context-window-planninglisted

# Skill: Context Window Planning Context window limits cause one of the most silent failures in AI coding tools: a request is truncated mid-input, the model completes it anyway, and the output is confidently wrong because half the code was never seen. This skill estimates demand, matches to a tier, and flags the overflow risk before it bites. ## The key principle: estimate relative demand, not absolute tokens Specific context-window sizes change frequently and belong in the dated knowledge bank (`../../knowledge/cross-tool-model-lineup-2026.md`) with `[verify-at-use]` markers. This skill works with **relative demand tiers** that remain stable even as the specific numbers churn: | Demand tier | Typical workload | |---|---| | Low | Single file or function; short chat turn; one-shot completion | | Medium | 2-5 files; a PR diff; a focused refactor with context | | High | Entire module or package; cross-repo project with many files | | Very high | Full repo scan; long autonomous agent run; many-turn conversation with large file contents | ## Step 1 — Estimate the task's context demand Walk through each input category and classify: 1. **Code payload** — how many files, LOC estimate, are imports/dependencies included? 2. **Conversation history** — is this a long multi-turn session, or a fresh request? 3. **Tool outputs** — does the agent run read test results, linting output, or API responses mid-run? 4. **Output space** — large generated artifacts (a full module, a test suit