HaoyueQin/dsh-better-reasoning-effort

DSH Better Reasoning Effort

English | 中文

Reasoning-effort and input-modality editing for third-party models in DeepSeek Harness, edited right inside the official Models page card — plus a quick reasoning-effort slider inside the official composer model menu (adapted from HanaAyane's dsh-reasoning-effort, see Acknowledgements).

Why

The llm-pi-ai adapter natively supports per-model reasoningEfforts and input declarations, but the official Models page editor deliberately keeps both fields out of reach. As a result, third-party models get no thinking-level picker in the composer, only the official DeepSeek API can set reasoning effort, hand-declared models are treated as text-only, and configuring any of this meant hand-writing settings.yaml blocks. This plugin brings both configuration surfaces back into the UI: edit inside the official model editor card, plus one-click auto-adapt.

Features

• In-page editor — an "Options" block appears under each model row's disclosure on the official Models page (reasoning effort, input modalities, endpoint compatibility), committing with the card's own Save; changes stay pending until then, and Cancel discards them with the card's fields.
• Auto-adapt — one click fills recommended levels, wire spellings and modalities from a built-in knowledge base (65 entries across 15 vendors — see Supported models), wire-protocol inference, and a same-origin probe of the provider's raw /models listing, every suggestion labeled by confidence; reference capacities show as read-only hints you copy yourself.
• Auto-fill — models with a curated official knowledge-base match are filled at boot and when added mid-session (opt out via autofill: false / modalityAutofill: false); unknown models stay unchanged until the user explicitly accepts Auto-adapt.
• Three intents — all levels off = unset (back to inheritance); only off = disable reasoning; levels armed = write the declaration.
• Catalog-safe writes — routes using DSH's official modelOverrides keep their installed catalog; only the edited model override is written.
• Composer slider — the official model menu's body is replaced by an upstream-style effort slider (drag / keyboard, optimistic commit with rollback), with one model row opening the official model list; the official trigger stays untouched, and model switches keep your level through a per-session memory chain (issue #4's per-model default effort outranks it across sessions).
• Per-model default effort — a "Default effort" picker on each model row, stored in the settings document; every new session starts the model there.
• Request headers & `user-agent` — a provider-card section edits the official headers field (masked, path-merged, Save-gated), and the plugin performs the user-agent override at the fetch layer per origin, because the official adapter reserves that name; same-origin /models probes are covered, conflicts are reported rather than guessed.
• Defensive injection — everything keys off the official page's DOM; if an official upgrade changes the structure, injection simply pauses and the official page is unaffected.
• Bilingual copy (中文 / English).

Supported models

The auto-adapt knowledge base carries 65 curated entries across 15 vendors (DeepSeek, OpenAI, Anthropic Claude, Gemini, Grok, Qwen, GLM, Kimi, Mistral, MiniMax, MiMo, Doubao, Hunyuan, Step, ERNIE — re-verified against official docs 2026-08/09, including vision-capable variants and no-effort-control families). The full table — match patterns, level → wire-spelling ladders, defaults, modalities, reference capacities — lives in docs/supported-models.md (generated from src/knowledge.ts, the authoritative source). Unlisted models are left unchanged by automatic fill; the user can still review a low-confidence Auto-adapt suggestion.

Install

Requires DeepSeek Harness `0.1.5-alpha.1` or later (per-line peer ranges; the 0.1.2-rc / 0.1.3-alpha lines are no longer supported — use plugin 0.3.7 there). Compiled and gated against 0.2.0-rc.2 and the 0.2.1-alpha line. The per-kernel seam re-checks behind this live in docs/compatibility-notes.md.

From npm, under the dsh web profile:

dsh plugin --profile web add dsh-better-reasoning-effort

DSH bundles pnpm 11.7.0, whose default minimum-release-age is 24 hours: an install started within a day of a release resolves to the newest version old enough and installs that one, while the plugin manager may still display the newer release. Name the version to install it without waiting — pnpm accepts a named version and records it as an exemption:

dsh plugin --profile web add dsh-better-reasoning-effort@0.5.2

The release page also carries the packed tarball, which skips both the registry and the release-age wait. It is prebuilt, so pnpm executes nothing and asks for no build approval. The URL names one version, so an update replaces it:

dsh plugin --profile web add \
  https://github.com/HaoyueQin/dsh-better-reasoning-effort/releases/download/v<version>/dsh-better-reasoning-effort-<version>.tgz

From GitHub source, for development and review rather than normal installs. pnpm runs the prepare hook to build lib/, and allows that only once an allowBuilds entry exists. The installer prints the exact key, which names the resolved commit: add it to the profile's pnpm-workspace.yaml and repeat the add. A new commit changes that key, so every release needs re-approval, and the commit-free key form needs a newer pnpm than the bundled 11.7.0:

dsh plugin --profile web add github:HaoyueQin/dsh-better-reasoning-effort

Or link a local checkout for development:

npm install && npm run build
dsh plugin --profile web add link:D:/Project/dsh-better-reasoning-effort

Restart dsh web and hard-refresh the browser.

Usage

1. Configure a third-party provider (API key etc.) on the official Models page.
2. Expand a model row: the editor block sits under the official capacity fields.

• Check levels (off / minimal / low / medium / high / xhigh / max) and fill the wire values (e.g. give high the spelling ultra, and the gateway receives ultra when you pick High in the composer);
• Toggle Image input under *Input modalities* to declare what the model accepts;
• Click Auto-adapt to fill recommended levels and modalities — reference capacities show up as read-only hints you copy into the official fields yourself;
• Any change is pending and lands when you press the card's own Save; Cancel (or a reload) discards it with the card's fields.

1. On a compatible protocol, the *Endpoint compatibility* section appears at the bottom — thinking budget field / vLLM priority on openai-completions, max_output_tokens handling on openai-responses.
2. All levels off + Save = unset the declaration; only off checked + Save = disable reasoning (false); *Clear declaration* + Save = back to inheriting the provider default.

Declared models are immediately selectable for reasoning effort in the composer, and image-declared models accept attachments end to end.

Configuration

Optional on the plugin's profile row (values shown are the defaults):

- insert:
    - id: dsh-better-reasoning-effort
      name: dsh-better-reasoning-effort
      config:
        autofill: true          # auto-fill undeclared models at boot
        modalityAutofill: true  # whether the boot fill also covers modalities
        probeTimeoutMs: 15000   # /models probe fetch timeout
        bootRetryDelaysMs: [1000, 2000, 4000, 8000, 16000, 30000]
        defaultGuard: true      # map effort-less calls on forced-thinking
                                # ladders to the vendor default

How it works

Browser (lib/client.js)                  Host (lib/index.js)
├─ DOM injector                          └─ Auto-fill
│   MutationObserver on the models page      settings/document-updated →
│   → mounts EffortEditor in each            invalidates the host cache;
│     model row's disclosure                 browser idle pass fills models
├─ Composer injection
│   MutationObserver on the document
│   → ComposerSlider (root pane)
├─ EffortEditor (React component)             (knowledge base + inference)
│   level checkboxes / wire values /
│   input-modality toggle /
│   auto-adapt (zoned suggestions) / committed with the card's Save
│   └─ writes settings.mutate (llm-pi-ai)

• suggestEfforts() in src/knowledge.ts is the knowledge base + inference engine — a pure function shared by host and browser.
• reconcile() in src/client/injection/models-page-editor.ts locates model rows and mounts the editor; src/client/index.ts assembles the browser half, one module per seam in src/client/injection/.
• createEditorApi() in src/client/ops.ts writes the declarations via settings.mutate, preserving every other row field and retrying once on a revision conflict.

Development

npm run typecheck   # tsc strict check on src
npm test            # vitest: knowledge / inference / autofill / DOM injection / writing
npm run build       # lib/*.js + lib/client.js (module-loader bundle)

Compiled and gated against the 0.2.0-rc.2 and 0.2.1-alpha.2 official packages; see docs/compatibility-notes.md for the per-kernel records.

Known limitations

• Injection depends on the official Models page's DOM (aria-label/class); an official upgrade may pause injection until adapted — the official page is unaffected meanwhile.
• The auto-adapt probe route answers loopback and IP-literal hosts only (the core /api Host-allowlist discipline without trustedHosts), and never follows redirects — a gateway listing its models only behind a 30x simply yields no endpoint evidence; Auto-adapt falls back to the knowledge base and protocol inference.
• reasoningEfforts declarations are suggestions — what an endpoint actually accepts is up to its docs; tweak in the UI. The knowledge base is not exhaustive; families without an effort ladder carry no entry at all.
• Endpoint-compatibility switches are never auto-filled by design: they describe a gateway, not a model.
• The modality vocabulary follows pi-ai's core (text / image today); wider gateway support (PDF, audio, video) is recorded per family until the core vocabulary grows.
• Name-heuristic modality advice (vision-flavored ids) is deliberately low-confidence and labeled as such.
• Self-hosted relays: auto-fill pins supportsDeveloperRole: false on routes no official host claims (some upstreams reject the developer role); explicit values are never overwritten.
• Forced-thinking models (ladders without off, e.g. GLM-5.3): effort-less calls map to the vendor default instead of sending thinking: disabled — set defaultGuard: false to restore raw behavior. A route-level reasoning default covers the same calls without the guard: the adapter resolves options.reasoningEffort ?? profile.reasoning, so a route that names one never reaches the off-equivalent (and the guard steps aside for it).
• The guard's activity is reported inside the request-header section (N effort-less calls rewritten, and the model of the most recent one), so an installed-but-idle fallback is distinguishable from one that is working; the same report is served at /dsh-better-reasoning-effort/headers-config.
• Credentials inside `headers` are not redacted on disk: the read-only view masks them, but the settings document still holds them in clear text — treat it like an API key.
• The request-header section's edit-state detection reads an unofficial signal (the official row exposes no data attribute for its editor state); if an official build renames that class root, the section stops appearing — never breaking the page.
• Only one `user-agent` rewrite should be active: sibling header plugins land on the same layer; the plugin detects and reports known ones, but the last writer on the wire wins.
• The request-layer takeover relies on the official adapter creating a fresh SDK client per request — guarded by an end-to-end test that fails loudly if that changes.

Acknowledgements

The composer slider is adapted from dsh-reasoning-effort by HanaAyane (MIT) — thank you for the original work and the codex-style effort control idea. This integration keeps the upstream session-selection contract and slider interaction, with deliberate changes: a white round thumb only (no chibi-runner knob), the official model seat never replaced, and placement on the 0.1.5-alpha+ line as a reduced re-implementation over the harness wire contract. If you used the upstream plugin, remove it to avoid two effort controls on the same seat:

dsh plugin --profile web remove dsh-reasoning-effort

Activity

License

MIT