- New skill: token-budget-report.md with MCP overhead and skill loading estimation model based on known file sizes and static overhead tables - Sprint-plan displays token report after planning complete - Sprint-close displays token report after lessons captured - Visual-output.md updated with token report format - .gitignore exception for token-budget-report.md (not API tokens) Helps identify which phases and components consume the most context budget. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
5.0 KiB
description
| description |
|---|
| Token consumption estimation model and display format for sprint workflows |
Token Budget Report
Purpose
Provides directional token consumption estimates at sprint workflow boundaries. Not exact — Claude Code doesn't expose token metrics — but accurate enough to identify which phases and components consume the most context budget.
When to Display
- End of
/sprint-plan(after all issues created) - End of
/sprint-close(after lessons captured) - On explicit request:
/sprint-status --tokens
Estimation Model
MCP Tool Definitions (Static Overhead)
These are loaded once at session start. Update this table if MCP servers change.
| MCP Server | Default Tools | Est. Tokens | With Module Filter |
|---|---|---|---|
| netbox | 182 | ~19,810 | ~4,500 (dcim,ipam,virt,extras) |
| gitea | 36 | ~4,785 | — |
| data-platform | 32 | ~3,458 | — |
| viz-platform | 20 | ~3,055 | — |
| contract-validator | 8 | ~1,048 | — |
| Total (default) | 278 | ~32,156 | — |
| Total (filtered) | ~139 | ~16,846 | — |
Skill Loading (Per Phase)
| Phase | Typical Skills Loaded | Est. Tokens |
|---|---|---|
Planning (/sprint-plan) |
mcp-tools-reference, label-taxonomy, sprint-planning, architecture-analysis, rfc-workflow | ~3,000–5,000 |
Execution (/sprint-start) |
mcp-tools-reference, branch-security, plan-then-batch | ~2,000–3,000 |
Review (/pm-review) |
mcp-tools-reference, review-checklist | ~1,500–2,500 |
Close (/sprint-close) |
mcp-tools-reference, sprint-lifecycle, lessons-learned | ~2,000–3,000 |
To get actual numbers: count characters in each skill file loaded during the phase, divide by 4.
Agent Overhead
| Agent | Model | Est. Overhead per Invocation |
|---|---|---|
| Planner | sonnet | ~500 tokens (frontmatter + system context) |
| Orchestrator | sonnet | ~500 tokens |
| Executor | sonnet | ~400 tokens |
| Code Reviewer | sonnet | ~400 tokens |
Display Format
+-- Token Budget Report -----------------------------------------------+
| Phase: [PLANNING / CLOSING] |
| Sprint: [Sprint Name] |
+----------------------------------------------------------------------+
| |
| MCP Overhead (session-wide): |
| netbox ........... ~4,500 tk (filtered: dcim,ipam,virt,extras) |
| gitea ............ ~4,785 tk |
| data-platform .... ~3,458 tk |
| viz-platform ..... ~3,055 tk |
| contract-valid ... ~1,048 tk |
| Subtotal ......... ~16,846 tk |
| |
| This Phase: |
| Skills loaded .... ~X,XXX tk (N files) |
| Agent overhead ... ~X,XXX tk (model: sonnet) |
| Command file ..... ~XXX tk |
| Subtotal ......... ~X,XXX tk |
| |
| Estimated Session Total: ~XX,XXX tk |
| Typical Claude Code Budget: ~200,000 tk |
| Estimated Usage: ~XX% |
| |
| Tip: Run `/doctor` for exact MCP overhead numbers. |
+----------------------------------------------------------------------+
Instructions for the Agent
When generating this report:
-
MCP Overhead: List the MCP servers from
.mcp.jsonand use the static token estimates from the table above. IfNETBOX_ENABLED_MODULESis set in the environment, use the filtered estimate for netbox. -
Skills Loaded: Count the skill files that were loaded (referenced in the command's "Skills Required" section). Estimate tokens by reading each file's character count and dividing by 4.
-
Agent Overhead: Use the table above based on which agent ran during this phase.
-
Totals: Sum the phase subtotal. Add to MCP overhead for session estimate.
-
Percentage: Express as a percentage of ~200,000 tokens (typical Claude Code context window).
Accuracy Disclaimer
These are estimates based on known file sizes and static overhead tables. Actual consumption varies based on:
- Conversation length
- Tool call results (especially large grep/read outputs)
- File content read during the session
- Number of agent invocations
For exact MCP overhead, run /doctor.
This report helps identify which components consume the most budget — not the precise total.