refactor: provider-name dispatch + fix agent finalization and rescue crash

[Iris-Ares]

Summary

• Pivot to first-party provider dispatch. The action now takes llm-provider (one of openai | anthropic | deepseek | google | groq | mistral | xai | openrouter | openai-compatible) and uses the matching @ai-sdk/* SDK. Each provider's own package handles structured outputs, tool calling, and auth correctly — no capability shims.
• Fix the step-budget miss. Added a prepareStep callback that restricts the final step to submit_report only and sets toolChoice: { type: 'tool', toolName: 'submit_report' }. The agent now finalizes reliably within budget.
• Fix the rescue crash. Dropped the @qraftr/json-repair first pass, the Lenient schema aliases, and normaliseAliases. Rescue is now one Output.object({ schema: VerifyReportSchema }) call — the provider's native structured-output path guarantees the exact Zod shape.
• Kept @ai-sdk/openai-compatible as the explicit escape hatch for self-hosted vLLM / Moonshot / Qwen, now with supportsStructuredOutputs: true set so it actually works.
• Fixed .gitignore so dist/index.js ships with the commit (the old dist/ ignore line contradicted the README and the comment right next to it).

Motivation

A prod run of Iris-Ares/pr-lens-github-actions@v0 failed with:

Warning: agent exhausted step budget without calling submit_report; attempting rescue from raw text
AI SDK Warning (byo.chat / ***): { type: "unsupported-setting", setting: "responseFormat",
  details: "JSON response format schema is only supported with structuredOutputs" }
Warning: PR-Lens failed: AI_NoObjectGeneratedError

Two compounding defects: the ToolLoopAgent never called submit_report within the 12-step budget, and the rescue crashed because the generic BYO provider was built without supportsStructuredOutputs — so Output.object({ schema }) silently sent a schema-less request.

Breaking change

Existing workflows must update:

 - uses: Iris-Ares/pr-lens-github-actions@v0
   with:
-    llm-base-url: https://api.openai.com/v1
+    llm-provider: openai
     llm-api-key:  ${{ secrets.LLM_API_KEY }}
-    llm-model:    gpt-4o
+    llm-model:    gpt-4.1

For self-hosted or non-first-party endpoints (Moonshot, local vLLM, Ollama):

with:
  llm-provider: openai-compatible
  llm-base-url: https://api.moonshot.cn/v1
  llm-api-key:  ${{ secrets.LLM_API_KEY }}
  llm-model:    kimi-k2.5

README has per-provider examples and a model-id table.

Test plan

• bun run typecheck — clean
• bun test — 47/47 pass
• bun run build — 376 modules, dist/index.js = 1.68 MB
• Bundle import smoke — loadConfig reaches the expected required-input error
• End-to-end live run against sindresorhus/ky#853 via Moonshot kimi-k2.5 (openai-compatible branch):
  • No responseFormat warning
  • 7 steps, submittedViaTool: true (well under the 12-step budget)
  • 15,388 tokens, valid 5-item checklist rendered
• CI self-test runs on this PR with an updated LLM_PROVIDER var (needs repo secret/var update before merging)
• Verify the new prepareStep force-finalization by running with STEP_BUDGET=2 (too small to finish) and confirming the single-call rescue completes cleanly

Files touched

• action.yml, src/config.ts — new llm-provider input + conditional base-URL validation
• src/ai/provider.ts — provider dispatcher (9 branches)
• src/ai/agent.ts — prepareStep + simplified rescue + V2→V3 type bump
• src/run.ts — pass llmProvider into buildModel
• package.json, bun.lock — add 8 new provider packages, drop @qraftr/json-repair
• README.md, .env.example, scripts/local-run.ts, .github/workflows/self-test.yml — docs + local tooling for new input shape
• .gitignore — un-ignore dist/
• dist/index.js — rebuilt bundle

🤖 Generated with Claude Code