prime-agent-dsh

A context sidecar for Prime Agent. It turns the active Prime branch into a searchable, rebuildable DeepSeek Harness (DSH) projection, exposes bounded context tools in Python, and reports provider cache use. Prime remains the only agent loop and the only session authority.

The pinned DeepSeek Harness 0.1.6-alpha.2 dependencies are developer previews. Review the security and compatibility notes before deployment.

Why use it?

• Search and inspect a long Prime session without placing the whole transcript in the next request.
• Pin immutable views, create private artifacts, and explicitly admit selected evidence.
• Give native RLM descendants bounded inherited evidence or an explicit expiring grant.
• See provider-reported turn and session cache rates in a native Prime widget.
• Keep failures non-blocking: projection errors leave the Prime request unchanged.

Quick start

Requirements, source installation, verification, upgrades, removal, and troubleshooting are in Getting started.

prime-agent package install npm:prime-agent-dsh

Restart Prime Agent after installation. Then run /dsh and try dsh_context.current() in IPython. Source and project-local installation instructions are in the getting-started guide.

Architecture invariants

1. Prime owns inference. Prime selects models, streams responses, authenticates, retries, and controls approvals.
2. Prime owns execution. Tools, IPython, skills, MCP, native rlm.spawn, and goals stay in Prime's loop.
3. Prime owns content and lifecycle. Prime JSONL is the only full-content session authority. Prime alone branches, compacts, and deletes sessions.
4. DSH is derived and fail-open. Its per-session projection and indexes are bounded, content-addressed, and rebuildable. DSH does not compact Prime context.
5. Scopes stay isolated. Each root and native RLM child gets its own scope. Sharing uses a bounded capsule or explicit grant, never a live transcript or parent authority.
6. Admission is explicit. Reading DSH data does not silently add it to model context. Printed or returned ctx.inject(...) output becomes a normal Prime tool result.

The package does not ship a DSH AgentLoop, provider wrapper, ACP delegation path, custom model loop, or independent conversation store. See Single-window cache architecture.

Feature status

Python API at a glance

ctx = dsh_context.current()
ctx.entries(last=10)
ctx.messages(last=10)
ctx.search("authentication", limit=20)
ctx.metrics

snapshot = ctx.snapshot()
selection = ctx.search("migration decision", limit=8)
artifact = ctx.artifact(selection, label="Migration evidence")
print(ctx.inject(selection, label="Relevant migration decisions"))

For grants and API rules, see Getting started and the dsh-context skill reference.

Configuration

Documentation

• Getting started
• Single-window cache architecture
• Durable context query
• Context spill
• Shadow telemetry validation
• Contributing
• Support
• Roadmap
• Changelog
• Code of Conduct

Project status

0.2.3 is a developer preview. The Prime integration, storage formats, and Python API may change before a stable release. The test suite covers the documented core paths, but this package is not a security boundary against another process running as the same OS user.

The current runtime target is Prime Agent 0.9.5 or newer (@earendil-works/pi-coding-agent >=0.86.1). See the roadmap for direction rather than release promises.

Development and validation

npm run typecheck
npm run lint
npm test
npm run test:python
npm run package:smoke
npm run release:check

See CONTRIBUTING.md before proposing a change. Preserve the Prime-only authority model.

License

MIT. Bundled dependency notices are in THIRD_PARTY_NOTICES.md.