The previous implementation used wrong flags (-p without value, --json,
--max-output-tokens) that don't exist in the real gemini CLI.
Correct invocation:
- Pass prompt as -p STRING value (not via stdin)
- Use --output-format json to get structured {response, stats} output
- Add --yolo to suppress interactive confirmation prompts
- Remove nonexistent --json and --max-output-tokens flags
- Parse `.response` field from JSON output, skipping MCP noise lines
- Extend timeout from 30s to 60s (agentic CLI is slower than raw API)
Smoke tested end-to-end: stdin HTML → summarize and --extract-json
both produce correct output via Gemini CLI.
- Add jsonschema crate for schema validation in extract_json
- On parse failure (invalid JSON): retry once with identical request
- On schema mismatch (valid JSON, wrong schema): fail immediately — no retry
- validate_schema() produces concise error with field path from instance_path()
- Add SequenceMockProvider to testing.rs for first-fail/second-success tests
- Fix env var test flakiness: mark env_model_override as ignored
- ProviderChain::default() order: Gemini CLI -> OpenAI -> Ollama -> Anthropic
- Add --llm-provider gemini arm to build_llm_provider() in noxa-cli
- Update unknown-provider error to mention gemini
- Update empty-chain error messages in CLI and MCP to mention gemini CLI
- Update MCP startup warn! to list gemini CLI as first option
