feat(deepseek): add DeepSeek Harness transcript provider

[tommy0103]

Summary

Adds a deepseek provider adapter that indexes DeepSeek Harness session artifacts ($DSH_HOME/sessions, default ~/.dsh/sessions) into the shared Obelisk schema, alongside claude, codex, kimi, and pi.

What changed

• packages/core/src/providers/deepseek.ts — the provider adapter:
  • Line-incremental like claude: mtime:lines cursor skips already-indexed lines (countMode: 'delta'), with a full reparse ('total') on first index or when a crash repair truncated the artifact below the indexed line count.
  • Subagent child sessions (header origin: 'subagent' + parentSession) project into their parent session as sidechain messages plus a subagent row.
  • user/message events with a non-user source kind map to is_meta (injected system prompts / skill reminders).
  • tool/call <-> tool/result correlation via callId; thinking (reasoning) blocks become content_type: 'thinking' messages ordered before the text message so assembleSessionDetail folds them correctly.
  • Stable message uuids deepseek:<sessionId>:<line> (plus :0/:1 suffixes for thinking/text) so raw() returns the original artifact line.
  • Respects $DSH_HOME; falls back to ~/.dsh/sessions.
• packages/core/src/zstd.ts — ports the frame scanner from @deepseek-ai/dsh-session-persistence-jsonl (MIT, attribution header kept) and decodes complete frames via built-in node:zlib (zstd), skipping a torn crash tail. No new runtime dependencies.
• builtins.ts / package.json — provider registration and exports.
• Tests — tests/deepseek-parse.test.mjs (golden: plaintext / zstd frames / torn tail / incremental delta resume / truncation fallback / subagent / discover / raw), tests/deepseek-session-detail.test.mjs (ADR-0007 conformance: parse -> assembleSessionDetail -> persist round-trip), plus updated registry and schema-doc assertions.
• Docs — README / SKILL / skill-doc source lists now include deepseek.
• Test-harness fixes — CLI test helpers clear a harness-exported DSH_HOME so CLI tests resolve the deepseek root against the temp HOME (the deepseek provider prefers $DSH_HOME over ~/.dsh).

Design notes

• DSH logs are single-carrier and append-only (unlike Codex), so the adapter does not need full-reparse; incremental resume is the norm.
• No inactive visibility handling: DSH compaction summarizes history but never supersedes branches (no tree semantics), so all messages stay visible, consistent with claude/codex.
• The ADR-0008 invocation-nonce self-identification works for DSH: DeepSeek Harness checkpoints the tool/call record to disk before the tool body runs (session-checkpoint-policy), so a CLI invocation's nonce record is indexed before the resolver runs. Verified end-to-end (overview().current.session_id resolves to the invoking DSH session in a live environment).

Verification

• npm run typecheck — clean.
• npm test — 473/473 pass.
• End-to-end: built against a real ~/.dsh/sessions artifact (zstd), searched, and confirmed the live DSH session is recognized as the invoking session via the nonce path.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.