Preface

When using the OpenAI-compatible gateway in DeepSeek Harness (DSH), the common practice is to manually write the llm-pi-ai provider lines in your profile and check subscription usage separately. Cline Pass operates via https://api.cline.bot/api/v1; the gateway may resolve to different upstreams behind the scenes, and responses may include anonymous tags. If you wish to use Cline Pass directly in the composer and see “who is actually doing the inference for this turn” along with plan usage in the UI, you need a dedicated route and accompanying UI, rather than manually checking logs every time.

Below is an introduction to the community plugin dsh-clinepass-deepseek (npm package name dsh-clinepass): maintained by BreakFree003, categorized as “Model Inference” in the SkillHub plugin directory. The project is provided under the MIT license. The README explicitly states it is a share-based snapshot, not intended for long-term maintenance, and does not guarantee follow-ups on upstream DSH interface drift. The GitHub repository currently has approximately 14 stars.

What is this

In short: Connect Cline Pass to DSH—it registers an OpenAI-compatible pi-ai route cline-pass. The API key is entered in Settings → Models → Cline Pass. The plugin does not rewrite any requests (does not touch body, headers, no retries, forwards SSE frames as-is); it only observes and displays.

Compared to 0.11.x, version 0.12.0 onwards emphasizes “pure pass-through + bypass observation”; the current package.json version is 0.12.7. Environment requirements: Node ≥ 20, DSH 0.1.7 and above. The host side uses only Node built-in modules, and the browser side loads React from the web shell’s static module table, resulting in no third-party dependencies and no build steps.

Core Features

1. Provider Capsule (Real Upstream Observation)

Displays at the bottom of the composer the actual upstream last read from the chat response for this route (e.g., Provider: deepseek). Data is written to a status file (default <DSH_HOME>/dsh-clinepass-status.json), exposed via the host route ?status=1; memory is maintained at two levels: session → model.

  • When the gateway returns the anonymous tag openai-compatible-private, the UI displays a warning style with a dedicated tooltip (relevant explanation for the contributor tier).
  • Displays Provider: -- when no conclusion has been observed yet.
  • Does not generate new conclusions when metadata usable for judgment cannot be read; it only counts as seen, and the capsule retains the previous result.

The server writes the conclusion before releasing the terminal frame; the capsule re-reads after the turn ends and refreshes approximately every 2 seconds while running.

2. Usage Card (Settings Page)

Displays three rolling windows on the Settings → Models → Cline Pass card: 5 hours / Weekly / Monthly. The gateway returns the percentage used (percentUsed), while the card displays the remaining amount (100 − percentUsed), along with a reset countdown, read time, and a refresh button.

The numbers are fetched by the host using your saved API key via GET https://api.cline.bot/api/v1/users/me/plan/usage-limits; the browser side cannot read credentials and only obtains the numbers via the same-origin path. The README notes that the magnitude of the accounting unit was calibrated around 2026-10-16 (for understanding only, not an official contract).

3. Model and Route Registration

The bundle or installer appends the llm-pi-ai cline-pass declaration to the profile (including model entries like DeepSeek V4.1 Flash, reasoningEfforts, and default reasoning: high). At runtime, provisionProfile() mainly aligns contextWindow / maxTokens; however, on DSH 0.1.7, this write-in is unreliable during cold starts, so do not rely solely on runtime registration.

See the README table for optional configuration items (e.g., upstream, provider, model, usage, statusFile, etc.). If not written, default values from withDefaults() are used.

Installation and Activation

Choose one of the following three methods, do not mix them; after installation, restart dsh (the profile is read only once at startup).

dsh plugin --profile web add github:BreakFree003/dsh-clinepass-deepseek
# Pin a tag: same as above, add #v0.12.7 at the end

The package declares dsh.bundle.patch, which appends the loader and llm-pi-ai lines to the profile manifest after installation; generally, no manual YAML writing is needed. Uninstall:

dsh plugin --profile web remove dsh-clinepass

Method B: Installer (Can work without pnpm)

Execute in the plugin directory:

node install.mjs                 # Installs to $DSH_HOME/profiles/web/plugins/dsh-clinepass by default
node install.mjs --dry-run       # Only prints the operations to be performed

Method C: Manual

Copy index.js, client.js, package.json, and cordis.patch.yml to <profile>/plugins/dsh-clinepass/, then write the loader line according to patch.example.yml in the repository.

Typical Usage

Fill in API Key

Open Settings → Models → Cline Pass, fill in the API key field, and save. Credentials are written to the DSH credential library, with the default key name CLINE_PASS_API_KEY (can be modified via the config item apiKeyEnv).

Select Model in Composer

After installation and restart, entries under Cline Pass will appear in the model list (e.g., cline-pass/deepseek-v4.1-flash). After sending a message, pay attention to whether the bottom Provider capsule and the usage bar on the settings page update.

When Maintaining llm-pi-ai Patches Yourself

If the profile already has an llm-pi-ai line, the installer may not overwrite it; you need to manually complete reasoningEfforts and reasoning: high in providers.cline-pass. Otherwise, the composer may lack thinking tiers or show an extra “Provider default” item. The README provides paste-able YAML snippets, including baseURL: https://api.cline.bot/api/v1 and the model list.

Advanced (Optional)

  • Mid-session system messages: Requires the profile model compat to work with patch-inhistory.mjs to modify the @deepseek-ai/dsh-llm-pi-ai gate; after upgrading node_modules, run --check to see if re-patching is needed.
  • Date snapshot model deepseek-v4.1-flash-1017 hard-pins DeepSeek upstream: Only write compat.vercelGatewayRouting.only: [deepseek] in the model line when pi-ai gates are met; on stable alias routes, this field is ignored by the gateway, resulting in different behavior—do not mix them.

Scenarios and Notes

Who it is for: Developers who are already using or intend to use a Cline Pass subscription, want to use pi-ai to unify model selection in the DSH web profile, and need to see the real upstream and plan usage in the interface.

Must Know:

  1. The plugin runs with the permissions of the current dsh process. Before installation, please read the GitHub source code and the MIT license to assess the trust boundary yourself.
  2. The project is a share-based snapshot and does not promise long-term maintenance; version pinning can be done via formats like #v0.12.7 at the end of the install command.
  3. It is synchronized with the pi-side sister project pi-clinepass-deepseek by release; the DSH plugin is one of the behavioral baselines (see RELEASING.md in the repo).
  4. SkillHub etc. are community directories and have no official affiliation with DeepSeek / High-Flyer; Cline gateway policies and quotas are subject to the Cline side.

Conclusion

dsh-clinepass-deepseek provides a pass-through cline-pass pi-ai route in DSH, integrating API key configuration, upstream Provider capsules, and the Cline Pass usage card into the same web settings and composer experience, while explicitly not rewriting the inference requests themselves.