Files
yvgude--lean-ctx/docs/contracts/tokenizer-translation-driver-v1.md
T
wehub-resource-sync 26382a7ac6
CI / Clippy (push) Failing after 15m13s
CI / Test (ubuntu-latest) (push) Failing after 16m1s
CI / Test (macos-latest) (push) Has been cancelled
CI / Test (windows-latest) (push) Has been cancelled
CI / Build (no embeddings / no ORT) (push) Has been cancelled
CI / Format (push) Has been cancelled
CI / Cookbook (Node) (push) Has been cancelled
CI / Pi Extension (Node) (push) Has been cancelled
CI / Rust SDK (lean-ctx-client) (push) Has been cancelled
CI / Embed SDK (lean-ctx-sdk) (push) Has been cancelled
CI / Python SDK (leanctx) (push) Has been cancelled
CI / Hermes Plugin (Python) (push) Has been cancelled
CI / SDK Conformance Matrix (push) Has been cancelled
CI / Coverage (push) Has been cancelled
CI / cargo-deny (push) Has been cancelled
CI / Adversarial Safety (push) Has been cancelled
CI / Benchmarks (push) Has been cancelled
CI / Output-Quality Gate (eval A/B) (push) Has been cancelled
CI / Documentation (push) Has been cancelled
CI / CI Green (push) Has been cancelled
JetBrains Plugin / Actionlint (push) Has been cancelled
CodeQL / Analyze (actions) (push) Has been cancelled
CodeQL / Analyze (javascript-typescript) (push) Has been cancelled
CodeQL / Analyze (rust) (push) Has been cancelled
JetBrains Plugin / Validation (push) Has been cancelled
JetBrains Plugin / Build (push) Has been cancelled
JetBrains Plugin / Test (push) Has been cancelled
Security Check / Security Scan (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 12:35:30 +08:00

1.8 KiB

Tokenizer-aware Translation Driver v1 (TokenizerTranslationDriverV1)

GitLab: #2310

LeanCTX nutzt “Translation” als letzten Optimierungsschritt für synthetische Context-Formate (z.B. TDD Signatures), damit Output-Tokens modell-/tokenizer-spezifisch reduziert werden können.

Ziele

  • Deterministisch: gleicher Input + gleicher model_key + gleiche Profile Policy ⇒ gleiches Translation Ruleset.
  • Safe defaults: Default verändert bestehende CRP/TDD Formate nicht (opt-in).
  • Tokenizer-aware: Rulesets können Unicode → ASCII tauschen, wenn Unicode im Ziel-Tokenizer teurer ist.
  • Verifier-safe: Translation darf keine File Paths oder Identifiers “kaputtoptimieren”.
  • Bounded: Translation läuft auf Tool-Outputs, aber wird für JSON-Ausgaben übersprungen (machine-readable bleibt exakt).

Aktivierung (Policy)

Translation wird über Profile gesteuert:

  • profile.translation.enabled = true|false
  • profile.translation.ruleset = "legacy" | "ascii" | "auto"

Default: enabled=false, ruleset="legacy".

Deterministische Ruleset Selection

ruleset="auto" verwendet den Model Key aus:

  • LEAN_CTX_MODEL (oder LCTX_MODEL)

Heuristik (v1):

  • OpenAI/GPT-family (model enthält gpt) → ascii
  • sonst → legacy

Überspringen von JSON Outputs

Tool-Outputs, die als JSON parsebar sind ({...} oder [...]), werden nicht verändert.

Messung / Bench

ctx_benchmark zeigt token_cost (o200k_base) vor/nach Ruleset-Translation für TDD/Signature Outputs.

Relevanter Code

  • Driver: rust/src/core/tokenizer_translation_driver.rs
  • Token cost oracle: rust/src/core/tokens.rs
  • Empirische safe rules: rust/src/core/neural/token_optimizer.rs
  • Verification safety: rust/src/core/output_verification.rs