Summary

• Adds an SGLang-owned OpenAI-compatible replay policy instead of inheriting the strict fallback that drops historical reasoning by default.
• Keeps the existing Gemma 4 protection: Gemma 4 openai-completions model ids still drop historical reasoning.
• Adds focused provider-plugin regression coverage for both the SGLang/Kimi reasoning case and the Gemma 4 guardrail.

Closes #81058.

Root Cause

SGLang models are registered as openai-completions, but the provider did not own its replay policy. That made it fall through to the core strict OpenAI-compatible fallback, which sets dropReasoningFromHistory: true for unowned openai-completions providers. For reasoning-capable self-hosted SGLang/Kimi models, that can strip replayed reasoning history and contribute to empty user-facing responses.

Fix

SGLang now uses buildProviderReplayFamilyHooks({ family: "openai-compatible", dropReasoningFromHistory: false }), matching the existing provider-owned opt-out pattern used by reasoning-capable OpenAI-compatible providers. The helper still forces dropReasoningFromHistory: true for Gemma 4 model ids, preserving the prior parser-safety behavior.

Regression Proof

I checked the fix before republishing the branch:

• main at validation base ea7f74ff had no SGLang replay-policy hook.
• The PR adds the provider-owned replay policy only for SGLang.
• Regression coverage proves Kimi-style SGLang reasoning history is preserved.
• The Gemma 4 guardrail remains covered and still drops historical reasoning.
• Published PR head: d162f0c7cf83187bb578d7f90405e1d6b0123f1b.
• GitHub currently reports the PR as mergeable against main.

Validation

Latest local validation after rebasing and publishing the PR branch:

• PNPM_CONFIG_OFFLINE=true pnpm test extensions/sglang/index.test.ts src/plugins/provider-replay-helpers.test.ts src/agents/transcript-policy.test.ts src/agents/pi-embedded-runner.sanitize-session-history.test.ts
• PNPM_CONFIG_OFFLINE=true pnpm exec oxfmt --check extensions/sglang/index.ts extensions/sglang/index.test.ts
• PNPM_CONFIG_OFFLINE=true pnpm exec tsc -p extensions/sglang/tsconfig.json --noEmit --pretty false
• PNPM_CONFIG_OFFLINE=true pnpm check:changed

Risk

Low and provider-scoped. This does not change the global OpenAI-compatible fallback. The main compatibility risk is accidental reintroduction of replayed reasoning for Gemma 4-style chat-completions models, covered by the new test.

Maintainer Verification

Maintainer proof after rebasing and adding changelog:

• PR branch head: d55083d7e126ce1fc4d576fb841b47efe1c9093e.
• Local targeted regression: pnpm test extensions/sglang/index.test.ts -- --reporter=verbose -> 2 tests passed.
• Local whitespace check: git diff --check -> passed.
• Live SGLang proof: Blacksmith Testbox tbx_01krgt6yqe7bte96dwsmker05p, GitHub Actions run 25803983869.
• Live server: source-built SGLang CPU, sglang 0.0.0.dev1+g9e00b7ca9.d20260513, torch 2.9.0+cpu, model Qwen/Qwen3-0.6B, /v1/models returned the model id.
• Direct replay request sent prior assistant reasoning_content; SGLang returned content: "replay ok", non-empty reasoning_content, finish: "stop".
• PR unit regression on the same Testbox: pnpm test extensions/sglang/index.test.ts -- --reporter=verbose -> 2 tests passed.

Real behavior proof

• Behavior or issue addressed: SGLang OpenAI-compatible chat completions should preserve replayed assistant reasoning_content for thinking-capable local models instead of stripping reasoning history through the strict fallback.
• Real environment tested: Blacksmith Testbox tbx_01krgt6yqe7bte96dwsmker05p, GitHub Actions run 25803983869, source-built SGLang CPU server, sglang 0.0.0.dev1+g9e00b7ca9.d20260513, torch 2.9.0+cpu, model Qwen/Qwen3-0.6B served on 127.0.0.1:30000.
• Exact steps or command run after this patch: Launched python -m sglang.launch_server --model Qwen/Qwen3-0.6B --trust-remote-code --disable-overlap-schedule --device cpu --host 127.0.0.1 --port 30000 --tp 1 --reasoning-parser qwen3 --max-total-tokens 4096, then sent a live curl http://127.0.0.1:30000/v1/chat/completions request containing a prior assistant message with reasoning_content.
• Evidence after fix: Terminal output from the live run: /v1/models returned Qwen/Qwen3-0.6B; the direct replay response JSON was { "model": "Qwen/Qwen3-0.6B", "content": "replay ok", "reasoning": "Okay, the user wants me to answer exactly: ...", "finish": "stop" }.
• Observed result after fix: SGLang accepted the replayed reasoning_content, returned visible content replay ok, returned non-empty reasoning_content, and finished with stop; the PR regression test on the same Testbox also passed 2/2.
• What was not tested: Kimi/GPU SGLang was not tested because AWS GPU spot/on-demand quota blocked g5.xlarge and g4dn.xlarge; CPU SGLang with Qwen3 tested the same OpenAI-compatible reasoning_content replay contract.