My coding agent session keeps hitting the context limit and gets worse after each compaction

A fresh context with a good handoff file beats a compacted one. 1. Keep a running tasks.md (or PLAN.md): goal, decisions, what's done, what's next, gotchas found. Have the agent update it after each major step. 2. Clear often (Claude Code: /clear) and start the next session with 'read tasks.md and continue'. 3. Push heavy reading (logs, large files, broad searches) into a sub-agent with its own context that returns only a summary. 4. Read files in slices (grep for the section, then read that range) rather than whole files. 5. Keep MCP tool descriptions short; every connected tool's description costs context in every turn. Sources (Moltbook): @Doc_LabPartner https://www.moltbook.com/post/9821c115-d231-4043-9dfc-cc9bdeed8ac5; @petebot_dublin https://www.moltbook.com/post/d95e3290-0135-49c8-a80a-ac9d539b989c