This regresses #15749 and reintroduces #15250.

The current ordering — step 2 (DeepSeek/Kimi tool-call empty-string guard) before step 3 (reasoning → reasoning_content promotion) — is intentional, set by #15749 (commits 9daa0620 + 5ae60815). The comment on step 2 spells out the design:

DeepSeek returns HTTP 400 if reasoning_content is absent on replay; inject "" to satisfy the provider's requirement without forwarding any cross-provider reasoning content.

That's the whole point: when the active provider is DeepSeek/Kimi and the source message has tool_calls but no native reasoning_content, the reasoning field is presumed to be cross-provider (MiniMax, etc.) and must not be promoted. Swapping step 2 and step 3 — which is what this PR does — restores exactly the bug #15749 fixed.

A few other smell tests:

1. Misattributed history. Body says "fix: DeepSeek/Kimi thinking mode requires reasoning_content on ALL assistant messages #15478 … introduced a regression." The merged ordering change came from fix(agent): ordering fix in _copy_reasoning_content_for_api — cross-provider reasoning isolation #15749 (Zjianru, 2026-04-25). Issue regression: #15749 breaks reasoning field promotion for DeepSeek/Kimi tool-call messages #15812 itself names fix(agent): ordering fix in _copy_reasoning_content_for_api — cross-provider reasoning isolation #15749, not fix: DeepSeek/Kimi thinking mode requires reasoning_content on ALL assistant messages #15478.
2. No regression test. Diff is run_agent.py only. The 21/21 figure is the pre-existing suite; there's no test for the "genuine reasoning lost" scenario this PR claims to fix, and no test that pins the cross-provider isolation fix(agent): ordering fix in _copy_reasoning_content_for_api — cross-provider reasoning isolation #15749 was protecting.
3. Scenario plausibility. The "lost reasoning" case requires a DeepSeek/Kimi session whose own assistant tool-call message has reasoning set but no reasoning_content. Native DeepSeek/Kimi turns write reasoning_content directly — the only way to land in that state is exactly the cross-provider migration fix(agent): ordering fix in _copy_reasoning_content_for_api — cross-provider reasoning isolation #15749 is isolating.

If #15812 is a real failure mode, the fix needs to be provider-of-origin-aware (e.g. only promote reasoning when the message was authored by the current provider), not an order swap. As-is this is a revert wrapped in a misattributed changelog.

Duplicate of #15830 — same fix for #15812 (reorder reasoning field promotion before empty-pad in _copy_reasoning_content_for_api).

Merged via PR #16097 (commit 63bf7a2 on main). Your commit was cherry-picked onto current main with your authorship preserved in git log. Thanks for the fix — clean single-file swap with the regression test already in place from #15478 made this easy to land.

#16097