Hermeslike-Mixagent-MoA

中文 · English

hermeslike-MoA

Hermes-style Mixture-of-Agents (MoA) request aggregator — runs every routed LLM call through a router → parallel reference advisors → aggregator streaming pipeline, with dual-provider failover and a V4P triple-gate cost guard.

hermeslike-moa is a DSH plugin (dsh.bundle). It registers a dsh-llm LlmAdapter (provider route deepseek-moa, plus go-moa when an OpenCode Go key is configured); every model request routed to those providers runs the MoA pipeline:

user request
  → V4F router (one call, JSON verdict: does the task warrant V4P?)
  → reference advisors ×N in parallel (V4F; Hermes-style advisor system prompt;
     tool calls/results folded into plain text; tool results head+tail preview;
     reference outputs redacted; synthetic user judgement instruction appended)
  → aggregator (model: explicit user choice first → router → V4P triple gate)
  → streaming passthrough (standard dsh-llm StreamChunk)

The three model entries

The model picker offers three models:

Any unknown model id (outside the three above) is treated as flash direct — it never silently triggers the expensive MoA path.

Highlights

• mix model category: the model picker now lists MoA Mix · 模型混合 (deepseek-v4-mix) — selecting it pins the full MoA pipeline (a "simulate selection" entry); the aggregator model still follows settings/router; flash/pro remain explicit aggregator choices.
• Hermes-faithful advisory view (lib/reference.js, modeled on Hermes moa_loop.py): drops the acting agent's system prompt, renders tool_calls as [called tool: name(args)], folds each tool result (head+tail preview, 4000-char budget) into the preceding assistant turn, emits zero tool-role messages (strict-provider safe), appends a synthetic user judgement instruction so the view ends on user.
• Redaction: centralized secret shapes (sk-/Bearer/JWT/private keys/DB URLs) + safe email/phone regexes (delimited, never mangling SHAs/IDs/dates).
• Context trim: rough token estimate, drop oldest frames, keep user-first invariant, keep trailing synthetic instruction.
• Dual channel: official (DeepSeek pay-as-you-go) + go (OpenCode Go subscription); go-preferred whole pipeline, auto-switch to official after 3 consecutive failures.
• V4P triple gate: budget multiple (look-ahead pool = F spend×ratio − P spend) + per-window cap + min interval.
• New-contract accounting: dsh-llm disjoint TokenUsage (inputTokens = uncached input; cache read/write separate), rolling-window cost.

Install

# bundle: enters the profile layer stack (restart web)
dsh plugin --profile web add github:beimianism/Hermeslike-Mixagent-MoA

# or local dev path (use your own)
dsh plugin --profile web add /path/to/hermeslike-moa

Public @deepseek-ai/* and cordis deps are intentionally undeclared (injected by the profile pnpm closure).

System settings (recommended)

The plugin ships its own web settings page (Settings → Hermeslike MoA, rendered by its client half — the official Models page disables third-party provider editors). Every field has a ⓘ hover tooltip, and each API-key field shows a live health diagnostic (keyHealth):

• Enable / disable the MoA aggregator.
• API keys (type them right here, robustness-first): the form has two write-only password fields — DeepSeek API key and OpenCode Go API key. Type a key directly in the settings UI and it takes effect (stored locally in settings.yaml, never echoed back, stripped on the wire); leave blank to fall back to the env refs (apiKeyEnv/goApiKeyEnv, resolved via credentials/environment). If both are empty: missing official key → requests fail 401; missing Go key → Go channel just disabled. Health shows "✓ valid / ✗ malformed reason" live.
• Model mix:
  • reference advisor count (1–8, default 3)
  • reference temperatures (in order)
  • aggregator model: auto (router decides) / flash / pro (V4P-gated) — the Hermes aggregator-slot equivalent
  • prefer Go channel: OpenCode Go subscription first; auto-switch to official after 3 consecutive failures
• V4P triple gate: budget multiple / per-window cap / min interval / degrade-to-flash on gate block
• Privacy: advisor-output redaction level
• API keys are configured in the credentials seam (referencing env var names like DEEPSEEK_API_KEY, never literal secrets)

The settings document holds the hermeslike-moa: section; the plugin row config: in cordis.patch.yml acts as the composition base (user overrides win).

Configuration (composition base)

These fields default from the plugin row config: in cordis.patch.yml; user settings override them:

Example:

- insert:
    - id: hermeslike-moa
      name: 'hermeslike-moa'
      config:
        apiKeyEnv: DEEPSEEK_API_KEY
        goApiKeyEnv: OPENCODE_API_KEY
        ratio: 0.5
        refCount: 3

Layout

hermeslike-moa/
├── package.json
├── cordis.patch.yml
├── index.js               # apply(): config parse + registerAdapter
├── lib/
│   ├── adapter.js         # MoaAdapter extends LlmAdapter (dual + failover)
│   ├── pipeline.js        # MoaPipeline: route → refs×N → aggregate → stream
│   ├── router.js          # V4F router + keyword fallback
│   ├── reference.js       # Hermes advisory view: render/truncate/redact/trim
│   ├── wire.js            # HTTP one-shot/stream + TokenUsage normalization
│   ├── ledger.js          # spend ledger + V4P triple gate
│   ├── prices.js          # price table
│   ├── prompts.js         # three system prompts
│   └── settings.js        # settings schema (web settings form) + resolution
└── README / README.en

License

MIT