dsh-router-standard

Task-aware reasoning-mode router preset: quantifies DeepSeek V4's measured "discontinuous, phase-transition" behavior along the persona axis onto stable regions (spec / react / weak), and injects the matching persona and tool surface on the first turn. Part of the dsh-routing-suite suite, and developed alongside dsh-super-injector (restart-free runtime injection).

📄 Research-history notes: docs/statement.md (errata statement) | docs/apology.md (apology letter)

---

Current Status (v1.19.1, 2026-08-24 · R&D line, not yet released)

Router Standard is on the "strict workflow" R&D line (refer to CHANGELOG; the version-line file is router-bootstrap-v34.mjs):

• Strict-phase workflow: completion-signal-driven promotion (0→1 requires clarification/plan, 1→2 requires plan lock, 2→3 requires delivery self-check), phase 0 mandates alignment (ambiguity first ask_user_question, complex tasks first todo_write); "tool name/text triggers phase skip" has been removed.
• Task echo: stageText now includes Task: <首条真实用户消息> — the model clearly sees what it is working on every turn, without losing focus.
• Progressive disclosure: phased unlocking + two-tier pre-release + direct-reach semantics + full open at delivery; tools_catalog full index / tools_help complete schema; all:true has been removed (no full-open exit, and unlocked-out tools do not enter the field of view).
• Standard-mode baseline: native direct calls (no PTC/run_code wrapping, no full SDK section, and the attention tax on the tool surface is substantially reduced).
• Built-in page verification: dev_page_check = screenshot + DOM smoke + console/pageerror/title/selector/scale; {js:…} mode = local JS engine (zero external node dependencies); external evidence is a first-class citizen.
• Description ⇄ behavior alignment: presentation=native self-check, phase copy states only the truth, platform facts (on win32, the Git Bash private shell seam is authoritative).
• Proactivity guidance: Proactivity standing section (self-check signals, not a stop command); the stress sensor has been retired.

Companion presets: router-react (v17) / router-spec (v10) (two execution presets based on standard mode, both supporting first-turn image reading); the router-pro line has been retired and removed. Full evolution: see CHANGELOG.md and docs/STANDARD-PLAN.md.

---

【Archived · v0.3 Era】 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:

routing mode	first request	thinking shape
standard (standard routing preset)	classified persona (spec/react/weak) + full prompt sections + banded first-turn tool surface	act per classified band: react produces directly, spec reads before editing, weak routes internally (close-range guidance each turn)
spec (spec routing preset)	classified persona (spec/react/weak) + full prompt sections	Thunderstorm deep thinking: an ultra-long first-turn chain of thought (101K of reasoning with 0 actions is its characteristic, not a defect)

Choice: install either of the two presets (Router Standard / Router Spec, see Usage). dev_router_status displays the current routing mode.

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:

band	mode	measured behavior
spec	0 – 0.19	stable plan-collective (We trajectories, let-me ≈ 0)
mixed	0.2 – 0.49	transition trap: unstable mixing of We/The/Let
react	0.5 – 1.0	stable doer (The/Let first-person, we ≈ 0) — 11 mode values behave alike

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

No-restart installation (recommended): after installing dsh-super-injector (see the suite's scripts/install-injector.ps1), changing preset code no longer requires renaming files / restarting:

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

Notes (hard-learned lessons):

1. ESM cache: the loader caches modules by URL — overwriting file contents in place does not take effect (code changes require dev_reload_preset or a filename change).
2. First session must be new: the routing mode is locked on the first request (path commitment); switching the GUI model mid-session or changing configuration does not affect a running session.
3. Subagents do not route: parentSession sessions skip the routing layer (community #5 fix), and shell-less subagents no longer crash.
4. Guidance injection channel (v0.3.0): near-field guidance goes through agent/pre-step, injecting one item per real user message (weak mode); from rc.6 on, session/event receives no events under standing scope, so guidance in older versions was dead.
5. First-turn real classification (issue #3/#13): first-turn routing reads the raw message text from agent/inbox/claimed and classifies it via classifyTask — the first turn is a real band (no more weak fallback).
6. Self-removal uninstall: dev_uninject_plugin --self=true can uninstall the injector itself (keeping the assembly chain, auto-restored on restart) — used to verify the installation closed loop.

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.