Context Refresh Runbook

Refresh on evidence, not discomfort

Refresh when authoritative facts changed, the active history contains rejected approaches, summaries no longer preserve key evidence, a model or tool changed, the workflow crosses a trust boundary, or a clean reviewer handoff is needed. Caching and provider-managed state can reduce repeated input; they do not make old context current or safe.

Before starting clean, reconcile tool calls and side effects. A missing response does not prove an external action failed. Check the system of record, operation identifier, or idempotency record before retrying.

Build a clean refresh package

Keep durable decisions and exact evidence separate from discarded exploration. Redact secrets and unnecessary personal data. Treat prior model output as untrusted unless it was independently verified.

Goal: [CURRENT OUTCOME]
Authority and permissions: [WHO MAY DECIDE OR ACT]
Current state: [VERIFIED SYSTEM-OF-RECORD FACTS]
Decisions: [CHOICE, OWNER, DATE, REASON]
Evidence: [SOURCE OR TEST FOR EACH MATERIAL CLAIM]
Completed actions: [ACTION ID AND VERIFIED POSTCONDITION]
Rejected approaches: [WHAT NOT TO REPEAT AND WHY]
Open risks: [UNRESOLVED UNCERTAINTY]
Sensitive data removed: [YES | EXPLAIN APPROVED EXCEPTION]
Next action: [ONE BOUNDED STEP]
Acceptance check: [HOW THE NEXT STEP WILL BE VERIFIED]

Expected result and verification

The expected result is a new working context that preserves the goal, authority, decisions, evidence, verified action state, open risk, and next step without preserving irrelevant transcript bulk.

Have a reviewer reconstruct the current state using only the refresh package and linked evidence. Compare volatile facts with the system of record, confirm no unresolved side effect can be repeated, and regression-test representative cases after changes to context selection, summarization, caching, retrieval, tools, or models.