Skip to content

Releases

DUO v0.6.3

Choose a tag to compare

@CZ-ZL CZ-ZL released this 08 Nov 03:33

DUO v0.6.3 provides a complete public path from a fixed download to a saved experiment report.

  • Concise English/Chinese Quickstarts, a read-only host check and a pinned host manifest.
  • Controller responsibilities separated into operation accounting, evaluation, selection and final assessment; public interfaces preserved.
  • Shared Simple baseline documentation and overview diagrams, independent of individual studies.
  • Continuous documentation, diagram and script-inventory checks alongside installation and behavioral tests.

Start with Quickstart or 中文快速开始. Download the plugin archive and SHA256SUMS below; verify before extracting. host-package.json prepares the documented isolated host. VERIFICATION.json binds the exact source, CI run, archive and acceptance scope.

Supported foundation: Linux, Node 24+, DSH 0.1.2-rc.1 and the declared Cordis 4.0.2 family. DSH remains an upstream preview. Providers are trusted in-process code. This release is not an npm publication, security certification or evidence that DUO outperforms Simple.

No paid experiment was run for this release. Earlier live starter evidence remains scoped to its original run; current checks cover the free product journey, compatibility, accounting, recovery and packaging. See release notes and QC coverage.

PUBLIC_ACCEPTANCE.json records the successful anonymous download, checksum and fresh free report journey on the published archive.

DUO 0.6.2 — Target/provider compatibility

Choose a tag to compare

@CZ-ZL CZ-ZL released this 16 Sep 04:30

DUO 0.6.2 — Target/provider compatibility

A Target could previously be paired with a persona-only model provider without a type-scope rejection during planning. This release declares the built-in providers' supported Target kinds and rejects incompatible combinations before work.

  • Discovery and preparation expose targetCompatibility; summary/full plans retain providers.targetCompatibility.
  • Wrong Generator/Executor kinds receive structured repair guidance before model work or cost reservation.
  • Legacy undeclared providers remain usable with support marked UNKNOWN. Evaluate-only does not inspect or require a Generator.
  • Persona remains a supported adapter. The custom example uses content snapshots and matching work providers, without persona fields.
  • Existing personas, snapshots, config context and historical experiments are preserved.

This is a developer-preview product correctness fix, not evidence of a quality or cost advantage. Target-kind declarations do not prove task compatibility; caller-owned providers still need validation. No npm publication.

v0.6.1 — Agent discovery and public installation

Choose a tag to compare

@CZ-ZL CZ-ZL released this 16 Sep 03:18

DUO is a DSH-native plugin for evaluating supported Agent targets and comparing candidate changes within an authorized budget. This developer-preview release adds task-oriented discovery and public installation to the previously reviewed product foundation; it does not establish an optimization advantage.

Install

Download dual-loop-dsh-plugin-0.6.1.tgz and SHA256SUMS from this release. In the download directory:

sha256sum -c SHA256SUMS
dsh plugin --profile YOUR_PROFILE add ./dual-loop-dsh-plugin-0.6.1.tgz
dsh --profile YOUR_PROFILE --dump-config

Use an authorized DSH profile. The repository root is a development package; the installable bundle is the dsh-plugin subpackage. After loading, ask your Agent to start with dualloop_describe; default preparation grants no execution or model spending. No npm publication.

Changes

  • Searchable DSH plugin metadata and bilingual task examples.
  • Existing describe/design tool descriptions explain when an Agent should use them.
  • Versioned prebuilt tarball and explicit subpackage catalog entry.
  • Previously reviewed baseline-quality/identity repairs, preparation controls, evidence semantics and public-source audit remain documented separately.

Verification

Code: 31a88508f5d8748ef166f816400b32e4a58500d2.
Exact product CI: https://github.com/CZ-ZL/duo/actions/runs/35050843877
This run passed: 340 native tests; 451 Python product tests (86 research deselected); fresh dependency installation, public examples, Slow evidence, baseline repair and 21 host profiles. The CI tarball matches the attached SHA256SUMS. See VERIFICATION.json for the exact scope.

Linux; tested Node 24.14.1, DSH 0.1.2-rc.1, Cordis 4.0.2 and pnpm 11.24.0. Native runtime needs neither Python nor a model account; real-model work requires separately configured providers and authorization. Custom targets need adapters; config support is partial. Trusted in-process providers and settled-checkpoint recovery boundaries apply.

Product acceptance is not method superiority. No new paid model experiment, automatic candidate adoption or cross-platform claim. Marketplace submission is a separate process, not an endorsement.

中文:这是 DUO 的公开开发预览安装包。新增面向 Agent 的发现说明、固定版本下载和校验方式;安装后先调用 dualloop_describe。不会自动付费或采用候选,也不宣称方法优于单环。

v0.6.0 — Slow Evidence Strategy

Choose a tag to compare

@CZ-ZL CZ-ZL released this 14 Sep 18:39

DUO 0.6.0 defines Slow as bounded evidence acquisition and decision policy, with explicit evaluate/basic/dual/auto presets. It repairs cumulative allowance admission and terminal selection, adds typed strategy contracts, and provides English and Chinese README guides.

本版完成 Slow evidence 产品语义、预算与选择一致性修复、代码整理和中英文使用入口。免费本地 provider 的产品验收与独立 Caller 使用不代表 DUO 方法优于合理单环;历史负结果保留。

The attached tarball is the reviewed installable DSH package. See CURRENT_STATUS.md, RELEASE_NOTES_v0.6.0.md and docs/slow-evidence-strategy/ACCEPTANCE.json for exact acceptance and limitations. Providers remain trusted in-process; recovery is bounded to settled unchanged checkpoints. Graph/Bayesian and method research are outside this release.

No package was published to npm. No candidate was automatically deployed. External model requests: 0; external model API fees: CNY 0.

Verified code commit: 5a45eb912dfdf093ffa344524482a32ba4cdea22.
Successful full GitHub CI: native 321, Python product 451, research 86 deselected, package/fresh official-registry installation and DSH acceptance PASS. CI and local tarballs are byte-identical.

The release tag includes the subsequent completion-documentation commit; no runtime/package/verification inputs changed. Independent Caller: 47 public CLI invocations, 48 settled native operations, 7 original scenarios plus one guide-only repack smoke. Calling Agent inference cost is separate and not measured by DUO.

Package SHA-256: 71364e3a6b9085c19763f0551a8f1d594c77c7545da2e2398b68b935f27ae2d3.

v0.5.0 — Product Foundation

Choose a tag to compare

@CZ-ZL CZ-ZL released this 14 Sep 15:26

DUO 0.5.0 — Product Foundation

A DSH-native evaluation/optimization component with a documented supported
boundary, installable tarball and public Agent workflow. The default bundle
exposes preparation and waits for providers. No fixture or model spend is
silently enabled.

This version unifies capabilities, moves Target identity/history projection out
of Core, adds evaluate/optimize preparation presets and consistent diagnostics,
and ships independent local examples for evaluation, optimization, BYO evaluator,
component replacement, warm start and a custom Target. Existing Cordis services,
DSH runtime, budget/receipt checks and legacy APIs are reused.

Product verification includes native and Python regression tests, real isolated
DSH profiles, actual tarball installation, public package workflows and a fresh
Calling Agent following only installed docs. Exact receipts and CI state are in
PRODUCT_ACCEPTANCE.json and docs/product-foundation/REPORT.md.

Upgrade by inspecting and re-planning; old checkpoint digests bind the old
implementation/provider graph. Fixture users must bind it explicitly. The
separately versioned Python protocol remains 0.1.0.

Known limits: Linux/Node24 and tested DSH snapshot; trusted in-process providers;
fetch configuration Target is partial; original-deadline settled checkpoint
recovery only; outer Caller fees are outside the native ledger. Examples measure
local text hygiene without independent final qualification. TypeScript semantic
compilation and Windows/macOS are not qualified. No automatic adoption or npm
registry publication.

RESEARCH NOT PROVEN: product acceptance does not establish a DUO quality or cost
advantage. Method research is paused and all earlier results remain unchanged.

Final product CI: https://github.com/CZ-ZL/duo/actions/runs/34860825068
Verified code: 2ac5451. Subsequent closure
changes are acceptance documentation only. Native 295 PASS; Python product
451 PASS /86 research deselected; 21 DSH scenarios; 55 public checks/80 steps;
fresh dependency install and actual installed-profile plan/run/report PASS.
Independent Caller: 43 public calls, six runs, no implementer intervention,
plus bounded final documentation/preset follow-up.