DSH Local Share

Local, privacy-first Markdown and self-contained HTML sharing for DeepSeek Harness Sessions. DSH Local Share is an independent community plugin.

简体中文

Turn a complete DSH Session into a reviewable document without uploading the conversation. DSH Local Share adds a Share action to the Web Session header, builds a local preview, and lets you copy the source or download a file.

• Markdown or one script-free, self-contained HTML file
• Safe GFM rendering for assistant headings, lists, tables, and code
• Best-effort redaction enabled on every dialog opening
• Stable, flicker-free preview before copy or download
• Tool names, bounded arguments, and outcomes are opt-in
• Tool result bodies and model reasoning are never exported
• No cloud service, account, public-link backend, or outbound request
• Zero model-visible text and zero model-token overhead

Assistant Markdown is rendered into a calm, readable document while the human prompt remains literal. This privacy-safe product mockup uses fictional data:

Quick start

DSH is currently a developer preview. Install the exact release into the Web profile, then start DSH:

dsh plugin --profile web add github:ChuanTianML/dsh-local-share#v0.3.0
dsh --profile web

Open a non-empty Session and select Share in its header. The safe defaults are Markdown, tool calls excluded, and redaction enabled.

To remove the plugin:

dsh plugin --profile web remove dsh-local-share

Upgrading from DSH Share 0.1.0

Version 0.2.0 and later use a distinct package and plugin id to avoid colliding with an unrelated community plugin named dsh-share:

dsh plugin --profile web remove dsh-share
dsh plugin --profile web add github:ChuanTianML/dsh-local-share#v0.3.0

See the privacy flow

The dialog starts safe. If redaction is turned off, copy and download stay locked until the user gives a fresh acknowledgement; reopening the dialog restores the safe defaults.

The recording uses the real Web application and an isolated DSH profile. The example Session contains benign demonstration text only.

Install with a Coding Agent

Yes. Installation, configuration, and verification use inspectable CLI commands and one YAML profile patch. Paste this request into a Coding Agent that has terminal access to the machine where DSH is installed:

Install DSH Local Share v0.3.0 into my DeepSeek Harness Web profile.

1. Detect the active DSH_HOME and dsh version. Do not modify another profile.
2. Inspect the repository package.json lifecycle scripts before installation.
3. Install github:ChuanTianML/dsh-local-share#v0.3.0 into profile web.
4. Preserve unrelated entries in profiles/web/cordis.patch.yml. Configure the
   dsh-local-share entry with maxEvents 20000, maxOutputChars 2000000, and
   maxToolArgumentChars 12000.
5. Run dsh --profile web --dump-config and prove those values are active.
6. Start the Web profile, open a non-empty Session, and verify that Share opens
   with Markdown selected, tool calls excluded, and redaction enabled.
7. Report every command run and any file changed. Never upload an exported
   Session or disable redaction without asking me first.

For an evaluation that must not touch an existing setup, tell the Agent to use a fresh temporary DSH_HOME for steps 1–6.

Host configuration

The defaults work for ordinary Sessions. For an explicit, auditable setup, add the complete config object to $DSH_HOME/profiles/web/cordis.patch.yml (~/.dsh/profiles/web/cordis.patch.yml when DSH_HOME is unset):

- id: dsh-local-share
  config:
    maxEvents: 20000
    maxOutputChars: 2000000
    maxToolArgumentChars: 12000

Harness profile patches replace the target config object rather than deeply merging individual fields, so writing all three limits is the clearest option. Confirm the final composition without starting the Web server:

dsh --profile web --dump-config

Session and output overflows fail visibly. Tool arguments alone may be truncated, with a warning in the preview. Invalid values fail when the plugin loads.

Privacy behavior

The default document contains only direct human prompts and visible assistant text in log order.

Redaction covers common credentials, authorization headers, secret-bearing environment assignments, email addresses, and absolute local paths. It is heuristic, not a guarantee. Always review the preview. If redaction is disabled, copy and download remain locked until a fresh risk acknowledgement is checked.

The preview runs in a sandboxed srcdoc iframe. Generated HTML has no scripts or external resources and carries a restrictive Content Security Policy. Visible assistant text is rendered as safe GFM; human prompts remain literal so the shared document preserves what the user actually entered.

Why this plugin instead of the Session log ZIP?

The official Session log export is the right lossless ZIP for diagnosis and migration. DSH Local Share produces a smaller, human-readable document for code review, issue reports, handoffs, and knowledge sharing. It intentionally omits data that a replay or forensic workflow would need.

Development

The verified Harness revision is 47f943859bef60e4160492346772ded9b24f765a. Keep the checkout isolated under .sandbox/harness:

corepack enable
pnpm install
git clone https://github.com/deepseek-ai/deepseek-harness.git .sandbox/harness
git -C .sandbox/harness checkout 47f943859bef60e4160492346772ded9b24f765a
pnpm --dir .sandbox/harness install --frozen-lockfile
pnpm --dir .sandbox/harness run build
pnpm run check

DSH_HARNESS_ROOT=/absolute/path/to/deepseek-harness selects another read-only development checkout. pnpm run check runs strict type checks, ESLint, unit and composition tests, Host/browser builds, coverage, and a package dry run. Built lib/ artifacts are committed because GitHub profile installs do not run a build step.

The complete product and security design is in docs/design.md.

Compatibility

Version 0.3.0 targets the DSH developer-preview API at the verified revision above. DSH does not yet promise stable external plugin compatibility; future Harness changes may require a new DSH Local Share release.

Security

See SECURITY.md. Never include a real credential or private Session in a public issue.

License

MIT