Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(pi): 载体缺失的切模也按目标快照收敛思考档位

Zhang Yunjin55397eec2179
diff --git a/docs/dev-rules/pi-harness.md b/docs/dev-rules/pi-harness.md
index 6b7b1c17fa..bf9eacd64e 100644
--- a/docs/dev-rules/pi-harness.md
+++ b/docs/dev-rules/pi-harness.md
@@ -91,12 +91,17 @@ Cindy 显式设置:models.json、`settings.json` 的 `transport:sse` 与 `retry.
漂移给新模型——实测推理模型在 off 下把思考写进正文)。思考意图由 renderer 的原子 selection 携带
(主进程只对 Pi 转成 `thinkingEnabled`,点选暂存与 deferred 重放一并回放);目标档位不在启动快照内时
按默认档/首档回落(同 `reconcilePiStartupEffort` 口径),目标声明不支持思考时明确 `off`,能力未知
-则不动。实现见 `packages/maker-core/src/agents/pi/index.ts` 的 `switchModel` 末尾,
+则不动。**载体缺失(旁路入口热切模型)时同样按目标快照收敛**:目标支持思考就按会话档位重新下发一次,
+否则保留现有档位——历史实现的「无意图就不动」会让这类切换静默丢思考通道(2026-10-06 实报:Orca lead
+切到 `commandcode/deepseek-v4.1-flash` 后 `usage.reasoning` 归零、推理整段写进正文,Pi 侧无任何
+`set_thinking_level` 记录,直到用户手动改一次档位才恢复)。实现见
+`packages/maker-core/src/agents/pi/index.ts` 的 `switchModel` 末尾,
回归见 `pi-provider-routing.test.ts` 的 thinking-level 用例。跨引擎切换提交时还会把这次选择写进
本机偏好镜像(普通 send 走 lazy-create 只读镜像、不经过 bootstrap),避免远端推送未回流时新建会话
-用旧值。已知未覆盖面:Bot 设置保存、
-agent/scheduler/IM 回滚等**旁路入口**热切 Pi 模型时没有 thinking 载体(`SessionRuntimeProfile`
-不含该字段),这些路径仍可能保留旧档位;需要补载体或按目标快照收敛后才能彻底消除。
+用旧值。**仍未覆盖**:Bot 设置保存、agent/scheduler/IM 回滚等**旁路入口**热切 Pi 模型时没有 thinking
+载体(`SessionRuntimeProfile` 不含该字段),按目标快照收敛能保证推理模型不丢通道,但无法区分
+「用户本想关思考」;要完全对齐用户意图仍需给这些入口补载体(本机偏好镜像已有
+`getThinkingEnabledFromMemory`,目前只在会话启动钩子接入)。
Pi 原生负责 threshold 与 overflow 压缩;
Cindy 消费 compaction 事件做 UI、usage、digest 投影,并只在本机原生自动压缩确定性失败后锁存
下一次发送前换窗。设置页的 Pi 百分比默认 90%(已有显式 override 保留),在每次启动或恢复
diff --git a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
index 782ab92d15..8785a44935 100644
--- a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
@@ -4142,7 +4142,7 @@ describe("Pi provider-aware model routing", () => {
await handle.close();
});
- it("does not touch the thinking level when the caller sends no thinking intent", async () => {
+ it("re-applies the session effort when a switch to a reasoning-capable model carries no intent", async () => {
const denseMap = {
minimal: "minimal",
low: "low",
@@ -4174,6 +4174,57 @@ describe("Pi provider-aware model routing", () => {
],
),
);
+ const handle = await agent.startSession({
+ sessionId: "thinking-converged-without-intent",
+ workingDir: cwd,
+ model: "model-a",
+ providerId: "native-a",
+ effort: "high",
+ });
+ const beforeSwitch = captured.requests.length;
+
+ await handle.setModel!("model-b", { providerId: "native-a" });
+
+ // 旁路入口(scheduler / IM 回滚 / 插件任务 / 会话恢复)不带 thinking 载体,也必须按会话
+ // 档位重新下发一次:历史实现的「没有意图就不动」让 Pi 保留旧档位,实测表现是切模后
+ // usage.reasoning 归零、推理整段写进正文(2026-10-06 Orca lead 切 commandcode 实报)。
+ const switched = captured.requests.slice(beforeSwitch);
+ expect(switched).toContainEqual({ type: "set_thinking_level", level: "high" });
+ await handle.close();
+ });
+
+ it("keeps the thinking level untouched when a switch without an intent targets a model without thinking levels", async () => {
+ const denseMap = {
+ minimal: "minimal",
+ low: "low",
+ medium: "medium",
+ high: "high",
+ xhigh: "xhigh",
+ max: "max",
+ } as const;
+ const agent = new PiAgent(
+ byomDeps(
+ async () => ({
+ providers: [
+ {
+ id: "native-a",
+ name: "Native A",
+ baseUrl: "http://a.test",
+ api: "openai-responses",
+ models: [
+ { id: "model-a", reasoning: true, thinkingLevelMap: { ...denseMap } },
+ { id: "model-b" },
+ ],
+ },
+ ],
+ env: {},
+ }),
+ [
+ { id: "model-a", displayName: "A", contextWindow: 200_000, efforts: ["low", "high"], defaultEffort: "low" },
+ { id: "model-b", displayName: "B", contextWindow: 200_000, efforts: [], defaultEffort: null },
+ ],
+ ),
+ );
const handle = await agent.startSession({
sessionId: "thinking-untouched-without-intent",
workingDir: cwd,
@@ -4185,7 +4236,7 @@ describe("Pi provider-aware model routing", () => {
await handle.setModel!("model-b", { providerId: "native-a" });
- // 老调用方(无 intent)保持既有语义:只切模型,不动 thinking level。
+ // 载体缺失只补「开」:目标自身没有档位时不猜、不写(要关思考的调用方会显式传 false)。
expect(
captured.requests
.slice(beforeSwitch)
diff --git a/packages/maker-core/src/agents/pi/index.ts b/packages/maker-core/src/agents/pi/index.ts
index f0167d2c72..41e270542f 100644
--- a/packages/maker-core/src/agents/pi/index.ts
+++ b/packages/maker-core/src/agents/pi/index.ts
@@ -6613,10 +6613,19 @@ export class PiAgent extends BaseAgent {
// 上一个模型的档位带过去(Pi 会把 thinkingLevelMap 里的 null 当成关闭
// reasoning)。快照未知时不猜、不动 Pi 现有档位。
//
+ // 载体缺失(scheduler / IM 回滚 / 插件任务 / 会话恢复等旁路入口不带
+ // thinkingEnabled)同样必须收敛:目标支持思考时按会话档位重新下发一次。历史上
+ // 「没有意图就不动」把这类切换留成了静默失效——2026-10-06 实报的 Orca lead 切到
+ // commandcode/deepseek-v4.1-flash 后 usage.reasoning 归零、推理整段写进正文,直到
+ // 用户手动改一次档位才恢复;Pi 侧没有任何 set_thinking_level 记录。载体缺失只影响
+ // 「开」:目标空档位时保持不动(Pi 会把 null 档位当关闭),显式 false 仍明确发 off。
+ //
// 失败只告警、不抛:模型切换已经成功,抛出会让上层回滚 route / 终止会话,反而
// 制造「Pi 已切模、宿主认为没切」的分裂。用户可在 UI 上重新选择。
- if (setOpts && typeof setOpts.thinkingEnabled === 'boolean') {
- const requestedEffort = setOpts.effort ?? mutableEffort ?? startupEffort ?? null;
+ const explicitThinking =
+ setOpts && typeof setOpts.thinkingEnabled === 'boolean' ? setOpts.thinkingEnabled : undefined;
+ {
+ const requestedEffort = setOpts?.effort ?? mutableEffort ?? startupEffort ?? null;
const nativeTarget = provider !== PI_PROVIDER_ID
? nativeProviderById.get(provider)?.models.find(
(candidate) => candidate.id === resolveNativeModelId(provider, model),
@@ -6629,7 +6638,7 @@ export class PiAgent extends BaseAgent {
: nextEffortSnapshot.length > 0;
let level: string | null = null;
let appliedEffort: Effort | null = null;
- if (!setOpts.thinkingEnabled || targetReasoning === false) {
+ if (explicitThinking === false || (explicitThinking === true && targetReasoning === false)) {
level = 'off';
} else if (targetReasoning === true) {
const usable = nextEffortSnapshot && nextEffortSnapshot.length > 0
@@ -6665,6 +6674,14 @@ export class PiAgent extends BaseAgent {
provider,
appliedEffort: reconciled,
});
+ } else if (explicitThinking === undefined) {
+ // 旁路入口(无 intent 载体)也要留痕:这类收敛是新补的能力,出现异常时
+ // 需要能把「谁把档位改回来了」和用户主动切档位区分开。
+ deps.logger.info('pi: thinking level re-applied after a model switch without an intent carrier', {
+ model,
+ provider,
+ appliedEffort: reconciled,
+ });
}
appliedEffort = reconciled;
level = effortToPiThinkingLevel(reconciled);