Add endpoint commands: responses, chat_completions, moderations, embeddings

[amaan-ai20]

Why

The Claude Code and OpenClaw plugins shell out to per-task commands (chat -m …, classify_iab, redact_pii, …). Each of those pins a model or checks --model against a list, so every model addition, rename, or removal needs a CLI release, and customers on an older or newer CLI than the plugin expects get broken skills.

This PR adds one stable command per API endpoint. They send whatever model the caller names and keep no model list. The plugins will pick the endpoint per skill and call these, so they stop depending on CLI model updates. login, status, and cost_savings stay as they are.

What's added

Command	Endpoint	Options	Prints
responses [text]	/v1/responses	-m (required), -i, --metadata, --body, --raw	output text (pretty JSON when it parses)
chat_completions [text] (alias chat-completions)	/v1/chat/completions	-m (required), -i → system message, --metadata, --body, --raw	choices[0].message.content
moderations [text]	/v1/moderations	-m (required), --body	full response
embeddings [text]	/v1/embeddings	-m (required), --body	full response

• Input comes from the positional argument, or from stdin when there is none, so large prompts stay off the command line and plugins can use a heredoc directly. Trailing newlines from stdin are dropped.
• --body <json> merges extra top-level fields (e.g. max_output_tokens). Fields set by the other options take precedence. This lets new API parameters be used without a CLI release too.
• Savings are recorded under the requested model, like every other command, so cost_savings and the periodic note keep working.
• Shared plumbing lives in src/lib/request.ts. Existing commands are untouched.
• Docs: README "Endpoints" section, DOCUMENTATION §4.16–4.19 plus a §5 note, and ADDING_COMMANDS. The model-sync skill now says never to add a model list to these commands or remove them.
• Version: 3.7.2 → 3.8.0. Merging publishes it.

Verification

• npm run lint, npm run build, and npm test pass (56 tests; 16 new in tests/endpointCommands.test.ts, with fetch mocked).
• Live smoke tests against api.zerogpu.ai from the built CLI, with a throwaway HOME:
  • responses … -m gliner-multi-pii-v1 --metadata '{"usecase":"redact","mask":"label"}': output identical to redact_pii
  • responses -m zlm-v1-iab-classify-edge with text on stdin via heredoc: works
  • chat_completions -m gliner2-base-v1 --metadata '{"usecase":"ner",…}': entities returned
  • responses -m gpt-oss-120b --body '{"max_output_tokens":64}' --raw: completed
  • moderations -m zlm-v1-moderation-edge: flagged; embeddings -m bge-small-en-v1.5: 384 dims
  • unknown model → Request failed with status 404. + body, exit 1; missing -m, empty stdin, invalid --body → clear error, exit 1
  • cost_savings --json showed every successful call recorded

Note: chat-completions "Say hi in three words." -m qwen3-30b-a3b-fp8 exited 1 with Response did not contain any message content. The model used its whole 2,000-token completion budget on reasoning and returned content: null. zerogpu chat -m qwen3-30b-a3b-fp8 fails the same way on that prompt, so it is model behaviour, not this change. --body '{"max_tokens":…}' is available if the plugins need more room.

Next

Once this is published, the Claude Code and OpenClaw plugins switch their inference skills to these commands and require zerogpu-cli >= 3.8.0.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: a578f8c6-09d3-4092-b1cf-eee109893c62

---

_{Comment @coderabbitai help to get the list of available commands.}