🧩 dsh-moa-native
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.
English | 中文说明
dsh-moa-native banner
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.
Quick start
- 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).
- Save (hot-reloads), then select a moa-native entry in the session model picker —
default or any recipe you saved.
- 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:
- F · full mirror — references see exactly what the acting model sees; KV prefix cache hit rates above 90%
- R · trimmed view — Hermes-style tool-result trimming with input budgets
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 | R trimmed view
"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
