HarnessMM DSH Plugin
Copyright 2026 Xiuxiu-6 contributors

This product is a derivative work of HarnessMM
(https://github.com/YuhuaJiang2002/HarnessMM), which is licensed under the
Apache License, Version 2.0. The upstream HarnessMM NOTICE reads:

    HarnessMM
    Copyright 2026 PhiAgent-0 contributors

    This repository extracts the HarnessMM implementation from the PhiAgent
    research workspace maintained by Yuhua Jiang. The Python package is named
    harnessmm; legacy phiagent imports and original protocol identifiers remain
    supported. Deployment paths and hostnames were replaced with placeholders
    for this public release.

    The project source retains its Apache License, Version 2.0 terms. Upstream
    agent runtimes, benchmarks, datasets, model weights and separately installed
    dependencies are governed by their respective licenses and access terms.
    No upstream model weights, benchmark datasets or native runtime checkouts
    are redistributed in this repository.

Changes from the upstream work
------------------------------

This plugin is not the HarnessMM research codebase. It is a DeepSeek Harness
Host plugin that re-implements the Guarded Decision Reuse (GDR) and Certified
Callback Substitution (CCS) contracts for a general-purpose harness. Relative
to upstream:

  * It is derived from the upstream modules
    `integrations/deepseek_harness/sar_v2.mjs` and
    `integrations/deepseek_harness/cot_jev.mjs` (the `ExecutableSubdag`
    controller and the v6 authored-bank contract), not the full repository.
  * The upstream token-based admission rules (slow-token priors, the
    `net > 1024` threshold, TRAIN-partition medians) are omitted. Admission
    is structural instead: a minimum literal chain length plus a public-byte
    cap plus deterministic receipt-guard satisfaction.
  * The upstream fast-selector path, which posts a one-token classification
    request to a loopback-only provider endpoint and accounts response
    logprobs, is omitted. This plugin never makes an extra model call: a
    request is either answered from an already-decided action or it is not.
  * Provider and model pinning (`options.model !== this.model`) is removed,
    because the harness can change models mid-session.
  * The Python research runners, benchmark controllers, recursive
    self-improvement (RSI) loops, configurations and datasets are not
    included and are not required by this plugin.
  * Persistence, session projection and the compaction-time controller message
    rewrite are not implemented; the authored plan stays in the conversation
    as ordinary assistant text.

No HarnessMM benchmark results, measured token savings, or published numbers
are reproduced here or claimed by this plugin.