Context Budget Guardrails Framework

Workflow

1. Read budget-thresholds.md to understand the current threshold configuration and how utilization is calculated (token count from session metadata).
2. Verify the hook reads context size from the correct field in the hook payload — field name varies by Claude Code version.
3. Test the soft-warning band: simulate a 75% utilization state and confirm the advisory message is injected without blocking execution.
4. Review mitigation-prompts.md — the mitigation prompt must instruct the model to summarize accumulated context and defer large reads.
5. Test the hard-guardrail band: confirm the hook blocks Read and Bash calls with large output potential but allows Write (to preserve in-progress work).
6. Check safety-fallbacks.md for the graceful-degradation path — if utilization cannot be determined, the hook must fail open (advisory only, no block).

Routing Boundaries

Use this skill for context-window budget guardrail hooks. Defer to session-continuity-ops when the concern is preserving session state across compaction rather than throttling tool use within a session.

References

• Context Budget Thresholds
• Mitigation Prompting
• Fallback Behavior