@bushizhanghan/dsh-context-budget 0.1.0 → 0.1.1

package/README.md

@@ -83,16 +83,19 @@ skill catalog: 59 skills, 5022 est tok per turn
 - **`estimateTokens` 是估计值，报告里一律标 `est`。** 这个插件要的是**比例**——同一除数在比值里抵消。需要精确数字时用 `request/context` 的 `contextWindow` 做分母。
 - **懒加载规则故意保守**：只标记"本轮从未调用且描述 ≥ 60 token"的技能。用过一次的技能永不标记；没有使用信号的技能不猜测它已死（它可能在别的会话里被用）。
 - **未识别的 `source.kind` 归入 `other` 而不是丢弃**，否则未来 DSH 新增注入面会让审计数字悄悄变小而没人发现。
+- **轮次靠 `turn/start` 累加，不能只读 `data.turn`。** `user/message` 和 `request/header` 根本不带 `turn` 字段（只有 `system/message`、`tool/*`、`turn/start` 带），只读它会把全部注入块和全部请求塞进第 0 轮——173 轮的会话会报成 37 轮。
+- **工具 schema 按请求累加，不按轮覆盖。** 同一轮里每个请求都会重发整套 schema，只留最后一份会把成本算成一份。
 
 ## 测试
 
 ```sh
-node --test test/
+node --test                              # 19 个单测 + 集成测试
 ```
 
 `test/engine.test.mjs` 一条规则一个测试（token 估计、CJK、分类映射、签名、重复判定、技能治理、工具分组、空会话不除零）。
 `test/integration.test.mjs` 用假 `ctx = { tools: { register } }` 启动插件，驱动全部 4 个工具打真实会话日志，
-并断言反向情况：不存在的会话/根目录必须返回 `ok: false`。
+并断言反向情况：不存在的会话/根目录必须返回 `ok: false`。集成测试会挑一个**多于一轮**的会话，
+因为重复注入是跨轮比较出来的，单轮会话没有重复可测。
 
 ## License
 

package/lib/engine.js

@@ -192,9 +192,21 @@ export function buildTurns(events) {
   // system/message carries the assembled system prompt, but only inside its
   // first turn is it new; the same text is resent every turn after. We record
   // it per turn with a signature so the repetition analysis can prove it.
+  // `user/message` and `request/header` never carry data.turn — only
+  // system/message, tool/* and turn/start do. Reading data.turn alone dropped
+  // every injected block and every request into synthetic turn 0, so a 173-turn
+  // session reported 37 turns and all 72 requests landed in one bucket.
+  let current = 0
   for (const ev of events) {
     const t = String(ev.type || '')
-    const turn = turnOf(ev)
+    const explicit = turnOf(ev)
+    if (t === 'turn/start') {
+      if (explicit != null) current = explicit
+      ensure(current)
+      continue
+    }
+    if (explicit != null) current = explicit
+    const turn = explicit ?? current
     if (t === 'system/message') {
       const text = contentText(ev.data?.message?.content)
       push(turn ?? 0, {
@@ -232,7 +244,10 @@ export function buildTurns(events) {
         // the model was actually shown.
         t.toolNames = tools.map((x) => x.name)
         t.toolChars = JSON.stringify(tools).length
-        t.toolTokens = estimateTokens(JSON.stringify(tools))
+        // Every request in the turn re-sends the whole schema set, so the cost
+        // is per request, not per turn; overwriting counted one copy and hid
+        // the rest (72 requests, one block).
+        t.toolTokens = (t.toolTokens || 0) + estimateTokens(JSON.stringify(tools))
         t.toolSignature = signature(tools.map((x) => `${x.name}:${JSON.stringify(x.parameters ?? {})}`).join('|'))
         t.reason = ev.data.reason ?? null
       }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bushizhanghan/dsh-context-budget",
-  "version": "0.1.0",
+  "version": "0.1.1",
   "description": "Audit where each turn's context budget goes: per-turn composition (AGENTS.md / skill catalog / tool schemas / runtime context / history), repeated injection of byte-identical content (pure waste), skill-catalog governance, and tool-schema inflation. Deterministic, no LLM. 上下文预算审计：每轮注入构成、重复注入浪费、技能目录治理、工具 schema 膨胀。",
   "type": "module",
   "main": "lib/index.js",
@@ -46,6 +46,6 @@
     }
   },
   "scripts": {
-    "test": "node --test test/"
+    "test": "node --test"
   }
 }