[FEATURE] Feature request: periodic lightweight context audit (cheaper model) during long sessions
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Problem
In long sessions on a large knowledge base (about 21,000 Markdown notes with instructions, decisions and reference material), Claude Code often overlooks constraints that are written down: instructions given earlier in the session, rules stored in project notes, decisions already made. I currently work around this with hooks and a scheduled nightly job, but nothing built in reminds the main model of what it is forgetting while it works.
Proposed Solution
Proposed solution
At a configurable interval (every N turns or tokens) or at meaningful checkpoints (before ending a turn, after a long tool run, after compaction), run a lightweight audit with a cheaper model (e.g. Haiku) over the conversation, the relevant project files and the current state. It would only look for forgotten constraints, unresolved requirements, inconsistencies and risks, and hand the main model a short checklist.
Why
Continuity and anticipation across long sessions, with a bounded token cost. Today's checkpoints are restore points only; they do not audit or remind.
Alternatives considered
Stop hooks that block on known checks (works, but only for failures I can already name), nightly batch jobs (not available during the session), a longer CLAUDE.md (dilutes the rules that matter).
Alternative Solutions
_No response_
Priority
Critical - Blocking my work
Feature Category
CLI commands and flags
Use Case Example
_No response_
Additional Context
_No response_