| name | ai-optimization |
| description | Fission engine for token-efficient coding (JS, TS, Node, Rust, Python, ML/AI): relevance scoring, hierarchical compression, strict token budgets, progressive disclosure, accuracy guardrails. Pair with fusion-sage for synthesis/surplus. Triggers: tokens, context window, prune, compress, Context Sage, large repo. |
ai-optimization (Context Sage)
Load rule: Formal SoT below. Lang playbooks → references/ only if needed. Expand references/english-procedure.md only if scoring/budget still ambiguous.
// Signature
Fission ≔ prune + compress + budget // this skill
Fusion ≔ synthesis + surplus // [fusion-sage](../fusion-sage/SKILL.md)
Score ∈ 0..100 // relevance
Budget ≔ context cap for *input* code/docs (not model CoT)
// Axioms
A1 Relevance ≻ Completeness // ~80% value from ~5–15% of code
A2 Hierarchy first // structure before bodies
A3 Language-native compression // AST/API idioms per lang
A4 Budget dictates form // lower score → denser form
A5 Progressive disclosure // expand only on request / auto-expand rules
A6 Correctness ⊥ tokens // never trade correctness for compression
A7 Evaluate(δ) ≔ (C, E, η) // keep tactics with Δ>0
Pair: fusion-sage · master-planner overlays · examples/overlays/.
Use / skip
| Use when | Skip when |
|---|
| repo ≳10k LOC · multi-file change | 1-file typo, full file already pasted |
| "tokens", "context", "Cursor slow/expensive" | pure design chat with no code load |
| implement / debug / refactor with code context | architecture-only → start fusion-sage |
| pasting partial code + "implement X" | |
Flow (internal — do not dump as essay)
intent → map (modules/API surface) → score → compress form → budget trim → act
| Step | Do |
|---|
| 1 Intent | goal ∈ {implement, debug, refactor, explain, test, migrate, optimize}; symbols; scope; constraints |
| 2 Map | lightweight module + public surface + recent files + arch style |
| 3 Score | table below (0–100) |
| 4 Form | score → compression tier |
| 5 Budget | hard cap 35% context for input code; reserve ~25% for reasoning+answer; drop lowest score first |
Relevance score
| Signal | Δ |
|---|
| symbol exact match | +40 |
| keyword in path/name | +20 |
| centrality (many importers) | +15 |
business-critical (auth, payment, core, model def) | +15 |
| recency (last ≤3 turns) | +10 |
test/ __pycache__/ node_modules/ dist/ target/ | −50 (waive if goal is debug/test — boost instead) |
Compression tiers
| Score | Form |
|---|
| >85 | full file stripped (noise out) |
| 60–85 | signature + critical blocks + docstring |
| 30–60 | public API + 1-line purpose + key edges |
| <30 | name only + "see prior context" |
Cross-lang strip rules
- Drop: licenses, generated markers, excess blank lines, repetitive getters
- Collapse standard validation/error-map to one sentence
- Use
... when name+signature+context makes body obvious
- ≤2 full function bodies per file unless score >90
Lang detail: references/typescript-optimizer.md · references/python-optimizer.md · references/rust-optimizer.md. Optional project overlay under .agents/skills/ai-optimization/references/ or examples/overlays/.
Accuracy guardrails (never violate)
| Rule | Detail |
|---|
| Never compress | auth, security, payments, secrets, permissions, migrations/schema, lockfile/deps edits, CI/E2E/build config, files you will edit, callers of changed symbols (sig+imports min) |
| Debug / flaky | full suspect files + config + related tests |
| Before edit | if body was summarized → read full file first |
| No invention | never invent APIs/patterns not seen |
| Done | state assumptions; run verify (type-check, lint, tests) |
Task overrides
| Task | Compression |
|---|
| explain / scout | summaries OK |
| implement | full types + full bodies for files edited |
| debug (esp. flaky/CI) | full suspects + config + tests |
| refactor / rename | full call graph for touched symbols |
Auto-expand (do not wait)
- Edge cases in summarized body (errors, auth, async/effect deps)
- User:
expand <symbol> · show full <file> · use whole project
- You would write "assuming standard pattern" without having read it
Red flags — stop summarizing
- Diff with no callers/tests noted
- Skip files required by
AGENTS.md / project rules
- One-line summary of complex control flow (
useEffect, retries, transactions)
Output protocol
🧠 Context Sage | Budget: Xk / Yk (Z%) | Relevance: R/100 | Files: N (forms…)
## Snapshot (2–4 lines: stack, key dirs, verify cmds)
## Context (structured, tiered)
## Action (minimal diff/answer + assumptions)
## Token note (expand <name> to deepen)
Hand off to fusion-sage when: architecture / long-term design · "make it better for the future" · 3+ related queries → suggest fusion pass.
IDE
| Host | Action |
|---|
| Cursor | symlink skill; optional assets/cursorrules-template.md → .cursor/rules/ai-optimization.mdc |
| Grok / others | install skill dir; description routing |
Self-improve (session)
After success: boost scores of symbols actually used; if user expanded a summary → lower compression next time for similar queries.
Done when: intent+scores applied; budget respected; guardrails held; verify cmds run for multi-file edits; fusion handoff noted when architecture.
Anti-patterns: dump whole repo "just in case" · compress auth/edit targets · invent unseen APIs · skip verify because "context was tight" · dual-load full English playbooks every turn.
Script helper: scripts/context-sage.py. English expansion: references/english-procedure.md.
