dsh-effort-router

English | 中文

A DSH (DeepSeek Harness) plugin: context-aware per-step reasoning-effort routing, plus reasoning levels for hand-declared llm-pi-ai custom models.

• Task-shape classification: trivial / standard / engineering / hard from the brief itself, not from tool-call counts
• Wish detector: a short delegation-style brief in the cue vacuum zone (帮我…, can you…) starts at the engineering floor — underspecified tasks get the benefit of the doubt
• Evidence escalation: tool errors, identical-argument retries, brief text volume, and recent tool-result volume lift the class (loudest signal wins, capped per task, sliding 32-step window)
• Hysteresis: upgrades apply immediately, downgrades wait for two consecutive quiet decisions
• Frugal signals: subagents, context pressure, or a session token budget demote one class
• Capability clamp: the scheduled tier maps onto the target model's advertised vocabulary (nearest rank, so gateway vocabularies like ultra work)
• Manual wins: a concrete level in the selector passes through verbatim; Auto hands the step to the scheduler

Why an Auto mask instead of a panel

The host's native model selector already is the per-session effort control. The plugin advertises an Auto level on every reasoning-capable model (injected into reasoning.efforts, resolved to a concrete wire level by the request interceptor before validation), so there is no second selector to keep in sync: Auto schedules, a concrete level passes through.

Custom models (llm-pi-ai)

Hand-declared custom models often ship no reasoningEfforts table, so the composer shows no effort selector at all. On startup and on settings changes the plugin scans the llm-pi-ai config and fills the default { low, high, max } table through a resolution gate: every table-less user-layer row is resolved with resolveModelInfo, and only models that resolve with NO thinking levels are filled — a model inheriting levels from its catalog is skipped, so catalog capability is never shadowed. An explicit table or an explicit false is never touched. The table deliberately has no off rung: an off selection followed by a tool call poisons the conversation for every later thinking-enabled request, and the router never routes to off either (you can still add off: null back per model in the editor — explicit intent is never overwritten). Models filled on explicit openai-completions routes also gain compat.supportsReasoningEffort: true, the switch that protocol needs to actually send the effort.

The per-model level → gateway wire-value mapping (e.g. high → ultra) is edited under Settings → Effort Router and written straight into the llm-pi-ai config, gated by the official schema, effective on the next request. pi-ai serializes the table natively; no plugin-side translation layer.

Settings

Settings → Effort Router:

• Enabled / default level / subagent default
• Bounds: allow max (on by default since 0.2.0 — hard tasks reach max; opt out here and maxFallback applies), max fallback, wish detector, poison recovery, trajectory hoisting, frugal demotion, context-pressure threshold, token budget
• Class → tier route table (trivial/standard default low, engineering high, hard max)
• Advanced: downgrade hysteresis, escalation thresholds (apply on plugin reload), custom cue dictionaries (strong / normal / cheap / wish)
• Custom-model editor: expand a provider, tick levels and enter wire values, or fill missing levels everywhere in one click

In-session visibility

• Composer badge (left of the model selector): the mode and effort of the session's last step (Auto · high), with class and reason in the tooltip. Read-only — control stays in the native selector.
• Trajectory hoisting (default on): each step's provider/model · effort is projected onto the NATIVE trajectory ledger rows (DOM projection anchored on data-trajectory-row-key — the assembly-layer route, no host package touched and no new view added). The anchor table is folded host-side from request/header + assistant/message + tool/call events and served on the state route. The chat flow itself stays annotation-free.

Safety and degradation

• The interceptor is one try/catch: any error passes the request through untouched (fail-open)
• Models without reasoning metadata never receive reasoningEffort — never an UNSUPPORTED_REASONING_EFFORT
• Automatic routing never emits off (thinking continuity on strict thinking-mode APIs). A history PROVEN poisoned — a reasoning-less tool-call turn exists AND the session's decision log recorded a manual off step — pins off with a warning until compaction removes the offending turn. Shape alone never pins: models like kimi-K3 emit no reasoning blocks before tool calls with thinking on (measured, 2026-10-10). Disable via poisonRecovery if your routes are all non-strict
• Session reads go through deriveMessages() + sessionProjections only; an empty sample raises an explicit warning instead of silently degrading

Authorship & maintenance

• Author: kimi-k3 (an AI model, designing and coding through DeepSeek Harness).
• Producer / repository owner: KEI.NEO (@neo805) — the human publisher responsible for this repository.
• Maintenance scope: a personal hobby (vibe-coding) project, published for sharing. No maintenance or updates are promised in response to external demand.
• Contribution policy: this repository has Issues disabled and does not accept Pull Requests. Forks are welcome under the MIT license.
• Disclaimer: the software is provided "as is", without warranty of any kind (see LICENSE). Evaluate it yourself before production use.

Thanks & license

MIT. The policy core (classification / escalation / hysteresis / tier mapping / signal collection / projection / store) is adapted from @neptune810/dsh-model-router (MIT © Neptune810); the capability layer and Auto-mask pattern follow dsh-thinking-levels (MIT © drscrewdriver); the custom-model default-fill idea comes from @hytime/dsh-thinking-effort (MIT © hytime).