Introduction¶
DSH’s plugin ecosystem allows extending a single LLM call into a more granular request pipeline. hermeslike-moa is a DSH plugin maintained by beimianism with an MIT license. It injects a Hermes-style Mixture-of-Agents (MoA) request aggregation pipeline into DSH’s LLM calls, featuring dual-channel failover and V4P triple-gate cost protection.
Overview¶
hermeslike-moa registers a dsh-llm LlmAdapter with a provider route of deepseek-moa; when configuring the OpenCode Go key, it appends go-moa. The model selector offers three model options:
deepseek-v4-mix: Forces the full MoA pipeline, including router, reference advisors ×N, aggregator, and streaming output.deepseek-v4-pro: Direct connection to V4 Pro.deepseek-v4-flash: Direct connection to V4 Flash.
Unknown model IDs are handled as direct connections to Flash to avoid silently triggering the expensive MoA.
Core Features¶
MoA Pipeline¶
Reference advisors execute in parallel, replicating the Hermes advisor view: tool calls/results are folded into plain text, head+tail are trimmed, no tool-role messages are emitted, and a synthetic user judgment instruction is appended at the end.
Desensitization and Context Trimming¶
The plugin provides centralized key shape masking and email/phone regex desensitization. Context trimming estimates the token budget, discarding the oldest frames, keeping user-first, and preserving the synthetic instruction at the tail.
Dual-Channel Failover¶
The plugin supports both official and Go channels. Go takes priority for the entire pipeline; automatic switch to the official pipeline occurs after three consecutive failures.
Cost Protection¶
The V4P Triple Gate includes budget multiplier, window limit, and minimum interval. When the gate blocks a request, it can downgrade to Flash; fallbackToFlash defaults to true, downgrading to V4F.
The plugin uses dsh-llm’s separated TokenUsage and rolling window cost accounting.
Settings Page¶
The plugin is semi-rendered in the web system settings page, featuring hover tooltips and key health diagnostics. Setting changes take effect in real-time without requiring a restart.
Installation and Enabling¶
- Install:
dsh plugin --profile web add github:beimianism/Hermeslike-Mixagent-MoA
Bundle installation enters the profile layer stack; restart web to take effect.
- Local Development Install:
dsh plugin --profile web add /path/to/hermeslike-moa
/path/to/hermeslike-moa is a placeholder path that needs to be replaced with the actual path.
- Public
@deepseek-ai/*andcordisdependencies are deliberately not declared; they are injected by the profile’s pnpm closure.
Typical Usage¶
- Select Model
Select deepseek-v4-mix to use the full MoA pipeline; select deepseek-v4-pro for a direct connection to V4 Pro; select deepseek-v4-flash for a direct connection to V4 Flash.
- Configure MoA Parameters
In the settings page under ‘Settings -> Hermeslike MoA’, enable/disable MoA aggregation, and configure reference advisor count, reference temperature, aggregation model, thinking intensity, reference fanout rhythm, show progress, and priority Go channel.
- Configure API Key
API keys can be filled directly with DeepSeek API Key and OpenCode Go API Key; leaving them blank falls back to apiKeyEnv/goApiKeyEnv environment variable references.
Key field values are stored locally in settings.yaml, not echoed back, and stripped from the wire. Missing official key causes a 401 request; missing Go key only disables the Go channel.
- Configure V4P Gate
Configure V4P Gate: budget multiplier, window limit, minimum interval, and downgrade to Flash when the gate blocks.
- Configure Composite Base Layer
In the plugin row config in cordis.patch.yml, set the default value for the composite base layer, with user settings taking priority. Setting documentation is stored in the hermeslike-moa: section of the settings documentation.
Applicable Scenarios and Notes¶
The following scenarios are a good match:
- Want to extend DSH’s LLM calls into multi-reference with aggregated output.
- Need official and Go dual channels with switching based on failure count.
- Need to limit V4P calls, budget, and intervals to avoid accidentally triggering high-cost paths.
The plugin runs with the current dsh process permissions. Source code and license should be checked before installation. The license is MIT. The directory page is an independent site with no official affiliation to DeepSeek or Huanfang.
Links¶
Directory page: https://www.skillhub.cn/plugins/beimianism/Hermeslike-Mixagent-MoA
GitHub: https://github.com/beimianism/Hermeslike-Mixagent-MoA