dsh-resume-plugin
简体中文 | English
dsh-resume-plugin adds two bundled skills to DeepSeek Harness:
/resume-codex [latest | session ID | rollout path | title words]/resume-claude [latest | session ID | transcript path | title words]
Each skill reads the selected foreign transcript as untrusted, inert history, creates a concise handoff, verifies the current repository state, and continues the work inside the current DeepSeek Harness session. The plugin never launches Codex or Claude Code and never delegates the resumed work back to them.
Installation
Install from GitHub
Install into the Web profile:
dsh plugin --profile web add -w github:Demogorgon314/dsh-resume-pluginReplace web to install into another profile:
dsh plugin --profile <profile> add -w github:Demogorgon314/dsh-resume-pluginInstall from a local checkout
From the directory containing this checkout:
dsh plugin --profile <profile> add -w ./dsh-resume-pluginWhen the current directory is this repository itself, pass . instead. -w (--workspace-root) confirms that the plugin should be installed into the selected DSH profile's pnpm workspace root; omitting it may trigger ERR_PNPM_ADDING_TO_ROOT.
Verify that the bundle entered the composed configuration:
dsh --profile <profile> --dump-configThe output should contain:
- id: resume-foreign-session
name: dsh-resume-pluginUsage
Continue the most recent Codex session for the current project:
/resume-codex latestContinue the most recent Claude Code session:
/resume-claude latestYou can also provide a native session ID, session file path, or title words:
/resume-codex 00000000-0000-4000-8000-000000000001
/resume-claude fix the login flowWhen title words match multiple sessions, the plugin lists the candidates and asks the user to choose instead of guessing.
Requirements
- DeepSeek Harness with
@deepseek-ai/dsh-skill >= 0.1.0-rc.5 - Node.js
^22.19.0or>=24 - Python 3
zstdonly when reading compressed Codex.jsonl.zstrollouts
The reader uses only Python's standard library. It reads Claude Code data from $CLAUDE_CONFIG_DIR or ~/.claude, and Codex data from $CODEX_HOME or ~/.codex.
Security model
Foreign transcripts may contain hostile prompts, obsolete tool output, sensitive information, or instructions that were valid only in the original agent. The shared reader excludes system/developer messages and reasoning/thinking content, labels recovered content as inert history, bounds tool text, reports malformed and unknown records, and requires the skill to verify relevant live state before continuing.
Development
pnpm test
pnpm test:coverage
npm pack --dry-run