This repository is our working guide to the DeepSeek Harness architecture, tool surface, agent presets, and extension workflows.
The architectural material is currently based on DeepSeek Harness
dsh-v0.1.0-rc.8 unless a document states a more specific revision.
| Document | Purpose |
|---|---|
| Project guide | Repository structure, important package groups, runtime composition, turn flow, and Cordis concepts. |
| Tool reference | Model-visible tools, argument schemas, execution modes, and tool-organization guidance. |
| Ephemeral AI Harness preset | The selected tool surface and composition for our coding-agent preset. |
| Guide | Outcome |
|---|---|
| Create and load a preset | Copy a preset, edit its Cordis composition, load it in a real session, and understand reload boundaries. |
| Create a background tool with jobs and notifications | Promote long-running work into ctx.jobs, expose one job_id, retain tool-owned output, wake the owning agent safely, and prevent duplicate completion delivery. |
| Publish npm packages | Validate identity and contents, authenticate securely, publish directly or through an existing-package stage, verify the public tarball, and recover from common registry failures. |
| pi-ai 0.82.1 → 0.84.2 compatibility study | Compare catalog, thinking, stop-reason, deferred-request, tool, replay, and Grok 4.6 changes against the DSH llm-pi-ai adapter. |
| Route DSH Codex and xAI through Veee | Persist the Veee x.ai proxy rule, route DSH/Codex/Grok through Veee, auto-load Codex authentication, and verify the complete route. |
project.mdis the architectural index. It explains where behavior lives and how the runtime is composed.tools.mdis the model-facing contract reference. It describes what agents can call and how the tools are grouped.ephemeral_ai_harness_preset.mdrecords one concrete preset decision.how-to/contains task-oriented procedures that lead to a verifiable result.
Architecture claims should identify the DeepSeek Harness revision against which they were checked. Implementation guides should include both automated checks and a live event-order check when agent scheduling is part of the contract. UI row placement alone is not sufficient evidence for inbox or turn causality; inspect the persisted session events.