DSH Adaptive Reasoning 0.2.0

Our standalone DeepSeek Harness plugin chooses reasoning automatically for each request. It uses the existing model. No per-prompt switches, classifier model, background inference, GPU warm-up, power-setting changes or recurring jobs.

This is a conservative first usable preview, tested with DSH 0.1.5-rc.1, Cordis 4.0.2, and the existing Qwen3.8 27B GSQ model. It recognises task families with local rules; it is not a trained or infallible difficulty estimator.

What happens automatically

Task	Reasoning	Tools
Clear rewrite, translation or proofreading with supplied text	Off	Hidden and blocked for that step
Short summary of supplied text	Low	Hidden and blocked for that step
Longer supplied-text summary	Medium	Hidden and blocked for that step
Ordinary explanation or comparison	Medium	Available
Prompt improvement	High	Text-only when the source prompt is supplied
Complex, consequential, ambiguous or unrecognised work	High	Available
Continuation after tools	At least medium	Restored
Tool failure in the current turn	High	Available

Prompt improvement deliberately retains high reasoning: faster modes invented requirements in evaluation. We did not promote that task family on latency alone. Clear text work gets a short instruction to preserve facts, uncertainty and scope. The plugin does not guarantee semantic correctness or validate every final answer.

The current bundle covers only augmentor-linux-product and mx-5090-tray/Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp-262k-dual. Other tasks and models pass through. Provider, model, sampling and output allowances are preserved.

Installation

Download the 0.2.0 preview or install its ready-made package:

dsh plugin --profile web add https://github.com/ManoloRemiddi/dsh-adaptive-reasoning/releases/download/v0.2.0/dsh-adaptive-reasoning-0.2.0.tgz --ignore-scripts --config.auto-install-peers=false

Configuration is required on other installations. The shipped allowlist is specific to the tested Augmentor preset and Qwen route. It does nothing for other presets/models until configured. See setup for your own model. This is a preview, not a universal automatic difficulty estimator.

To package a source checkout instead:

npm pack
dsh plugin --profile web add /absolute/path/dsh-adaptive-reasoning-0.2.0.tgz --offline --ignore-scripts --config.auto-install-peers=false

The tarball needs no runtime dependency downloads: it consumes services from the installed DSH host. CLI installation adds its bundle once. New bundle manifests are picked up on the next DSH start; installing the package does not restart DSH. Finish running tasks before restarting. No per-prompt adjustment is needed after activation. The local deployment record is under ignored outputs/install-state.json.

To remove:

dsh plugin --profile web remove dsh-adaptive-reasoning --config.ignore-scripts=true --config.auto-install-peers=false

Or set enabled: false for the plugin row and reload. Removing/disabling the plugin restores normal reasoning selection and tools. It never rewrites the saved provider default. The native picker still displays the saved selection, not an Auto badge; request/header records the actual effort used.

Evidence without background work

Every participating model request records a reason in adaptive-reasoning/decision, then timing and character counts in adaptive-reasoning/measurement. These events copy no prompt or answer text. They measure time from request preparation, not from clicking Send, and do not measure electricity. Cancelled requests may lack a completed measurement.

To read a summary from existing logs:

npm run report

The report inspects the 100 most recently modified session files, performs no inference and runs only when requested. Task mix and server load affect the results.

Configuration

• enabled: boolean, default true.
• textOnly: boolean, default true.
• presets: exact preset allowlist; empty means inactive.
• routes: exact provider/model pairs with an efforts mapping for off, low, medium, high. The existing Qwen maps high to xhigh.

Automatic mode owns per-request effort on allowlisted routes. Explicit requests to think deeply remain high. Unsupported effort strings fail DSH's adapter validation before inference. Unrecognised languages retain high reasoning.

Engineering and validation

Uses supported DSH lifecycle/request hooks, an assembly transform and a tool guard. Tools are restored on the next step; invented actions in a text-only step are blocked, then the normal loop can recover with more reasoning. Known PTC and structured-output contributions opt out of tool filtering. No automatic replay of completed actions or permission relaxation is introduced.

npm test
npm run test:integration
npm pack
node test/install-proof.mjs

The integration test uses the installed DSH loop and pi-ai adapter against a local deterministic HTTP endpoint. It tests sequential and concurrent sessions, wire-level effort, tool blocking/recovery, telemetry, and clean removal. It creates no persistent user chats. Set DSH_INSTALL_ROOT if DSH is installed elsewhere. The install proof uses and removes a disposable DSH home.

An explicitly opted-in, finite live suite is available with ADAPTIVE_LIVE_TEST=1 npm run test:live. It sends five synthetic writing tasks and one matched baseline to the already-running loopback model through the real DSH loop, with no actual tools registered. Private historical experiments and live outputs are excluded from the distributable package.

See 0.2 validation, initial research, and changes.

Support and related plugins

Report issues with your DSH version, model and a synthetic example. Do not include credentials or private session logs. Browse the DeepSeek Harness Plugins collection.