dsh-research-atlas

Baigong Research Atlas — DeepSeek Harness plugin

中文项目说明 · English project overview · Source

Native Cordis tools for persistent research tasks, evidence and stage handoffs across executors. This is a community plugin, not a DeepSeek first-party package. It calls the existing Python CLI; it does not duplicate the research core or use the model HTTP bridge.

Validated host: @deepseek-ai/dsh / dsh-tools 0.1.5-rc.3, Cordis 4.0.2, Schemastery 3.18.2, Node 24.18.0, Windows. Other Harness versions and operating systems are unverified. Harness is a developer preview: pin the tested host and re-run the tests before upgrading.

Install

Prerequisites: a trusted checkout of baigong_research_atlas with its Python environment prepared according to the project instructions, and the tested Harness host already installed. The plugin contains ready-to-run JavaScript, with no build or installation scripts. It uses the host's declared peers; no separate model SDK is needed.

From the repository root:

dsh --profile research-atlas --from-default-profile web --help
dsh plugin --profile research-atlas add ./plugins/dsh-research-atlas
dsh --profile research-atlas --dump-config

Alternatively, install the release tarball with dsh plugin --profile research-atlas add ./dsh-research-atlas-0.1.0.tgz. A release tarball contains the adapter, not the Python checkout. Do not install the repository root as an npm package.

The bundle is disabled on installation. In $DSH_HOME/profiles/research-atlas/cordis.patch.yml, configure and enable the row:

- id: baigong-research-atlas
  disabled: false
  config:
    projectRoot: '/absolute/path/to/baigong_research_atlas'
    python: '/absolute/path/to/its/python'
    configPath: config/research_assistant_v0001.json
    allowWrites: false
    timeoutMs: 60000
    maxOutputBytes: 8388608

On Windows use quoted forward-slash paths, e.g. C:/your-checkout/.venv/Scripts/python.exe. Paths are deployment settings, never model arguments. The project must contain provenance/source_manifest.json and the existing skill CLI. Initialize a new research ledger explicitly using that CLI's initialize operation before using task operations. The plugin never initializes or migrates a database automatically.

The first command creates a fresh profile from the shipped web template without starting a conversation. Run dsh --profile research-atlas --host 127.0.0.1 --no-open to start its local web surface. Alternatively, install the bundle into an existing web profile and configure that profile's patch. A base-only profile created by plugin add has tools but no interactive surface. The profile selects the host model; this plugin does not change it. Normal Harness model use may incur the host Provider's fees.

Set allowWrites: true only when you intend to let this profile create tasks and publish explicit stage deliveries. This deployment authorization exposes the write tool. Harness policies can still deny calls; this plugin does not grant or bypass them. Core source_ids, revision checks, package validation and scoped paths remain enforced. This is trusted local process integration, not an OS sandbox.

Tools and requests

Both tools accept operation, data_json (a JSON-encoded object), and optional dry_run. They return the unchanged successful Python JSON object. Python and transport failures become Harness tool errors, not successful values. describe lists all core operations; the plugin only exposes the allowlists below.

Tool	Operations
research_atlas_read	describe, find_projects, find_tasks, resume_task, get_delivery_status, search_evidence, read_evidence, list_proposals
research_atlas_write (opt-in)	create_project, create_task, begin_run, submit_delivery, retry_archive, prepare_ingestion, index_revision

Examples:

{"operation":"find_tasks","data_json":"{\"query\":\"counterexample\"}"}
{"operation":"resume_task","data_json":"{\"task_id\":\"task_...\"}"}
{"operation":"submit_delivery","data_json":"{\"package\":\"package_...\"}","dry_run":true}

Read the research-continuity skill before creating delivery packages. Call resume_task, then begin_run with the exact current revision and honest executor identity. A stage package in the configured inbox contains artifacts, a hash-checked manifest and READY written last. Prepare it using trusted local file tools or manual exchange, then explicitly submit it. The adapter never invents evidence or writes a package on behalf of the caller. Preserve project/task IDs when switching agents or models.

Only submit_delivery and prepare_ingestion support mutation dry-run. Publishing locally does not imply indexing completed; query delivery status. index_revision is an explicit SQLite/DuckDB operation. No AI, vector or Neo4j invocation is added by this plugin. AI permission/start/generation and unrestricted strategy execution are intentionally outside this adapter's operation allowlist. Existing project AI capabilities still use Action AI Controller. External Harness inference is owned/audited by Harness, not automatically by that controller.

Temporary request files are created under the checkout's runtime/dsh-request-* and removed after child exit; research outputs follow the configured core paths. Source text returned to Harness may enter its model context and session logs: authorize sources accordingly. Python inherits a minimal OS environment without Provider keys or application tokens. Timeouts, cancellation or child failure during writes can leave an unknown outcome; inspect the ledger/delivery status before retrying. The adapter never retries automatically. Trusted local concurrent filesystem modification is outside its isolation guarantee.

Verify and roll back

The repository test uses the real installed Cordis and ToolRuntime, plus the real Python CLI, with synthetic data only:

$env:ATLAS_TEST_ROOT = (Get-Location).Path
$env:ATLAS_TEST_PYTHON = (Resolve-Path .venv/Scripts/python.exe).Path
node plugins/dsh-research-atlas/test_plugin_v0001.js

Standalone Node tests require the declared peers and @deepseek-ai/dsh-system-prompt from the same tested host to be resolvable. Tests do not install them. Results remain under test-output/dsh-plugin-*. This tests plugin execution and synthetic executor transitions; it does not claim fresh live model inference. See the acceptance report.

Disable the row or remove the bundle:

dsh plugin --profile research-atlas remove dsh-research-atlas

Unload removes registered tools and terminates/waits for owned child processes. Removal does not erase research data. Code rollback uses the prior Git commit/release; reconcile interrupted writes separately. This external plugin is registered in Cordis, not by running the local Action development-registration writers. No action.db or original baigong engineering data is changed during installation.

Provenance and community

Apache-2.0, Ruixin Yang / David7583. Existing baigong provenance remains in the parent repository; see its NOTICE, CITATION.cff and provenance/ directory.

Primary interface references: tool authoring, packaging, contribution policy.

DeepSeek's Community plugins link leads to the GitHub dsh-plugin topic. Topic association provides community discovery, not first-party endorsement or incorporation into the core repository. No npm registry publication is claimed by a GitHub release.