DeepSeek Harness Plugin

DDL095/dsh-moa-native

Stars ★ 1 Downloads (30d) 2,567 Category AGI Architecture Exploration Added 2026-10-29 npm dsh-moa-native

Native generation-time Mixture-of-Agents: every turn, reference models run the same session context in parallel as pre-run drafts and one aggregator executes with full tooling; config-as-model recipes, F mirror / R trimmed views, enforced subagent gating (dual-shape session indexing), derived per-recipe vision (vision refs relay image reads to text-only aggregators), tiered cadence with separate steering and patrol rosters plus frozen reuse between beats, degeneration guard dropping degenerate reference output, anonymized model face (failure notices carry no provider identity), and cache-friendly guidance placement (the rehearsal record is pinned where it was produced instead of appended to every request, so the cached prefix covers it and each step pays only for what was added since - on by default across every cadence).

Install

# from npm (prebuilt)

dsh plugin --profile web add dsh-moa-native

# from a prebuilt release tarball

dsh plugin --profile web add "https://github.com/DDL095/dsh-moa-native/releases/download/v4.5.0/dsh-moa-native-4.5.0.tgz"

# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)

dsh plugin --profile web add github:DDL095/dsh-moa-native

Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).

README

Native Mixture-of-Agents for DeepSeek Harness — generation-time fusion: reference models produce pre-run drafts in parallel every turn; the single acting model executes with full tooling.


What it solves

A single model has blind spots. In long sessions you want a few models reviewing the situation behind every reply — without switching tools, copy-pasting, or giving up tool use. dsh-moa-native internalizes that loop: picking one model entry runs the whole MoA pipeline for the session.

user message ──► refs consulted in parallel (private drafts, mutually unaware)
                      │
                      ▼
                guidance injected as a trailing message (never pollutes session
                history, never breaks KV prefix caching)
                      │
                      ▼
                aggregator (acting model) executes with full tooling ──► reply

Recommended mode (advanced)

F mirror + 3-4 heterogeneous blank-injection refs + a single aggregator + the parallel-plan discipline — one glance:

Recommended mode

  • Every ref uses the __blank__ injection (zero instruction, pure continuation — pure model-family signal, no persona noise);
  • The aggregator uses agg_forrefsblank (candidate action plans: where a draft agrees with your own judgement it is worth more weight, disagreements are judged independently, all actions executed by the aggregator itself);
  • The factory default stays Hermes-aligned (review-style refs with the default injection) — start there, switch to the recommended combo once familiar.

Full part-by-part, parameter-by-parameter walkthrough: architecture.svg.

Tiered rhythm (v4.4.0)

Heavy models steer every m steps, flash models patrol every n steps, and every other step reuses the last beat — free while that beat is still cached (a cache miss refills with the patrol roster).

Tiered rhythm

  • Cadence is the fourth tier — user_turn (once per user turn, cheapest), per_iteration (every step), every_n×N (skip N steps after each fan-out, so the period is N+1 steps), and tiered (the rhythm below);
  • Two roles, two periods: the steering beat asks heavy models (@max / @xhigh) whether the direction is still right — every m steps; the patrol beat asks flash refs whether the execution has drifted — every n steps;
  • Non-beat steps are frozen: the previous beat's guidance is reused verbatim, at zero ref cost;
  • One beat, one job: when a step is due for both, steering wins and the patrol beat is suppressed;
  • Where the rosters live: cadence_tiered.steering (the heavy models) and cadence_tiered.patrol (the flash models — omit it to keep patrol following refs); under the tiered cadence a recipe's own refs are just the patrol fallback pool;
  • The model never sees the rhythm itself: beat vocabulary, failures and scale stay host-side (trace + console only) — the injected text carries no beat words, no failure notices and no counts (block count remains a first-order proxy, as the design doc states honestly).

Install

dsh plugin --profile web add dsh-moa-native

Restart the instance and refresh the browser. The config file lives at <DSH_HOME>/profiles/<profile>/moa-native.config.json — each profile keeps its own lineup (per-profile isolation since v4.5.3; earlier releases always read the web profile's file, so a Desktop-profile panel could overwrite your web config).

Quick start

  1. Settings → MoA-Native console: pick 3-4 reference routes (flash tiers work well — fast and cheap) and one aggregator (the model you want to keep). This step is mandatory on first install — the factory lineup is only an example and must point at your own configured providers/models. Since references are framed as the aggregator's own pre-run drafts, you can skip a ref that shares the aggregator's model (saves one route). Seeing a model tagged (表外 / off-catalog) in a dropdown? That entry is referenced by the current config but absent from this environment's model catalog — typically a lineup migrated from another profile/machine, or the factory example pointing at providers you don't have. The console's top warning banner lists all off-catalog entries at once; re-pick those routes from in-catalog models and the banner clears itself.
  2. Save (hot-reloads), then select a moa-native entry in the session model picker — default or any recipe you saved.
  3. Chat normally. Fusion happens every turn; /moaR /moaF temporarily switch the reference view (lineup, aggregator, and cadence stay with the current entry; the console shows the live slash preview). Subagent sessions stream straight through the aggregator by default (zero fanout) — switch via the subagent gating dropdown (7 tiers: 3/5/8/10/15/20/30).

Highlights

Config as model — default and every recipe appear as model entries on the provider page; a recipe is a full preset (lineup, mode, cadence, params).

Three-level model picker (identical to the session UI) — provider → model → effort for every route.

Injection modules — each route's trailing note is a visible, editable text module: built-in default (Hermes-style), blank (zero injection), or any custom one. No hidden personas, no output-length contracts.

Two reference modes — the default is F; pick R when the input bill matters more than perfect fidelity:

  • F · full mirror (default) — references see exactly what the acting model sees; KV prefix cache hit rates above 90%. Costs the most: every reference route bills the full system + tools + history on every beat, so the input spend scales with conversation length. The recommended lineup (F + blank-injection flash refs + tiered cadence + anchored guidance) exists precisely to keep this bill low — flash-tier refs, frozen reuse between beats, −36% uncached tokens vs the trailing form.
  • R · trimmed view — Hermes-style tool-result trimming; drops the system prompt, folds tool results. The economy lane. Its overall window is per-model dynamic (4.5.2): each route's budget comes from its own registered context window — (window × 0.9 − 8192) × 1.5 chars, the whole fleet flying on the minimum, so a 64k ref tightens the view and a 1M ref sees more. Tool results truncate at Hermes' own 4,000 chars (ref_tool_result_budget); ref_input_char_budget, when set, caps the dynamic budget. Unknown windows ride the conservative 128k baseline (warned once; fill contextWindow in DSH-MoA-model-registry.json to enable exact adaptation).

Subagent gating — MoA applies to the main session by default; subagent sessions pass straight through the aggregator, with an optional every-N-steps upgrade. Design rationale: SUBAGENT-GATING-DESIGN.md.

Run traces — four-level drill-down: sessions (workspace-isolated) → turns → per-model cards → full advisory text. Cold-start backfill from daily zstd archives; inputs are never persisted.

Fail-loud — a failing reference route is reported as-is, never silently replaced.

Theory positioning

Comparison with the MoA paper (arXiv:2406.04692), Hermes /moa, and community dsh-moa plugins — inheritance, superseding forms, and honest boundaries: THEORY-POSITION.md.

Configuration

~/.dsh/profiles/web/moa-native.config.json (panel saves hot-reload):

{
  "refs": [                                 // direct lineup; or use named recipes
    "provider-a/model-flash@high",          // string = default injection
    { "model": "provider-b/model-x@max", "injection": "__blank__" } // object = explicit injection module
  ],
  "aggregator": "provider-c/main-model@max",
  "aggInjection": "agg_forrefsblank",       // parallel-plan discipline (recommended); default agg_default
  "mode": "F",                              // F full mirror (default, zero info gap, highest input cost) | R trimmed view (economy)
  "ref_tool_result_budget": 4000,           // R only: per-tool-result truncation budget (chars; Hermes lineage default 4000)
  "ref_input_char_budget": null,            // R only: cap on the per-model dynamic budget (unset = purely dynamic; window-derived)
  "cadence": "user_turn",                   // user_turn | per_iteration | {every_n: 2} | "tiered"(see above)
  "subagent_policy": "passthrough",         // subagent gating: passthrough | {every_n: 5} | full
  "injections": { "critic": "Your first duty is to scrutinize…" },
  "recipes": { "my-lineup": { "refs": ["..."], "aggregator": "...", "mode": "F" } },
  "prices": { "provider-a/model-flash": { "in": 0.2, "out": 0.8 } },  // optional cost estimates
  "trace": true
}

HTTP endpoints

Endpoint Description
GET /moa-native/panel.json Config + model tables
GET /moa-native/trace.json?limit=&ws= Run traces
GET /moa-native/trace-archive.json?date=YYYYMMDD Daily archive read
GET /moa-native/efforts.json?provider=&model= Effort tiers

Design boundaries

  • guidance is a per-request injection and never enters session history — refs originals are invisible in the session UI by design (see the trace panel instead);
  • references are pure advisers with no tool execution; the acting model decides what to adopt;
  • complementary to "competitive draft" MoA (multiple full answers then a verdict): this plugin targets per-turn continuous quality lift.

License

MIT

Content from the project README on GitHub ↗

Links

More in this category

View the whole category →

Community comments

Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.