deepseek-harness-jev

English | 简体中文 | 中文功能与实测网站

Native DeepSeek Harness (DSH) plugin integrating TypeSafe Jev as a System One decision layer.

deepseek-harness-jev connects DeepSeek Harness (DSH) to Jev by TypeSafe AI for agent skill and file selection, task supervision, shared-finding corrections, tool-output filtering, single-operation approval assistance, and historical stage navigation. Its 12 features are individually configurable from one Jev settings page and are all disabled by default.

The main model continues to plan, generate answers, and call native tools. The plugin automatically invokes enabled Jev judgments at DSH extension points for skill catalogs, agent lifecycle, tool results, and approvals, then applies results according to each feature. DSH configures the main model; Jev has a separate connection. Integration uses public Cordis / DSH plugin APIs without modifying the host source.

This is an independent community project, not an official DeepSeek or Jev release. It is an early-stage plugin tested with DSH 0.1.7-rc.2; its APIs and model judgments are not a correctness guarantee.

The Chinese feature website explains each DSH integration point, the information sent to Jev, and the observed test cases and limits.

What is included?

The following features are in main. Every feature is independently disabled by default. Installing the package does not enable them.

Example feature settings from an earlier build. The screenshot shows nine switches and user-selected states; current main includes the twelve features listed above, all disabled on a fresh installation.

All features share a connection, profile-scoped settings, decision records, and operation receipts. Most agent-facing features target live Web root sessions; correcting a child agent does not enable every feature inside that child.

Feature branches are not all included in main. Tool-output filtering is included in main; codex/jev-tool-output-admission preserves its development snapshot. Native web execution is on codex/jev-native-web-execution and is paused; ordinary-site effectiveness has not passed acceptance. Historical split branches preserve earlier work. See branch status before switching branches; this table always describes main.

Stage navigation

The Stage navigation tab follows Trajectory in the Web client. It reads recorded turns without changing the original Session.

The independent Stage navigation switch is off by default. It controls the tab's visibility: enabling shows the page without calling Jev, and disabling hides it, cancels unfinished classification, and retains saved results. Analysis starts only when the user selects a completed turn or requests the session's unanalyzed completed turns.

Each classification covers one complete DSH model step: its recorded reasoning, text, all tool calls, and paired results. Jev selects one of six stages, mixed, or unknown; adjacent equal labels merge only within the same turn. The page keeps a multi-turn directory beside the original steps and exposes the actual classification input and answer. Labels and confidence do not establish tool success or classification accuracy. See the package reference for input, storage, and failure behavior.

Install through the Web UI (recommended)

If you already use DSH 0.1.7-rc.2 Web, install directly from the GitHub repository URL. No source checkout, manual packaging, or npm login is required.

1. Open Plugins in the sidebar → Add plugin.
2. Paste the GitHub URL below into Package name or address, then click Install.
3. Click Enable now after installation. Restart the current profile only if DSH says it will load on the next start.
4. Open Jev, configure the endpoint, model, and API key, then enable the individual features you need.

https://github.com/luobosibing2/deepseek-harness-jev

Paste the repository URL into “Package name or address”, then click Install.

Enabling the package does not enable its 12 Jev features; they remain off by default. Installation applies to the Host profile serving the current Web UI. The Host needs pnpm and access to GitHub. The repository includes the plugin entry and prebuilt files, so installation does not compile source on the user's machine or require an npm registry publication.

The GitHub entry provides the main features, not experimental branches. The older v0.1.0 release does not include the newly integrated log filters. Build from source below only when changing or building the code yourself.

Install from source (developers)

Use the following steps when modifying or building the plugin yourself. Existing DSH Web users can install using the GitHub URL above.

Build requirements

• Node.js 24.11 or later is recommended; the publication build is checked on Node 24.14.1.
• pnpm 11.7.0 available on PATH.
• DeepSeek Harness CLI 0.1.7-rc.2. The plugin pins the corresponding DSH peers and Cordis 4.0.4; newer versions are not automatically supported.
• A configured main-model provider in DSH, plus your own Jev-compatible System One endpoint and credentials.

If needed, install the tools:

npm install --global pnpm@11.7.0 @deepseek-ai/dsh@0.1.7-rc.2

Build the package

Build a .tgz from source, then install it through the Web UI or official CLI.

git clone https://github.com/luobosibing2/deepseek-harness-jev.git
cd deepseek-harness-jev
pnpm install --frozen-lockfile --ignore-scripts
pnpm run build
mkdir -p dist
pnpm -C packages/jev pack --pack-destination "$PWD/dist"

Create a separate trial profile with the CLI

You can also paste the built tarball's absolute path into an existing Web plugin manager. The CLI method below creates a separate trial environment.

Use a new, unused profile name for a first trial; the example uses jev. Initialize it from the Web template before adding the plugin:

dsh --profile jev --from-default-profile web --dump-default-config > /dev/null
dsh plugin --profile jev add ./dist/dsh-jev-plugin-0.1.0.tgz
dsh --profile jev

The first command creates the Web profile without launching it. Adding a plugin to a brand-new profile without this step initializes only the base configuration, not the Web application. The plugin's bundle patch is applied by the official installer; no manual host-source changes are needed.

Open the authenticated Web address printed by DSH. Configure your main model through DSH, then open the plugin's Jev page.

Configure Jev

1. Set the full System One endpoint, for example https://api.typesafe.ai/v1/systemone.
2. Set the model, for example jev-latest.
3. Choose a DSH credential reference, save the connection, and save your API key using the page's credential control. Do not put a key in source files or a repository URL.
4. Review the timeout, then enable only the features you need.
5. Inspect Decision records for input, answers, attempts, and actual adoption or execution receipts.

The main agent's provider and the Jev judgment connection are separate. A credential marked “configured” is not a successful connectivity test. Connection tests and enabled judgments make requests to your provider.

Selection defaults are 5 skill summaries, at most 40 glob matches eligible for ranking, and 12 displayed ranked paths. A larger glob skips Jev rather than silently judging only the first 40. Supervision defaults are a drift check every 6 completed model steps and a pause after 3 native goal rounds without progress. These values can be changed without enabling the features.

Long-log and test-log admission have independent switches, both off by default. Generic command logs start at 6,000 Unicode code points and recognized test logs at 4,000. The default omit-probability threshold is 0.8 and the judgment wait limit is 4 seconds. The settings page exposes these and the other admission budgets without enabling either feature.

Behavior and limitations

• Reminders are advisory. Drift and instruction guidance do not block or cancel tools, and do not force the main model to comply.
• Completion checks in a fixed 16-slot diagnosis. With scripted initial errors, real Jev judgments and DeepSeek supplements met the core requirements in 6/6 enabled error cases; 2/2 accurate enabled controls were judged complete without a supplement. See the new results. Completion checks review recorded evidence and do not independently run tests. This diagnosis does not establish a natural-task error rate or average coding benefit. In an earlier long coding trial, the default evidence budget cropped the original requirement, leading to zero adopted judgments and zero automatic supplements before manual Retry/Cancel; see the historical report.
• Approvals remain single-operation. Workspace approval neither changes the session's sandbox mode nor overrides fixed host checks. approve can supply allowed-once; unauthorized or unknown returns to the original human approval flow. Technical failures retain manual Retry/Cancel.
• Shared corrections have a limited scope. They process already-shared reports and messages, not every agent's private exploration. Automatic delivery targets the live root agent and its active, continuable direct children. Duplicate corrections can still arise when the same finding appears in different report forms.
• Judgment success is not action success. The ledger distinguishes an answer, its adoption, permission issuance, and execution results.
• Log admission keeps an original reference. It changes only eligible model-visible tool text after execution; DSH's immediate spill, tool output limits, and later context compaction still apply. An isolated real-profile build reduced one 8,510-character log by 75.7%. A neutral 180-test run reached the test-log judge but stayed complete because its omit probabilities were below 0.8; that run does not establish test-log reduction effectiveness. See the tool-output admission report.
• Validation is scoped. Deterministic tests establish integration. Limited real-service examples do not establish general semantic accuracy. See validation notes.

Enabled features send the relevant task context or operation data to the configured judgment endpoint. Exact judgment inputs and answers are stored in the profile's local plugin records; model-visible effects use normal DSH session records. Keep runtime records and credentials private. Public source history excludes personal QA screenshots and raw session captures.

Updating or removing the plugin

For an existing profile, rebuild and pack, then install the new tarball with dsh plugin --profile jev add <new-tarball-path> and restart that profile. Do not rerun --from-default-profile on an existing profile. Use a new tarball filename for a changed build of the same package version and check the installed contents when validating an update.

Disable individual features in the Jev page. For package removal, consult dsh plugin --help for the CLI version you have installed. Removing or switching the package can remove branch-specific features; keep a profile backup before replacing an experimental branch build.

Development

The project is named deepseek-harness-jev; its internal package and import identifier remains @dsh-jev/plugin, matching existing profile plugin configurations.

The repository root is the GitHub install entry; packages/jev retains development sources. pnpm run build also regenerates runtime/; commit these generated files when releasing source changes.

Maintainers can use the DeepSWE paired-evaluation runner for coding tasks and the glob-ranking pipeline for fixed path-selection cases. Both use isolated DSH/Pier trials and keep model execution separate from offline checks and reports. The Chinese evaluation guide covers the coding-task workflow.

The completion and skill evaluation entry provides four named scenarios, including completion-recovery-16, explicit resource manifests, new-batch preparation, keyless preflight, paid execution, and offline reports. The new fixed diagnosis and historical results remain separate from validation of the maintained runner. A rerun records the runner commit and source hashes; dependencies and credentials remain local prerequisites.

The public glob experiment records six synthetic cases and 12 real DeepSeek/Jev trials: four successful Jev judgments returned 69 scores, while zero and 41 candidates bypassed ranking. It retains quality failures, recovery checks, source-read counts, and estimated costs. The historical run used a pinned earlier plugin artifact; the maintained pipeline does not imply the current main was rerun or that general task success improved.

pnpm run typecheck
pnpm run build
pnpm exec vitest run packages/jev/tests/host.test.ts packages/jev/tests/wire.test.ts

Run the focused tests for the feature you change. Do not enable real-provider experiments or use someone else's credentials without explicit authorization. Development fixtures and tests are excluded from the installable tarball.

• Package reference and consumer API
• Workspace-approval integration tests
• Workspace-approval QA cases
• Tool-output admission report
• Branch status
• Validation notes

DSH integration points

License and acknowledgements

MIT; see LICENSE. Package-level third-party licenses are included in THIRD_PARTY_NOTICES.md.

The feature research was inspired by Mu. This project implements DSH plugins against public extension points; it does not ship a modified DeepSeek Harness, Mu, or Cua runtime. DeepSeek Harness, its Typert tooling, and Zod retain their respective notices.