🚚 dsh-claude-move

Migrate Claude Code, Codex, OpenCode and Hermes into DeepSeek Harness — copy sessions, memories, skills, instructions and slash commands as resumable DSH sessions, copy-only and approval-gated.

Keep your Claude Code history when you move: one install, resumable sessions, live sync with a running Claude Code, and a four-source migration wizard.

English · 简体中文 · Español · Português · हिन्दी

Compatibility

• Targets dsh 0.1.1-rc.2 (web profile); peer dependencies require >=0.1.0-rc.8 <0.2.0. Node ^22.19 || >=24.
• Last verified against a fresh tarball install: real scan, real batch import (idempotent re-import), workspace attach and persistence artifacts confirmed; macOS/Linux covered by the CI matrix.

Compatibility matrix (public seams only)

What you get

1. Auto-discovery — claude_scan locates the Claude data root ($CLAUDE_CONFIG_DIR, fallback ~/.claude) and indexes every project/session, memory, skill, global CLAUDE.md and settings.json, with incremental caching and parallel scanning (scanConcurrency).
2. Full-fidelity import — import_claude turns transcripts into balanced, resumable DSH sessions (turn/start → step/start → user/message → assistant/message → tool/call → tool/result → step/end → turn/end), repairs interrupted tool calls, and stream-imports transcripts larger than maxTranscriptBytes in chunks.
3. One claudecode workspace — every imported session lands in a dedicated workspace (default $DSH_HOME/claudecode); workspaceMode: 'per-project' restores one-workspace-per-project grouping.
4. Copy-only & incremental — nothing on either side is moved, rewritten, or deleted; re-running appends only the new turns (force: true saves an extra full copy under a new id).
5. Personal context, always fresh — memories injected as a live prompt section, Claude skills registered as real DSH skills (global + project-level), global + project CLAUDE.md injected early.
6. Four-source migration wizard — /move plus move_detect / move_preview / move_run migrate Claude Code, Codex, OpenCode and Hermes, approval-gated and idempotent (move.json).
7. Web panel & commands — /claude-import-all, /resume-claude, /claude-move-reset, and a floating migration panel.

Four-source migration wizard

/move              # one-shot wizard: detect → preview → execute → report (all four sources)
move_detect        # scan Claude Code / Codex / OpenCode / Hermes
move_preview       # per-item plan: new | unchanged | changed | conflict (with diff) | unsupported
move_run           # execute behind the approval gate; conflict resolution:
                   #   skip | overwrite | rename | merge  (default skip — never guesses)

• Sources — Claude Code (~/.claude), Codex (~/.codex), OpenCode (data + config roots), Hermes (skills/memory roots); each source has its own parser + mapper.
• Mapping — memories/instructions → append-only managed sections in the DSH global AGENTS.md (one marked section per item); skills → real DSH skills (SKILL.md bundles copied verbatim, other formats converted); slash commands → registered DSH commands (rebuilt from move.json after a restart); sessions → resumable DSH sessions (the same importers as phase 1).
• Idempotent — every applied plan is recorded in $DSH_HOME/claude-move/move.json (digest / targetDigest / appliedAt); re-runs skip unchanged items and force re-applies them.
• Approval-gated — a run that would write anything asks ctx.approval first; anything but allowed-once means zero writes.

Quick start

# 1. install the bundle into your profile
dsh plugin --profile web add "github:PerryLink/dsh-claude-move#master"

# or from npm (published releases)
dsh plugin --profile web add dsh-claude-move

# 2. restart and verify the row
dsh --profile web --dump-config | grep -A4 'id: claude-move'

Then, in any DSH session, run one command:

/claude-import-all      # scan → copy every Claude session → report

No DSH restart is needed after importing — refresh the open Web page once and click any imported session to continue.

Install & uninstall

• git channel (latest master): dsh plugin --profile web add "github:PerryLink/dsh-claude-move#master" — pure ESM, no prepare or allowBuilds step.
• npm channel (published releases): dsh plugin --profile web add dsh-claude-move.
• tarball channel: npm pack in this repo, then dsh plugin --profile web add ./dsh-claude-move-<version>.tgz.
• uninstall: remove the claude-move row from the profile's bundles and restart dsh. Imported sessions stay in DSH's data directory; the plugin only writes its cache ($DSH_HOME/claude-move/) and the claudecode workspace folder, and never touches Claude source data.

What gets migrated

~/.claude (read-only)
 ├─ projects/*/*.jsonl  ──→  resumable DSH sessions, grouped in one "claudecode" workspace (default)
 ├─ projects/*/memory/  ──→  live system-prompt memory section (re-read per request)
 ├─ skills/**           ──→  real DSH skills
 └─ CLAUDE.md + settings ──→  early prompt section + config suggestions (never auto-applied)

Usage

Call the tools in any session with the plugin mounted:

claude_scan                          # full scan (incremental cache)
claude_scan { path: "~/.claude/projects/<slug>" }   # partial scan
claude_scan { refresh: true }        # skip cache, rescan everything
claude_scan { projectsLimit: 10, sessionsLimit: 5, fields: "brief" }  # trim output

import_claude { path: "~/.claude/projects/<slug>/<sessionId>.jsonl" }  # one session
import_claude { path: "~/.claude/projects" }        # directory (recursive)
import_claude { path: "all" }                       # everything
# Re-run any time: unchanged files are skipped, grown transcripts append only the new turns.
# Files over maxTranscriptBytes are stream-imported in chunks (no memory ceiling).
import_claude { path: "...", force: true }          # fresh full copy (previous copy kept)

Commands (user-triggered, no model turn):

/claude-import-all                # one-shot: scan → import everything → report → inject into the current session
/resume-claude latest             # continue the most recent Claude session
/resume-claude <sessionId>        # by source session id or import-<src> id
/resume-claude <keyword>          # match titles; multiple matches are listed, never guessed
/claude-move-reset                # reset the plugin cache (bookmarks + import map); imported sessions are kept

Web panel: a floating migration panel with the project/session tree, status badges (not imported / imported / imported-with-new-turns / source missing / directory missing / git dirty), keyword filter, paged rendering, per-session "Import & continue" + "Open session" + "Refresh session list", batch import with a live progress bar and cancel, and a cache-reset button. Texts follow the browser language (zh/en). Served through the plugin's own /api/claude-move/* JSON routes on the public ctx.webServer seam.

After importing

You do not need to restart DSH. Imports land durably through the public sessionPersistence service the moment they complete:

• The server-side lists (session.list / workspace.list RPCs, the CLI, any new page load) show the imported sessions under the claudecode workspace immediately.
• The panel refreshes the already-open page's session list itself and offers an Open session button per imported session.
• Imported sessions can be opened, read, and resumed right away — /resume-claude, or click the session in the list. Re-running the import at any time syncs only the new turns into the same sessions.

Configuration

All optional, overridable in cordis.yml.

Tools & surfaces

Permissions & data

• Permissions: the workshop manifest declares filesystem:read and filesystem:write.
• Reads ~/.claude (transcripts, memories, skills, CLAUDE.md, settings.json) — strictly read-only — and the project directories it imports into.
• Writes DSH session logs via the public sessionPersistence service (create + append only, never delete/rewrite/archive), workspace-registry records, its cache under $DSH_HOME/claude-move/, and the claudecode workspace folder.
• Never modifies Claude source files, touches other applications' data, or accesses the network. No credentials are read or transmitted.

Security boundaries

• Source files are read-only; DSH logs are append-only (create + append only).
• External transcripts are untrusted input — nothing in them is executed; system/developer/thinking content never enters the resume handoff.
• Public services only — sessionPersistence / workspaceRegistry / tools / commands / systemPrompt / skills / webServer; no engine or UI changes.
• Secrets reported by position only (file:line:kind); permission/permission-mode/queue-operation records are counted, not imported.
• Wizard writes are approval-gated — anything but allowed-once means zero writes.

Known limitations

• Titles come from custom-title/ai-title/first prompt; Claude summary records are reported but not mapped to DSH compaction nodes (synthesizing a valid compaction transaction would fabricate its seq range and checkpoint message).
• thinking blocks are kept as reasoning content but never enter the resume handoff.
• Interrupted tool calls are repaired with a synthetic error result (never dropped), reported as repaired.synthesized.
• Permission-class records are counted, not imported; DSH permission-preset suggestions are generated in reports.
• On hosts without a streaming fs.streamText surface, transcripts larger than maxTranscriptBytes fail loudly instead of partial import.
• In workspaceMode: 'per-project', sessions whose source directory was deleted still import, but workspace attach fails (left ungrouped; workspace.attached: false plus a reason). The default claudecode workspace does not depend on the source directory.
• If a transcript was truncated or reset in place (fewer turns than the recorded import), re-import skips it and reports sourceShrunk; use force: true for a fresh full copy.
• The Web panel is a zero-build floating panel driven by the plugin's own JSON routes; it does not use the shell's internal UI slot system.

Model Experience

• The model-facing surface is the two tools' descriptions/schemas and their outputs: claude_scan returns the structured index, import_claude returns per-file summaries with positions of warnings. Tool results are themselves logged tool/result events, so everything is reconstructable.
• No hidden model-facing text; memory/CLAUDE.md sections are registered on ctx.systemPrompt (prompt assembly, rebuildable from the session log).

Troubleshooting

• Row not effective: dsh --profile <p> --dump-config should print # == dsh-claude-move; re-run dsh plugin --profile <p> add ....
• Web boots but hangs silently: new profiles initialized by dsh plugin add contain only dsh-base — add @deepseek-ai/dsh-web-app to dsh.profile.bundles. Installing into the existing web profile needs nothing.
• Panel routes 404: they are served only when enableWebPanel: true and a web server is composed; check the boot log for FAILED fibers.
• Import fails with "transcript 过大": raise maxTranscriptBytes or import that file individually.
• Import succeeded but the sidebar shows no new session: the page was already open — click the panel's refresh button (or reload the page) once. No DSH restart is ever needed.
• Logs: boot failures print to the dsh console; the plugin logs [claude-move]-prefixed errors for workspace/import-map issues.

Attribution (open-source components)

This project is licensed under the Apache License 2.0; the following MIT-licensed components retain their own licenses (full text in THIRD_PARTY_NOTICES.md):

• Conversion core vendored from Nwflower/dsh-chat-import (MIT).
• Discovery conventions & safety model from Demogorgon314/dsh-resume-plugin (MIT).
• Memory/skills injection & frontmatter parsing patterns from YYTbit/dsh-plugin-claude-bridge (MIT).

Development

npm install   # peer deps: @deepseek-ai/dsh-tools@>=0.1.0-rc.8, @deepseek-ai/cordis, schemastery
npm test      # node --test test/*.test.mjs

CI runs the full suite on Node 22 across Linux/macOS/Windows via GitHub Actions (test.yml).

Topics

deepseek-harness, dsh-plugin, claude-code, migration, session-import, resume

Contributors

• @PerryLink — creator and maintainer: the import pipeline, the four-source migration wizard, the Web panel, docs, CI/CD and releases.
• @OLDnana1 — root-cause analysis of the interrupted tool-call corruption that made imported sessions permanently return HTTP 400 on resume.
• @GooodWei — identified README.md (and any description-less .md) being misregistered as a skill, which broke DSH's skill load.

PerryLink DSH Plugin Family

This project is one of the DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:

License

Apache License 2.0 © 2026 dsh-claude-move contributors