andurilcode Context Compressor
Finite channel (context window) must carry maximum decision-relevant signal. Score fragments by entropy and redundancy, cut or compress accordingly.
Vocabulary
| Info Theory | Context Engineering |
|---|---|
| Channel / Capacity | Context window / window minus system prompt overhead |
| Entropy | Information per token |
| Self-information | How much a fragment reduces agent uncertainty |
| Redundancy | Tokens carrying no new info given other context |
| Lossy compression | Summarize at lower fidelity to save tokens |
| Lossless compression | Reformulate more densely, same facts |
Phase 1 — Budget
Calculate: window - system_prompt - tools - history - output_reserve = available. Compare available vs. candidate context:
| Regime | Ratio | Strategy |
|---|---|---|
| Abundant | >3x | Include all. Stop. |
| Comfortable | 1.5-3x | Cut obvious redundancy |
| Tight | 0.5-1.5x | Active compression |
| Starved | <0.5x | Aggressive triage. Every token justifies itself. |
Phase 2 — Entropy Scoring
Score every candidate fragment on three dimensions:
Self-Information (0-4): 0=derivable from training, 1=agent would guess right, 2=meaningfully reduces uncertainty, 3=changes behavior substantially, 4=without this agent produces wrong output.
Redundancy: Unique / Partial overlap / Mostly redundant / Fully redundant.
Token Efficiency: Dense / Normal / Bloated / Wasteful.
Decision rules:
- Self-Info 0 -> CUT always
- Self-Info 4 -> KEEP always at full fidelity
- Self-Info 1 + Mostly/Fully redundant -> CUT
- Self-Info 2-3 + Bloated -> COMPRESS
- Self-Info 2-3 + Unique + Dense -> KEEP
Output as table: Fragment | Tokens | Self-Info | Redundancy | Efficiency | Action.
Phase 3 — Compress
Lossless (Self-Info 3+, low efficiency): Telegram rewrite, structure extraction, reference elimination, dedup merge.
Lossy (budget Tight/Starved, lossless insufficient): Fidelity reduction (drop rationale, keep rule), scope narrowing (only task-relevant subsection), abstraction raising (drop detail, keep default), probabilistic pruning (cut edge-case-only context, accept risk).
Always log lossy losses:
Fragment: [name] ([before] -> [after] tokens)
Kept: [what survived]
Lost: [what was cut]
Risk: [what agent may get wrong]
Phase 4 — Assemble & Verify
Order by descending self-information (critical first — attention degrades).
Verify:
- Fits budget? Total tokens <= available.
- Agent could complete task with ONLY this context?
- All Self-Info 4 fragments present at full fidelity?
- Any cut fragment that would change agent behavior?
Calibration
- No task = no compression. Always audit against a specific task.
- 80/20. ~20% of context carries ~80% of signal. Find it, keep it full fidelity.
- Compression taxes flexibility. Leave slack for multi-turn needs.
- When in doubt, keep. Wrong output costs more than extra tokens.
- Empirics > theory. Compare agent output with full vs. compressed context when possible.