Ingestion Compression (RAG offloading)

The main context is append-only within a session — once you read a 20-page document, those tokens stay until the session is compacted. You cannot selectively unload them. So the rule is: never let large source material enter the main context in the first place. Instead, read it in a subagent (which has its own throwaway context) and bring back only a synthesis.

When to apply

• The user points you at a reference document to understand (article, report, spec, transcript, PDF, long .md).
• Skip when: (a) it's a code file you must edit — you need the real bytes; (b) the user explicitly says "ingest / read the full X" (the override); (c) it's small (under the threshold).

Procedure

1. Check size first. bash scripts/doc-size.sh <file> (threshold ~1500 words ≈ 2000 tokens). Under threshold → read directly, done. Over → continue.
2. Dispatch a subagent (Agent / Task tool) to read the full document. The subagent has a separate context window; the raw document never touches yours.
3. The subagent returns a synthesis in the .synth.md shape below — substance + map + pointer — sized to the requested fidelity.
4. Optionally persist it next to the original as <name>.synth.md so future reads (this session or later) use the compressed version. Return the synthesis to the main thread; that's all that enters your context.
5. Drill down on demand. If you later need a detail the synthesis dropped, use the map to read just that slice of the original (offset/grep) — never re-ingest the whole thing.

Fidelity levels (pick per task)

• Gist (~1%) — 5–10 bullets, the thesis and conclusions only.
• Structured extract (~5–10%) — claims, numbers, named entities, decisions, caveats. Default.
• Full outline (~15%) — section-by-section, preserves structure and most facts.

The .synth.md shape (three parts — always include all three)

---
source: /abs/path/to/original.md
original_tokens: ~12000   synth_tokens: ~600   fidelity: structured-extract
---
## Synthesis        # the SUBSTANCE — key concepts an agent needs to act
- …
## Map              # the NAVIGATION — where to drill down in the original
- §1 framing — lines 1–40
- §3 the tariff figures — lines 210–340   ← exact numbers here
## Pointer          # the ADDRESS — already in frontmatter `source:`; restate if helpful

• Synthesis = the compressed content (what enters context).
• Map = a section index so missing details are cheap to fetch from the original.
• Pointer = the path back to the full source for targeted re-reads.

Caveats

• Synthesis is lossy. Keep the pointer + map so exact facts are recoverable. Don't aggressively compress where precise detail is load-bearing (specs, legal, code).
• Works on files, not pastes. If the user pastes a long document into the chat, it's already in context — too late to offload; note that and suggest a file next time.
• A subagent's value is isolation, not deletion — it prevents the load, it can't undo one that already happened.

Pairs with

The context-budget skill (same repo) governs the static always-on tier; this skill governs the dynamic per-session intake. Static + dynamic = full context hygiene.