Baigong Research Atlas: persistent research continuity tools for Harness

[David7583]

Baigong Research Atlas: native research continuity tools

Sharing dsh-research-atlas, a community Cordis plugin for continuing a research project across agents and AI clients while keeping the same project/task IDs, evidence provenance, revision history and stage handoff.

• Repository and installation: https://github.com/David7583/baigong_research_atlas/tree/main/plugins/dsh-research-atlas
• Release tarball: https://github.com/David7583/baigong_research_atlas/releases/tag/dsh-plugin-v0.1.0
• Acceptance evidence: https://github.com/David7583/baigong_research_atlas/blob/main/docs/harness_plugin_acceptance_20260927.md

The plugin registers research_atlas_read and an opt-in research_atlas_write with ctx.tools.register. It forwards allowlisted operations to an existing local Python core: find/resume a task, begin a run, submit a hash-checked stage package, inspect delivery status, and explicitly index a published revision. It does not use private client conversation history as project identity.

Installation starts disabled; deployment paths are configured by the user. Read-only is the default, and writes require allowWrites: true. Harness tool policies still apply. Model SDKs, Provider keys and permission management are not exposed through these tools. This is trusted local integration, not an OS sandbox; authorized evidence can enter the host model context/logs.

Validated with Harness/dsh-tools 0.1.5-rc.3, Cordis 4.0.2, Node 24.18.0 on Windows. Checks include 27 existing Python tests, actual Cordis/ToolRuntime calls, full dsh CLI bundle loading, local and tarball installation, unload/removal and a four-stage synthetic executor roundtrip. No new live model inference is claimed by these plugin tests.

Apache-2.0, Ruixin Yang / David7583. Existing baigong attribution and source provenance are retained. The repository follows the dsh-plugin discovery convention; this is not a first-party package or a request to merge code into the core repository.

[UmeshCode1]

This is a much-needed primitive for agentic research workflows. Decoupling experiment identity and evidence provenance from ephemeral client chat histories is essential once an investigation spans multiple sessions or model providers.

A couple of architectural questions from building empirical benchmark pipelines:

1. Cross-Model Schema Drift & Handoffs:
   When switching between different frontier models (e.g., DeepSeek-V3, Claude, Gemini) across handoff stages, have you observed issues with agents mutating or dropping metadata keys in the stage packages? Does dsh-research-atlas enforce a strict JSON-schema validator before registering research_atlas_write?

2. Stage Package Checkpointing:
   For long-running evaluation steps (e.g., multi-epoch training or hyperparameter sweeps), does the hash-checked stage package support intermediate checkpoint delta commits, or does it require atomic whole-package submissions?

3. CLI / Headless Integration:
   Is there a lightweight Python or CLI headless entry point to query the Atlas status without launching the full dsh web interface?

Really solid design with the Cordis tool registration and default read-only safety boundary. Looking forward to testing dsh-research-atlas!

[David7583]

Thanks again for taking the time to look through the project and for the thoughtful questions. They map very closely to the reliability and integration boundaries I’m currently working through, so your feedback is genuinely useful.
One versioning note first: the public baigong / research_atlas repositories currently reflect an earlier backend baseline, roughly around v0019. My local development baseline has moved to v0021, and some of the capabilities below have only been validated there and are not published yet.

1. Schema drift and handoffs
   We have tested both model and DeepSeek Harness integration, but not systematic handoffs across DeepSeek, Claude and Gemini. We observed semantic drift—such as losing query scope or generating an unsuitable query—rather than confirmed stage-package metadata loss. The backend validates structured query contracts and source provenance, but I cannot yet confirm strict JSON Schema validation specifically before research_atlas_write. Feeding execution errors back to the agent for correction is also still incomplete.
2. Checkpointing
   The ingestion/vector backend supports incremental batch commits, durable checkpoints and idempotent recovery. We verified recovery after a forced process crash and resumed real vector computation after a CUDA failure without restarting the entire job. However, this does not establish delta-commit support for Atlas stage packages; that interface still needs verification.
3. Headless integration
   The local backend has a Python CLI and a localhost HTTP status endpoint that work without opening the web UI. A dedicated Atlas status command independent of dsh web is not yet verified.
   The current prototype demonstrates persistent data, provenance and recoverable execution. General cross-model handoffs and a complete agent execution loop remain work in progress.
   I’m currently validating this on a laptop while also in the middle of a job search, so the broader cross-model and long-running test matrix is still pending. I’d rather be precise about the current validation boundary than overstate support.
   Your questions are especially useful because they separate three issues I had partially treated together before: schema invariants across model handoffs, stage-level checkpoint semantics, and a clean headless integration boundary.
   If you do test dsh-research-atlas against an empirical benchmark pipeline, I’d be very interested in the failure modes you encounter. That kind of feedback would be extremely useful for deciding which of these boundaries should be hardened first.
   Research Atlas:
   https://github.com/David7583/baigong_research_atlas
   DeepSeek Harness plugin release:
   https://github.com/David7583/baigong_research_atlas/releases/tag/dsh-plugin-v0.1.0
   Baigong base repository:
   https://github.com/David7583/baigong
   Thanks again for the careful review and for pointing out these boundaries. I really appreciate the level of detail in your feedback.