[qirh]

Use current model examples and documented preset mappings. Correct search budgets, document inputs and limits, and image-search mappings. Require an API key for model catalog checks.

The leftover scanner flags Sonar image and pdf_url parameters. The scanner and grader accept nested web_search token budgets while rejecting request-level max_tokens, including when both appear on the same line or in nearby objects.

Verified scanner and grader regressions, both reference migrations, and a typed SDK generation request with a current model, nested search budgets, and an explicit output limit. Legacy fixtures remain unchanged.

[qirh]

Changes needed before merge.

[P2] The new guidance recommends tools[web_search].max_tokens, but verification.md and evals/migrate-sonar-to-agent-api/check.sh reject max_tokens globally. Reproduction: the raw-http reference passes; adding only valid nested max_tokens/max_tokens_per_page makes the mandatory grader fail. Restrict the prohibition to legacy top-level generation fields, permit nested search budgets, and add positive/negative grader cases.

[P3] Align verification.md's remaining no-auth statement with the updated model-catalog guidance.

[swatikumar-pplx]

LGTM — mappings match the docs and all links return 200. Nits: request-mapping.md:131 the 72 MiB limit applies to all /v1/responses bodies, not just documents; models-and-presets.md:18 sonar-reasoning→low isn't in the migration guide; integration-styles.md:59 still uses a "low" example.