feat: add interactive command mode, :help/:usage commands, and token usage tracking

[seagochen]

📝 Description

This PR adds several new interactive features to the PicoClaw CLI agent:

1. Interactive Command Mode (:cmd / :pico): Users can now switch between AI chat mode and a direct shell command execution mode. Command mode supports cd with working directory tracking, cross-platform shell execution (sh on Unix, PowerShell on Windows), and persists the working directory across commands.

2. AI-Assisted Command Mode (:hipico): From within command mode, users can invoke one-shot AI assistance. The current working directory is injected into the system prompt so the AI resolves paths relative to the cmd working directory.

3. :help Command: A printInteractiveHelp() function that displays detailed usage for all modes and commands, available in any interactive mode.

4. :usage Command: Displays current model name, max tokens, temperature, and accumulated session token statistics (prompt tokens, completion tokens, total requests).

5. Token Usage Tracking: AgentInstance now tracks TotalPromptTokens, TotalCompletionTokens, and TotalRequests via atomic counters, accumulated after each LLM call via AddUsage().

6. Refactoring: Renamed printHelp → printCliHelp (top-level CLI help) and printInteractiveHelp (interactive mode help) to eliminate function name collisions.

🗣️ Type of Change

• 🐞 Bug fix (non-breaking change which fixes an issue)
• ✨ New feature (non-breaking change which adds functionality)
• 📖 Documentation update
• ⚡ Code refactoring (no functional changes, no api changes)

🤖 AI Code Generation

• 🤖 Fully AI-generated (100% AI, 0% Human)
• 🛠️ Mostly AI-generated (AI draft, Human verified/modified)
• 👨‍💻 Mostly Human-written (Human lead, AI assisted or none)

🔗 Related Issue

📚 Technical Context (Skip for Docs)

• Reference URL: https://claude.ai/code/session_01PoHBfe7eKmhHvVF12W3xZ2
• Reasoning: Improve developer experience by providing a built-in terminal-like command mode alongside the AI chat interface, making PicoClaw more versatile as a day-to-day assistant.

🧪 Test Environment

• Hardware:
• OS:
• Model/Provider:
• Channels:

📸 Evidence (Optional)

Click to view Logs/Screenshots

☑️ Checklist

• My code/docs follow the style of this project.
• I have performed a self-review of my own changes.
• I have updated the documentation accordingly.

[nikolasdehor]

This is a 1068-line PR that adds interactive command mode, file editing, token tracking, and several UX features. It is doing too many things at once and would benefit from being split into separate PRs (token tracking, command mode, edit command, ls formatting). Several significant issues:

Security Issues:

1. :edit bypasses workspace restriction: The resolveEditPath function allows absolute paths (/etc/passwd) and ~/ expansion to the real home directory (not the workspace). This means from any channel (WhatsApp, Telegram, Discord), a user can read and write arbitrary files on the host. The cd command correctly restricts absolute paths to the workspace, but :edit does not.

2. :edit -m creates files with MkdirAll: Combined with the above, a user can create arbitrary directories and files anywhere on the filesystem.

3. Shell command execution via command mode: executeCmdMode calls agent.Tools.ExecuteWithContext which should go through the ExecTool safety guard, but the command mode concept itself needs careful thought about access control — especially when exposed over multi-user channels like Telegram groups.

Correctness Issues:

4. Global mutable state cmdWorkingDir: The package-level var cmdWorkingDir with an init() function is only used by the CLI executeShellCommand but is global state. If the agent loop is used from the gateway (non-CLI), this is unused. More importantly, the per-session working directory in loop.go (via getSessionWorkDir/setSessionWorkDir) and the global cmdWorkingDir are two separate tracking mechanisms for the same concept.

5. Duplicated logic: The mode switching logic (interactiveMode and simpleInteractiveMode) is nearly identical — 100+ lines of duplicated switch/case blocks. Extract a shared handler function.

6. : commands intercepted globally: handleCommand now intercepts ALL messages starting with : across ALL channels. If a Telegram user sends :) or :thinking:, it will return "Unknown command: :)". This breaks normal chat usage. The : prefix should only be recognized in specific contexts (CLI interactive mode), not in the agent loop's global command handler.

7. ensureLsLong silently modifies user commands: Injecting -l into ls commands changes the output format. This is surprising behavior — if someone types ls *.go they expect filename-only output, not a long listing. The emoji formatting is a nice touch but should be opt-in.

Missing:

8. No tests: 1068 lines of new code with zero test coverage. At minimum, editLineOp, ensureLsLong, isPermString, resolveEditPath, handleCdCommand, and shortenHomePath should have unit tests.

9. No documentation: The PR adds significant user-facing features but does not update the README or docs.