From 12ae360dc7f6bfac77573d07c9e3a875720d6db6 Mon Sep 17 00:00:00 2001 From: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com> Date: Tue, 23 Jun 2026 10:57:30 -0300 Subject: [PATCH] =?UTF-8?q?docs(compression):=20declare=20Phase=204=20laye?= =?UTF-8?q?rs=20=E2=80=94=20Output=20Styles,=20adaptive=20dial,=20per-requ?= =?UTF-8?q?est=20control=20(#4801)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The README compression section listed the 9 input engines but not the Phase 4 layers now in production: - Output Styles (output-axis steering: terse-prose / less-code / terse-cjk, lite/full/ultra) - adaptive context-budget dial (reserve-output|percentage|absolute ยท floor|replace-autotrigger|off) - per-request x-omniroute-compression precedence + the offline eval harness Also bumped the highlights range to v3.8.35, expanded the compression feature bullet, and marked the GUIDE's Phase 4 row Shipped (was 'Planned' โ€” it's merged on v3.8.35). Co-authored-by: Diego Rodrigues de Sa e Souza Co-authored-by: Claude Opus 4.8 (1M context) --- README.md | 19 ++++++++++++++++--- docs/compression/COMPRESSION_GUIDE.md | 2 +- 2 files changed, 17 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index 03508d9da8..150f630fd5 100644 --- a/README.md +++ b/README.md @@ -287,12 +287,12 @@ Result: 4 layers of fallback = zero downtime -> Recent highlights from **v3.8.20 โ†’ v3.8.33**. Full history in [`CHANGELOG.md`](CHANGELOG.md). +> Recent highlights from **v3.8.20 โ†’ v3.8.35**. Full history in [`CHANGELOG.md`](CHANGELOG.md). - **๐Ÿค– One-command CLI/agent setup** โ€” a dedicated `setup-*` command configures each coding tool to route through OmniRoute (Claude Code, Codex, Cline, Continue, Cursor, Roo Code, Kilo Code, Crush, Goose, Qwen Code, Aider, OpenCode, Gemini CLI); `omniroute launch` / `omniroute launch-codex` are zero-config launchers. โ†’ [CLI Integrations](docs/guides/CLI-INTEGRATIONS.md) - **๐Ÿ›ฐ๏ธ Remote mode** โ€” drive a remote OmniRoute from any machine with scoped access tokens (`omniroute connect` / `omniroute contexts` / `omniroute tokens`). โ†’ [Remote Mode](docs/guides/REMOTE-MODE.md) - **๐Ÿงญ Smarter auto-routing** โ€” OpenRouter-style `auto/:` combos (e.g. `auto/coding:fast`, `auto/reasoning:pro`), live Arena-ELO + models.dev model intelligence, per-step account allowlists, provider-wildcard combo steps, nested combo-ref execution, sticky weighted selection, and `web_search`-aware routing to a configured model. โ†’ [Auto-Combo](docs/routing/AUTO-COMBO.md) -- **๐Ÿ—œ๏ธ Pluggable compression** โ€” an async compression pipeline with Compression Studios, a stable LLMLingua-2 ONNX engine, RTK, delegated Anthropic Context Editing, and a unified settings panel with named profiles + an active-profile selector as the single source for per-engine on/off + level. โ†’ [Compression](docs/compression/COMPRESSION_ENGINES.md) +- **๐Ÿ—œ๏ธ Pluggable compression** โ€” an async pipeline of **9 composable engines** with Compression Studios, an LLMLingua-2 ONNX engine and a heuristic/SLM two-tier **Ultra**, RTK, delegated Anthropic Context Editing, **Output Styles** (output-axis steering: terse-prose / less-code / terse-CJK), an **adaptive context-budget dial** (escalate only as far as needed to fit the context window), per-request `x-omniroute-compression` control, an opt-in offline eval harness, and a unified panel with named profiles + an active-profile selector. โ†’ [Compression](docs/compression/COMPRESSION_ENGINES.md) - **๐Ÿ•ต๏ธ Transparent MITM decrypt (TPROXY)** โ€” capture & translate traffic from CLIs that ignore proxy env vars, with a per-SNI certificate authority and a trust-store installer. โ†’ [MITM/TPROXY](docs/security/MITM-TPROXY-DECRYPT.md) - **๐Ÿ’ธ Cost telemetry everywhere** โ€” `X-OmniRoute-*` cost/usage headers on every endpoint (including media), a non-token cost engine, a cache-HIT `X-OmniRoute-Cost-Saved` header, and per-key USD spend quotas. โ†’ [API Reference](docs/reference/API_REFERENCE.md) - **๐Ÿง  Memory you control** โ€” opt-in int8 vector quantization (Qdrant + sqlite-vec), memory off by default, and a per-request `x-omniroute-no-memory` header. โ†’ [Memory](docs/frameworks/MEMORY.md) @@ -538,7 +538,20 @@ average = 1 โˆ’ (1 โˆ’ 0.80) ร— (1 โˆ’ 0.46) = 89.2% range = 78.4 โ€“ 94.6% ``` -Code blocks, URLs, JSON and structured data are **always protected** by the preservation engine. Auto-trigger compression by token threshold, or assign a compression pipeline per routing combo. +Code blocks, URLs, JSON and structured data are **always protected** by the preservation engine. + +### ๐ŸŽš๏ธ Beyond the engines โ€” output styles, the adaptive dial & per-request control + +The 9 engines above shrink what goes **in**. Three more layers shape **how**, **when**, and what comes **out**: + +- **๐Ÿช„ Output Styles** *(output-axis steering)* โ€” inject deterministic, cache-safe response-shaping instructions; combinable, each at `lite` / `full` / `ultra` intensity. Adding a style is a one-line registry entry: + - **Terse prose** โ€” drop filler / articles / hedging; keep technical substance exact. + - **Less code** โ€” "lazy senior dev" YAGNI: smallest working change, no unrequested scaffolding. + - **Terse CJK (ๆ–‡่จ€)** โ€” classical-Chinese ultra-terse style (locale-gated to `zh`). +- **๐ŸŽฏ Adaptive context-budget** *(the dial)* โ€” instead of one on/off token threshold, escalate the cheapest, most-lossless engines only as far as needed to **fit the model's context window**. Policy: `reserve-output` (default, model-aware) ยท `percentage` ยท `absolute`. Mode: `floor` (guarantee fit) ยท `replace-autotrigger` (your explicit choice wins) ยท `off` (legacy threshold). +- **๐ŸŽ›๏ธ Where compression is decided** *(precedence, high โ†’ low)* โ€” per-request `x-omniroute-compression` header โ€บ routing-combo override โ€บ active named profile โ€บ adaptive / auto-trigger โ€บ panel default โ€บ off. The applied plan echoes back in the `X-OmniRoute-Compression: ; source=` response header. + +Auto-trigger by token threshold, flip on the adaptive dial, pin a named profile, set a one-off per request, or assign a pipeline per routing combo โ€” whichever fits the workload. An opt-in offline **eval harness** (`npm run eval:compression`) scores fidelity vs. savings on a pinned corpus before you promote a change. ๐Ÿ“– [`COMPRESSION_GUIDE.md`](docs/compression/COMPRESSION_GUIDE.md) ยท [`RTK_COMPRESSION.md`](docs/compression/RTK_COMPRESSION.md) ยท [`COMPRESSION_ENGINES.md`](docs/compression/COMPRESSION_ENGINES.md) diff --git a/docs/compression/COMPRESSION_GUIDE.md b/docs/compression/COMPRESSION_GUIDE.md index 5655aea48e..1665cfef7e 100644 --- a/docs/compression/COMPRESSION_GUIDE.md +++ b/docs/compression/COMPRESSION_GUIDE.md @@ -278,7 +278,7 @@ Every compressed request includes stats in the server logs: | Phase 1 | Off, Lite | โœ… Shipped | | Phase 2 | Standard, Aggressive, Ultra | โœ… Shipped | | Phase 3 | RTK, Stacked, Compression Combos | โœ… Shipped | -| Phase 4 | Per-model adaptive, ML-based pruning | ๐Ÿ—“๏ธ Planned | +| Phase 4 | Output Styles, SLM-tier Ultra, adaptive context-budget, eval harness | โœ… Shipped | ---