feat: add /tokens /context /cost commands to show token usage and context size #224

Feature Request: Token Usage & Context Visibility

It would be very useful to have visibility into token consumption and context size directly from the CLI, similar to Claude Code's /context and /cost commands.

Motivation

Currently, Deep Code already stores token usage data internally in sessions-index.json. According to the docs, each session entry tracks:

"token 用量 and 活跃 token 数"

However, there is no way to view this information from the CLI interface. Users have no visibility into:

• How many tokens they have consumed in the current session
• How much of the context window is being used
• What is consuming the most tokens (system prompt, conversation history, skills, MCP tools)

Proposed Commands

/tokens – Show current session token consumption

Displays the token usage breakdown for the current session:

• Input tokens (this turn)
• Output tokens (this turn)
• Thinking tokens (if thinking mode is enabled)
• Total tokens (entire session)
• Estimated cost (based on model pricing)

/context – Show context window breakdown

Displays what is consuming the context window:

• System prompt size (tokens)
• Conversation history size (# of messages, tokens)
• Skills loaded and their contribution
• Active MCP tools and their descriptions size
• Total context window usage (e.g., 15K / 128K / 1M)

/cost – Show cumulative session cost

Shows the estimated cost of the current session based on model pricing:

• Per-turn cost breakdown
• Total session cost
• Thinking vs non-thinking token cost

Why This Matters

1. Cost awareness – Helps users monitor spend when using paid API models (especially with reasoning/thinking tokens which add significant cost)
2. Debug context limits – When approaching context window limits, users can identify the culprit (bloated history, too many skills, etc.)
3. Optimization feedback – Allows users to measure the impact of prompt optimization, skill trimming, or MCP cleanup
4. Parity with alternatives – Tools like Claude Code already offer /context and /cost; adding these would reduce friction for users migrating to or evaluating Deep Code

Implementation Notes

• The data already exists in ~/.deepcode/projects/<project-code>/sessions-index.json which stores "token 用量" per session entry
• For real-time context window tracking, the token counts could be computed from the messages being sent to the model API
• A lightweight tokenizer or cached counts per message would suffice — no need for full tokenization on every request

This is an open-ended feature request. Happy to help with implementation or testing.