[sky11023]

问题

Agent 回合的 MessageRequest 从不显式下发采样参数，多数推理服务端按 temperature=1.0 处理——随机性最大，且实测会略降 MTP 类引擎的投机接受率（TensorFold 10-30 性能报告②：默认 1.0 vs 低温的接受率差异）。

修复（2026-11-02 拍板）

acp_server.rs 的 Agent 回合请求构造处：temperature: None → Some(DEFAULT_AGENT_TEMPERATURE)，拍板值 0.3（兼顾确定性与多样性），以具名常量+钉测试防漂移。

• 只影响 Agent 回合；缓存预热、翻译、压缩等内部请求维持各自既有值（预热显式 off reasoning，翻译非交互）。
• 提供方固定采样契约不受影响：apply_kimi_code_fixed_sampling 等按 provider 剥离 temperature 的既有逻辑在序列化之后仍生效。
• 副作用说明：本地引擎卡片的 LMM_TEMPERATURE 旋钮改的是服务端缺省，显式请求值会覆盖它——需要引擎侧自定义采样时应在模型配置层显式设置（未来引入用户级采样配置后由用户值优先）。

验证

• cargo check -p codewhale-tui 绿；cargo fmt --check 绿；clippy 无新增（基线存在与本改动无关的历史 too_many_arguments 项）。
• 焦点测试 agent_turn_default_temperature_is_the_ratified_value 通过；序列化契约（Some(t) → body.temperature）由 client 侧既有 Kimi 固定采样用例覆盖。
• 基点 cb577ba55（= PR fix(tui): widen local-endpoint SSE open timeout to 180s #130 的 180s 超时修复，独立合并互不依赖）。

DCO：Signed-off-by 已带。

评审追补（r1 处置说明，见评论区）

• 拍板溯源：LMM_TEMPERATURE 是品悟模型商店卡片脚本侧的引擎旋钮（随部署注入推理服务环境，定义见品悟仓 docs/模型卡片脚本规范.md §环境变量；本仓库不携带该规范，属跨仓引用）。2026-11-02 拍板记录在品悟侧任务书（双机性能优化②：品悟默认下发 temperature=0.x）。
• 过渡期无用户级覆盖：已知并接受——品悟侧的用户级采样配置在同一批次规划中，落地前显式值不可配；引擎侧自定义采样的既有路径（模型配置层显式设置）仍可用，仅"缺省跟随引擎"的用法在 agent 回合被本 PR 覆盖。维持者已知悉。
• 合并顺序：fix(tui): widen local-endpoint SSE open timeout to 180s #130 先合，本 PR rebase 后 diff 将只剩温度改动（评审已注明）。

[asto18089]

First formal review — #132 feat(tui): default agent-turn temperature to 0.3 (stacked on #130)

Review of commit 9725d2cc (the second commit; the first is #130's cb577ba5, reviewed separately). One change requested — see MAJOR. Stacked note: this PR contains #130 in full. Merge order must be #130 → #132; any change to #130 (including its MINOR items) requires this branch to rebase, and after #130 merges a rebase would let this PR's diff show only the temperature change.

Per-file findings

crates/tui/src/acp_server.rs — temperature: None → Some(DEFAULT_AGENT_TEMPERATURE)

• Scope verified: acp_server.rs:1995 is the only production construction site of the agent-turn MessageRequest (other sites are translation, cache-warmup copy, reviewer, and tests). The reviewer/guardian path already pins Some(0.0), so explicit sampling on agentic paths has precedent; translation/compaction keep server defaults. Scope is consistent.
• Rejecting routes are protected, so Some(0.3) cannot leak where sampling is forbidden: the Anthropic adapter guards with anthropic_model_rejects_sampling (opus-4-7/4-8, *-5 families, fable, mythos); the Chat adapter strips temperature/top_p for exact direct-Moonshot-K3 and Kimi-Code membership routes; Responses writes temperature only when Some. Verified on base.
• Justification: "servers treat absent temperature as 1.0" is accurate for OpenAI-compatible stacks (and their typical self-hosted reimplementations). The MTP/speculative-acceptance claim is attributed to an internal report I cannot verify from this repo — plausible, and clearly attributed, so acceptable.
• MINOR (unverifiable cross-reference): LMM_TEMPERATURE appears nowhere in CodeWhale or pinvou3-app (grepped both). The comment's claim that the local-engine card knob sets the server-side default and that an explicit request value overrides it is standard API semantics, but the knob itself cannot be verified from this codebase. Please link the engine-side definition or the ratification record (issue/report) in the PR description so the "2026-11-02 拍板" is traceable.
• MINOR (no escape hatch): once this lands, agent-turn sampling is hard-coded with no in-repo override until the promised user-level sampling config exists. Users who tuned sampling via engine-side defaults silently lose it on agent turns. The comment discloses this — good — but please confirm maintainers explicitly accept the interim state.
• nit: the commit body is empty apart from the sign-off; the rationale lives only in code comments. A two-line body would help git log archaeology.

crates/tui/src/acp_server.rs — agent_turn_default_temperature_is_the_ratified_value

• MAJOR: assert_eq!(DEFAULT_AGENT_TEMPERATURE, 0.3) is a pure tautology — it asserts a constant equals a literal and provides zero regression protection (it can only fail if someone edits the constant, which any diff would show). Please replace it with a behavioral assertion, e.g. build the ACP agent-turn request (or capture the wire body via the existing wiremock harness in client.rs tests) and assert body["temperature"] == 0.3 for a vanilla route, plus the negative space: no temperature on a sampling-rejecting route (anthropic_model_rejects_sampling hit) and on the Kimi-Code stripping route. That would also fix the comment's claim below.
• MINOR: the test comment says the Some(t) → body.temperature serialization contract is covered by "client 侧 Kimi 固定采样用例" — those tests assert the opposite direction (Some(0.3) is stripped on that route). Pass-through is covered elsewhere (e.g. the wire assertion of temperature == 0.25 on the neighbor route); cite that instead.
• nit: the test fn is inserted before use super::*; in mod tests. Compiles fine, unconventional placement — move it below the imports.

CI

No workflow runs exist for head 9725d2cc (runs list, check runs, and combined status all empty — first-contribution approval gate; zero runs in status=waiting). Nothing for me to approve via the API at my permission level — maintainers, please approve/trigger the workflow run for both this PR and #130.

Conclusion

• MAJOR: tautological pinned test — replace with a wire-level assertion of the new default plus the rejecting-route negative cases.
• MINOR: LMM_TEMPERATURE citation unverifiable in-repo (link it); no interim user override for the hard-coded 0.3 (confirm acceptance); test-comment cites the wrong covering test; empty commit body.
• The behavior change itself is well-contained and correctly guarded on all wire paths. Requesting changes only for the test; once replaced and #130 has settled, this is ready. Merge order: #130 first, then this (rebased).

[asto18089]

Second review — rationale re-verification of the 0.3 default (2026-11-04)

The first review provisionally accepted the justification ("servers treat absent temperature as 1.0 … acceptable"). That line is superseded. The rationale was re-verified from primary sources — vendor API docs, official model cards and generation_config.json files, serving-stack source code, and the coding-agent ecosystem — via seven investigation lanes (four external, three in-repo). The premise of the change does not hold for either endpoint class this PR serves. The standing MAJOR (tautological test) from round 1 remains open.

What omitting temperature actually does (all verified 2026-11-03/04)

Endpoint	Omitted →	Vendor recommendation	Explicit value
OpenAI (GPT-5.x / o-series)	1.0	1.0 — the only supported value ("Only the default (1) value is supported")	non-1.0 → 400 on reasoning models
Anthropic Messages (2026-10-21)	1.0	1.0 — temperature/top_p/top_k removed: sending any of them "returns a 400 error, even at the parameter's default value"; omitted requests "sample with their defaults, including a temperature of 1.0"	deprecated → 400
Google Gemini 3	1.0	1.0 — "we strongly recommend keeping the temperature parameter at its default value of 1.0"; below 1.0 "may lead to unexpected behavior, such as looping or degraded performance"; migration guide says to remove explicit low values	warned against
DeepSeek API	1.0	V3.2 (final) ships generation_config.json temperature: 1.0, top_p: 0.95; SGLang's official V3.2 recipe: "temperature = 1.0, top_p = 0.95 (as recommended by DeepSeek)"; temperature "has no effect in thinking mode". (Legacy task table recommends 0.0 for coding/math — pre-thinking-era guidance for old non-thinking models)	accepted
Qwen3 (DashScope)	0.7 non-thinking / 0.6 thinking	= card recommendations (thinking 0.6/0.95, non-thinking 0.7/0.8); "DO NOT use greedy decoding"	accepted
Zhipu GLM-4.6	1.0	1.0 ("we recommend using a sampling temperature of 1.0", + top_p 0.95 / top_k 40)	accepted
Moonshot Kimi (k2.6+/K3/k2.7-code)	n/a — parameter absent from the current OpenAPI spec	fixed server-side (1.0 thinking/code, 0.6 k2.6 non-thinking); "Do not pass temperature explicitly when calling these models"	error on any explicit value
MiniMax M2/M2.1	recommended 1.0	1.0 / top_p 0.95 / top_k 40; M3 ignores explicit values entirely	ignored
vLLM ≥ 0.8.0	model's generation_config.json (--generation-config auto is the default since vllm#12622); 1.0 is only the last-resort fallback	the model author's shipped values	client value overrides generation config
SGLang ≥ 0.5.4	model generation config (--sampling-defaults model, the default); 1.0 fallback	same	client value wins
Ollama (Sep 2026+) / llama.cpp (Nov 2025+)	GGUF-embedded / generation_config defaults; 0.8 only when the model ships none	same	client value wins

Sources: Anthropic migration 2026-10-21, Gemini 3 guide, DeepSeek parameter settings, DashScope defaults, GLM-4.6 card, Kimi models overview, MiniMax-M2.1 card, vLLM OpenAI-compat docs, SGLang sampling params, Ollama PR #16471, llama.cpp PR #17120.

Why the 0.3 rationale fails on both endpoint classes

1. Cloud: "omitted → 1.0" is true, but 1.0 is the vendors' recommendation for current flagships — not a pathology to be corrected client-side. Gemini explicitly warns that sub-1.0 values degrade reasoning tasks; Anthropic removed the knob entirely; OpenAI reasoning models reject non-1.0. For older Anthropic models on unguarded routes, the PR now sends 0.3 where the server previously applied the per-model default — exactly what the Gemini migration guide tells integrators to stop doing.
2. Local: "most inference servers treat missing temperature as 1.0" is not the current behavior of the stacks this PR's own first commit targets. vLLM ≥ 0.8.0, SGLang ≥ 0.5.4, and current Ollama/llama.cpp all prefer the model's shipped sampling defaults (that is what generation_config.json is for); 1.0 is only a last-resort fallback. An explicit client value overrides those vendor recommendations on every stack — so Some(0.3) forces, e.g., a correctly-served Qwen3-thinking from its recommended 0.6 down to 0.3, and DeepSeek-V3.2-Exp from 0.6 down to 0.3. The first commit of this very PR treats local endpoints as first-class; the second commit overrides their tuned defaults.
3. No vendor recommends 0.3. The recommendation landscape is: 1.0 for the current flagship generation (OpenAI, Anthropic, Gemini, DeepSeek V3.2, GLM-4.6, Kimi thinking/code, MiniMax), or 0.6–0.7 for the reasoning-era minority (Qwen3, DeepSeek-Exp/V3.1/R1, Kimi K2-Instruct). 0.3 matches no published recommendation on any route it reaches.

What survives scrutiny

• The MTP direction has one public corroboration: llama.cpp documents that with rejection-sampling verification "acceptance falls away as temperature rises" (PR #27694). But vLLM/SGLang MTP drafters are greedy by default (temperature-independent), and no vendor or stack publishes "keep temperature at X for accept rate" guidance — DeepSeek's own V3.2 recipe is 1.0/0.95 despite MTP. The cited TensorFold 10-30 report remains unverifiable from outside.
• The wire guards hold: verified that Some(0.3) is stripped on Kimi-Code/direct-K3 membership routes (client/chat.rs:923-968) and anthropic_model_rejects_sampling routes (client/anthropic.rs:734-748); it reaches the wire on plain Chat Completions and older-Anthropic routes, and on Responses routes only for DeepSeek (client/responses.rs:140-147).

In-repo evidence against a hard-coded client default

• This codebase already lived the failure mode: compaction once hard-coded 0.3, it leaked onto the Kimi Code membership route and failed the whole compaction pass; the fix removed the fixed value (compaction now sends None, pinned by tests at compaction.rs:2329/2660). PR #132 reintroduces the same pattern one path over.
• Scope inconsistency: commit 9725d2cc0 touches only the ACP server path. The interactive TUI's own main turn (crates/core/src/request.rs:76), subagents, RLM, one-shot exec, and MCP chat all still send None — so the ratified "agent turns need stable output" premise is not applied to the product's own agent loop, only to external ACP clients.
• LMM_TEMPERATURE is unverifiable: zero hits in both CodeWhale and pinvou3-app (grepped this round). The comment's override interaction is standard semantics (client value always wins), but the knob itself is not in any repo this PR touches — please link the engine-side definition or the ratification record. And since no user/model-level sampling config exists in-repo (no config field, no env var), any engine-side default is silently overridden with no escape hatch.
• App-side prior art mandates omission: the pinvou3 settings-layer image probe deliberately omits temperature — "kimi-for-coding 等模型只接受默认值,显式传会 400" (settings.rs:811, test-pinned) — consistent with Kimi's current "do not pass explicitly" contract.
• Ecosystem direction supports omission over hardcoding: Codex CLI omits temperature, OpenAI Agents SDK defaults to None, LangChain moved 0.7 → omitted, Claude Code has no sampling knob, and Gemini CLI is the notable (aging) exception that still sends explicit 0 on its main loop.

Requested changes

1. Primary: revert temperature: Some(DEFAULT_AGENT_TEMPERATURE) back to None. Omitting is the only choice that yields the vendor-recommended value on every verified endpoint — cloud and local — and it matches where the API ecosystem is heading (parameter removal/rejection). If the "2026-11-02 拍板" decision survives this evidence, it needs re-ratification by maintainers with these sources on record.
2. Alternative: if a deterministic agent-turn default is genuinely wanted, implement it as an explicit opt-in (model-level config field, the already-promised user-level sampling config), not a silent hardcode — and do not pick 0.3, which sits off every vendor's recommendation.
3. Unchanged from round 1: replace the tautological constant test with wire-level assertions (0.3 present on a vanilla route; absent on the Anthropic-rejecting and Kimi stripping routes), fix the test comment's wrong cross-reference, and move the test below use super::*.
4. Stacking: #130 has advanced to ca079796c while this PR still carries the old cb577ba5 — after #130 merges, rebase so the diff shows only the temperature change. CI remains empty on 9725d2cc0 (first-contribution gate) — maintainers, please still approve/trigger the workflow runs for #130 and #132.

[sky11023]

r1 处置（head 97dd00aef）

MAJOR（同义反复钉测试）：已替换为行为级三路由断言 agent_turn_temperature_survives_vanilla_and_is_stripped_on_sampling_fixed_routes——构造抽成 agent_turn_message_request（生产唯一 agent-turn 构造点），普通路由经生产 build_chat_wire_body 断 wire body temperature==0.3；Kimi-Code 会员路由经生产 apply_kimi_code_fixed_sampling 断剥除；Anthropic 拒绝采样族经新抽的 apply_anthropic_sampling 断两个采样参数都不带（含非拒绝族对照臂）。通道全部走生产函数，零重抄。

这个测试当场抓到一个真 bug：f32 采样参数直扩 f64 后，wire 上是 0.30000001192092896（17 位尾数）——旧的同义反复测试永远抓不到。已在序列化接缝修复：新增 sampling_wire_value（经 f32 最短十进制表示中转），chat 与 anthropic 两处落 wire 统一走它；既有 anthropic 测试钉的期望值恰是伪影值（Some(f64::from(0.7f32))=0.699999988079071），已更新为干净 0.7 并注明原委。

MINOR：①LMM_TEMPERATURE 溯源与②过渡期无覆盖的接受声明已写入 PR 正文（见「评审追补」节）；③测试注释的错误引用随测试重写消除；④提交 body 已补（rationale + r1 处置摘要）。测试位置已移到 use super::* 之后。

验证：agent_turn 1/1、sampling 7/7、anthropic 71/71、kimi 124/124；fmt ✅。clippy -D warnings 的 6 条 too_many_arguments 为仓内既知存量（context_report/engine/turn_loop/lib.rs，与本 PR 文件零交集）。

堆叠：维持 #130 → 本 PR 的合并顺序；#130 的 MINOR 项有任何改动我会随之 rebase。

[sky11023]

r1 处置（head 97dd00aef）

MAJOR（同义反复钉测试）：已替换为行为级三路由断言 agent_turn_temperature_survives_vanilla_and_is_stripped_on_sampling_fixed_routes——构造抽成 agent_turn_message_request（生产唯一 agent-turn 构造点），普通路由经生产 build_chat_wire_body 断 wire body temperature==0.3；Kimi-Code 会员路由经生产 apply_kimi_code_fixed_sampling 断剥除；Anthropic 拒绝采样族经新抽的 apply_anthropic_sampling 断两个采样参数都不带（含非拒绝族对照臂）。通道全部走生产函数，零重抄。

这个测试当场抓到一个真 bug：f32 采样参数直扩 f64 后，wire 上是 0.30000001192092896（17 位尾数）——旧的同义反复测试永远抓不到。已在序列化接缝修复：新增 sampling_wire_value（经 f32 最短十进制表示中转），chat 与 anthropic 两处落 wire 统一走它；既有 anthropic 测试钉的期望值恰是伪影值（Some(f64::from(0.7f32))=0.699999988079071），已更新为干净 0.7 并注明原委。

MINOR：①LMM_TEMPERATURE 溯源与②过渡期无覆盖的接受声明已写入 PR 正文（见「评审追补」节）；③测试注释的错误引用随测试重写消除；④提交 body 已补（rationale + r1 处置摘要）。测试位置已移到 use super::* 之后。

验证：agent_turn 1/1、sampling 7/7、anthropic 71/71、kimi 124/124；fmt ✅。clippy -D warnings 的 6 条 too_many_arguments 为仓内既知存量（context_report/engine/turn_loop/lib.rs，与本 PR 文件零交集）。

堆叠：维持 #130 → 本 PR 的合并顺序；#130 的 MINOR 项有任何改动我会随之 rebase。