dsh-layered-memory 0.8.3 → 0.8.5

package/README.en.md

@@ -10,7 +10,7 @@
 [简体中文](README.md) · [Latest release](https://github.com/JunNanLYS/dsh-layered-memory/releases/latest) · [Report issues](https://github.com/JunNanLYS/dsh-layered-memory/issues)
 
 [![npm version](https://img.shields.io/npm/v/dsh-layered-memory?color=6f83ff&style=flat-square&label=npm)](https://www.npmjs.com/package/dsh-layered-memory)
-[![DSH 0.1.0-rc.8](https://img.shields.io/badge/DSH-0.1.0--rc.8-8b5cf6?style=flat-square)](https://github.com/deepseek-ai/deepseek-harness)
+[![DSH 0.1.1-rc.2](https://img.shields.io/badge/DSH-0.1.1--rc.2-8b5cf6?style=flat-square)](https://github.com/deepseek-ai/deepseek-harness)
 [![MIT License](https://img.shields.io/badge/license-MIT-536990?style=flat-square)](LICENSE)
 
 </div>
@@ -21,7 +21,7 @@ Requires Node ≥ 22.16. Two invocation styles — the `npx` prefix can replace
 any command below:
 
 ```bash
-# Option 1: run the official CLI directly via npx (no pre-installed dsh; version can be pinned, e.g. dsh-layered-memory@0.8.2)
+# Option 1: run the official CLI directly via npx (no pre-installed dsh; version can be pinned, e.g. dsh-layered-memory@0.8.4)
 npx -y @deepseek-ai/dsh plugin --profile web add dsh-layered-memory
 
 # Option 2: with the dsh CLI installed (dsh is a pnpm forwarder; npm i -g pnpm first if missing)
@@ -129,6 +129,15 @@ trajectory view):
 - **Control**: the pill next to the mode selector in the input bar (`Memory · Auto`);
   clicking opens a macOS-style sliding picker above — release to snap to the nearest
   mode; adapts to light/dark themes;
+- The lower half of the popover is a **per-session info area**: recall hits
+  (hit/searched turns plus cumulative items), batching progress (this session's
+  slice x/effective threshold; the off mode shows parked slices instead), memories
+  produced for this session, and session message count — plus status lines for
+  anomalies (storage degraded / vector search unavailable) and a global summary
+  (pending distill count, last distill time). Data comes from the
+  `dsh-memory/session-stats` endpoint (in-memory registries + an indexed COUNT,
+  zero file I/O), adaptively polled while open (2s busy / 5s idle) and stopped on
+  close;
 - Each session's choice is persisted by sessionId to `session-modes.json`, surviving
   restarts/session restore; stacks with the global switches (global is the master gate);
   L2/L3 are fully family-isolated — content never leaks across families.
@@ -144,39 +153,50 @@ trajectory view):
 
 ## Measured Comparison (DSH-MemBench: Automated Benchmark)
 
-Screenshots show what the plugin looks like — this section answers "**what does enabling it actually buy you?**" with measured numbers from an **automated benchmark** ([`bench/`](./bench/), one command to reproduce). Method: the same scenario bank with verbatim-identical inputs runs in **Group A (memory on)** and **Group B (memory off)**, 3 repetitions each, merged; environment DeepSeek official `deepseek-v4-flash`, plugin 0.8.0, Windows; taxonomy adapted from [LongMemEval](https://github.com/xiaowu0162/longmemeval) / [LoCoMo](https://snap-research.github.io/locomo/) / [AMB](https://github.com/vectorize-io/agent-memory-benchmark).
+Screenshots show what the plugin looks like — this section answers "**what does enabling it actually buy you?**" with measured numbers from an **automated benchmark** ([`bench/`](./bench/), one command to reproduce). Method: the same scenario bank with verbatim-identical inputs runs in **Group A (memory on)** with 3 merged repetitions and **Group B (memory off)** with 1 repetition (a memory-off long task burns multiples of the tokens per scenario — a cost guardrail); the dialog track now runs Group A only (memory-off probes in independent sessions cannot succeed, so the control carries no information — retired). Dialog-track environment: DeepSeek official `deepseek-v4-flash`, plugin 0.8.5 (judge same-source as tested; every answer archived for manual audit), Windows; taxonomy adapted from [LongMemEval](https://github.com/xiaowu0162/longmemeval) / [LoCoMo](https://snap-research.github.io/locomo/) / [AMB](https://github.com/vectorize-io/agent-memory-benchmark), with the extended probe types and lifecycle track informed by [MemoryAgentBench](https://arxiv.org/abs/2507.05257) / [GoodAI LTM](https://github.com/GoodAI/goodai-ltm-benchmark) / BEAM.
 
-### Dialog track (15 scenarios × 6 probe types × 3 reps = 270 questions/group): does it remember correctly
+> The dialog track below is the **fresh 0.8.5 baseline** (fixed plugin + corrected judging criteria); the workflow-track numbers remain the archived 0.8.3 run (the bank has since grown to 8 scenarios with a prospective-memory addition — re-run pending).
+
+### Dialog track (20 scenarios × 10 probe types × 3 reps = 420 questions): does it remember correctly
+
+> 0.8.5 baseline (Group A data; the dialog-track B arm is retired — Group A only).
 
 <p align="center">
   <img src="./assets/readme/bench-dialog.svg" width="100%"
-       alt="DSH-MemBench dialog track, Group A vs Group B bar chart: overall accuracy A (memory on) 92.6% (250/270) vs B (memory off) 17.8% (48/270); per probe type, 45 each — extraction A 45/45 vs B 3/45, multi-hop A 45/45 vs B 0/45, temporal A 43/45 vs B 0/45, knowledge updates A 31/45 vs B 0/45, scene recall A 41/45 vs B 0/45, abstention both 45/45 with 0 fabricated">
+       alt="DSH-MemBench dialog track accuracy chart (Group A, memory on): overall accuracy 95.2% (400/420); six core probe types, 60 questions each — extraction 58/60, multi-hop 60/60, temporal 56/60, updates 55/60, scene recall 52/60, abstention 60/60 with 0 fabricated; four extended probe types, 15 each — accretive completion 15/15, update chains 15/15, event ordering 14/15, paraphrase 15/15">
 </p>
 
-**Dual-channel recall** (Group A): passive injection hit rate **75.1%** (the answer's key points appear in the recall injection, 169/225); most of the rest the model recovered by **actively calling the memory tools** — 84 questions with active queries, **60 rescued by tools**. The end-to-end 92.6% is the composite of both channels plus model utilization. With the memory store accumulating across scenarios for the whole run, 144 probe injections carried other scenarios' memories (honestly counted) — yet overall accuracy held at 92.6%: interference resistance under a growing store, measured.
+**Dual-channel recall** (Group A): passive injection hit rate **78.1%** (the answer's key points appear in the recall injection, 281/360); most of the rest the model recovered by **actively calling the memory tools** — 106 questions with active queries, **75 rescued by tools**. The end-to-end 95.2% is the composite of both channels plus model utilization. With the memory store accumulating across scenarios for the whole run, 295 probe injections carried other scenarios' memories (honestly counted) — yet accuracy actually *rose* from 92.8% (early, small store) to 97.7% (late, largest store), and offline flooding with 600 extra synthetic records moved retrieval recall@5 by only −2.8pp: interference resistance under a growing store, measured.
 
-### Workflow track (4 scenarios × 3 reps, real tool sandbox): does it do it right, and cheaper
+**Layered weaknesses**: offline retrieval metrics (recall@5, controlled replay) total 73.3%, with event ordering at 0% and scene recall at 50% — end-to-end still 93%+ thanks to model robustness over adjacent injected memories. **Efficiency triangle** (the cost of memory): injections add no latency (injected turns respond 210ms *faster* on average), recall text is ~10.3% of per-turn input, and the whole distillation pipeline costs ≈2727 input / 240 output tokens per captured message (1172 calls, 0 failures).
+
+### Workflow track (archived 0.8.3 · 7-scenario edition · Group A ×3 / Group B ×1, real tool sandbox): does it do it right, and cheaper
 
 <p align="center">
   <img src="./assets/readme/bench-workflow.svg" width="100%"
-       alt="DSH-MemBench workflow track, Group A vs Group B: task completion A 24/33 (72.7%) vs B 11/33 (33.3%); cost comparison (Group B as the full-bar baseline) — steps 125 vs 186 (B +49%), tool calls 184 vs 296 (B +61%), input tokens 1.30M vs 1.86M (B +43%); asks-user-for-help A 0 vs B 3; login scenario input tokens A 241k vs B 453k (+88%)">
+       alt="DSH-MemBench workflow track, Group A vs Group B: probe-phase completion A 59/69 (85.5%) vs B 10/23 (43.5%); cost comparison (Group B as the full-bar baseline, per-scenario means) — steps 24.3 vs 41.4 (B +70%), tool calls 37.7 vs 62.1 (B +65%), input tokens 266k vs 1.81M (B 6.8×); style-convention scenario probes A 12/12 vs B 0/4; long-task input tokens per scenario A 266k vs B 1.81M">
 </p>
 
-**Login scenario close-up** (credentials exist only in memory; the site is a local service with unforgeable tokens): Group A completed all three runs **in a single turn each** (6/6, 241k input); Group B had to **ask the user for credentials every time** (3 asks, double the turns) and still finished only 5/6, at 453k input — **+88%**. This is one of memory's core values: **what it saves is not task difficulty, but pointless round-trips and re-teaching**.
+**Probe-phase completion 85.5% vs 43.5% (+42pp)**: both groups have live context during teach/change phases — the probe phase (continuation task in a fresh session) is the pure memory window. Group A scored a perfect 12/12 on all three new probe archetypes (workflow knowledge update / twin-runbook disambiguation / style-convention continuity), consistent across all three reps; Group B scored **0/4** on style-convention probes (naming/structure/thousands-separator/footer conventions exist only in memory — they cannot be explored out of the sandbox), while on the workflow-update scenario it can reverse-engineer the procedure by reading the script (discrimination limited by sandbox affordances, honestly noted).
+
+**Long-task cost: Group B burns 6.8× Group A's input tokens per scenario** (1.81M vs 266k) — without memory the agent advances by re-exploring, and under a high reasoning effort it even builds its own projects to probe what a one-line script convention would have done; output tokens 3× (46.2k vs 15.4k), steps +70%. This is memory's core value: **what it saves is not task difficulty, but pointless round-trips and re-exploration**.
 
 ### Methodology & reproduction
 
 ```bash
-node bench/harness/run.mjs --arm A --repeats 3 --provider deepseek-official --model deepseek-v4-flash
-node bench/harness/run.mjs --arm B --repeats 3 --provider deepseek-official --model deepseek-v4-flash   # dialog track
-node bench/harness/run.mjs --track workflow --arm A/B --repeats 3 ...                                  # workflow track
+node bench/harness/run.mjs --arm A --repeats 3 --provider deepseek-official --model deepseek-v4-flash   # dialog track (Group A only)
+node bench/harness/run.mjs --track workflow --arm AB --repeats 3 ...                                  # workflow track (A/B arms in parallel)
+node bench/harness/run.mjs --track lifecycle --arm A ...                                              # lifecycle track (gating/off/rebuild/forget)
 node bench/harness/report.mjs --latest [dialog|workflow]                                               # aggregate report
+node bench/harness/retrieval-metrics.mjs <runDir> --flood 200,600                                     # retrieval metrics + flooding curve
 ```
 
-- Scoring: programmatic `contains-all` plus an LLM judge against key points (every answer and verdict is preserved in `result.json` for human audit); workflow completion is verified programmatically from produced files and their contents;
-- Metrics come from provider-reported usage (input with cache-hit split) and session-event folding; the **steady-state cache rate** excludes each session's first request (A 88.7% vs B 85.4% — memory injection does not hurt caching);
-- Regression use: run before/after a plugin change and diff with `compare.mjs` (environment header check + Group-B control-drift warning);
-- Limitations (stated honestly): single machine, 3 merged runs; the judge model is the same as the tested model; the scenario bank is author-built (biased toward memory-advantage scenarios — reproduce it yourself); the tool audit flags out-of-sandbox access (agents occasionally probed the user home dir in tests; this benchmark's answers never exist in the real memory store, so the numbers are unaffected).
+- Scoring: programmatic `contains-all` plus an LLM judge against key points (every answer and verdict is preserved in `result.json` for human audit); for stale-bearing probes (updates/update-chains/forget) an old value only fails when stated *as the current answer*, and abstention probes allow citing real adjacent facts while denying the asked point; workflow completion is verified programmatically from produced files and their contents (four check kinds: positive / forbidden-word / must-not-exist / exists);
+- Metric surface: beyond the per-type accuracy table (6 core + 4 extended types), reports automatically include **offline retrieval metrics** (recall@5 / injection precision / stale leakage), the **efficiency triangle** (injection latency differential / injection share / distillation accounting per captured message), **scale-position analysis** (accuracy & contamination vs store growth), and the lifecycle-track section (family-gating matrix / off-mode dual assertions / rebuild fidelity / forget requests);
+- Live progress: running the benchmark auto-starts a local progress panel and opens the browser (`--no-panel` to disable) — per-arm scenario/phase/message-level progress, heartbeat & activity freshness (distinguishes "stuck" from "process died"), and cumulative cost as it accrues;
+- Metrics come from provider-reported usage (input with cache-hit split) and session-event folding; the steady-state cache rate excludes each session's first request (0.8.5 baseline: 89.1% — memory injection does not hurt caching);
+- Regression use: run before/after a plugin change and diff with `compare.mjs` (environment header check including git SHA + Group-B control-drift warning + retrieval-metric comparison);
+- Limitations (stated honestly): single machine; Group A ×3 merged, Group B ×1 (cost guardrail — noisier); judge vs tested model: same model in the 0.8.5 dialog baseline, heterogeneous in the archived workflow run (glm-5.3 judging v4-flash); the scenario bank is author-built (biased toward memory-advantage scenarios — reproduce it yourself); sandbox-file affordances partially leak procedures (Group B can reverse-engineer by reading scripts — discrimination limits honestly noted); dual-tier tool audit (strict violation voids the scenario / loose heuristic flags only), with 0 violations measured on both sides.
 
 Full reports and per-question data: [`bench/baseline/`](./bench/baseline/).
 
@@ -235,11 +255,12 @@ the bundle layer appends and causes `duplicate loader entry id` startup failure)
 | `embedding.proxy` | `''` | Three-state download proxy: `''` (default) = auto-detect proxy env vars (`HTTPS_PROXY`/`ALL_PROXY` etc., honoring `NO_PROXY`); `none` = disable, always direct; any other value = proxy URL (e.g. `http://127.0.0.1:7890`). Direct connections to the mirror are intermittently unreachable on some networks (connect timeouts and poisoned bytes have both been observed) — keep the default auto-detection on machines with a proxy |
 | `llm.provider/model` | empty | Static distillation route (deployment pin): when **both** fields are set the route is locked, outranking the settings-page selection and the default model (deployments can force distillation onto a specific route); when empty the route follows "settings-page selection → default model". At runtime, switch among **configured providers** (including custom ones added in dsh Settings → Models) via the "distillation model" picker in Settings → Memory → Overview — effective immediately, no restart needed |
 | `llm.maxTokens` | `65536` | Fallback output cap for non-layered calls. Each distillation stage has its own budget (extraction 16k / dedup 8k / L2 32k / L3 16k; auto ×4 when the reasoning effort is high/xhigh/max, so thinking can't starve the text budget); the per-layer budgets are runtime-adjustable in Settings → Memory → Overview → distillation parameters (empty/0 = built-in defaults) |
-| `llm.reasoningEffort` | empty | Distillation reasoning effort: empty = **auto** (resolved from model capability: the model's default tier, else `high`); an explicit value (`off`/`none`/`low`/`medium`/`high`/`xhigh`/`max`) is only sent when the model declares support — effort vocabularies differ across providers (deepseek accepts `off`, OpenAI-style APIs use `none`, models that declare no tiers get nothing), and unsupported tiers degrade to not-sending with a one-time warning; output budgets auto-×4 at high/xhigh/max. Switchable at runtime in Settings → Memory → Overview; the selectable tiers follow the current model live |
+| `llm.reasoningEffort` | empty | Distillation reasoning effort: empty = **auto** (resolved from model capability: the model's default tier, else `high`); an explicit value (`off`/`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`) is only sent when the model declares support — effort vocabularies differ across providers (deepseek accepts `off`, OpenAI-style APIs use `none`, models that declare no tiers get nothing), and unsupported tiers degrade to not-sending with a one-time warning; output budgets auto-×4 at high/xhigh/max. Switchable at runtime in Settings → Memory → Overview; the selectable tiers follow the current model live |
 | `llm.temperature` | `0.3` | Distillation temperature |
 | `llm.maxInputChars` | `700000` | Input character budget per distillation call (over-budget L1 inputs are chunked automatically); runtime-adjustable in Settings → distillation parameters → input budget (empty/0 = follow this value) |
 | `llm.timeoutMs` | `120000` | Per-call distillation timeout (ms) |
 | `tools` | `true` | Whether to register model-callable memory tools |
+| `benchControl` | `false` | Register the in-process bench control service (rebuild trigger / session-mode setting / distillation usage snapshot — used by the benchmark's lifecycle track). Off by default — zero surface in production deployments; do not enable casually |
 
 ## Storage Layout
 

package/README.md

@@ -10,7 +10,7 @@
 [English](README.en.md) · [最新发行版](https://github.com/JunNanLYS/dsh-layered-memory/releases/latest) · [反馈问题](https://github.com/JunNanLYS/dsh-layered-memory/issues)
 
 [![npm version](https://img.shields.io/npm/v/dsh-layered-memory?color=6f83ff&style=flat-square&label=npm)](https://www.npmjs.com/package/dsh-layered-memory)
-[![DSH 0.1.0-rc.8](https://img.shields.io/badge/DSH-0.1.0--rc.8-8b5cf6?style=flat-square)](https://github.com/deepseek-ai/deepseek-harness)
+[![DSH 0.1.1-rc.2](https://img.shields.io/badge/DSH-0.1.1--rc.2-8b5cf6?style=flat-square)](https://github.com/deepseek-ai/deepseek-harness)
 [![MIT License](https://img.shields.io/badge/license-MIT-536990?style=flat-square)](LICENSE)
 
 </div>
@@ -20,7 +20,7 @@
 需要 Node ≥ 22.16。两种调用方式任选（`npx` 前缀可替换下面任何 `dsh` 命令）：
 
 ```bash
-# 方式一：npx 直接跑官方 CLI（无需预装 dsh；可 pin 版本，如 dsh-layered-memory@0.8.2）
+# 方式一：npx 直接跑官方 CLI（无需预装 dsh；可 pin 版本，如 dsh-layered-memory@0.8.4）
 npx -y @deepseek-ai/dsh plugin --profile web add dsh-layered-memory
 
 # 方式二：已装 dsh CLI（dsh 是 pnpm 转发器，未装 pnpm 时先 npm i -g pnpm）
@@ -110,6 +110,11 @@ npx tsc src/smoke.ts --outDir dist-smoke --module nodenext --moduleResolution no
 </p>
 
 - **控件**：输入栏内、模式选择器右侧的 pill（`记忆·自动`），点击在上方浮出档位滑块深浅主题自适应；
+- 悬浮板下半部是**会话信息区**：召回命中（命中/检索轮次与累计条数）、攒批进度
+  （本会话切片 x/生效阈值；关闭档显示挂起切片数）、本会话产出记忆条数、会话消息数，
+  外加异常状态行（存储降级 / 向量检索不可用）与全局摘要（待蒸馏条数、上次蒸馏时间）；
+  数据走 `dsh-memory/session-stats` 端点（纯内存注册表 + 索引 COUNT，零文件 I/O），
+  打开期间自适应轮询（忙 2s / 静 5s），关闭即停；
 - 每会话的选择按 sessionId 持久化到 `session-modes.json`，重启/恢复会话不丢；
   与全局开关叠加（全局是总闸）；L2/L3 完全分类，分类内容不渗透。
 
@@ -124,39 +129,50 @@ npx tsc src/smoke.ts --outDir dist-smoke --module nodenext --moduleResolution no
 
 ## 实测对比（DSH-MemBench：自动化基准）
 
-图文回答"长什么样"，这一节用**自动化基准**的实测数字回答"**开了到底有什么用**"（[`bench/`](./bench/)，一条命令可复现）。方法：同场景库、逐字相同输入，**A 组（记忆开）** 与 **B 组（记忆关）** 各跑 3 次取合并值；环境 DeepSeek 官方 `deepseek-v4-flash`、插件 0.8.0、Windows；题型设计借鉴 [LongMemEval](https://github.com/xiaowu0162/longmemeval) / [LoCoMo](https://snap-research.github.io/locomo/) / [AMB](https://github.com/vectorize-io/agent-memory-benchmark)。
+图文回答"长什么样"，这一节用**自动化基准**的实测数字回答"**开了到底有什么用**"（[`bench/`](./bench/)，一条命令可复现）。方法：同场景库、逐字相同输入，**A 组（记忆开）跑 3 次取合并值，B 组（记忆关）跑 1 次**（无记忆的长任务每场景要吞数倍 token，成本护栏）；对话赛道只跑 A 组（B 组会话独立无记忆必然失败，对照无信息量，已下线）。对话赛道环境：DeepSeek 官方 `deepseek-v4-flash`、插件 0.8.5（判卷与被测同源，答案原文全部留痕可人工复核）、Windows；题型设计借鉴 [LongMemEval](https://github.com/xiaowu0162/longmemeval) / [LoCoMo](https://snap-research.github.io/locomo/) / [AMB](https://github.com/vectorize-io/agent-memory-benchmark)，扩展题型与生命周期赛道参照 [MemoryAgentBench](https://arxiv.org/abs/2507.05257) / [GoodAI LTM](https://github.com/GoodAI/goodai-ltm-benchmark) / BEAM。
 
-### 对话赛道（15 场景 × 6 题型 × 3 次 = 270 题/组）：答得准吗
+> 对话赛道为 **0.8.5 新基线**（修复版插件 + 修正后的判卷口径）；工作流赛道数字仍为 0.8.3 存档（0.8.5 起场景库扩至 8 个，新增前瞻记忆场景，重跑待做）。
+
+### 对话赛道（20 场景 × 10 题型 × 3 次 = 420 题）：答得准吗
+
+> 0.8.5 基线（A 组数据；对话赛道 B 组已下线，只跑 A 组）。
 
 <p align="center">
   <img src="./assets/readme/bench-dialog.svg" width="100%"
-       alt="DSH-MemBench 对话赛道 A/B 对照条形图：A 组（记忆开）总准确率 92.6%（250/270），B 组（记忆关）17.8%（48/270）；分题型各 45 题——抽取 A 45/45 对 B 3/45、多跳 A 45/45 对 B 0/45、时序 A 43/45 对 B 0/45、知识更新 A 31/45 对 B 0/45、场景回忆 A 41/45 对 B 0/45、拒答两组均 45/45 且 0 编造">
+       alt="DSH-MemBench 对话赛道准确率图（A 组·记忆开）：总准确率 95.2%（400/420）；核心六题型各 60 题——抽取 58/60、多跳 60/60、时序 56/60、更新 55/60、场景回忆 52/60、拒答 60/60 且 0 编造；扩展四题型各 15 题——增量积累 15/15、连锁更新 15/15、事件排序 14/15、同义改写 15/15">
 </p>
 
-**召回双通道**（A 组）：被动注入召回率 **75.1%**（该题要点出现在召回注入中，169/225），其余多数由模型**主动调用记忆工具**查回——84 题主动查询、**60 题靠工具兜底答对**；端到端 92.6% 是两通道 + 模型利用的合成结果。记忆库跨场景全程累积下，探针召回注入混入其他场景记忆 144 次（已如实计数），总准确率仍稳在 92.6%——抗干扰能力经受住了膨胀记忆库的考验。
+**召回双通道**（A 组）：被动注入召回率 **78.1%**（该题要点出现在召回注入中，281/360），其余多数由模型**主动调用记忆工具**查回——106 题主动查询、**75 题靠工具兜底答对**；端到端 95.2% 是两通道 + 模型利用的合成结果。记忆库跨场景全程累积下，探针召回注入混入其他场景记忆 295 次（已如实计数），总准确率反而前段 92.8% → 后段 97.7%——抗干扰能力经受住了膨胀记忆库的考验（离线灌水再灌 600 条合成噪声，检索层 recall@5 也只降 2.8pp）。
 
-### 工作流赛道（4 场景 × 3 次，真实工具沙箱）：做得对、做得省吗
+**分层看短板**：检索层离线指标（recall@5 受控复现）总 73.3%，其中事件排序 0%、场景回忆 50%——端到端仍 93%+ 靠的是注入邻近记忆后模型的鲁棒性；**效率三角**（记忆的开销）：注入非但不加延迟（注入轮响应比无注入轮平均快 210ms）、注入占每轮输入约 10.3%，蒸馏全链路摊到每条捕获消息 ≈2727 输入 / 240 输出 token（1172 次调用 0 失败）。
+
+### 工作流赛道（0.8.3 存档 · 7 场景版 · A 组 3 次 / B 组 1 次，真实工具沙箱）：做得对、做得省吗
 
 <p align="center">
   <img src="./assets/readme/bench-workflow.svg" width="100%"
-       alt="DSH-MemBench 工作流赛道 A/B 对照图：任务完成度 A 组 24/33（72.7%）对 B 组 11/33（33.3%）；成本对比（B 组为满格基准）——步骤 125 对 186（B +49%）、工具调用 184 对 296（B +61%）、输入 token 1.30M 对 1.86M（B +43%）；向用户求助 A 0 次对 B 3 次；登录场景输入 token A 241k 对 B 453k（+88%）">
+       alt="DSH-MemBench 工作流赛道 A/B 对照图：探针段完成度 A 组 59/69（85.5%）对 B 组 10/23（43.5%）；成本对比（B 组为满格基准，每场景均值）——步骤 24.3 对 41.4（B +70%）、工具调用 37.7 对 62.1（B +65%）、输入 token 266k 对 1.81M（B 6.8 倍）；风格规范场景探针 A 12/12 对 B 0/4；长任务每场景输入 token A 266k 对 B 1.81M">
 </p>
 
-**登录场景特写**（凭据只存在于记忆，站点为本地服务、令牌不可伪造）：A 组三次全部**一轮直取**完成（6/6，241k 输入）；B 组每次都要**反问用户要凭据**（3 次求助、双倍轮次）才完成 5/6，输入 453k——**+88%**。这正是记忆的核心价值之一：**省掉的不是任务难度，是无谓的往返与重复教学**。
+**探针段完成度 85.5% vs 43.5%（+42pp）**：教学/变更段两组都有现场上下文，探针段（新会话延续任务）才是纯记忆窗口——A 组三个新考法场景（流程知识更新 / 双胞胎消歧 / 风格规范延续）全部 12/12 满分且三轮一致；B 组在风格规范场景探针 **0/4**（命名/结构/千分位/页脚约定只存记忆，沙箱探不出来），在流程更新场景则能靠读脚本逆向（判别力受沙箱可供性限制，已如实标注）。
+
+**长任务成本：B 组每场景输入 token 是 A 组的 6.8 倍**（1.81M vs 266k）——无记忆时 agent 靠重新探索前进，high 思考档下甚至会自建工程去探测本可用一条脚本约定完成的流程；输出 token 3 倍（46.2k vs 15.4k）、步骤 +70%。这正是记忆的核心价值：**省掉的不是任务难度，是无谓的往返与重复探索**。
 
 ### 方法论与复现
 
 ```bash
-node bench/harness/run.mjs --arm A --repeats 3 --provider deepseek-official --model deepseek-v4-flash
-node bench/harness/run.mjs --arm B --repeats 3 --provider deepseek-official --model deepseek-v4-flash   # 对话赛道
-node bench/harness/run.mjs --track workflow --arm A/B --repeats 3 ...                                  # 工作流赛道
+node bench/harness/run.mjs --arm A --repeats 3 --provider deepseek-official --model deepseek-v4-flash   # 对话赛道（只跑 A 组）
+node bench/harness/run.mjs --track workflow --arm AB --repeats 3 ...                                  # 工作流赛道（A/B 双组并行）
+node bench/harness/run.mjs --track lifecycle --arm A ...                                              # 生命周期赛道（门控/off/rebuild/遗忘）
 node bench/harness/report.mjs --latest [dialog|workflow]                                               # 汇总报告
+node bench/harness/retrieval-metrics.mjs <runDir> --flood 200,600                                     # 检索层指标 + 灌水曲线
 ```
 
-- 判分：`contains-all` 程序判 + 判卷模型按要点判（答案原文与判分理由全部留痕 `result.json` 可人工复核）；工作流完成度为产物文件 + 关键内容程序化校验；
-- 指标全部来自供应商上报 usage（输入含缓存命中拆分）与会话事件折叠；**稳态缓存率**剔除每会话首请求（A 88.7% vs B 85.4%——记忆注入不伤缓存）；
-- 回归用途：改插件前后各跑一遍，`compare.mjs` 出对比表（环境头校验 + B 组对照组漂移告警）；
-- 局限（诚实声明）：单机 ×3 次合并、判卷模型与被测模型同源、作者自建场景库（倾向记忆优势场景，欢迎自行复现）；工具审计会标记越界访问（实测 Agent 偶发探测用户主目录，本基准场景答案不存于真实记忆库、数字不受影响）。
+- 判分：`contains-all` 程序判 + 判卷模型按要点判（答案原文与判分理由全部留痕 `result.json` 可人工复核）；带 stale 的题（更新/连锁/遗忘）"旧值当作现状陈述"才判负、拒答题允许引用真实背景解释"不知道被问点"；工作流完成度为产物文件 + 关键内容程序化校验（四型判据：正检查/禁词/产物缺席/存在性）；
+- 指标面：准确率总表（6 核心 + 4 扩展题型）之外，自动产出**检索层离线指标**（recall@5 / 注入精度 / 作废泄漏）、**效率三角**（注入开销差分 / 注入占比 / 蒸馏记账摊到每消息）、**规模位置分析**（库容膨胀下的准确率/污染）与生命周期赛道专属节（分族门控矩阵 / off 双断言 / rebuild 保真 / 遗忘）；
+- 实时进度：跑基准时自动拉起本地进度面板并打开浏览器（`--no-panel` 关闭）——A/B 双臂场景/阶段/消息粒度进度、心跳与活动新鲜度（直判"卡住 vs 进程挂了"）、累计成本随跑随涨；
+- 指标全部来自供应商上报 usage（输入含缓存命中拆分）与会话事件折叠；稳态缓存率剔除每会话首请求（0.8.5 基线：89.1%——记忆注入不伤缓存）；
+- 回归用途：改插件前后各跑一遍，`compare.mjs` 出对比表（环境头校验含 gitSha + B 组对照组漂移告警 + 检索层指标对比）；
+- 局限（诚实声明）：单机；A 组 ×3 合并、B 组 ×1（成本护栏，噪声更大）；判卷与被测模型：对话 0.8.5 基线同源、工作流存档跑为异构（glm-5.3 判 v4-flash）；作者自建场景库（倾向记忆优势场景，欢迎自行复现）；沙箱文件的可供性会部分泄露流程（B 组可读脚本逆向，判别力受限处已如实标注）；工具审计双档（严格违规判负/宽松提示），实测双方 0 违规。
 
 完整报告与逐题数据：[`bench/baseline/`](./bench/baseline/)。
 
@@ -252,11 +268,12 @@ ONNX 量化 **CPU 推理**——无需 API Key，数据不出本机）。本地
 | `embedding.proxy` | `''` | 模型下载代理三态：`''`（默认）= 自动探测代理环境变量（`HTTPS_PROXY`/`ALL_PROXY` 等，尊重 `NO_PROXY`）；`none` = 禁用强制直连；其他值 = 代理 URL（如 `http://127.0.0.1:7890`）。镜像直连在国内网络间歇不可达（直连超时与污染字节交替出现过），开代理的机器建议保持默认自动探测 |
 | `llm.provider/model` | 空 | 蒸馏模型静态路由（部署 pin）：provider 与 model **双字段齐**时锁定蒸馏路由，优先于设置页的运行时选择与默认模型（部署可强制蒸馏走指定路由）；留空则跟随"设置页选择 → 默认模型"。运行时可在设置页 → 记忆 → 概览的"蒸馏模型"选择器从**已配置的供应商**（含 dsh 设置 → 模型里添加的自定义供应商）中切换，即时生效无需重启 |
 | `llm.maxTokens` | `65536` | 未分层调用的兜底输出总闸。各蒸馏层有独立预算（抽取 16k / 去重 8k / L2 32k / L3 16k；思考档 high/xhigh/max 时自动 ×4，防 reasoning 吃光预算），分层预算可在设置页 → 记忆 → 概览 → 蒸馏参数运行时调整（留空/0 = 跟随内置默认） |
-| `llm.reasoningEffort` | 空 | 蒸馏思考档位：空串 = **自动**（按模型能力解析：模型默认档 → `high`）；显式值（`off`/`none`/`low`/`medium`/`high`/`xhigh`/`max`）仅在该模型声明支持时发送——跨供应商 effort 词汇表不同（deepseek 认 `off`，OpenAI 系是 `none`，未声明档位的模型不传），不支持的档位自动降级为不传并告警一次；思考档 high/xhigh/max 时输出预算自动 ×4。运行时在设置页 → 记忆 → 概览切换，可选档位表跟随当前模型实时显示 |
+| `llm.reasoningEffort` | 空 | 蒸馏思考档位：空串 = **自动**（按模型能力解析：模型默认档 → `high`）；显式值（`off`/`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`）仅在该模型声明支持时发送——跨供应商 effort 词汇表不同（deepseek 认 `off`，OpenAI 系是 `none`，未声明档位的模型不传），不支持的档位自动降级为不传并告警一次；思考档 high/xhigh/max 时输出预算自动 ×4。运行时在设置页 → 记忆 → 概览切换，可选档位表跟随当前模型实时显示 |
 | `llm.temperature` | `0.3` | 蒸馏温度 |
 | `llm.maxInputChars` | `700000` | 单次蒸馏输入字符预算（超限的 L1 输入自动分块抽取）；运行时可在设置页 → 蒸馏参数 → 输入预算调整（留空/0 = 跟随本值） |
 | `llm.timeoutMs` | `120000` | 单次蒸馏调用超时（ms） |
 | `tools` | `true` | 是否注册模型可调用的记忆工具 |
+| `benchControl` | `false` | 注册 bench 控制服务（进程内 rebuild 触发/会话档位设置/蒸馏用量快照，供基准 lifecycle 赛道）。默认关——生产部署零表面积，勿随意开启 |
 
 ## 日志与故障排查
 

package/assets/readme/bench-dialog.svg

@@ -1,87 +1,70 @@
 <svg xmlns="http://www.w3.org/2000/svg"
-     width="1200" height="690" viewBox="0 0 1200 690"
+     width="1200" height="630" viewBox="0 0 1200 630"
      role="img" aria-labelledby="benchDialogTitle benchDialogDesc">
-  <title id="benchDialogTitle">DSH-MemBench 对话赛道：A/B 两组准确率对照</title>
-  <desc id="benchDialogDesc">同题同输入对照：A 组（记忆开）总准确率 92.6%（250/270），B 组（记忆关）17.8%（48/270）。分题型（每题型 45 题）：抽取 A 45/45 对 B 3/45；多跳 A 45/45 对 B 0/45；时序 A 43/45 对 B 0/45；知识更新 A 31/45 对 B 0/45；场景回忆 A 41/45 对 B 0/45；拒答两组均 45/45、0 编造。</desc>
+  <title id="benchDialogTitle">DSH-MemBench 对话赛道：A 组（记忆开）准确率</title>
+  <desc id="benchDialogDesc">A 组（记忆开）总准确率 95.2%（400/420，20 场景 × 10 题型 × 3 次）。核心六题型各 60 题：抽取 58/60、多跳 60/60、时序 56/60、更新 55/60、场景回忆 52/60、拒答 60/60 且 0 编造；扩展四题型各 15 题：增量积累 15/15、连锁更新 15/15、事件排序 14/15、同义改写 15/15。对话赛道 B 组已下线（会话独立、无记忆必然失败，对照无信息量）。</desc>
 
-  <rect width="1200" height="690" rx="26" fill="#0D1526"/>
+  <rect width="1200" height="630" rx="26" fill="#0D1526"/>
 
   <g font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','PingFang SC','Microsoft YaHei',sans-serif">
 
-    <!-- ── 标题与图例 ── -->
-    <text x="64" y="66" font-size="26" font-weight="700" fill="#EFF3FA">对话赛道：同题同输入 · A/B 对照准确率</text>
-    <text x="64" y="94" font-size="18" fill="#8296B3">15 场景 × 6 题型 × 3 次 = 270 题/组（每题型 45 题）· 条长 = 答对率，数值 = 答对题数</text>
+    <text x="64" y="66" font-size="26" font-weight="700" fill="#EFF3FA">对话赛道：记忆开 · 准确率</text>
+    <text x="64" y="94" font-size="18" fill="#8296B3">20 场景 × 10 题型 × 3 次 = 420 题（核心六题型各 60 题、扩展四题型各 15 题）· 条长 = 答对率，数值 = 答对题数</text>
 
     <rect x="842" y="50" width="18" height="18" rx="5" fill="#FFD15D"/>
-    <text x="868" y="65" font-size="18" fill="#C2CDDE">A 组 · 记忆开</text>
-    <rect x="842" y="76" width="18" height="18" rx="5" fill="#4E6DA8"/>
-    <text x="868" y="91" font-size="18" fill="#C2CDDE">B 组 · 记忆关</text>
+    <text x="868" y="65" font-size="17" fill="#C2CDDE">核心六题型</text>
+    <rect x="966" y="50" width="18" height="18" rx="5" fill="#6FB1FF"/>
+    <text x="992" y="65" font-size="17" fill="#C2CDDE">扩展四题型</text>
 
-    <!-- ── 网格线（0/25/50/75/100%）── -->
     <g stroke="#1C2A47" stroke-width="1">
-      <line x1="260" y1="122" x2="260" y2="614"/>
-      <line x1="435" y1="122" x2="435" y2="614"/>
-      <line x1="610" y1="122" x2="610" y2="614"/>
-      <line x1="785" y1="122" x2="785" y2="614"/>
-      <line x1="960" y1="122" x2="960" y2="614"/>
+      <line x1="260" y1="120" x2="260" y2="525"/>
+      <line x1="435" y1="120" x2="435" y2="525"/>
+      <line x1="610" y1="120" x2="610" y2="525"/>
+      <line x1="785" y1="120" x2="785" y2="525"/>
+      <line x1="960" y1="120" x2="960" y2="525"/>
     </g>
     <g font-size="17" fill="#8296B3" text-anchor="middle">
-      <text x="260" y="640">0%</text>
-      <text x="435" y="640">25%</text>
-      <text x="610" y="640">50%</text>
-      <text x="785" y="640">75%</text>
-      <text x="960" y="640">100%</text>
+      <text x="260" y="552">0%</text>
+      <text x="435" y="552">25%</text>
+      <text x="610" y="552">50%</text>
+      <text x="785" y="552">75%</text>
+      <text x="960" y="552">100%</text>
     </g>
 
-    <!-- ── 总准确率（强调行）── -->
-    <text x="240" y="176" font-size="21" font-weight="700" fill="#EFF3FA" text-anchor="end">总准确率</text>
-    <rect x="260" y="148" width="648" height="26" rx="6" fill="#FFD15D"/>
-    <text x="922" y="169" font-size="22" font-weight="700" fill="#FFD15D">92.6%（250/270）</text>
-    <rect x="260" y="182" width="124" height="26" rx="6" fill="#4E6DA8"/>
-    <text x="396" y="203" font-size="22" font-weight="700" fill="#8FA6C6">17.8%（48/270）</text>
+  <text x="240" y="149" font-size="18" font-weight="700" fill="#EFF3FA" text-anchor="end">总准确率</text>
+  <rect x="260" y="130" width="667" height="25" rx="6" fill="#FFD15D"/>
+  <text x="939" y="149" font-size="17" font-weight="600" fill="#FFD15D">400/420（95.2%）</text>
+  <text x="240" y="184" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">抽取</text>
+  <rect x="260" y="165" width="677" height="25" rx="6" fill="#FFD15D"/>
+  <text x="949" y="184" font-size="17" font-weight="600" fill="#FFD15D">58/60（96.7%）</text>
+  <text x="240" y="219" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">多跳</text>
+  <rect x="260" y="200" width="700" height="25" rx="6" fill="#FFD15D"/>
+  <text x="972" y="219" font-size="17" font-weight="600" fill="#FFD15D">60/60（100.0%）</text>
+  <text x="240" y="254" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">时序</text>
+  <rect x="260" y="235" width="653" height="25" rx="6" fill="#FFD15D"/>
+  <text x="925" y="254" font-size="17" font-weight="600" fill="#FFD15D">56/60（93.3%）</text>
+  <text x="240" y="289" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">更新</text>
+  <rect x="260" y="270" width="642" height="25" rx="6" fill="#FFD15D"/>
+  <text x="914" y="289" font-size="17" font-weight="600" fill="#FFD15D">55/60（91.7%）</text>
+  <text x="240" y="324" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">场景回忆</text>
+  <rect x="260" y="305" width="607" height="25" rx="6" fill="#FFD15D"/>
+  <text x="879" y="324" font-size="17" font-weight="600" fill="#FFD15D">52/60（86.7%）</text>
+  <text x="240" y="359" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">拒答（0 编造）</text>
+  <rect x="260" y="340" width="700" height="25" rx="6" fill="#FFD15D"/>
+  <text x="972" y="359" font-size="17" font-weight="600" fill="#FFD15D">60/60（100.0%）</text>
+  <text x="240" y="394" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">增量积累</text>
+  <rect x="260" y="375" width="700" height="25" rx="6" fill="#6FB1FF"/>
+  <text x="972" y="394" font-size="17" font-weight="600" fill="#6FB1FF">15/15（100.0%）</text>
+  <text x="240" y="429" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">连锁更新</text>
+  <rect x="260" y="410" width="700" height="25" rx="6" fill="#6FB1FF"/>
+  <text x="972" y="429" font-size="17" font-weight="600" fill="#6FB1FF">15/15（100.0%）</text>
+  <text x="240" y="464" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">事件排序</text>
+  <rect x="260" y="445" width="653" height="25" rx="6" fill="#6FB1FF"/>
+  <text x="925" y="464" font-size="17" font-weight="600" fill="#6FB1FF">14/15（93.3%）</text>
+  <text x="240" y="499" font-size="18" font-weight="400" fill="#C2CDDE" text-anchor="end">同义改写</text>
+  <rect x="260" y="480" width="700" height="25" rx="6" fill="#6FB1FF"/>
+  <text x="972" y="499" font-size="17" font-weight="600" fill="#6FB1FF">15/15（100.0%）</text>
 
-    <line x1="64" y1="226" x2="1136" y2="226" stroke="#263A5C" stroke-width="1.5"/>
-
-    <!-- ── 分题型行 ── -->
-    <!-- 抽取 -->
-    <text x="240" y="286" font-size="19" fill="#C2CDDE" text-anchor="end">抽取</text>
-    <rect x="260" y="254" width="700" height="18" rx="5" fill="#FFD15D"/>
-    <text x="974" y="269" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">45/45</text>
-    <rect x="260" y="278" width="47" height="18" rx="5" fill="#4E6DA8"/>
-    <text x="319" y="293" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">3/45</text>
-
-    <!-- 多跳 -->
-    <text x="240" y="348" font-size="19" fill="#C2CDDE" text-anchor="end">多跳</text>
-    <rect x="260" y="316" width="700" height="18" rx="5" fill="#FFD15D"/>
-    <text x="974" y="331" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">45/45</text>
-    <text x="272" y="355" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">0/45</text>
-
-    <!-- 时序 -->
-    <text x="240" y="410" font-size="19" fill="#C2CDDE" text-anchor="end">时序</text>
-    <rect x="260" y="378" width="669" height="18" rx="5" fill="#FFD15D"/>
-    <text x="943" y="393" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">43/45</text>
-    <text x="272" y="417" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">0/45</text>
-
-    <!-- 知识更新 -->
-    <text x="240" y="472" font-size="19" fill="#C2CDDE" text-anchor="end">知识更新 *</text>
-    <rect x="260" y="440" width="482" height="18" rx="5" fill="#FFD15D"/>
-    <text x="756" y="455" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">31/45</text>
-    <text x="272" y="479" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">0/45</text>
-
-    <!-- 场景回忆 -->
-    <text x="240" y="534" font-size="19" fill="#C2CDDE" text-anchor="end">场景回忆</text>
-    <rect x="260" y="502" width="638" height="18" rx="5" fill="#FFD15D"/>
-    <text x="912" y="517" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">41/45</text>
-    <text x="272" y="541" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">0/45</text>
-
-    <!-- 拒答 -->
-    <text x="240" y="596" font-size="19" fill="#C2CDDE" text-anchor="end">拒答 *</text>
-    <rect x="260" y="564" width="700" height="18" rx="5" fill="#FFD15D"/>
-    <text x="974" y="579" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">45/45</text>
-    <rect x="260" y="588" width="700" height="18" rx="5" fill="#4E6DA8"/>
-    <text x="974" y="603" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">45/45</text>
-
-    <!-- ── 脚注 ── -->
-    <text x="64" y="672" font-size="16" fill="#8296B3">* 知识更新：改口后仍答旧值记 0 分（直接考核 L1 去重更新）；拒答：没发生过的事编造即 0 分（两组均 0 编造）。</text>
+    <text x="64" y="600" font-size="16" fill="#8296B3">DSH-MemBench 0.8.5 基线 · DeepSeek v4-flash · 3 次重复合并 · 场景库跨场景累积（越靠后的场景记忆干扰越大）</text>
   </g>
 </svg>

package/assets/readme/bench-workflow.svg

@@ -2,7 +2,7 @@
      width="1200" height="660" viewBox="0 0 1200 660"
      role="img" aria-labelledby="benchWorkflowTitle benchWorkflowDesc">
   <title id="benchWorkflowTitle">DSH-MemBench 工作流赛道：A/B 两组完成度与成本对照</title>
-  <desc id="benchWorkflowDesc">真实工具沙箱（4 场景 × 3 次）探针会话合计：任务完成度 A 组 24/33（72.7%）对 B 组 11/33（33.3%）；成本对比以 B 组为满格基准，A 组相对更省——步骤 125 对 186（B +49%）、工具调用 184 对 296（B +61%）、输入 token 1.30M 对 1.86M（B +43%）；向用户求助 A 0 次对 B 3 次；登录场景输入 token A 241k 对 B 453k（+88%）。</desc>
+  <desc id="benchWorkflowDesc">真实工具沙箱（7 场景，A 组 3 轮 / B 组 1 轮）探针段合计：任务完成度 A 组 59/69（85.5%）对 B 组 10/23（43.5%）；成本对比以 B 组为满格基准（每场景均值），A 组相对更省——步骤 24.3 对 41.4（B +70%）、工具调用 37.7 对 62.1（B +65%）、输入 token 266k 对 1.81M（B 6.8 倍）；风格规范场景探针 A 12/12 对 B 0/4；向用户求助 A 组 3 轮共 1 次、B 组 1 轮 1 次。</desc>
 
   <rect width="1200" height="660" rx="26" fill="#0D1526"/>
 
@@ -10,70 +10,70 @@
 
     <!-- ── 标题与图例 ── -->
     <text x="64" y="66" font-size="26" font-weight="700" fill="#EFF3FA">工作流赛道：真实工具沙箱 · 做得对，也更省</text>
-    <text x="64" y="94" font-size="18" fill="#8296B3">4 场景 × 3 次 · 探针会话合计 · 成本行以 B 组为满格基准，蓝徽章 = B 组多付</text>
+    <text x="64" y="94" font-size="18" fill="#8296B3">7 场景 · A 组 3 轮 / B 组 1 轮（成本护栏）· 探针段合计 · 成本行以 B 组为满格基准</text>
 
     <rect x="842" y="50" width="18" height="18" rx="5" fill="#FFD15D"/>
     <text x="868" y="65" font-size="18" fill="#C2CDDE">A 组 · 记忆开</text>
     <rect x="842" y="76" width="18" height="18" rx="5" fill="#4E6DA8"/>
     <text x="868" y="91" font-size="18" fill="#C2CDDE">B 组 · 记忆关</text>
 
-    <!-- ── 任务完成度 ── -->
-    <text x="64" y="152" font-size="20" font-weight="700" fill="#EFF3FA">任务完成度</text>
-    <text x="186" y="152" font-size="17" fill="#8296B3">（产物程序化校验）</text>
+    <!-- ── 探针段完成度 ── -->
+    <text x="64" y="152" font-size="20" font-weight="700" fill="#EFF3FA">探针段完成度</text>
+    <text x="216" y="152" font-size="17" fill="#8296B3">（新会话延续任务，产物程序化校验）</text>
 
-    <rect x="260" y="170" width="509" height="26" rx="6" fill="#FFD15D"/>
-    <text x="783" y="191" font-size="22" font-weight="700" fill="#FFD15D">24/33 · 72.7%</text>
-    <rect x="260" y="206" width="233" height="26" rx="6" fill="#4E6DA8"/>
-    <text x="505" y="227" font-size="22" font-weight="700" fill="#8FA6C6">11/33 · 33.3%</text>
+    <rect x="260" y="170" width="598" height="26" rx="6" fill="#FFD15D"/>
+    <text x="872" y="191" font-size="22" font-weight="700" fill="#FFD15D">59/69 · 85.5%</text>
+    <rect x="260" y="206" width="304" height="26" rx="6" fill="#4E6DA8"/>
+    <text x="578" y="227" font-size="22" font-weight="700" fill="#8FA6C6">10/23 · 43.5%</text>
 
     <line x1="64" y1="258" x2="1136" y2="258" stroke="#263A5C" stroke-width="1.5"/>
 
     <!-- ── 成本对比（B = 满格基准）── -->
     <text x="64" y="290" font-size="20" font-weight="700" fill="#EFF3FA">成本对比</text>
-    <text x="166" y="290" font-size="17" fill="#8296B3">（B 组 = 满格基准，条越短越省）</text>
+    <text x="166" y="290" font-size="17" fill="#8296B3">（每场景均值，B 组 = 满格基准，条越短越省）</text>
 
     <!-- 步骤 -->
     <text x="240" y="348" font-size="19" fill="#C2CDDE" text-anchor="end">步骤</text>
-    <rect x="260" y="318" width="376" height="18" rx="5" fill="#FFD15D"/>
-    <text x="648" y="333" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">125</text>
+    <rect x="260" y="318" width="329" height="18" rx="5" fill="#FFD15D"/>
+    <text x="603" y="333" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">24.3</text>
     <rect x="260" y="342" width="560" height="18" rx="5" fill="#4E6DA8"/>
-    <text x="834" y="357" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">186</text>
+    <text x="834" y="357" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">41.4</text>
     <rect x="980" y="321" width="136" height="36" rx="10" fill="#315DA6"/>
-    <text x="1048" y="345" font-size="19" font-weight="700" fill="#FFD15D" text-anchor="middle">B +49%</text>
+    <text x="1048" y="345" font-size="19" font-weight="700" fill="#FFD15D" text-anchor="middle">B +70%</text>
 
     <!-- 工具调用 -->
     <text x="240" y="430" font-size="19" fill="#C2CDDE" text-anchor="end">工具调用</text>
-    <rect x="260" y="400" width="348" height="18" rx="5" fill="#FFD15D"/>
-    <text x="620" y="415" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">184</text>
+    <rect x="260" y="400" width="340" height="18" rx="5" fill="#FFD15D"/>
+    <text x="614" y="415" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">37.7</text>
     <rect x="260" y="424" width="560" height="18" rx="5" fill="#4E6DA8"/>
-    <text x="834" y="439" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">296</text>
+    <text x="834" y="439" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">62.1</text>
     <rect x="980" y="403" width="136" height="36" rx="10" fill="#315DA6"/>
-    <text x="1048" y="427" font-size="19" font-weight="700" fill="#FFD15D" text-anchor="middle">B +61%</text>
+    <text x="1048" y="427" font-size="19" font-weight="700" fill="#FFD15D" text-anchor="middle">B +65%</text>
 
     <!-- 输入 token -->
     <text x="240" y="512" font-size="19" fill="#C2CDDE" text-anchor="end">输入 token</text>
-    <rect x="260" y="482" width="391" height="18" rx="5" fill="#FFD15D"/>
-    <text x="663" y="497" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">1.30M</text>
+    <rect x="260" y="482" width="82" height="18" rx="5" fill="#FFD15D"/>
+    <text x="356" y="497" font-size="18" font-weight="700" fill="#FFD15D" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">266k</text>
     <rect x="260" y="506" width="560" height="18" rx="5" fill="#4E6DA8"/>
-    <text x="834" y="521" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">1.86M</text>
+    <text x="834" y="521" font-size="18" font-weight="700" fill="#8FA6C6" font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,monospace">1.81M</text>
     <rect x="980" y="485" width="136" height="36" rx="10" fill="#315DA6"/>
-    <text x="1048" y="509" font-size="19" font-weight="700" fill="#FFD15D" text-anchor="middle">B +43%</text>
+    <text x="1048" y="509" font-size="19" font-weight="700" fill="#FFD15D" text-anchor="middle">B +580%</text>
 
     <!-- ── 底部特写卡片 ── -->
     <rect x="64" y="544" width="512" height="72" rx="14" fill="#15223A" stroke="#263A5C" stroke-width="1.5"/>
-    <text x="88" y="574" font-size="17" fill="#8296B3">向用户求助（探针会话）</text>
+    <text x="88" y="574" font-size="17" fill="#8296B3">风格规范场景 · 探针段（约定只存记忆，沙箱探不出来）</text>
     <text x="88" y="600" font-size="20" font-weight="700">
-      <tspan fill="#FFD15D">A 0 次</tspan>
-      <tspan fill="#8FA6C6"> · B 3 次（每轮要凭据）</tspan>
+      <tspan fill="#FFD15D">A 12/12</tspan>
+      <tspan fill="#8FA6C6"> · B 0/4</tspan>
     </text>
 
     <rect x="624" y="544" width="512" height="72" rx="14" fill="#15223A" stroke="#263A5C" stroke-width="1.5"/>
-    <text x="648" y="574" font-size="17" fill="#8296B3">登录场景 · 输入 token（凭据只存在记忆里）</text>
+    <text x="648" y="574" font-size="17" fill="#8296B3">长任务成本 · 每场景输入 token（B 无记忆重新探索）</text>
     <text x="648" y="600" font-size="20" font-weight="700">
-      <tspan fill="#FFD15D">A 241k</tspan>
-      <tspan fill="#8FA6C6"> · B 453k</tspan>
+      <tspan fill="#FFD15D">A 266k</tspan>
+      <tspan fill="#8FA6C6"> · B 1.81M</tspan>
     </text>
     <rect x="1040" y="578" width="72" height="26" rx="8" fill="#315DA6"/>
-    <text x="1076" y="597" font-size="18" font-weight="700" fill="#FFD15D" text-anchor="middle">+88%</text>
+    <text x="1076" y="597" font-size="18" font-weight="700" fill="#FFD15D" text-anchor="middle">6.8×</text>
   </g>
 </svg>