[PR #301] [CLOSED] docs: update Xenon to v0.8.2 — SWE-bench 40.0% instance-level, same-model A/B +6.7pp #325

New issue

Closed

opened 2026-10-26 09:48:01 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 09:48:01 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/awesome-deepseek-agent/pull/301
Author: @xianyu-sheng
Created: 7/18/2026
Status: ❌ Closed

Base: main ← Head: main

📝 Commits (10+)

b477978 docs: add OmniAgent integration guide (EN + zh-CN)
557f8b1 fix: correct 3 factual inaccuracies in guides
ad64851 fix: 28 -> 27 skills (accurate count)
39484e9 docs: add GitHub repo link to guides
b165798 fix: credentials.yaml format (flat dict, not nested providers)
100de4b rename: OmniAgent → Xenon + add DeepSeek cache optimization highlight
8030085 feat: add Vision Bridge (multimodal for DeepSeek) to Xenon guide
f1880c4 docs: align Xenon guide with verified V4 support
01c12be docs: update Xenon guide for v0.7.0
b801c91 docs: update Xenon guide for v0.7.2

📊 Changes

4 files changed (+185 additions, -0 deletions)

View changed files

📝 README.md (+1 -0)
📝 README.zh-CN.md (+1 -0)
➕ docs/xenon.md (+93 -0)
➕ docs/xenon.zh-CN.md (+90 -0)

📄 Description

Summary

Updates the English and Chinese integration guides for Xenon v0.8.2, an open-source terminal coding agent with verified DeepSeek V4 support.

What is new in v0.8.2

SWE-bench Official Evaluation

Xenon scored 40.0% instance-level / 45.8% cell-level on SWE-bench_Lite (30 instances × 5 engines, official harness, single attempt, cheapest model tier, reproducible). Same-model A/B (v0.8.0 → v0.8.2, DeepSeek V4 Flash) isolates the framework gain at +6.7pp instance / +11.0pp cell.

Key results:

Metric	Value
Cells	150 (30 instances × 5 engines)

Engine-agnostic design: ReAct, Plan-Execute, Reflection, and their combinations all benefit from the same Evidence Runtime infrastructure. Plan-engine patch output improved by +50% after the framework update.

Evidence Runtime

Five-stage online verification chain that prevents LLMs from claiming "I fixed it" when they actually did not:

EvidenceGate — checks context completeness before execution
FixBindingGate — patches must bind to root cause, not symptom
FactBindingGate — tracks every tool call across the engine lifecycle
Evidence Ledger — structured audit trail of all operations
Enforce — verifies sufficient evidence before allowing side effects

Regression tested: 6/6 protection, zero false-kills. FileClaimGate caught a real hallucination (model claimed "modified docstring.py" but edit tool actually failed).

Other v0.8.2 improvements

Cache Rails: 97.35% warm input-token hit rate, 93.03% cost savings
7 engines, 9 tool families, 12 command groups, 26 built-in tools
MCP protocol support with pluggable transports
Online-chain regression verification framework
2110 tests, all passing

Verification

Published v0.8.0 wheel and source distribution pass Twine metadata checks
Full offline suite: 2110 passed; Ruff, compileall, and package builds passed
SWE-bench full report: evals/results/SWEBENCH_REPORT.md
Regression verification: 6/6 baseline cells protected, zero regression
Previous live checks also verified V4 Pro reasoning_effort and native tool-call protocol continuity

Files changed

File	Purpose
docs/xenon.md	Updated English v0.8.2 installation and DeepSeek guide
docs/xenon.zh-CN.md	Updated Chinese guide
README.md	English tool index entry
README.zh-CN.md	Chinese tool index entry

Checklist

Published v0.8.2 release and reproducible install command
SWE-bench official evaluation: 40.0% instance / 45.8% cell, same-model A/B +6.7pp/+11.0pp, fully reproducible
Evidence Runtime: 5-stage online verification chain, 6/6 zero regression
Current V4 model names, reasoning effort, and native tool continuity
Provider-reported cache evidence and Cache Rails documented
2110 tests passing
English and Chinese guides in one PR
Alphabetical README entry

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/deepseek-ai/awesome-deepseek-agent/pull/301 **Author:** [@xianyu-sheng](https://github.com/xianyu-sheng) **Created:** 7/18/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `main` --- ### 📝 Commits (10+) - [`b477978`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/b4779786af209e5ced218108471f2c604ca00905) docs: add OmniAgent integration guide (EN + zh-CN) - [`557f8b1`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/557f8b1ba4a6391750a5e527d0c82687fc0946de) fix: correct 3 factual inaccuracies in guides - [`ad64851`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/ad64851d2a8f07bfaa925b081c33d59a17c35476) fix: 28 -> 27 skills (accurate count) - [`39484e9`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/39484e9209c6fb4afe694a2bfe39d2b3ea66d705) docs: add GitHub repo link to guides - [`b165798`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/b16579875e253f10c3d96ac5ee4756c115c9f685) fix: credentials.yaml format (flat dict, not nested providers) - [`100de4b`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/100de4bdf8e293deef2899d558f4fb80a7ebda0a) rename: OmniAgent → Xenon + add DeepSeek cache optimization highlight - [`8030085`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/8030085d8cedc581385f467c32ba9614982a00c6) feat: add Vision Bridge (multimodal for DeepSeek) to Xenon guide - [`f1880c4`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/f1880c4ebdee1e94d0023051ca11b081a7f6d87c) docs: align Xenon guide with verified V4 support - [`01c12be`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/01c12bec1031c3d1b39ec2b84fe328420028c6d3) docs: update Xenon guide for v0.7.0 - [`b801c91`](https://github.com/deepseek-ai/awesome-deepseek-agent/commit/b801c9158aea5d5b5f90ce6e0a70615c9c1fc3d5) docs: update Xenon guide for v0.7.2 ### 📊 Changes **4 files changed** (+185 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `README.md` (+1 -0) 📝 `README.zh-CN.md` (+1 -0) ➕ `docs/xenon.md` (+93 -0) ➕ `docs/xenon.zh-CN.md` (+90 -0) </details> ### 📄 Description ## Summary Updates the English and Chinese integration guides for [Xenon v0.8.2](https://github.com/xianyu-sheng/Xenon/releases/tag/v0.8.2), an open-source terminal coding agent with verified DeepSeek V4 support. ## What is new in v0.8.2 ### SWE-bench Official Evaluation Xenon scored **40.0% instance-level / 45.8% cell-level** on SWE-bench_Lite (30 instances × 5 engines, official harness, single attempt, cheapest model tier, reproducible). Same-model A/B (v0.8.0 → v0.8.2, DeepSeek V4 Flash) isolates the framework gain at **+6.7pp instance / +11.0pp cell**. Key results: | Metric | Value | |--------|-------| | Cells | 150 (30 instances × 5 engines) | | Resolved (cell) | 33/72 = 45.8% | | Resolved (instance) | 12/30 = 40.0% | | Model | DeepSeek V4 Flash (single tier) | | A/B (v0.8.0 → v0.8.2) | +6.7pp instance / +11.0pp cell | | Cost | < ¥0.1 per instance with Flash | Engine-agnostic design: ReAct, Plan-Execute, Reflection, and their combinations all benefit from the same Evidence Runtime infrastructure. Plan-engine patch output improved by +50% after the framework update. ### Evidence Runtime Five-stage online verification chain that prevents LLMs from claiming "I fixed it" when they actually did not: 1. EvidenceGate — checks context completeness before execution 2. FixBindingGate — patches must bind to root cause, not symptom 3. FactBindingGate — tracks every tool call across the engine lifecycle 4. Evidence Ledger — structured audit trail of all operations 5. Enforce — verifies sufficient evidence before allowing side effects Regression tested: 6/6 protection, zero false-kills. FileClaimGate caught a real hallucination (model claimed "modified docstring.py" but edit tool actually failed). ### Other v0.8.2 improvements - Cache Rails: 97.35% warm input-token hit rate, 93.03% cost savings - 7 engines, 9 tool families, 12 command groups, 26 built-in tools - MCP protocol support with pluggable transports - Online-chain regression verification framework - 2110 tests, all passing ## Verification - Published v0.8.0 wheel and source distribution pass Twine metadata checks - Full offline suite: 2110 passed; Ruff, compileall, and package builds passed - SWE-bench full report: evals/results/SWEBENCH_REPORT.md - Regression verification: 6/6 baseline cells protected, zero regression - Previous live checks also verified V4 Pro reasoning_effort and native tool-call protocol continuity ## Files changed | File | Purpose | | --- | --- | | docs/xenon.md | Updated English v0.8.2 installation and DeepSeek guide | | docs/xenon.zh-CN.md | Updated Chinese guide | | README.md | English tool index entry | | README.zh-CN.md | Chinese tool index entry | ## Checklist - [x] Published v0.8.2 release and reproducible install command - [x] SWE-bench official evaluation: 40.0% instance / 45.8% cell, same-model A/B +6.7pp/+11.0pp, fully reproducible - [x] Evidence Runtime: 5-stage online verification chain, 6/6 zero regression - [x] Current V4 model names, reasoning effort, and native tool continuity - [x] Provider-reported cache evidence and Cache Rails documented - [x] 2110 tests passing - [x] English and Chinese guides in one PR - [x] Alphabetical README entry --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 09:48:01 +00:00