dsh-import-agents: bring pi / opencode / codex / claude-code history & agents into dsh

[Chang-Tong]

Sharing a plugin I built: dsh-import-agents — import sessions, chat history, and agents from pi (pi-coding-agent), opencode, codex, and claude-code into DeepSeek Harness (dsh).

Why

I work across several coding agents. dsh is where I want my history to live, but past conversations were scattered across pi / opencode / codex / claude-code JSONL files and an opencode SQLite DB. This plugin pulls them all in:

• Imported sessions become real dsh sessions (written in the exact on-disk format the dsh JSONL persistence backend uses, checksummed zstd frames) — they appear in the session list, are fully browsable, and can be resumed: the model continues with the complete history (text, reasoning, tool calls) as context.
• pi agents / mode prompts and opencode agents become discoverable dsh skills under $DSH_AGENTS_HOME/skills, with metadata.source / metadata.kind in the frontmatter.
• Imported sessions attach to a workspace matching their original cwd instead of piling up ungrouped.

In the GUI

• Sync button — a one-click button in the composer tool row runs the full import (/import-all) and shows the result inline.
• Migration prompt — when a new top-level session starts, it asks once (per project) whether to migrate pending sessions/agents; decisions are remembered, so it never nags twice.
• Slash commands — /import-pi, /import-opencode, /import-agents, /import-all, /attach-workspaces, with options like --limit N, --project <substr>, --since, --no-tools.

Install

cd ~/.dsh/profiles/web
pnpm add dsh-import-agents
# append to cordis.patch.yml:
#   - insert:
#       - id: import-pi-opencode
#         name: dsh-import-agents

Restart dsh web, refresh the page — the Sync button appears in the composer tool row. That's it.

Links

• Repo: https://github.com/Chang-Tong/dsh-import-agents (MIT, zero runtime deps — only Node built-ins + dsh platform modules)
• npm: dsh-import-agents v0.2.1
• 中文文档: https://github.com/Chang-Tong/dsh-import-agents/blob/main/README.zh.md

Verified against the real dsh backend (list / load / prepare round-trip), idempotent re-imports with stable ids (pi-<uuid>, oc-<id>, codex-<id>, claude-<id>).

Feedback welcome — especially if you also hop between multiple coding agents 😄

[Chang-Tong]

If you think this should be part of dsh itself, join the proposal discussion: #1359 (asking maintainers whether to make it a first-party plugin / built-in). Feedback on scope, storage contract, or sources most welcome there.

[Chang-Tong]

Update — v0.2.4: one-command install via the official dsh plugin add

The plugin now declares a dsh.bundle, so it installs and auto-activates through the official installer — no manual cordis.patch.yml editing:

dsh plugin --profile web add dsh-import-agents

Restart dsh web and refresh the page. Highlights since the original post:

• v0.2.4 — dsh plugin add one-command install; README rewritten (Apache-style, EN/中文 with a one-click language switcher); clean Docker demo screenshots
• v0.2.3 — English-ized plugin output (locale-aware Sync button), trajectory screenshot
• Covers pi / opencode / codex / claude-code: resumable sessions, restored tool-call trajectories, agents → dsh skills, idempotent re-imports

Repo: https://github.com/Chang-Tong/dsh-import-agents · npm: dsh-import-agents@0.2.4

[Chang-Tong]

Update — fixed in v0.2.7 🎉

Two fixes shipped in 0.2.7 (now live on npm):

• The Sync button now calls commands.execute with the required three business arguments — the previously missing images argument made dsh reject the call (expected 3 business argument(s) plus an optional AbortSignal, got 2), so the button always failed.
• Re-runs are now safe: the backend registers a session in memory on create but only materializes it through a ~200 ms write-batching window, so a quick re-run (or a migration prompt racing the Sync button) could abort /import-all with session "…" already exists in this backend. Those rejections are now counted as skips, and duplicate candidate ids (e.g. one codex rollout file per turn) are imported once.

Update with: dsh plugin --profile web add dsh-import-agents (or pnpm add dsh-import-agents).

[vshulcz]

Writing the imported sessions in dsh's own on-disk format, checksummed frames and all, is the part that makes this genuinely useful rather than a viewer — resuming an imported session with its full history as context is something an index cannot offer.

Worth naming the fork in the road for anyone choosing, because we took the other branch and the trade is real.

Importing converts: the history then exists twice, the copy is a snapshot as of the import, and fidelity depends on the mapping from each source format into dsh's. Re-running it later has to reconcile what changed on both sides. In exchange you get first-class sessions — resume, browse, everything dsh already does.

Indexing in place does not convert: the source files stay the only copy, sessions that grow after the fact are re-read incrementally, and nothing has to be reconciled. In exchange there is no resumable dsh session at the end — you get search and recall over what happened, not a conversation to continue.

Which one is right depends on whether the goal is continue that session or answer from what that session established. Both are legitimate; they are not substitutes.

One thing you will hit either way, from maintaining parsers for twenty of these formats: they drift without notice. opencode moved its SQLite schema, Cursor keeps CLI transcripts and IDE chats in different stores, Codex changed its session directory layout. We ended up with synthetic fixtures per format and a conformance test that fails when a real file stops matching the fixture, which is the only reason drift gets noticed at parse time rather than as a silent gap in someone's history months later.

Ours is https://github.com/vshulcz/deja-vu — the dsh side is dsh plugin add dsh-deja. Happy to compare notes on any format's edge cases; the parser registry with paths and record shapes per harness is public.