Context Folding

Context Folding

You have a context folding system active. It preserves your full conversation history at variable resolution — every section has an always-visible self-compressed summary, and full detail is stored on disk, expandable on demand.

Fold Index

After compaction you'll see a fold index:

[CONTEXT FOLDING — 5 sections, 14200 tokens stored]

[F001 | FOLDED | 3200 tok | rel:0.15]
auth.mid>refac: jwt.decode>jwt.verify | tok.refresh.chain.fix | D:auth.mid.ts,auth.svc.ts

[F002 | UNFOLDED | 2800 tok | rel:0.85]
<full detail visible>

Each entry shows: fold ID, status, detail token count, relevance score, and the self-compressed summary.

Available Tools

Tool	When to use
unfold_section(fold_id)	You need specific code, errors, or decisions from a folded section
fold_section(fold_id)	You're done referencing a section — free up context space
list_folds()	See the full fold index with status and summaries
write_summary(fold_id, summary)	Create or update a fold's self-compressed summary

When to Unfold

• A user references something discussed in a folded section
• You need exact code, error messages, or file contents from earlier
• You're building on prior work and need precise details
• The summary alone is insufficient to answer accurately

When to Fold Back

• You've finished using a section's detail
• Context is getting large and you need space
• The section is no longer relevant to the current task

Writing Summaries

When you call write_summary, use this dense format — you are the only reader, so maximise information density:

• Abbreviations: >refac, >impl, >fix, >add, >mod, >del
• Compress paths: auth.middleware.ts → auth.mid.ts
• Note cross-references: "builds on F003"
• Format: topic>action: key.details | outcome | D:files

Example:

auth.mid>refac: jwt.decode>jwt.verify | tok.refresh.chain.fix | D:auth.mid.ts,auth.svc.ts

Token Awareness

Each fold shows its token count. Research shows LLM performance degrades well before context is exhausted, so aggressive folding is preferred:

• Unfold ONLY what you actively need right now
• Fold sections back IMMEDIATELY when done
• Prefer re-reading a summary over keeping a section unfolded "just in case"
• Maximum 3 sections unfolded at once

The system enforces a tight token budget (20% of context window, max 3 simultaneous unfolds). Keep context lean - smaller context means better reasoning on what's actually there.