#6685: fix: suppress thinking leak for Synthetic reasoning models

## Summary Fixes #6442 When using Synthetic provider with reasoning-capable models (e.g. Kimi K2.5), thinking text leaks into user-facing output as plain text. This happens because pi-ai's Anthropic provider sends no `thinking` parameter when `thinkLevel` is `"off"` (the default), and Synthetic/Kimi defaults to emitting thinking as plain text content blocks. - Adds `createThinkingDisabledWrapper()` in `extra-params.ts` that injects `thinking: { type: "disabled" }` into the Anthropic API payload via the `onPayload` callback - Only applies when: `thinkLevel === "off"`, `model.reasoning === true`, and `model.api === "anthropic-messages"` - No effect on actual Anthropic models (Claude already handles absent `thinking` param correctly; `disabled` is a valid no-op) - No effect on non-anthropic-messages providers (Venice, OpenRouter, Ollama all use `openai-completions`/`openai-responses`) ## Test plan - [x] New unit tests for `createThinkingDisabledWrapper` (4 tests covering inject, no-overwrite, non-reasoning skip, non-anthropic skip) - [x] `pnpm lint && pnpm build && pnpm test` pass - [x] Manual test: sent messages via Synthetic/Kimi K2.5 with `thinkLevel: "off"` — thinking no longer leaks into user output - [x] Manual test: sent messages with `--thinking medium` — model reasons, structured thinking blocks properly filtered from output 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This PR prevents “thinking” text from leaking into user-visible output when using Anthropic-compatible providers (notably Synthetic/Kimi) with reasoning-capable models and `thinkLevel: "off"`. It does so by adding a `createThinkingDisabledWrapper()` stream wrapper that injects `thinking: { type: "disabled" }` into the Anthropic Messages payload (via `onPayload`) when the model is marked `reasoning: true` and uses the `anthropic-messages` API, and then applying that wrapper during embedded runs when thinking is disabled. Unit tests cover injection, non-overwrite, and skip conditions for non-reasoning / non-anthropic APIs. <h3>Confidence Score: 4/5</h3> - This PR is likely safe to merge; it’s a small, targeted payload mutation with tests. - The change is narrowly scoped (only wraps streamFn when thinkLevel is off and model is reasoning + anthropic-messages) and has unit tests for core behavior. Main remaining risk is subtle onPayload chaining/ordering or models/providers where forcing `thinking: disabled` has unintended effects, but the wrapper avoids overwriting existing thinking settings. - src/agents/pi-embedded-runner/extra-params.ts and src/agents/pi-embedded-runner/run/attempt.ts <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->