#217 Make PR reviewer provider-switchable (OpenRouter/Qwen default)

[hubertgajewski]

Summary

Closes #217. Replaces the hardcoded Claude Sonnet reviewer with a provider-switchable setup. Defaults to Qwen3.6 Plus via OpenRouter (Anthropic-compatible /v1/messages endpoint); flip AI_REVIEW_PROVIDER=anthropic in repo Variables to revert.

• New repo variables: AI_REVIEW_PROVIDER (openrouter default | anthropic), AI_REVIEW_MODEL (optional model override per provider).
• Renames the enable-gate CLAUDE_REVIEW → AI_REVIEW so the name is vendor-neutral.
• Resolver step in claude-code-review.yml validates that the required secret is non-empty before the action runs — prevents the Anthropic key from being sent to openrouter.ai if OPENROUTER_API_KEY is missing.
• New secret requirement: OPENROUTER_API_KEY (documented in .env.example and README.md).

Changes

• .vars.example — rename CLAUDE_REVIEW → AI_REVIEW; add AI_REVIEW_PROVIDER, AI_REVIEW_MODEL.
• .env.example — add OPENROUTER_API_KEY placeholder.
• .github/workflows/claude-code-review.yml — resolver step branches on provider, sets ANTHROPIC_BASE_URL, picks the right secret, and passes the right --model.
• README.md — rename gate in catalog + table, add rows for the two new variables, document OPENROUTER_API_KEY.

Rollout prerequisites (done before merge)

• OPENROUTER_API_KEY secret created
• AI_REVIEW=true repo variable created; CLAUDE_REVIEW deleted
• AI_REVIEW_PROVIDER=openrouter and AI_REVIEW_MODEL=qwen/qwen3.6-plus set

Test plan

Design superseded by #220. The qwen/qwen3.6-plus path failed pre-flight (Claude Code SDK client-side slug validation) on the first live run. Follow-up #220 / #221 / #222 replaced AI_REVIEW_PROVIDER + AI_REVIEW_MODEL with direct ANTHROPIC_BASE_URL + ANTHROPIC_DEFAULT_{SONNET,HAIKU}_MODEL pass-through. Most of this test plan is moot — the vars it references no longer exist.

• This PR itself triggers claude-code-review.yml with the new defaults — confirm the Resolve reviewer provider settings step outputs provider=openrouter, model=qwen/qwen3.6-plus, base_url=https://openrouter.ai/api/v1
• anthropics/claude-code-action@v1 step completes without a 401/404 (Failed pre-flight on run 24405785979; root cause addressed in [bug] PR reviewer fails with "model not found" when ANTHROPIC_BASE_URL points to OpenRouter #220/#220 Fix PR reviewer — ANTHROPIC_DEFAULT_*_MODEL + Bearer auth via direct vars #221/#220 Pass proxy key via anthropic_api_key to satisfy action pre-flight #222.)
• A formal gh pr review verdict is posted on this PR by the Qwen-backed reviewer (Not on this PR. First Qwen verdict on PR #145 Add quality-metrics MCP server #219 came after the [bug] PR reviewer fails with "model not found" when ANTHROPIC_BASE_URL points to OpenRouter #220/#220 Fix PR reviewer — ANTHROPIC_DEFAULT_*_MODEL + Bearer auth via direct vars #221/#220 Pass proxy key via anthropic_api_key to satisfy action pre-flight #222/#223 Sharpen PR reviewer prompt — imperative framing for autonomous agents #224/#223 Require binary reviewer verdict — --comment reserved for drafts only #225 chain.)
• Temporarily flip AI_REVIEW_PROVIDER=anthropic, push a trivial commit, confirm model=claude-sonnet-4-6, base_url= (empty), review still works (Var removed by [bug] PR reviewer fails with "model not found" when ANTHROPIC_BASE_URL points to OpenRouter #220.)
• Temporarily set AI_REVIEW_MODEL=qwen/qwen3-coder, confirm resolver honors the override (Var removed by [bug] PR reviewer fails with "model not found" when ANTHROPIC_BASE_URL points to OpenRouter #220.)
• Restore production settings (AI_REVIEW_PROVIDER=openrouter, AI_REVIEW_MODEL empty or qwen/qwen3.6-plus) (Vars removed by [bug] PR reviewer fails with "model not found" when ANTHROPIC_BASE_URL points to OpenRouter #220.)
• grep -rn CLAUDE_REVIEW (excluding .claude/worktrees/) returns no hits

Assumptions flagged for verification

1. anthropics/claude-code-action@v1 forwards the ANTHROPIC_BASE_URL env var into the Claude Code runtime.
2. OpenRouter's /v1/messages endpoint accepts qwen/qwen3.6-plus (Anthropic-compatible routing).
3. Qwen via OpenRouter supports Anthropic-style tool use (required for mcp__github_inline_comment__create_inline_comment and Bash(gh pr review:*)).

If any of these fail during the Test plan checks, the fallback is a custom Python script (scripts/pr-review.py) following the self-healing.py dispatch pattern.