Spec-kit commands consume significant portion of context window in every session #1401
Description
Problem
When spec-kit is initialized, slash commands are added to .claude/commands/ (or other agent's equivalent locations).
If others work like Claude Code, these commands are loaded into the agent's context window for every chat session, regardless of whether spec-kit is being actively used.
This creates a significant "context tax" that reduces the available working space for actual development tasks.
Context Window Consumption
Based on /context output in Claude Code, spec-kit commands consume approximately 18.6k tokens:
| Command | Tokens |
|---|---|
| speckit.checklist | 4.2k |
| speckit.specify | 3.1k |
| speckit.implement | 3.1k |
| speckit.clarify | 2.8k |
| speckit.analyze | 1.8k |
| speckit.tasks | 1.5k |
| speckit.constitution | 1.2k |
| speckit.plan | 690 |
| speckit.taskstoissues | 218 |
Impact Across Supported Platforms
The ~18.6k token consumption impacts others in this way:
| Platform | Context Window | spec-kit Consumption |
|---|---|---|
| Cursor (default chat) | ~20k tokens | ~93% |
| Windsurf (typical) | 32-100k tokens | ~19-58% |
| GitHub Copilot (standard) | 64k tokens | ~29% |
| GitHub Copilot (VS Code Insiders) | 128k tokens | ~14.5% |
| Codex CLI (OpenAI) | 192-272k tokens | ~7-10% |
| Claude Code | 200k tokens | ~9.3% |
| Cursor (max mode) | 200k tokens | ~9.3% |
| Amazon Q Developer CLI | 200k tokens | ~9.3% |
| Amp (Sourcegraph) | 200-432k tokens | ~4-9% |
| Gemini CLI | 1M tokens | ~1.9% |
| Roo Code | model-dependent | varies |
The impact is most severe on:
- Cursor default chat (~20k): spec-kit would consume nearly the entire context
- GitHub Copilot standard (64k): nearly a third of available context
- Windsurf: varies by model, but often significant
Why This Matters
Context window space is a finite and valuable resource. Every token consumed by spec-kit commands is a token unavailable for:
- Code context: Less room for file contents, reducing the agent's ability to understand the codebase
- Conversation history: Earlier parts of the conversation get truncated sooner
- Agent reasoning: Less space for the agent to think through complex problems
Users who have spec-kit installed but aren't actively using it for a particular task are paying this cost with no benefit.
Possible Mitigation
One simple option might be specify disable / specify enable commands that move command files to/from a backup location (e.g., .claude/commands-disabled/), allowing users to toggle spec-kit out of context when not in use.
However, I imagine there may be better architectural solutions the maintainers could consider.
My Environment
- Claude Code with Max plan (200k context)