context-escalationlisted

# Context Escalation Policy A discipline for handling context pressure deliberately rather than reactively. Inspired by DCI-Agent-Lite's escalating context-manager but adapted to Claude Code's harness, where `/compact`, `handoff`, `/save-session`, and the Agent tool already exist as separate primitives. ## The principle Context limits are inevitable in long sessions. Default reflex — "I'll compact when I hit the wall" — loses information unnecessarily and creates a hard cliff. The DCI-Agent-Lite pattern that's worth borrowing: a sliding policy of progressively-more-aggressive interventions, applied as the context fills. Apply the **cheapest** intervention that still preserves what matters. Escalate only when the cheaper option is no longer enough. ## Five levels | Level | When | Action | Cost (information loss) | |---|---|---|---| | **L1 — Truncate** | A single tool output is large (find dump, long log, multi-file diff) but the rest of context is fine | Re-run with stricter filters: `head -50`, `grep ... | head`, `wc -l` first to size, `--limit` flags. Or summarize the output verbally before the next turn. | Near zero — the raw blob wasn't load-bearing | | **L2 — Drop redundant reads** | You've Read the same file twice, or have multiple iterations of the same diff visible | Acknowledge: "I have what I need from X.py, the line numbers above are authoritative" — implicit hint that the older read can be forgotten | Low — the latest copy survives | | **L3 — Fork to subagent