feat(backthink): 安全域穿透安静时段 + 长静默不拦心跳 + 启动即思考（裁决 1.8/07.4.2/07.4.3/07.4.4/07.4.5）
执行方案阶段 2 · P2-2（`DevDocs/audit/_plan.md`）。四条裁决： - 1.8/07.4.2：`think()` 不再走忽略 importance 的 `disturbAllowed`（已删），**接回 `Decide` 级联**； 安静时段只挡常规推送、**critical 安全域穿透**；`DecisionInput` 增 `ConclusionPush`/`Heartbeat`。 睡眠仍绝对不打扰（裁决只豁免安静时段，此点已在设计与报告写明）。 - 07.4.3：长静默（>2h）分支对**心跳类**放行（常规事件仍需 ≥high），刚聊过天 10min 内扣留不变。 - 07.4.4：`tickLoop` **启动立即先思考一轮** + 拟人化启动消息；启动轮与心跳**同门同限频** （重启风暴不会刷屏），ctx 已取消则跳过。 - 07.4.5：核对一致（`THINK_INTERVAL_SEC` 默认 900s 与 design/20 §0 权威口径一致）⇒ **代码零改动**。 验证：go build 绿；backthink 包 vet/测试绿；`-count=20`（抗 flake）+ `-race` 绿； linux/amd64 + darwin/arm64 交叉编译绿；gofmt 干净；文件 ≤500 行、无 >80 行函数。 变异验证（三条，均带还原证明）：① 安静时段改回"全拦" → 安全域穿透断言 FAIL（首次只删 critical 分支时仍绿——`critical >= high` 会从 high 分支漏过，故真实变异必须是"全拦"）； ② 删心跳豁免 → 长静默心跳断言 FAIL；③ 删启动思考 → 启动即思考断言 FAIL。 未验证：无真实模型凭据 ⇒ 启动轮问候与 `[我决定:…]` 解析仅经 mock + 真 loop 验证，端到端 LLM 未验。
This commit is contained in:
AskaEth committed
1 parent
a35bcb8aa9
commit
afef75f640
7 files changed
+468
-176
No files matched your search
@@ -6,9 +6,11 @@ import (
|
||||
"time"
|
||||
)
|
||||
|
||||
// Importance is the estimated severity of an event that woke a thinking pass. It
|
||||
// is the primary positive signal in the interrupt decision: a low-importance
|
||||
// heartbeat almost never interrupts, while a critical safety event normally does.
|
||||
// Importance is the estimated severity of an event that woke a thinking pass.
|
||||
// Under design/20 §2 P2 it no longer decides *whether* she wants to speak (her
|
||||
// conclusion does); it decides the hard-gate exemptions: a critical safety event
|
||||
// punches through work/busy and quiet hours (裁决 07.4.2), and a high event still
|
||||
// passes quiet hours (the pre-existing cascade semantics).
|
||||
type Importance int
|
||||
|
||||
const (
|
||||
@@ -18,7 +20,8 @@ const (
|
||||
ImportanceNormal
|
||||
// ImportanceHigh is a notable event (a reminder the user set firing).
|
||||
ImportanceHigh
|
||||
// ImportanceCritical is a safety-critical event that may override work/busy.
|
||||
// ImportanceCritical is a safety-critical event that punches through
|
||||
// work/busy and quiet hours (裁决 07.4.2 安全域豁免).
|
||||
ImportanceCritical
|
||||
)
|
||||
|
||||
@@ -38,14 +41,10 @@ func (i Importance) String() string {
|
||||
}
|
||||
}
|
||||
|
||||
// highBoredomThreshold is the boredom (0..1) above which a normal daytime event
|
||||
// is worth interrupting a bored, otherwise idle user. It is a const so no magic
|
||||
// number appears in the decision path.
|
||||
const highBoredomThreshold = 0.7
|
||||
|
||||
// longIdleThreshold is the gap after the owner's last message beyond which the
|
||||
// activity gate only lets a genuinely important event through (a long-away owner
|
||||
// should not be woken for trivia). Compile-time so no magic number appears.
|
||||
// activity gate only lets a genuinely important event through — except heartbeat
|
||||
// passes, which are never held by a long-silent gap (裁决 07.4.3). Compile-time
|
||||
// so no magic number appears.
|
||||
const longIdleThreshold = 2 * time.Hour
|
||||
|
||||
// UserState describes the user's current condition as observed from device
|
||||
@@ -62,7 +61,9 @@ type UserState struct {
|
||||
// Busy reports whether the user is marked busy (calendar / do-not-disturb).
|
||||
Busy bool
|
||||
// Boredom ranges 0..1. Higher boredom means the user is idle and more open
|
||||
// to a gentle interruption.
|
||||
// to a gentle interruption. Under design/20 §2 P2 the push intent is her own
|
||||
// conclusion, so this dial no longer gates the decision; it is kept as a
|
||||
// telemetry signal her judgment already absorbs.
|
||||
Boredom float64
|
||||
}
|
||||
|
||||
@@ -90,6 +91,12 @@ type DecisionInput struct {
|
||||
LastActivity time.Time
|
||||
// HasActivity reports whether LastActivity is meaningful.
|
||||
HasActivity bool
|
||||
// ConclusionPush 是她自己思考结论的推送意愿（design/20 §2 P2）：「想不想说」
|
||||
// 由她的结论决定，本决策只回答「此刻能不能说」。
|
||||
ConclusionPush bool
|
||||
// Heartbeat 标记本轮是否为心跳类触发（周期心跳或启动轮）。裁决 07.4.3：
|
||||
// 长期静默段不拦心跳——主人久未回应时，主动关心仍可照常发出。
|
||||
Heartbeat bool
|
||||
}
|
||||
|
||||
// Decision is the outcome of Decide.
|
||||
@@ -100,36 +107,37 @@ type Decision struct {
|
||||
Reason string
|
||||
}
|
||||
|
||||
// Decide applies the interrupt rule from design/12 §2 (and design/11 §6) with
|
||||
// "不打扰 (do not interrupt) as the first principle" as its bias, then softens it
|
||||
// with the activity-aware gate (design/12 §6). It is a pure method on Backthink
|
||||
// only so it can read the configured quiet-hours window; it does not touch state.
|
||||
// Decide 是生产路径的打扰决策（design/12 §2 + design/20 §2 P2 + 裁决 07.4.2/07.4.3）。
|
||||
// 「想不想说」交给她的结论（ConclusionPush），本方法只判「此刻能不能说」，评估顺序
|
||||
// 刻意 fail-closed（宁可不打扰）：
|
||||
//
|
||||
// baseDecide's evaluation order is deliberately fail-closed (prefer not
|
||||
// interrupting):
|
||||
// 0. 她的结论要求不推送 → 不推送。
|
||||
// 1. 用户在睡 → 绝不打扰。
|
||||
// 2. 上班/忙碌 → 仅 critical 安全事件可打扰。
|
||||
// 3. 安静时段（深夜窗口）→ 挡常规推送；critical 安全事件穿透（裁决 07.4.2，
|
||||
// 与 design/12 §6 策略五同口径），high 沿袭原级联口径照常放行。
|
||||
// 4. 其余时段 → 放行（是否值得说已由她的结论回答）。
|
||||
//
|
||||
// 1. A sleeping user is never interrupted.
|
||||
// 2. A working or busy user is interrupted only by a critical event.
|
||||
// 3. Inside the quiet-hours (deep-night) window only high/critical events
|
||||
// interrupt; low/normal events are deferred.
|
||||
// 4. In the daytime, high/critical events always interrupt; a normal event
|
||||
// interrupts only when the user is quite bored; a low event never does.
|
||||
//
|
||||
// A would-push decision is then run through the activity gate, which holds a push
|
||||
// when the owner just chatted (fresh conversation) or has been away a long while
|
||||
// and the event is not genuinely important.
|
||||
// 通过硬门后再过活动软门：刚聊过天或主人久未回应时酌情扣下；心跳类触发在
|
||||
// 长期静默段不拦（裁决 07.4.3，主动关心照常发出）。频率门（PushGate）在推送器
|
||||
// 内部单独生效，不在此处。
|
||||
func (b *Backthink) Decide(in DecisionInput) Decision {
|
||||
d := b.baseDecide(in)
|
||||
if !d.ShouldPush {
|
||||
if !in.ConclusionPush {
|
||||
return Decision{ShouldPush: false, Reason: "conclusion requested no push"}
|
||||
}
|
||||
if d := b.baseDecide(in); !d.ShouldPush {
|
||||
return d
|
||||
}
|
||||
if !b.activityAllowed(in.Import, in.LastActivity, in.HasActivity, in.At) {
|
||||
return Decision{ShouldPush: false, Reason: "proactive push held: " + d.Reason}
|
||||
if held, why := b.activityHeld(in); held {
|
||||
return Decision{ShouldPush: false, Reason: "proactive push held: " + why}
|
||||
}
|
||||
return d
|
||||
return Decision{ShouldPush: true, Reason: "proactive push approved"}
|
||||
}
|
||||
|
||||
// baseDecide is the do-not-interrupt cascade of design/12 §2.
|
||||
// baseDecide 是不打扰硬门级联（design/12 §2）。critical 安全事件走安全域分支：
|
||||
// 工作忙碌与安静时段都放行（裁决 07.4.2 安全域豁免）；睡眠仍是绝对不打扰。
|
||||
// 白天是否「值得说」不再由重要性/无聊度判定——那是她的结论（design/20 §2 P2）
|
||||
// 已经回答的问题。
|
||||
func (b *Backthink) baseDecide(in DecisionInput) Decision {
|
||||
s := in.UserState
|
||||
if s.Sleeping {
|
||||
@@ -139,78 +147,49 @@ func (b *Backthink) baseDecide(in DecisionInput) Decision {
|
||||
return Decision{ShouldPush: false, Reason: "user is working/busy; not interrupting"}
|
||||
}
|
||||
if b.inQuietHours(in.At) {
|
||||
if in.Import >= ImportanceCritical {
|
||||
return Decision{ShouldPush: true, Reason: "quiet hours but critical safety event bypasses"}
|
||||
}
|
||||
if in.Import >= ImportanceHigh {
|
||||
return Decision{ShouldPush: true, Reason: "quiet hours but event is important"}
|
||||
}
|
||||
return Decision{ShouldPush: false, Reason: "quiet hours; not interrupting"}
|
||||
}
|
||||
switch in.Import {
|
||||
case ImportanceCritical, ImportanceHigh:
|
||||
return Decision{ShouldPush: true, Reason: "high/critical importance during daytime"}
|
||||
case ImportanceNormal:
|
||||
if s.Boredom >= highBoredomThreshold {
|
||||
return Decision{ShouldPush: true, Reason: "normal importance and user is bored"}
|
||||
}
|
||||
return Decision{ShouldPush: false, Reason: "normal importance; not worth interrupting"}
|
||||
default:
|
||||
return Decision{ShouldPush: false, Reason: "low importance; not interrupting"}
|
||||
}
|
||||
return Decision{ShouldPush: true, Reason: "worth interrupting"}
|
||||
}
|
||||
|
||||
// activityAllowed is the activity-aware soft gate (design/12 §6). It widens the
|
||||
// do-not-interrupt principle to the conversation's own rhythm:
|
||||
// activityHeld 是活动感知软门（design/12 §6 + 裁决 07.4.3）。它把「不打扰」延伸到
|
||||
// 对话自身的节奏，返回是否扣下与简短理由：
|
||||
//
|
||||
// - Within RecentChatCooldown of the owner's last message, the conversation is
|
||||
// fresh — a non-critical push is held so she is not interrupted mid-chat.
|
||||
// - After longIdleThreshold without a message, the owner has been away — only a
|
||||
// genuinely important event is worth waking them for.
|
||||
// - Anywhere in between, and when activity is unknown, the decision is not
|
||||
// softened (fall through).
|
||||
// - 最近 RecentChatCooldown 内主人刚说过话 → 对话正热，非 critical 扣下，
|
||||
// 不在聊天中途插话；
|
||||
// - 超过 longIdleThreshold 未说话 → 主人久未回应，仅真正重要的事件值得唤醒；
|
||||
// 心跳类触发（含启动轮）例外照常放行（裁决 07.4.3：长期静默段不拦心跳，
|
||||
// 主动关心仍可发出）；
|
||||
// - 其间与活动状态未知 → 不软化，照硬门结论通过。
|
||||
//
|
||||
// Activity is a soft signal: the hard do-not-disturb gate is separate, and the
|
||||
// frequency gate (PushGate) applies separately inside the pusher.
|
||||
func (b *Backthink) activityAllowed(imp Importance, lastActivity time.Time, hasActivity bool, at time.Time) bool {
|
||||
if !hasActivity {
|
||||
return true
|
||||
// 活动是软信号：硬门（baseDecide）独立生效，频率门（PushGate）在推送器内部生效。
|
||||
func (b *Backthink) activityHeld(in DecisionInput) (bool, string) {
|
||||
if !in.HasActivity {
|
||||
return false, ""
|
||||
}
|
||||
since := at.Sub(lastActivity)
|
||||
since := in.At.Sub(in.LastActivity)
|
||||
if since <= b.cfg.RecentChatCooldown {
|
||||
return imp >= ImportanceCritical
|
||||
if in.Import >= ImportanceCritical {
|
||||
return false, ""
|
||||
}
|
||||
return true, "owner just chatted"
|
||||
}
|
||||
if since >= longIdleThreshold {
|
||||
return imp >= ImportanceHigh
|
||||
}
|
||||
return true
|
||||
}
|
||||
|
||||
// disturbAllowed is the hard "do not disturb" safety gate for an autonomous push
|
||||
// (design/12 §2 first principle). Unlike Decide it deliberately ignores trigger
|
||||
// importance — which design/20 P2 replaces with her own judgment — and only blocks
|
||||
// when it is genuinely inappropriate to interrupt: the user is asleep, working or
|
||||
// busy, or it is inside the quiet-hours window. The frequency gate (PushGate)
|
||||
// applies separately inside the pusher.
|
||||
func (b *Backthink) disturbAllowed(at time.Time, s UserState) bool {
|
||||
if s.Sleeping || s.Working || s.Busy {
|
||||
return false
|
||||
}
|
||||
return !b.inQuietHours(at)
|
||||
}
|
||||
|
||||
// interruptReason is the short human-readable note recorded/logged when no push
|
||||
// goes out (or as the pre-push intent). It keeps the panel self-explanatory:
|
||||
// either she concluded no push, the do-not-disturb gate refused a push she
|
||||
// wanted, or the activity gate held it.
|
||||
func (b *Backthink) interruptReason(push bool, at time.Time, s UserState, imp Importance, lastActivity time.Time, hasActivity bool) string {
|
||||
switch {
|
||||
case !push:
|
||||
return "conclusion requested no push"
|
||||
case !b.disturbAllowed(at, s):
|
||||
return "proactive push blocked by do-not-disturb"
|
||||
case !b.activityAllowed(imp, lastActivity, hasActivity, at):
|
||||
return "proactive push held by activity gate"
|
||||
default:
|
||||
return "proactive push approved"
|
||||
if in.Heartbeat {
|
||||
return false, ""
|
||||
}
|
||||
if in.Import >= ImportanceHigh {
|
||||
return false, ""
|
||||
}
|
||||
return true, "owner away a long while"
|
||||
}
|
||||
return false, ""
|
||||
}
|
||||
|
||||
// inQuietHours reports whether at falls inside the configured do-not-disturb
|
||||
|
||||
Reference in new issue
Block a user