DeepSeek Harness and external Agent coordination

Qiongli can be loaded into DeepSeek Harness as a Cordis bundle containing the canonical qiongli and no-qiongli Skills and native Full MCP. The coordinating Host can also prepare and collect bounded proposals from Codex, Claude Code, DeepSeek Harness and Antigravity CLI. The Host executes and supervises the child process; Qiongli validates the result against the original task and sources.

Workflow entries in DeepSeek Harness

New bundles expose the same 22 Skill entries as Codex: qiongli, no-qiongli and 20 qiongli-<workflow> entries, including qiongli-paper-read, qiongli-lit-review, qiongli-study-design and qiongli-paper-write. The existing canonical workflow wrapper generator supplies identical entry bytes for both Hosts; every wrapper points to the shared Skill and its workflow. Descriptions come from the canonical workflows. Adding a workflow through that owner adds its entry to both projections on regeneration.

The bundle's skills.json provider catalog registers each entry separately, with modelInvocable and userInvocable both enabled. DSH can therefore list the entries on its Skill surfaces and load one by name, such as ctx.skills.get('qiongli-paper-read'); the model-facing interface is skill({name: 'qiongli-paper-read'}). See the official Skill contract. The entry's directory is its resource base, so its shared relative paths resolve. DSH's normal project/user override priority remains in force; verify that the winning entries belong to provider qiongli when checking an installation.

The earlier two-entry bundle remains verifiable, but it does not acquire these entries automatically. Regenerate and install the new bundle into the selected profile, then check the actual catalog and load every entry. Packaging checks compare the complete entry sets and bytes with Codex and reject changed or missing wrappers/catalogs even when their receipt hashes are rewritten.

Build and load the DeepSeek bundle

For normal installation, use the released npm bundle described in Plugin setup. For a custom development bundle, use the existing native content exporter and archive verifier. Build from a clean source commit; use that same commit for the binary and content export. From the repository root, with the current native binary built:

QIONGLI_NATIVE_SOURCE_COMMIT="$(git rev-parse HEAD)" cargo run --locked \
  --manifest-path packages/qiongli-native/Cargo.toml -p qiongli \
  --example export_marketplace_content -- /absolute/new-content-directory
python3 tooling/scripts/native_marketplace_plugins.py \
  --content-dir /absolute/new-content-directory --out-dir /absolute/new-bundle-directory \
  --version 2.3.0 --commit "$(git rev-parse HEAD)" \
  --binary packages/qiongli-native/target/debug/qiongli \
  --target aarch64-apple-darwin --platform deepseek

Use the actual executable version and target; Linux x64/ARM64 and Windows x64 layouts are also supported by the generator. Each output directory must be new. Existing source/byte/target/receipt checks remain enforced. Default packaging continues to build only Codex and Claude; no release or upload is performed.

Extract the generated archive. Its plugins/qiongli-macos-arm64 directory contains package.json and the bundle patch (the name differs by target/channel). Install that directory with DeepSeek's own plugin manager into the chosen profile, which requires its documented pnpm dependency:

dsh plugin --profile web add /absolute/extracted/plugins/qiongli-macos-arm64
dsh --profile web --dump-config
dsh web

The bundle adds Skills and one mcp-qiongli row; it does not select a model, change authentication or replace the profile's other configuration. Use only one Qiongli bundle per profile. Inspect the actual Skill catalog and MCP tool list, then call qiongli_config_status to verify the live connection. For the current shared MCP schemas, DeepSeek uses route platform unknown and Host family other-local, with actual versions and observed component states. A loaded Plugin is not project-write approval. Removal uses the same Host manager: dsh plugin --profile web remove dsh-qiongli-macos-arm64.

For released 2.1+ versions, the standard qiongli npm package also contains this bundle. Desktop users enter [email protected] in Add plugin with Official npm registry; local users run dsh plugin --profile desktop add [email protected] (substitute the intended profile). The official manager also accepts the local npm .tgz path. This retains all 22 Codex-equivalent entries and Full MCP while reusing the package's platform-selected executable. No global CLI installation is required. See installation and upgrade steps.

Dispatch and collect an external proposal

The coordinator obtains its handoff and authenticated source reads from an existing Full MCP run. Save that handoff and a JSON packet containing only scope and sourceText; the source snapshot must be authorized for the selected external Host. Run:

qiongli agent claude prepare --handoff handoff.json --packet packet.json --json

Replace claude with codex, deepseek or antigravity as requested. Execute the returned argv using the current Host's execution tool, feed stdin unchanged, and merge any returned env overrides into only that child process. Use an authorized isolated working directory, a bounded deadline and separate stdout and stderr files. Wait for the actual process exit. Then collect:

qiongli agent claude collect --handoff handoff.json --packet packet.json \
  --events stdout.json --status completed --exit-code 0 --json

Supply the observed status and exit code, never the model's interpretation. Failed, cancelled, timed-out, truncated, mismatched and unsuccessful streams cannot become completed delegation results. Do not copy a result between Host adapters. The coordinator reconciles findings, rechecks current source bindings and submits through its own MCP process. Existing preview/approval/CAS owns all research writes. No adapter changes the user's saved model, reasoning effort or account; no direct provider API, Agent daemon, persistent resume or silent fallback is introduced.

Adapter	Required protocol	Per-run constraint
Codex	codex exec --json with thread/turn events	Ephemeral read-only sandbox
Claude Code	claude --print --output-format json success envelope and session ID	Built-in tools and MCP disabled; no saved session
DeepSeek	dsh --profile headless --json session, turn-end and final events	DSH_PERMISSION_MODE=read-only
Antigravity	agy --input-format stream-json --output-format stream-json	One input message; plan mode; slash expansion disabled

These are proposal restrictions, not complete isolation from configured hooks or all Host tools. Tool inventory and policy remain Host-owned. A permission prompt, missing login or unavailable protocol is a visible gap; never auto-approve it or silently substitute a different model/Host. Reap interrupted processes before starting an explicitly authorized fresh run.

Protocol references and limits

The DeepSeek protocol was checked against official source 477b4f420553e8a52c2fbccc464d7561b239c443, including the headless runner, bundle contract and Skill provider. The September 24 npm @deepseek-ai/[email protected] observation supported bundles but its headless help did not expose --json. That build cannot satisfy the external collector. Plain-text output is deliberately insufficient to identify a completed run.

On September 30, the installed macOS ARM64 Desktop CLI 0.2.0-rc.2 accepted the current Qiongli bundle through dsh plugin. In isolated profiles, repeat installation preserved the profile, both Skills loaded, 32 Full MCP tools were registered, and qiongli_config_status returned ok; activation exited 0. Removal preserved the other bundles. This version also exposes headless --json. Actual headless model execution and installation into the user's profile remain separate checks; the execution plan records the exact source and observations.

The other protocol references are the official Claude Code CLI and Antigravity headless interface. Protocol checks and synthetic tests do not establish installed-Host, real research, two-Host collaboration or cross-platform release acceptance. Exact observations and remaining checks belong to the current execution plan.