feat(orchestrator): timeouts — per-bucket/tier timeout profiles

Port ORCHESTRATOR_TIMEOUT_PROFILES, SUB_AGENT_TIMEOUT_PROFILES, and builtin-planning timeouts from the TS ai-runtime-config / prompt-optimizer / model-profiles trio into crates/op-orchestrator/src/timeouts.rs. - Timeouts { hard, no_text, first_text } + PromptBucket enum - bucket_for(len): Short(<2200) / Medium(<4200) / Long - orchestrator_timeouts, sub_agent_timeouts (Basic-tier clamps), builtin_planning_timeouts - apply_profile_to_timeouts (multiplier scaling, Math.round parity) - CallRequest gains no_text_timeout/first_text_timeout: Option<Duration> (all existing literals updated to None)
2026-05-23 01:59:27 +08:00 · 2026-05-23 01:59:27 +08:00 · e416ee1cf6
parent 56c99f5aa8
commit e416ee1cf6
4 changed files with 387 additions and 0 deletions
--- a/crates/op-orchestrator/src/lib.rs
+++ b/crates/op-orchestrator/src/lib.rs
@ -17,6 +17,7 @@ pub mod plan;
 pub mod plan_normalize;
 pub mod retry;
 pub mod style_guide_context;
+pub mod timeouts;
 pub mod types;
 pub mod variables;

--- a/crates/op-orchestrator/src/prompt.rs
+++ b/crates/op-orchestrator/src/prompt.rs
@ -85,6 +85,8 @@ pub fn build_orchestrator_prompt(
                    provider: req.provider.clone(),
                    timeout: PLANNING_TIMEOUT,
                    abort,
+                    no_text_timeout: None,
+                    first_text_timeout: None,
                },
                forced_style_guide_name: Some(cp.selected_style_guide_name),
                mode,
@ -121,6 +123,8 @@ pub fn build_orchestrator_prompt(
                    provider: req.provider.clone(),
                    timeout: PLANNING_TIMEOUT,
                    abort,
+                    no_text_timeout: None,
+                    first_text_timeout: None,
                },
                forced_style_guide_name: None,
                mode,
@ -164,6 +168,8 @@ pub fn build_subagent_prompt(
        provider: req.provider.clone(),
        timeout: SUBAGENT_TIMEOUT,
        abort,
+        no_text_timeout: None,
+        first_text_timeout: None,
    }
 }

--- a/crates/op-orchestrator/src/timeouts.rs
+++ b/crates/op-orchestrator/src/timeouts.rs
@ -0,0 +1,374 @@
+//! 超时档位表 —— S3b-1b Task A2。
+//!
+//! Port of:
+//! - `ai-runtime-config.ts` — `PROMPT_TIMEOUT_BUCKETS`,
+//!   `ORCHESTRATOR_TIMEOUT_PROFILES`, `SUB_AGENT_TIMEOUT_PROFILES`
+//! - `orchestrator-prompt-optimizer.ts` — `getOrchestratorTimeouts`,
+//!   `getBuiltinPlanningTimeouts`, `getSubAgentTimeouts`
+//! - `model-profiles.ts` — `applyProfileToTimeouts`
+
+#![allow(dead_code)]
+
+use crate::model_profile::ModelTier;
+use std::time::Duration;
+
+// ── 从 TS `ai-runtime-config.ts` 逐字搬来的毫秒常量 ────────────────────────
+
+/// 规划 prompt 桶阈值 1 —— `PROMPT_OPTIMIZER_LIMITS.longPromptCharThreshold`。
+/// `< SHORT_THRESH` → Short。
+const SHORT_THRESH: usize = 2200;
+
+/// 规划 prompt 桶阈值 2 —— `PROMPT_TIMEOUT_BUCKETS.mediumPromptMaxChars`。
+/// `SHORT_THRESH <= len < MEDIUM_THRESH` → Medium; `>= MEDIUM_THRESH` → Long。
+const MEDIUM_THRESH: usize = 4200;
+
+// Orchestrator timeout profiles (ms) — verbatim from ORCHESTRATOR_TIMEOUT_PROFILES
+const ORCH_SHORT_HARD_MS: u64 = 300_000;
+const ORCH_SHORT_NO_TEXT_MS: u64 = 150_000;
+const ORCH_SHORT_FIRST_TEXT_MS: u64 = 300_000;
+
+const ORCH_MEDIUM_HARD_MS: u64 = 420_000;
+const ORCH_MEDIUM_NO_TEXT_MS: u64 = 210_000;
+const ORCH_MEDIUM_FIRST_TEXT_MS: u64 = 420_000;
+
+const ORCH_LONG_HARD_MS: u64 = 600_000;
+const ORCH_LONG_NO_TEXT_MS: u64 = 300_000;
+const ORCH_LONG_FIRST_TEXT_MS: u64 = 600_000;
+
+// Sub-agent timeout profiles (ms) — verbatim from SUB_AGENT_TIMEOUT_PROFILES
+// Base: hardTimeoutMs=420_000, noTextTimeoutMs=210_000 (SUB_AGENT_TIMEOUT_BASE)
+const SA_SHORT_HARD_MS: u64 = 420_000;
+const SA_SHORT_NO_TEXT_MS: u64 = 210_000;
+const SA_SHORT_FIRST_TEXT_MS: u64 = 420_000;
+
+const SA_MEDIUM_HARD_MS: u64 = 600_000;
+const SA_MEDIUM_NO_TEXT_MS: u64 = 300_000;
+const SA_MEDIUM_FIRST_TEXT_MS: u64 = 600_000;
+
+const SA_LONG_HARD_MS: u64 = 900_000;
+const SA_LONG_NO_TEXT_MS: u64 = 480_000;
+const SA_LONG_FIRST_TEXT_MS: u64 = 900_000;
+
+// Basic-tier sub-agent clamps — from getSubAgentTimeouts:
+//   noTextTimeoutMs  = Math.min(noTextTimeoutMs,  45_000)
+//   firstTextTimeoutMs = Math.min(firstTextTimeoutMs, 75_000)
+const SA_BASIC_NO_TEXT_MAX_MS: u64 = 45_000;
+const SA_BASIC_FIRST_TEXT_MAX_MS: u64 = 75_000;
+
+// Builtin planning base timeouts — from getBuiltinPlanningTimeouts
+const BUILTIN_HARD_MS: u64 = 60_000;
+const BUILTIN_NO_TEXT_MS: u64 = 30_000;
+const BUILTIN_FIRST_TEXT_MS: u64 = 30_000;
+
+// Basic-tier builtin planning floors (Math.max raises):
+//   hardTimeoutMs       = Math.max(hard,       150_000)
+//   noTextTimeoutMs     = Math.max(no_text,     75_000)
+//   firstTextTimeoutMs  = Math.max(first_text,  75_000)
+const BUILTIN_BASIC_HARD_MIN_MS: u64 = 150_000;
+const BUILTIN_BASIC_NO_TEXT_MIN_MS: u64 = 75_000;
+const BUILTIN_BASIC_FIRST_TEXT_MIN_MS: u64 = 75_000;
+
+// ── 公开类型 ─────────────────────────────────────────────────────────────────
+
+/// 一次 LLM 调用的三路超时。
+///
+/// - `hard` — 整体硬截止
+/// - `no_text` — 从连接建立到第一个 text chunk 的超时(非 abort 中途超时)
+/// - `first_text` — 从第一个 text chunk 到真正"文字内容"出现的超时
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Timeouts {
+    pub hard: Duration,
+    pub no_text: Duration,
+    pub first_text: Duration,
+}
+
+impl Timeouts {
+    fn from_ms(hard_ms: u64, no_text_ms: u64, first_text_ms: u64) -> Self {
+        Self {
+            hard: Duration::from_millis(hard_ms),
+            no_text: Duration::from_millis(no_text_ms),
+            first_text: Duration::from_millis(first_text_ms),
+        }
+    }
+}
+
+/// prompt 长度对应的桶 —— 决定采用哪组超时档。
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum PromptBucket {
+    /// `< 2200` 字符
+    Short,
+    /// `2200 <= len < 4200` 字符
+    Medium,
+    /// `>= 4200` 字符
+    Long,
+}
+
+/// 根据 prompt 字节长度确定桶。
+///
+/// Port of the bucket selection in `getOrchestratorTimeouts` /
+/// `getSubAgentTimeouts`:
+/// ```ts
+/// if (len < PROMPT_OPTIMIZER_LIMITS.longPromptCharThreshold) Short
+/// else if (len < PROMPT_TIMEOUT_BUCKETS.mediumPromptMaxChars)  Medium
+/// else                                                          Long
+/// ```
+pub fn bucket_for(prompt_len: usize) -> PromptBucket {
+    if prompt_len < SHORT_THRESH {
+        PromptBucket::Short
+    } else if prompt_len < MEDIUM_THRESH {
+        PromptBucket::Medium
+    } else {
+        PromptBucket::Long
+    }
+}
+
+// ── 公开函数 ─────────────────────────────────────────────────────────────────
+
+/// 规划阶段超时 —— port of `getOrchestratorTimeouts(promptLength, model)`.
+///
+/// `tier` 只影响通过 `apply_profile_to_timeouts` 注入的 `timeout_multiplier`。
+/// 本函数不接 model_id,调用方提前 resolve 好 multiplier 传入即可。
+/// 注:TS 原函数最后调用 `applyProfileToTimeouts` —— 此处分两步
+/// (`orchestrator_timeouts` 取基值,再由调用方做 `apply_profile_to_timeouts`)
+/// 保持函数接口纯净。
+pub fn orchestrator_timeouts(prompt_len: usize) -> Timeouts {
+    match bucket_for(prompt_len) {
+        PromptBucket::Short => Timeouts::from_ms(
+            ORCH_SHORT_HARD_MS,
+            ORCH_SHORT_NO_TEXT_MS,
+            ORCH_SHORT_FIRST_TEXT_MS,
+        ),
+        PromptBucket::Medium => Timeouts::from_ms(
+            ORCH_MEDIUM_HARD_MS,
+            ORCH_MEDIUM_NO_TEXT_MS,
+            ORCH_MEDIUM_FIRST_TEXT_MS,
+        ),
+        PromptBucket::Long => Timeouts::from_ms(
+            ORCH_LONG_HARD_MS,
+            ORCH_LONG_NO_TEXT_MS,
+            ORCH_LONG_FIRST_TEXT_MS,
+        ),
+    }
+}
+
+/// sub-agent 超时 —— port of `getSubAgentTimeouts(promptLength, model)`.
+///
+/// 基础档位由 prompt 长度决定;Basic tier 对 `no_text` / `first_text`
+/// 施加上限(TS: `Math.min(..., 45_000)` / `Math.min(..., 75_000)`)。
+pub fn sub_agent_timeouts(prompt_len: usize, tier: ModelTier) -> Timeouts {
+    let mut t = match bucket_for(prompt_len) {
+        PromptBucket::Short => Timeouts::from_ms(
+            SA_SHORT_HARD_MS,
+            SA_SHORT_NO_TEXT_MS,
+            SA_SHORT_FIRST_TEXT_MS,
+        ),
+        PromptBucket::Medium => Timeouts::from_ms(
+            SA_MEDIUM_HARD_MS,
+            SA_MEDIUM_NO_TEXT_MS,
+            SA_MEDIUM_FIRST_TEXT_MS,
+        ),
+        PromptBucket::Long => {
+            Timeouts::from_ms(SA_LONG_HARD_MS, SA_LONG_NO_TEXT_MS, SA_LONG_FIRST_TEXT_MS)
+        }
+    };
+    if tier == ModelTier::Basic {
+        t.no_text = t
+            .no_text
+            .min(Duration::from_millis(SA_BASIC_NO_TEXT_MAX_MS));
+        t.first_text = t
+            .first_text
+            .min(Duration::from_millis(SA_BASIC_FIRST_TEXT_MAX_MS));
+    }
+    t
+}
+
+/// 内建规划超时(Compact/Minimal 模式的轻量规划调用)——
+/// port of `getBuiltinPlanningTimeouts(model)`.
+///
+/// 基础值固定;Basic tier 用 `Math.max` 提高下限
+/// (Basic 模型较慢,需要更多时间思考但要禁 thinking)。
+pub fn builtin_planning_timeouts(tier: ModelTier) -> Timeouts {
+    let mut t = Timeouts::from_ms(BUILTIN_HARD_MS, BUILTIN_NO_TEXT_MS, BUILTIN_FIRST_TEXT_MS);
+    if tier == ModelTier::Basic {
+        t.hard = t.hard.max(Duration::from_millis(BUILTIN_BASIC_HARD_MIN_MS));
+        t.no_text = t
+            .no_text
+            .max(Duration::from_millis(BUILTIN_BASIC_NO_TEXT_MIN_MS));
+        t.first_text = t
+            .first_text
+            .max(Duration::from_millis(BUILTIN_BASIC_FIRST_TEXT_MIN_MS));
+    }
+    t
+}
+
+/// 按模型的 `timeout_multiplier` 缩放三路超时 —— port of
+/// `applyProfileToTimeouts` (the timeout-scaling part only).
+///
+/// `multiplier == 1.0` 为 no-op。四舍五入到整毫秒(`Math.round` 对齐)。
+pub fn apply_profile_to_timeouts(t: Timeouts, multiplier: f64) -> Timeouts {
+    if (multiplier - 1.0).abs() < f64::EPSILON {
+        return t;
+    }
+    Timeouts {
+        hard: mul_ms(t.hard, multiplier),
+        no_text: mul_ms(t.no_text, multiplier),
+        first_text: mul_ms(t.first_text, multiplier),
+    }
+}
+
+fn mul_ms(d: Duration, m: f64) -> Duration {
+    let ms = (d.as_millis() as f64 * m).round() as u64;
+    Duration::from_millis(ms)
+}
+
+// ── 测试 ──────────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // ── bucket_for ────────────────────────────────────────────────────────────
+
+    #[test]
+    fn bucket_short_below_2200() {
+        assert_eq!(bucket_for(0), PromptBucket::Short);
+        assert_eq!(bucket_for(2199), PromptBucket::Short);
+    }
+
+    #[test]
+    fn bucket_medium_at_2200_below_4200() {
+        assert_eq!(bucket_for(2200), PromptBucket::Medium);
+        assert_eq!(bucket_for(4199), PromptBucket::Medium);
+    }
+
+    #[test]
+    fn bucket_long_at_and_above_4200() {
+        assert_eq!(bucket_for(4200), PromptBucket::Long);
+        assert_eq!(bucket_for(9999), PromptBucket::Long);
+    }
+
+    // ── orchestrator_timeouts ─────────────────────────────────────────────────
+
+    #[test]
+    fn orchestrator_timeouts_short_bucket() {
+        let t = orchestrator_timeouts(100);
+        assert_eq!(t.hard, Duration::from_millis(300_000));
+        assert_eq!(t.no_text, Duration::from_millis(150_000));
+        assert_eq!(t.first_text, Duration::from_millis(300_000));
+    }
+
+    #[test]
+    fn orchestrator_timeouts_medium_bucket() {
+        let t = orchestrator_timeouts(3000);
+        assert_eq!(t.hard, Duration::from_millis(420_000));
+        assert_eq!(t.no_text, Duration::from_millis(210_000));
+        assert_eq!(t.first_text, Duration::from_millis(420_000));
+    }
+
+    #[test]
+    fn orchestrator_timeouts_long_bucket() {
+        let t = orchestrator_timeouts(5000);
+        assert_eq!(t.hard, Duration::from_millis(600_000));
+        assert_eq!(t.no_text, Duration::from_millis(300_000));
+        assert_eq!(t.first_text, Duration::from_millis(600_000));
+    }
+
+    // ── sub_agent_timeouts ────────────────────────────────────────────────────
+
+    #[test]
+    fn sub_agent_timeouts_short_full_tier() {
+        let t = sub_agent_timeouts(100, ModelTier::Full);
+        assert_eq!(t.hard, Duration::from_millis(420_000));
+        assert_eq!(t.no_text, Duration::from_millis(210_000));
+        assert_eq!(t.first_text, Duration::from_millis(420_000));
+    }
+
+    #[test]
+    fn sub_agent_timeouts_medium_standard_tier() {
+        let t = sub_agent_timeouts(3000, ModelTier::Standard);
+        assert_eq!(t.hard, Duration::from_millis(600_000));
+        assert_eq!(t.no_text, Duration::from_millis(300_000));
+        assert_eq!(t.first_text, Duration::from_millis(600_000));
+    }
+
+    #[test]
+    fn sub_agent_timeouts_long_full_tier() {
+        let t = sub_agent_timeouts(5000, ModelTier::Full);
+        assert_eq!(t.hard, Duration::from_millis(900_000));
+        assert_eq!(t.no_text, Duration::from_millis(480_000));
+        assert_eq!(t.first_text, Duration::from_millis(900_000));
+    }
+
+    #[test]
+    fn sub_agent_timeouts_basic_tier_clamps_no_text_and_first_text() {
+        // Short bucket base: no_text=210_000, first_text=420_000
+        // Basic clamp: no_text ≤ 45_000, first_text ≤ 75_000
+        let t = sub_agent_timeouts(100, ModelTier::Basic);
+        assert_eq!(t.hard, Duration::from_millis(420_000)); // hard unclamped
+        assert_eq!(t.no_text, Duration::from_millis(45_000));
+        assert_eq!(t.first_text, Duration::from_millis(75_000));
+    }
+
+    #[test]
+    fn sub_agent_timeouts_basic_tier_long_bucket_still_clamped() {
+        // Long bucket base: no_text=480_000, first_text=900_000
+        let t = sub_agent_timeouts(5000, ModelTier::Basic);
+        assert_eq!(t.hard, Duration::from_millis(900_000)); // hard unclamped
+        assert_eq!(t.no_text, Duration::from_millis(45_000));
+        assert_eq!(t.first_text, Duration::from_millis(75_000));
+    }
+
+    // ── builtin_planning_timeouts ─────────────────────────────────────────────
+
+    #[test]
+    fn builtin_planning_full_tier_base_values() {
+        let t = builtin_planning_timeouts(ModelTier::Full);
+        assert_eq!(t.hard, Duration::from_millis(60_000));
+        assert_eq!(t.no_text, Duration::from_millis(30_000));
+        assert_eq!(t.first_text, Duration::from_millis(30_000));
+    }
+
+    #[test]
+    fn builtin_planning_standard_tier_base_values() {
+        let t = builtin_planning_timeouts(ModelTier::Standard);
+        assert_eq!(t.hard, Duration::from_millis(60_000));
+        assert_eq!(t.no_text, Duration::from_millis(30_000));
+        assert_eq!(t.first_text, Duration::from_millis(30_000));
+    }
+
+    #[test]
+    fn builtin_planning_basic_tier_floors_raised() {
+        // Basic: Math.max(hard, 150_000), Math.max(no_text, 75_000),
+        //        Math.max(first_text, 75_000)
+        let t = builtin_planning_timeouts(ModelTier::Basic);
+        assert_eq!(t.hard, Duration::from_millis(150_000));
+        assert_eq!(t.no_text, Duration::from_millis(75_000));
+        assert_eq!(t.first_text, Duration::from_millis(75_000));
+    }
+
+    // ── apply_profile_to_timeouts ─────────────────────────────────────────────
+
+    #[test]
+    fn apply_multiplier_scales_all_three() {
+        let base = Timeouts::from_ms(300_000, 150_000, 300_000);
+        let scaled = apply_profile_to_timeouts(base, 2.0);
+        assert_eq!(scaled.hard, Duration::from_millis(600_000));
+        assert_eq!(scaled.no_text, Duration::from_millis(300_000));
+        assert_eq!(scaled.first_text, Duration::from_millis(600_000));
+    }
+
+    #[test]
+    fn apply_multiplier_one_is_noop() {
+        let base = Timeouts::from_ms(300_000, 150_000, 300_000);
+        let scaled = apply_profile_to_timeouts(base.clone(), 1.0);
+        assert_eq!(scaled, base);
+    }
+
+    #[test]
+    fn apply_multiplier_rounds_to_ms() {
+        // 210_000 * 1.5 = 315_000 (exact)
+        let base = Timeouts::from_ms(210_000, 210_000, 210_000);
+        let scaled = apply_profile_to_timeouts(base, 1.5);
+        assert_eq!(scaled.hard, Duration::from_millis(315_000));
+    }
+}
--- a/crates/op-orchestrator/src/types.rs
+++ b/crates/op-orchestrator/src/types.rs
@ -41,6 +41,12 @@ pub struct CallRequest {
    pub provider: Option<String>,
    pub timeout: Duration,
    pub abort: AbortFlag,
+    /// 从请求开始到收到第一个文本 chunk 的超时;`None` 表示不设。
+    /// Port of `noTextTimeoutMs` in the TS timeout profiles.
+    pub no_text_timeout: Option<Duration>,
+    /// 从第一个文本 chunk 到"真正内容"出现的超时;`None` 表示不设。
+    /// Port of `firstTextTimeoutMs` in the TS timeout profiles.
+    pub first_text_timeout: Option<Duration>,
 }

 /// 流元素 —— 区分文本与思考,与 TS 的 text/thinking/error 三分