根据有界证据记录限定范围的工作流规则与纠错提醒,提供本地设置并支持可选 Brain Hub 集成。
安装
# Release 预构建包
dsh plugin --profile web add "https://github.com/Missher12/dsh-missher-evolution/releases/download/v0.7.0/dsh-missher-evolution-0.7.0.tgz"
# GitHub 源码(首次需按提示配置 allowBuilds 构建授权后重试)
dsh plugin --profile web add github:Missher12/dsh-missher-evolution
装任何插件都等于在你的机器上跑第三方代码,权限和你本人一样大——能读你的文件、用你的凭据、访问网络,工具审批管不到它。GitHub 来源的插件还会在安装时执行构建脚本——pnpm 默认拦截,所以安装可能停在 ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED 或 ERR_PNPM_IGNORED_BUILDS;dsh 会打印出需要添加的确切键名,把它加进该 profile 的 pnpm-workspace.yaml 的 allowBuilds 下,重跑一次即可装上。放行构建本身就是一次信任判断:请只安装可信来源,并尽量锁定 commit(github:owner/repo#sha)。
README
该插件的 README 只有英文版本。
dsh-missher-evolution is a prebuilt Cordis bundle for DeepSeek Harness Desktop
>=0.3.8, using native lifecycle hooks or the optional Brain Hub service. Version 0.7.0 retains current eventAt-based session logs and evidence-bound registered methods, and belongs to the unified MSE release. It learns only bounded workflow categories, fixed preferences,
closed outcomes, counters, hashes, and verified Chinese instructions. It does not
train model weights or rewrite Harness, the plugin, or user projects.
It is the first production reference adapter for @missher/evolution-sdk. Cordis lifecycle events
are translated by the Harness adapter, while selection, Capture, attribution, persistence, and
maintenance run through the host-neutral SDK. The SDK is bundled into this plugin's generated Host
code, so Harness users do not install or configure a second package.
The same JavaScript tarball runs on macOS Intel, macOS Apple Silicon, and Windows x64. There is no install-time build, native binary, external daemon, Python runtime, messaging connector, or remote analytics service.
Install
Public distribution: v0.7.0. The release tarball is byte-identical to the unified MSE 0.7.0 Harness artifact. See AGENT_INTEGRATION.md before installing on a new computer. Marketplace submission is separate from release publication; a release alone does not make this plugin searchable in DSH Market.
Install the prebuilt public archive after verifying the release checksum:
dsh plugin --profile <profile> add https://github.com/Missher12/dsh-missher-evolution/releases/download/v0.7.0/dsh-missher-evolution-0.7.0.tgz
Close the target host before changing an existing installation, and preserve a
matched backup of the plugin, Profile configuration and $DSH_HOME/missher-evolution.
Do not overwrite the experimental 0.1.2 review branch's data: its approval schema
is incompatible with this unified release. Different Profiles in the same DSH_HOME
still share evolution state; a separate Profile alone does not isolate that data.
Build and verify a local release tarball from this directory:
pnpm install --frozen-lockfile
pnpm run build
pnpm pack --pack-destination ./dist
node scripts/verify-package.mjs ./dist/dsh-missher-evolution-0.7.0.tgz
Install that tarball into the Harness profile that should use it:
dsh plugin --profile <profile> add <absolute-path-to>/dsh-missher-evolution-0.7.0.tgz
Harness activates the bundle for that profile. Open Harness Settings and select “Missher Evolution” to inspect status, enable or disable learning, review rules, or start a protected reset.
Here “review” means inspect rules. Version 0.7.0 does not expose per-rule
reviewRule approval/revocation or a restore Remote; do not promise those APIs.
Verify
Inspect the composed profile without starting an agent:
dsh --profile <profile> --dump-config
The output should contain one dsh-missher-evolution bundle layer and one
enabled missher-evolution Host entry. node scripts/verify-package.mjs <tgz>
prints fixed JSON containing ok, file count, byte count, and SHA-256.
Runtime verification should use the Settings snapshot. Recall has one owner: when missherBrain
exists it offers scoped handles; otherwise native agent/pre-step uses session.header.cwd and
credits only its context message actually appended to the session. Rejected/canceled context is not
an injection. eventAt is scanned within a bounded tail, with legacy event arrays still supported.
Candidate evidence can create a Trial, but Active promotion additionally
requires three supported treatment successes, two supported control observations, no treatment
contradiction, and at least 0.15 observed uplift. These minimum samples do not establish statistical
significance. Instruction rewrites start a new evidence epoch, and partial acceptance cannot be used
to evaluate the trial. The Settings snapshot distinguishes recorded tasks from causal evidence.
Registered Methods
Accepted source-date and unknown-value checks can now retain bounded failure/repair case metadata. Maintenance proposes a compatible registered method using the existing AI budget or a clearly identified deterministic fallback. A separate fixed held-out evaluator tests failures, correct outputs and non-applicable inputs before the method can enrich an eligible local rule. These offline results never award Active status, real-task successes or measured model uplift.
After a trusted first failed file check, an existing next agent/pre-step continuation
can receive one fixed repair hint. It remains within the user's original authority,
does not itself write files, and requires a fresh native check. Read-only requests,
Stop, closed turns, changed evidence, removed rules or exhausted budgets suppress the
hint. There is no strict final-delivery blocker and no automatic final-turn restart.
The Brain path also supports this bounded continuation without duplicating rule recall.
Portable templates use the SDK's explicit local operator interface. Only registered method IDs transfer; target scope, validation and reputation never come from the source host. Each target needs its own relevant evidence and local evaluation.
Data
Each Harness home owns an independent local directory:
$DSH_HOME/missher-evolution/
It contains schema-validated state, a bounded closed-event audit, a short-lived cross-process lock, and validated backups. Durable data never contains raw user messages, assistant responses, thinking, tool arguments, tool results, absolute paths, URLs, email addresses, credentials, or provider/model route names. Project identity is represented only by a SHA-256 scope hash; controlled semantic identifiers drive recall and duplicate consolidation.
The state is not shared with any other agent host. Model review uses only rule metadata and the most recent in-memory foreground route, and is skipped when no route exists. Store, classifier, lock, or advisor failures are fail-open and do not reject a normal Harness step.
Reset
In Settings, choose Reset and complete both confirmations. The Remote requires
the exact current state revision and the literal confirmation RESET; stale or
malformed requests are rejected. Reset clears learned rules and counters while
preserving the current enabled preference.
Backup
A reset creates and validates a backup before replacing state. Scheduled
maintenance also creates a backup before applying expiry, retirement, or a
bounded advisor rewrite. The Settings response identifies the reset backup;
backup contents remain local under $DSH_HOME/missher-evolution/backups/.
Do not edit or copy a live state directory between profiles. Stop Harness before performing any explicit offline backup or restore operation.
Uninstall
Remove the plugin from one profile with:
dsh plugin --profile <profile> remove dsh-missher-evolution
Uninstall removes the profile dependency and bundle layer. It intentionally
preserves $DSH_HOME/missher-evolution so uninstall cannot erase learned data
or unrelated profile/session content. Delete that directory separately only
after Harness is stopped and any desired backup is verified.
Limitations
- The baseline is DeepSeek Harness Desktop
>=0.3.8;missherBrainis optional. Compatibility is established by actual tests, not version comparison alone. Its packaged dsh interfaces remain within the declared>=0.1.0-rc.5 <0.2.0peer range. - Only direct foreground user turns are learned; subagents, internal work, plugin messages, scheduled work, and tool continuations are filtered.
- The plugin recognizes closed preference and experience allowlists and does not copy arbitrary correction text into durable rules.
- It improves future instructions and workflow reuse; it does not fine-tune or replace the underlying DeepSeek model.
- It does not import or synchronize state from another agent host.
- No process continues after Harness exits.
- Native smoke uses the real profile CLI plus controlled adapter events. It is not a measurement of autonomous model task improvement or a claim that every future host API is compatible.
Development
pnpm test
pnpm exec tsc -p tsconfig.json --noEmit
pnpm exec tsc -p tsconfig.client.json --noEmit
Official distribution
Source remains MIT. Existing official distribution policy and trademark policy continue to apply. Public availability does not grant third parties permission to present mirrors as Missher-endorsed releases. The public source export contains the Harness adapter and its bundled MIT TypeScript core, with provenance in SOURCE_PROVENANCE.json. It does not contain other host adapters or any live learning state.
链接
同类插件
Q00/ouroboros#integrations/dsh-plugin★ 5825
通过 DSH MCP 客户端挂载 Ouroboros 的纯配置包,在 DSH 中提供 36 个涵盖需求访谈、Seed、执行、评估与演化流程的工具。
huangruiteng/loopx#dsh-loopx-plugin★ 5818
LoopX——面向长周期 Agent 的提供商中立、本地优先状态内核与控制平面:在 DeepSeek Harness 执行层之上持久化 Goal、Todo、门禁、证据、配额、恢复与交接状态;插件负责引导安装 CLI 与技能、准入有界的同会话续跑,并为精确绑定的工作循环提供本地 GoalBar。
chuspeeism/dashi-taskboard#deepseek-harness★ 3070
把当前已安装并运行中的 Codex Taskboard 嵌入 DeepSeek Harness 侧边栏,并通过 Launcher 运行时描述文件连接,而不是使用固定端口。
NanmiCoder/dsh-agent-teams★ 1596
AgentTeams 多智能体团队。
tong-io/tongflow#dsh-tongflow★ 1020
基于 TongFlow 的“片场”插件,用于图片、配音、音乐与视频制作:agent 为每个资产生成 TongFlow 工作流文件(.tongflow.json)并通过 TongFlow 插件执行,内嵌工作流画布,按镜头/角色/take 组织项目,附漫剧模板;以 @tongflow 开头的会话进入 Studio 界面。
EthanYoQ/AI-Novel-Writer#dsh-ai-novel-writer★ 824
安装专用 AI 小说创作预设与工作台:提供带修订号的本地项目资产、紧凑侧边工作台,以及需要原生审批的逐文件变更。
社区评论
评论公开保存在 GitHub Discussions。加载评论会连接 GitHub 和 Giscus;发表内容需要 GitHub 账号。