@dsh-external/dsh-session-search

Cross-agent session search plugin for DeepSeek Harness — directly scan past conversations from dsh, Codex, Claude Code, pi, and OpenCode without creating a derived database.

中文

Capabilities

Search observes the current source artifacts on each call and uses a case-insensitive literal substring match. English, Chinese, punctuation, and short queries all follow the same rule. Relevance is intentionally basic: the message with the most occurrences represents each session, with recency as the tie-breaker.

Supported session sources

All sources are read-only: session files are never modified, and the plugin creates no database, index, or persistent cache. OpenCode's own database is opened read-only as a source artifact.

Install (marisa / dshx)

git clone https://github.com/dsh-external/dsh-session-search.git
dshx install dsh-session-search ./dsh-session-search

Or install directly from a git URL. The plugin is mounted into ~/.dsh/config.yaml and takes effect on the next dsh web/TUI start (hot with Web HMR).

Manual mount (without dshx)

# ~/.dsh/config.yaml
- insert:
    - id: dsh-session-search
      name: '/absolute/path/to/dsh-session-search/lib/index.js'
      config:
        sources: { dsh: true, codex: true, claude: true, pi: true, opencode: true }
        maxResults: 10
        readWindow: 10

Configuration

Development

./scripts/build.sh   # compile src → lib with the dsh checkout's tsc (lib/ committed)
node tests/smoke.mjs # smoke test: scan real sessions and search

peerDependencies are provided by the host dsh's node_modules (cordis, @deepseek-ai/dsh-tools, @deepseek-ai/dsh-system-prompt, …); the build script creates the symlinks.

Implementation notes

• Parsers are defensive: oversized, corrupt, or unreadable files are skipped without failing the whole scan.
• dsh session logs are concatenated-frame zstd streams (one independent frame per appended event batch): frame boundaries are located with a structural scan, then decoded one frame at a time; a torn final frame (interrupted write) is skipped on that call.
• sources is applied before discovery. For canonical plain JSONL sources, safe single-line queries use rg --fixed-strings --ignore-case only to select candidate files; escaped queries skip this optimization. Every candidate is parsed and matched again, so raw metadata never becomes a result.
• Search processes one parsed session at a time and retains only bounded Top-K hits. DSH concatenated Zstd frames are decompressed and consumed one frame at a time instead of materializing the complete decoded log.
• agent_session_read tries the source-native artifact id first and falls back to a source-scoped scan only when an external format does not expose a direct mapping. OpenCode selects the exact source row.
• Every call observes current source state. There is no index lifecycle or stale derived state; a broad DSH search must still decompress the complete DSH corpus.
• Matching is literal and case-insensitive. Results group by session, rank the strongest message by occurrence count, and use timestamps for stable tie-breaking.

License

BSD-3-Clause