Feature Request: Per-skill token usage tracking

Feature Request: Per-Skill Token Usage Tracking

Submitted by: Eric H. Howard

Problem

When using Claude Code skills (e.g., /implement-work-item, /commit, custom skills), there is no way to see how many tokens a specific skill invocation consumed. This makes it difficult to:

1. Understand cost per skill — Which skills are expensive vs. cheap to run?
2. Optimize skill prompts — Without usage data, it's hard to know if a skill rewrite reduced token consumption
3. Budget and plan — Teams evaluating Claude Code for workflows need to estimate per-task costs
4. Debug context exhaustion — When a skill runs out of context, knowing where tokens went (main conversation vs. subagents vs. tool results) would help diagnose the issue

Current Behavior

• /cost shows cumulative session cost (API billing only; subscription plans just show "You are currently using your subscription")
• /context shows current context window fill as a snapshot, not cumulative consumption
• Subagent token usage (launched via Task tool) is not surfaced to the parent conversation at all
• No per-skill or per-invocation breakdown exists

Proposed Solution

Add a token usage summary that tracks per-skill invocations, ideally including:

| Metric | Description |
|--------|-------------|
| Skill name | Which skill was invoked |
| Main context tokens | Input + output tokens in the primary conversation |
| Subagent tokens | Total tokens consumed by Task tool subagents spawned during the skill |
| Tool result tokens | Tokens from tool call results (file reads, bash output, etc.) |
| Total tokens | Sum of all above |

This could be exposed via:

• A new /usage command with per-skill breakdown
• An addition to /cost output
• A summary line printed when a skill completes (e.g., "Skill completed. Tokens used: 45,230 (main: 32,100, subagents: 13,130)")

Additional Context

This came up while running a complex multi-phase skill (implement-work-item) that orchestrates codebase exploration, architecture design, implementation, code review, git operations, and ADO/GitHub integration. The skill spawns multiple subagents and consumes significant context, but there's no visibility into the total cost of the run. This is especially relevant for teams evaluating Claude Code for production workflows where per-task cost tracking matters.