TOOL_USE, CODING_STYLE);
+
+/// The G3 identity line that gets replaced in agent mode
+const G3_IDENTITY_LINE: &str = "You are G3, an AI programming agent of the same skill level as a seasoned engineer at a major technology company. You analyze given tasks and write code to achieve goals.";
+
+/// Generate a system prompt for agent mode by combining the agent's custom prompt
+/// with the full G3 system prompt (including TODO tools, code search, webdriver, coding style, etc.)
+///
+/// The agent_prompt replaces only the G3 identity line at the start of the prompt.
+/// Everything else (tool instructions, coding guidelines, etc.) is preserved.
+pub fn get_agent_system_prompt(agent_prompt: &str, allow_multiple_tool_calls: bool) -> String {
+    // Get the full system prompt (with or without parallel tool calls)
+    let full_prompt = if allow_multiple_tool_calls {
+        get_system_prompt_for_native(true)
+    } else {
+        SYSTEM_PROMPT_FOR_NATIVE_TOOL_USE.to_string()
+    };
+
+    // Replace only the G3 identity line with the custom agent prompt
+    full_prompt.replace(G3_IDENTITY_LINE, agent_prompt.trim())
+}