dsh-router-standard

⚠️⚠️⚠️ 重要勘误与道歉（必读） ⚠️⚠️⚠️

我不需要被造神，也不配被造神

这篇 README 顶部必须放这段话：我错了，而且错得很有代表性。

• 论文不撤回，但其中理论解释部分（双吸引子假设 A1–A4 及"god/ghost"、"自路由不可能"等强归因）已正式标注作废。
• 心路历程：我最初把 "We need / Let me" 的差异当成"官方刻意设计的双模式"；后来才意识到那更可能是一条原生深度路径 + 一条后压未收敛的极简路径之间的断层/断裂带。这个断层本身就像一层路由层——我们实际做的是把它当路由层用，实现了 "Let me / We need" 自路由。
• 我们做对了一件事，也请记住这件事：利用后训练的一个小缺陷（断层/断裂带），我们实现了 V4 Flash 能力的可复现提升——这不是我聪明，是那个缺陷恰好可被工程利用。
• Pro 是另一场完全不同的硬仗：雷霆大思考、工具面敏感、内部路由不稳定。我已经找到方向（黑盒 logprobs / 嵌入向量层逆向 / 语义锚点指纹），初步数据已在仓库。这条路我会继续用实测走，而不是用叙事走。

📄 完整勘误声明：docs/statement.md ｜ 道歉函：docs/apology.md

当前状态（v1.8.0，2026-08-22 · 研发线，尚未发布）

Router Standard 已进入「渐进式披露套装」研发线（五轮真实会话实测驱动；发布与否以研发定稿为准）：

• 渐进披露：四阶段解锁 + 两档预放（阶段 0 即用 write/edit）+ 直达语义（调用哪档工具就跳到哪档）+ 交付全量开放；
• 二级注册表：tools_catalog（全量索引 + 阶段标记 + 参数名/类型/上限速览）/ tools_help（完整 schema）；
• PTC 底座：run_code 阶段化 SDK（39K → 阶段可见签名），工具面注意力税大幅下降；
• 页面验证内置：dev_page_check = 截图 + DOM smoke + console/pageerror/title/selector/scale；{js:…} 模式 = 本地 JS 引擎（语法检查+纯逻辑单测，零外部 node 依赖）；
• 描述 ⇄ 行为对齐：presentation=code|native 自检、阶段文案只说真话、平台事实（win32 仅 pwsh、node 入 PATH）；
• 压力感应：真实事件通道 + 模型自适应 + 自检信号语义清楚（不是停手命令）。

配套预设：router-react（v17）/ router-spec（v10）（基于标准模式的两大执行预设，均支持 v4-flash-vision-exp 首轮读图——read_image 首轮面）。router-pro 线已退役删除。 完整演进：见 CHANGELOG.md 与 docs/STANDARD-PLAN.md。

【历史文档 · v0.3 时代】 Task-aware reasoning-mode router for DeepSeek Harness. One preset, two routing modes (v0.2.0 naming), plus the measured three-band axis behind them:

选择：安装两个预设之一（Router Standard / Router Spec，见 Usage）。 dev_router_status 显示当前路由模式。

This is a research artifact. It encodes a measured property of DeepSeek V4 Pro / V4 Flash: model behavior along the persona axis is not a continuum — it collapses into a few stable regions separated by phase transitions. The router therefore quantizes to the stable regions instead of pretending the axis is continuously tunable.

v0.3.0 — real-assembly-chain fixes

v0.2.x shipped routing logic that was validated against bare-API probes but was broken on the REAL DeepSeek Harness assembly chain. v0.3.0 fixes all of it, verified against @deepseek-ai/dsh-agent-loop (0.1.0-rc.7) event ordering:

• First-turn routing actually works (issue #13): the loop claims the inbox BEFORE assembling the system prompt, and inbox.claim() emits the agent-scoped agent/inbox/claimed event synchronously — the router captures the first REAL user message there (source.kind === 'user' only), so the first request is classified instead of unconditionally falling into weak. (The captured text is CLASSIFIED, not fed to bandOf raw — the old capture path silently mapped every captured message to the spec band.)
• Near-field guidance moved to agent/pre-step (issues #34/#36/#55): session/event never fires inside agent-plane presets (dsh-scope filters it out of entry-local realms), so the old inbox re-append never delivered guidance — and wherever it did fire, the next-step append forced a SECOND model request per user message (the 2× API-call spike). The guide is now inserted into decision.messages at agent/pre-step: same request as the user message, near-field, cache-neutral, zero extra round-trips.
• Fixes: missing extractText/bandOf imports in both bootstrap files (#11) — the session/event handlers crashed with ReferenceError whenever they did fire; sessionMode ignoring plugin-origin messages when pinning the band; router.test.mjs import path; preset.yml YAML quoting (#53); subagent-session skip (#5); session-selected model from assembled.variables (#9); the RL-standard mode of the spec preset now returns the assembly untouched after the first tool/call (#44).
• New: router.integration.test.mjs replays the real claim → assemble → pre-step ordering against the actual bootstrap code.

What it does

router-standard: reads the session's first REAL user message, classifies the task (build → react / fix → spec / ambiguous → weak), and on the first model request injects the matching persona while keeping the full prompt sections; the first-turn core tool surface follows the band (spec=read/edit/glob/grep, react=read/write/edit, weak=read/write/edit, each plus the platform shell). Weak-band sessions also get a near-field routing guide in the SAME request as every real user message.

router-spec: same routing core with the deep-think-first branding; keeps the v0.2.0 dual-mode code path (routerMode), so a copy configured with routerMode: standard still gets the RL-interface first turn (RL sentence + shell/str_replace_editor) with full sections restored after the first durable tool/call.

After the first durable tool call the full Standard catalog is exposed and the router stops touching anything. The mode is derived from durable session events, so resume/reload keeps it. The plan-mode prompt section is preserved, so plan boundaries do not reset the model's focus.

The three measured behavior bands

Fine-grained probing (21 mode points × n=2, official API, reasoning_effort=max) on V4 Pro shows behavior along the persona axis collapses into three bands:

V4 Flash is threshold-like (0–0.5 all spec side, jumps at 0.75+). The numeric dev_router_mode interface is kept, but it quantizes to the three bands — the transition band is never selected automatically.

Why: dual-attractor RL policy

Evidence across projects (see docs/paper.md and docs/experiments.md):

• The same model reaches top-band scores under spec conditions on a maintenance benchmark (Project2: minimal 99/96, anchored 98/99) and under react/code conditions on a greenfield build task (Mario: 10/10), while the wrong mode scores 91 / 6 respectively — a ~10-point swing from prompt conditioning alone ("god/ghost duality").
• Persona is the dominant trigger (one-sentence swap flips the trajectory); tool-schema surface is a secondary condition; catalog text in a user message has no effect.
• Behavior is path-committed: once anchored, expanding the tool catalog perturbs at most one reasoning block and never flips the mode.
• Intermediate personas are out-of-distribution (training-distribution gap), which is the measured unstable band.

The model cannot self-route: P3 (same persona, task swap → trajectory unchanged), P5 (router personas → doer attractor absorbs the instruction) and P8 (domain-overlap scan) show the only internal-routing window is a WEAK persona + few-shot routing instruction (lean, not flip; discrimination +2.3..+3.3). There is no reward signal for switching modes mid session, and the behavior phase transition means the model commits on the first request. Mode selection must come from outside — a human (the "streamer"), a heuristic classifier, or a learned router. This preset is the automated version of that external routing.

Usage

Two presets (v0.3.0; the router-pro line was retired): install one or more under ~/.dsh/.agent-presets/:

# 标准路由预设（RL 接口还原，默认推荐）
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-standard'
Copy-Item -Recurse .\preset\router-standard $target

# spec 路由预设（深度思考优先）
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-spec'
Copy-Item -Recurse .\preset\router-spec $target

免重启安装（推荐）：装好 dsh-super-injector 后（见套件 scripts/install-injector.ps1），改预设代码不再需要换文件名/重启：

dev_reload_preset router-standard   # 预设热更新：?v=N query 绕 ESM 缓存，新会话立即用新代码

注意事项（实测血泪）：

1. ESM 缓存：loader 按 URL 缓存模块——原地覆盖文件内容不生效（改代码必须 dev_reload_preset 或换文件名）。
2. 首次会话必须新开：路由模式在首个请求锁定（路径承诺），中途切 GUI 模型/ 改配置不影响已运行会话。
3. 子代理不路由：parentSession 会话跳过路由层（社区 #5 修复），shell-less 子代理不再崩溃。
4. 引导注入通道（v0.3.0）：近场引导走 agent/pre-step，每个真实用户消息 注入一条（weak 模式）；rc.6 起 session/event 在 standing scope 收不到事件， 旧版本引导是死的。
5. 首轮真实分类（issue #3/#13）：首轮路由读 agent/inbox/claimed 的原始 消息文本并经 classifyTask 分类——首轮即真实 band（不再 weak 兜底）。
6. 自举卸载：dev_uninject_plugin --self=true 可卸载注入器自身（保留 装配链，重启自动恢复）——用于验证安装闭环。

Restart DSH (or install via the suite script for zero-touch), start a new session, pick Router Standard (experimental) (RL-interface, think-act loops), Router Spec (experimental) (deep-think-first, the long first-turn chain is the point) or Router Pro (V4 Pro measured optimum).

• dev_router_status — current mode, band, persona, core tools, override state
• dev_router_mode <spec|weak|mixed|react|0-100|0.0-1.0|auto> — explicit mode (numeric inputs quantize to the three bands)
• dev_mode_subagent <spec|react|balanced> <task> — run one task in a DIFFERENT reasoning mode inside a fresh isolated context (its own system prompt), leaving the current trajectory untouched. Mode isolation is the only reliable way to change modes mid-session: mid-session persona switches invalidate the whole prefix cache, tail personas are ineffective (P6), and the native subagent inherits this persona.

One preset, auto-matched per model. There is no Pro/Flash split to configure: personaFor(mode, modelId) reads the session's model route and selects the measured optimum automatically — Pro → w6c (spec sentence + classify instruction, no anchors; 24/24 = 100% routing, P24), Flash → w7 + recall/anti-runaway anchors (96% routing; 100% single-task completion, P23). The model is fixed at the first request (path commitment), so the persona is locked for the session; switching the GUI model starts a new session with the matching configuration.

Depth-adaptive guidance (v20, thinking efficiency). Per-message guidance is dispatched by task complexity (isComplexTask: length or architecture keywords):

• simple tasks → fast-convergence guide (P30: 1 step, zero waste);
• complex tasks → decision-closure deep guide: "Think deeply about the architecture, edge cases, and integration points. Do not spend reasoning on the environment or tooling. Produce when your information is complete. End each reasoning block with a decision or an information need." — P30: depth +12% AND faster convergence (8.0 vs 8.3 steps), 3/3 completion.
• Rumination (environment suspicion / re-confirmation) is suppressed by the anti-runaway anchor: measured 0.0-0.3% of reasoning tokens.

Tests

node --test router.test.mjs   # 11 tests: classification, bands, personas, plan-section survival

Files

• preset/agent.cordis.yml — full rc.6 Standard composition + router row
• preset/router-core.mjs — pure routing logic (zero deps, unit-testable)
• preset/router-bootstrap.mjs — Cordis plugin (zero external imports)
• router.test.mjs — unit tests
• docs/paper.md — the theory + experiments write-up
• docs/experiments.md — full data tables

Evidence & attribution

• Trajectory trigger matrix, dual-model matrices, and the 21-point phase probe: dsh-probe (this repo's sibling scripts live in the paper's appendix tables).
• Project2 evaluation data: xiaobright/modeltest (V4.1b, frozen) — minimal 99/96, standard 91, PTC 92, anchored-standard 98/99.
• Two-phase anchoring preset: xiaobright/dsh-anchored-standard (MIT). The router's first-turn anchoring is a plugin-level port of its tool-bootstrap mechanism.
• DeepSeek Harness official minimal preset snapshot (sends the exact RL prompt and schemas test) — the spec persona and the RL-alignment claim.

License

MIT. preset/agent.cordis.yml derives from the DeepSeek Harness Standard preset (MIT); original attribution in NOTICE.