fix(agent-runtime): backfill reasoning_content for relayed DeepSeek models

[L4XB]

Refs #223 (the backfill part, "B" in the report)

Problem

In @earendil-works/pi-ai convertMessages() (dist/api/openai-completions.js) the reasoning_content: "" backfill is guarded by

if (compat.requiresReasoningContentOnAssistantMessages && model.reasoning && assistantMsg.reasoning_content === undefined)

model.reasoning is false for a DeepSeek model reached through a relay (not in the models.dev catalogue, no thinkingLevels), so even with the compat flag set the backfill never ran. A session whose history contains some assistant turns with thinking and some without was sent as a mix; as the report measured, DeepSeek rejects the mix with HTTP 400 and accepts all-or-none.

Change

The repo already patches pi-ai (patches/@earendil-works__pi-ai@0.85.1.patch), so the patch is extended (via pnpm patch / pnpm patch-commit, hence the pnpm-lock.yaml hash change):

• after the message loop in convertMessages(), if compat.requiresReasoningContentOnAssistantMessages is set and any assistant message ends up with a non-empty reasoning_content, every other assistant message gets reasoning_content: "".
• The existing per-message backfill for models flagged reasoning is untouched, and a history without any thinking is left exactly as before (no reasoning_content at all).

Not included: the report's part "A" (treating any model whose id contains deepseek as a DeepSeek target). deepseekRequestCompat() in packages/shared/src/provider-presets.ts already matches on the provider key or a deepseek.com base URL; widening that to arbitrary relay base URLs by model id is a product decision I did not want to take here.

Tests

packages/agent-runtime/src/reasoning-content-backfill.test.ts calls the exported convertMessages with a reasoning: false model and the DeepSeek compat:

• one assistant turn with thinking plus one without → ["let me think", ""];
• no thinking anywhere → [undefined, undefined].

Verified locally: vitest run src/reasoning-content-backfill.test.ts (2 passed), tsc -p tsconfig.json --noEmit in packages/agent-runtime, pnpm install applies the updated patch.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[vercel]

@L4XB is attempting to deploy a commit to the vastsa's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vastsa]

Reviewed and merged as commit 370f076. The change is scoped to the patched pi-ai OpenAI-completions replay path: when DeepSeek-compatible history contains a non-empty reasoning_content on any assistant turn, missing assistant turns receive an empty value, while all-no-thinking histories remain unchanged. Validation passed: focused regression 2/2, agent-runtime typecheck, all 353 agent-runtime tests, and protocol smoke E2E 18/18 (2 live-model cases skipped because credentials were not configured). CI JS and Rust checks were green; the Vercel status requires deployment authorization and was not a code failure.