Agentic Mode

LLMs explore your codebase and propose edits

Agentic mode gives LLMs tools to explore your codebase, then propose file edits with permission prompts — like Claude Code, but for any LLM.

Usage

The TUI runs in agentic mode by default. Just type your task:

fix the bug in src/utils.ts
add error handling to api/routes.ts
create a hello world script

Set the agent for your tasks:

/agent claude    # Use Claude
/agent gemini    # Use Gemini
/agent ollama    # Use local Ollama

See Model Selection to configure which model each agent uses.

Available Tools

Tool	Description
view	Read file contents with line numbers
glob	Find files by pattern (e.g., **/*.ts)
grep	Search file contents with regex
bash	Execute shell commands
write	Create or overwrite files
edit	Search and replace in files

Approval Modes

Control how permissions and file edits are handled with /approval-mode:

Mode	Permissions	File Edits	Use Case
Default	Ask for each	Diff review	Normal development
Plan	Ask for each	Show plan only	Planning/reviewing
Accept	Ask for each	Auto-apply	Faster iteration
YOLO	Auto-approve	Auto-apply	Full trust, max speed

Permission Prompts

In Default and Plan modes, every tool call shows a permission prompt:

• Allow — Execute this tool call
• Allow from directory — Auto-approve reads from this directory
• Allow all reads — Auto-approve all file reads
• Allow all edits — Auto-approve all file writes for this session
• Deny — Skip this tool call
• Esc — Cancel entire operation

”Allow all edits” persists across messages in the same session — useful when you trust the agent to make multiple changes.

Trust Management

Pre-approve directories so the agent can read files without prompts:

/trusted                  # List trusted directories
/trusted add ./src        # Trust the src folder
/trusted add .            # Trust current directory
/trusted remove ./src     # Remove trust

Files in trusted directories are auto-approved for read operations.

Live Tool Activity

When the LLM uses tools, you see:

• ● Green = completed
• ● Yellow = running
• ● Red = error
• ● Gray = pending

Press Ctrl+S to expand/collapse full output.

How It Works

1. Describe the task

   Type your request in the TUI

2. LLM explores

   The LLM uses view, glob, grep to understand your codebase

3. Permission prompts

   Each tool call shows a prompt — you approve or deny

4. LLM proposes edits

   The LLM uses write/edit to propose changes

5. Review diff

   You see a diff preview and approve or reject each change

6. Apply changes

   Approved changes are written to disk

Consensus → Agentic Workflow

Run consensus first, then continue with the agent:

/consensus claude,gemini "best approach for auth"
# Choose "Continue"
# Type your follow-up — has consensus context

The consensus result is automatically injected as context.

Agent Support

Agent	Agentic Support
Claude	Full support — uses our tool system
Ollama	Full support — uses our tool system
Gemini	Partial — auto-reads files (bypasses read prompts)
Codex	Safe mode — backup/rollback for file changes
Mistral	Experimental — inconsistent tool support

Codex Safe Mode

Codex runs in “safe mode” by default. Since Codex’s exec mode writes files directly (bypassing our tool system), PuzldAI:

1. Backs up source files before Codex runs
2. Runs Codex with workspace-write sandbox
3. Detects changes by comparing files after execution
4. Applies approval mode:
   • Plan mode → rolls back all changes
   • Accept/YOLO → keeps changes
   • Default → shows notification, keeps changes

Codex changes are applied immediately, then rolled back if rejected. This differs from Claude/Ollama where you approve before changes are written.