Fixes #281.

The bug

DeepSeek's thinking mode is on by default, and once a turn has called a tool the API requires that turn's reasoning_content back on every subsequent request. Without it the next call is a 400, which is what the issue reports.

eko never sends it — not because of a provider gap, but because the reasoning is discarded before the provider ever sees it:

• callLLM handles reasoning-start / reasoning-delta / reasoning-end only to emit thinking stream events for the UI. The array it returns is built from streamText and toolParts alone, so the reasoning is gone the moment the stream ends.
• convertAssistantContent, which turns that array into { role: "assistant", content }, then filters explicitly to text and tool-call.

The provider layer was ready the whole time. @ai-sdk/openai-compatible walks an assistant message's parts and does exactly what DeepSeek needs:

case "reasoning": { reasoning += part.text; break; }
...
...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),

So the part only had to survive eko's own assembly of the turn.

The fix

reasoning-end records the block, carrying providerMetadata through as providerOptions — Anthropic's signature rides in there, and a thinking block replayed without its signature is rejected too, so the same fix covers extended thinking. thinkText resets afterwards so a turn with several reasoning blocks keeps them apart instead of concatenating them into the second one.

Both return paths now go through a single collected(), which puts reasoning ahead of text and tool calls. That also means the abort path returns the same shape rather than a second hand-built array that could drift.

LLMAssistantPart names the widened union once instead of repeating it at a dozen signatures.

What the widened union means for existing consumers

Adding a third part type to an array others already read is the risky half, so each consumer was checked rather than assumed:

Verification

npx jest test/llm/reasoning-round-trip.test.ts — 6 passed. The suite is offline: it drives callLLM with a hand-built ReadableStream of provider chunks, so it needs no API key.

Each test was checked against a mutation, so none of them passes for an unrelated reason:

npx tsc --noEmit reports only the pre-existing agent-llm.ts:47 providerOptions error, which is present on a clean checkout of main as well (verified by stashing). npx rollup -c builds.

Not covered

The alternative the issue also mentions — disabling thinking mode for DeepSeek — is not done here. Sending the reasoning back is the behaviour-preserving half: it keeps the model's thinking available to both the user and the model, and it fixes Anthropic's signature case at the same time. A switch to turn thinking off is a configuration feature, not a bug fix, and belongs in its own change.

Not verified against the live DeepSeek API — I have no key for it. What is verified is that the reasoning part now reaches the provider layer, and that @ai-sdk/openai-compatible turns that part into reasoning_content.