[pjain1]

Fixes AI chat failing on current models with the existing AI drivers, found while building the Bedrock driver (#10069) and from a report that every chat failed on Claude Sonnet 5.5, Opus 5.5, and Haiku 5.5 through the claude connector. The reasoning and prompt caching changes follow the approach in #10074.

Driver fixes (first commit)

• claude: only send temperature if configured (Opus 4.7+ rejects sampling parameters, and Sonnet/Haiku 5.5 reject non-default values), and keep $defs in tool schemas (BetaToolInputSchemaParam dropped every keyword except type/properties/required, so $ref params like time_range/where pointed at nothing). Empty text blocks are skipped in requests and responses.
• gemini: only send temperature if configured (Google recommends the default 1.0 for Gemini 3 and warns lower values can cause looping).
• openai: no driver change. Its default temperature of 0.1 is accepted by the default model (gpt-5.2) and is already omitted when reasoning_effort is set, so the docs now say to set reasoning_effort for models that reason by default.
• Behavior change: when not configured, Claude and Gemini models now sample at their own default temperature instead of 0.1, including Rill-managed AI on those drivers.

Opt-in prompt caching for Claude (second commit)

• New prompt_caching option on the claude connector (off by default) marks the tools, system prompt, and conversation as cache breakpoints. Rill-managed AI can enable it with RILL_ADMIN_CLAUDE_PROMPT_CACHING.
• Cache writes (billed at 1.25x input) are reported as a new CompleteResult.CacheWriteInputTokens, passed through the admin Complete API, and emitted as a cache_write_input_tokens metric. Billing will need to price this new metric before enabling caching for managed AI.

Claude reasoning within the tool loop (third commit)

• Anthropic requires thinking blocks to be passed back unmodified within a tool-use turn. New Reasoning block on rill.ai.v1.ContentBlock (text, signature, or redacted data).
• Reasoning is kept in memory until the tool loop completes (not added to the session). It's removed when the history is truncated and on the final no-tools iteration, since thinking blocks are bound to the history and tools they were produced with (removing all of them is allowed).
• The admin Complete API removes reasoning from responses unless the client sets include_reasoning, so older runtimes (e.g. older Rill Developer versions on managed AI) keep working.
• The gemini driver and the metrics view / canvas generators skip reasoning blocks instead of erroring.

Tests: drivers are tested against fake servers through their real SDKs (request body assertions); the tool loop, truncation, and admin stripping have unit tests. The key tests fail against the code without the corresponding fix. Not yet run against the real APIs.

Not in scope: moving the openai driver to the Responses API (the newest OpenAI models only support tool calling there).

Checklist:

• Covered by tests
• Ran it and it works as intended
• Reviewed the diff before requesting a review
• Checked for unhandled edge cases
• Linked the issues it closes
• Checked if the docs need to be updated. If so, create a separate Linear DOCS issue
• Intend to cherry-pick into the release branch
• I'm proud of this work!

---

Developed in collaboration with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[nishantmonu51]

👍 , LGTM

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[nishantmonu51]

This review was posted by an AI agent (OpenHands).

No material correctness, security, or design issues found in this head. Reviewed driver conversion, reasoning lifetime and truncation, admin backward compatibility, and cache-write token propagation.

Focused Go tests passed in runtime/drivers/claude, runtime/drivers/gemini, runtime/drivers/openai, runtime/ai, and admin/server, covering temperature, schema preservation, caching, reasoning round-trips, truncation, and legacy response stripping. All 16 current GitHub checks passed. Provider requests were exercised through fake servers; I did not validate against live model APIs.

✅ APPROVED

LLM profile: unknown · Model: gpt-6-astra