[Proposal]: Add explicit OpenAI-compatible and Responses text providers

Summary

Add an explicit custom text-provider contract while preserving AI Gateway as the default:

• --provider openai-compatible for Chat Completions
• --provider openai-responses for Responses streaming
• --base-url, --api-key-env, optional --models-url, and existing --model
• equivalent AI_CLI_* environment variables
• fixed, non-secret attribution headers from PAPERCLIP_AGENT_ID / PAPERCLIP_RUN_ID
• fail closed for image, video, speech, and transcription in custom modes
• redact configured keys, bearer auth, and URL userinfo from errors

The existing package metadata declares Apache-2.0, but the repository and npm tarball omit the license text. The proposal also adds the canonical Apache-2.0 LICENSE to the repository and packed CLI.

Motivation

Self-hosted routers and local model gateways need a supported text path that cannot silently fall back to Vercel AI Gateway. API-key environment indirection keeps keys out of command history, and custom model listing should use the selected endpoint rather than contacting Gateway.

Implementation notes

A reviewed reference implementation is available in valkyriweb/ai-cli PR 1:
https://github.com/valkyriweb/ai-cli/pull/1

Exact reviewed head: 7803a8f7a53fa37d113adb0b469c76c988166165

The provider selection is centralized in src/lib/provider.ts; model discovery dispatches by provider; Responses mode uses streaming because some compatible routers require stream: true. Gateway code paths and defaults are unchanged.

The bundled ai-cli skill is extended rather than adding a competing skill, with a focused Agent Skill authoring reference.

Testing notes

• deterministic fetch tests assert exact /models, /chat/completions, and /responses URLs
• assert allowlisted headers and no ai-gateway.vercel.sh request in custom modes
• unsupported modality and redaction tests
• full typecheck, format, lint, tests, and build
• real synthetic Responses smoke through a self-hosted router using a cheap text model
• independent exact-head review and TruffleHog scan are clean

Risks or tradeoffs

Custom modes intentionally support text only. OpenAI-compatible model metadata is less rich than Gateway metadata, so pricing/provider endpoint details are unavailable. Responses mode streams internally and collects the final text to preserve current CLI output semantics.

Co-contributor credit

Yes, credit @valkyriweb.

Checklist

• I searched existing issues for duplicates.
• I understand pull requests are limited to collaborators and maintainers will handle implementation.

Start with src/lib/provider.ts and the reviewed reference implementation at commit 7803a8f7a53fa37d113adb0b469c76c988166165. Trace provider selection, model discovery, custom text requests, headers, redaction, and modality checks, then review the deterministic fetch tests and full typecheck, lint, test, and build commands. Done means both providers work without Gateway fallback, unsupported modalities fail closed, secrets are redacted, and the Apache-2.0 license is included in the repository and packed CLI.