Proposal: add a version-gated DeepSeek Harness transcript source to SkillOpt-Sleep #225

Motivation

DeepSeek Harness (DSH) now persists event-sourced sessions, but SkillOpt-Sleep has no DSH transcript source. The recently merged OpenCode source provides a useful precedent for adding a narrow, privacy-conscious harvester independently from an execution backend or plugin.

DSH is currently a developer preview (0.1.0-rc.5) and its session header is version 0, so this integration should be explicitly experimental and fail closed rather than claim forward compatibility.

Proposed MVP

• Add an explicit --source dsh and --dsh-session-root; do not add DSH to auto discovery yet.
• Read sessions locally and read-only, accepting only the tested DSH schema/version.
• Keep root-session visible user/assistant text, short tool names, working directory, timestamps, and immutable session feedback records.
• Exclude reasoning/chunk events, tool arguments/results, file payloads, request/provider/account metadata, child sessions, and SkillOpt-created replay sessions.
• Normalize accepted sessions into the existing SessionDigest contract and preserve the current redaction boundary.
• Add sanitized fixtures and tests for valid sessions, corruption, unknown versions, filtering, redaction, CLI/config routing, and documentation.

Design question

Would maintainers prefer the smallest direct importer for version-0 JSONL (with a documented policy for DSH's default zstd files), or a versioned sanitized export produced by a thin DSH plugin? Direct import is smaller and has no plugin prerequisite; an exporter is a more durable format boundary and can include DSH's separate message-feedback store.

Non-goals

• A DSH execution backend or tool-aware agentic replay
• A DSH UI/plugin bundle
• Automatic skill adoption
• Recovering message-level feedback from DSH's sidecar when using raw session logs
• Compatibility with unknown future DSH session versions

Acceptance criteria

• Harvesting is local, read-only, and requires no DSH model login or running daemon.
• Unknown or malformed schemas are skipped with a clear warning; no partial permissive parse.
• since, project, limit, child-session, and replay-session filters have tests.
• Sensitive event payloads never enter SessionDigest or fixtures.
• Existing harvest sources remain unchanged and the full test/documentation checks pass.