DSH Safe Runtime

Production-oriented safe execution infrastructure for agent harnesses. DeepSeek Harness is the first adapter, not the protocol domain model.

Architecture

Agent / Subagent
      |
      v
Capability Broker       -- who may act?
      |
      v
Transactional Runtime   -- where do effects happen?
      |
      v
Acceptance Engine       -- did required checks pass?
      |
      v
Evidence / AVP Bridge   -- can the result be proved?
      |
   COMMIT / ROLLBACK

Non-negotiable boundaries

1. packages/protocol MUST NOT depend on @deepseek-ai/*.
2. Core semantics MUST NOT be inferred from the DeepSeek Harness adapter.
3. Harness-specific event payloads MUST remain inside packages/adapter-dsh.
4. v1 workspace transactions cover workspace filesystem effects only.
5. Tool-level policy MUST NOT be described as isolation of arbitrary in-process plugins.
6. Unknown or unsupported guarantees fail closed; they never silently degrade to PASS.

Maturity pipeline

DRAFT -> SPECIFIED -> SCHEMA_COMPLETE -> TCK_COMPLETE ->
REFERENCE_IMPLEMENTED -> ACCEPTANCE_AUDITED -> READY

Current milestone

M0 + M1 baseline:

• repository/governance foundation;
• runtime-independent normative protocol;
• JSON Schema Draft 2020-12 contracts;
• positive/negative fixtures;
• protocol TypeScript projection;
• architecture boundary checks.

See docs/roadmap.md.