dsh-honcho

Persistent memory for DeepSeek Harness, powered by Honcho.

dsh forgets everything when a session ends. This plugin gives it memory that doesn't: what you're building, how you like to work, what you decided last week and why. It reads the same ~/.honcho/config.json as the other Honcho integrations, so there is one place to configure all of them — and pointing two of them at the same workspace gives them one shared memory.

Install

dsh plugin --profile <name> add @honcho-ai/dsh-honcho

Then put your API key and name in ~/.honcho/config.json:

{
  "peerName": "your-name",
  "auth": { "apiKey": "${HONCHO_API_KEY}" },
  "hosts": {
    "dsh": { "workspace": "dsh" },
  },
}

Get a key at app.honcho.dev. HONCHO_API_KEY in the environment also works on its own — the config file is only needed to change defaults.

What it does

Injects what Honcho knows at the start of a session — your profile, a summary of this project's session so far, and the conclusions Honcho has drawn that are relevant to what you just asked. One API call, shaped to a character budget, refreshed as you work.

Captures every turn. User and assistant messages go to Honcho in the background, debounced, and flushed at turn boundaries, before compaction, and on shutdown. Secrets are redacted first.

Gives the model three tools:

/honcho shows status and a link to the session in the Honcho dashboard — see Commands below.

Configuration

Everything behavioral lives in ~/.honcho/config.json under hosts.dsh — the same file claude-honcho, codex-honcho, and the other integrations read. Root holds identity and connection; the host block holds behavior.

{
  "peerName": "your-name",
  "workspace": "honcho",
  "baseUrl": "https://api.honcho.dev", // bare host or …/v3 both fine
  "timeoutMs": 30000,
  "auth": { "apiKey": "${HONCHO_API_KEY}" },
  "enabled": true, // global kill switch
  // sessions may also sit here as a fallback; the host block wins

  "hosts": {
    "dsh": {
      "workspace": "dsh",
      "aiPeer": "dsh", // defaults to the host name
      "observationMode": "unified", // unified | directional
      "sessionStrategy": "per-directory", // see Sessions below
      "sessionPeerPrefix": true, // session names are <peer>-<dir>
      "sessions": { "/path/to/repo": "pinned-session-name" }, // pin a session
      "injection": {
        "sessionStart": ["directives", "summary", "peerCard"], // + representation
        "perTurn": ["userContext", "dialectic"], // userContext = representation + card
        "tools": true,
        "searchTopK": 10,
        "searchMaxDistance": 0.6,
        "maxConclusions": 15, // how many conclusions Honcho RETURNS
        "maxRenderedConclusions": 4, // how many survive filtering into the prompt
        "contextTokens": 1500,
        "cadence": { "dialectic": 5, "ttlSeconds": 300 },
        "dialectic": {
          "reasoning": "low", // minimal | low | medium | high | max
          "maxChars": 600,
        },
      },
      "capture": {
        "saveMessages": true,
        "saveToolUse": false, // one-line summaries of tool activity
        "writeFrequency": "async", // async | sync
        "noisePatterns": [], // additive to the built-in secret patterns
      },
      "messageUpload": {
        "maxUserTokens": 6000,
        "maxAssistantTokens": 6000,
      },
      },
    },
  },
}

Injection components

The two menus differ in cadence, not in what they can carry.

injection.sessionStart is injected once when a session opens: directives (guidance on using memory), summary, peerCard, representation.

injection.perTurn refreshes as you work:

• userContext — a fresh, prompt-scoped peer context blob: representation + peer card, retrieved using your current message as the search query so recall is associative rather than merely recent. It is a bundle, so it supplies both regardless of what sessionStart names. To get one without the other, name it in sessionStart and set perTurn: [] — at the cost of per-turn refresh.
• dialectic — a reasoned answer about you, run every cadence.dialectic turns and shaped by injection.dialectic. Nothing waits on it after the first turn, so a late answer reaches the next one.

Components the canonical schema names but this plugin does not implement — briefing, assistantContext, sessionContext — are reported at startup rather than silently dropped, as are schema keys it does not act on (showContents, statusline, globalOverride, granular observation, multiUser) and any key renamed since an earlier version. Nothing here is accepted and quietly ignored.

The plugin's own cordis.yml config carries plumbing only — configPath, apiKeyRef, host, enabled. Set host to run a credential-isolated profile ("dsh_work") against the same install.

Sharing memory with other integrations

Each integration defaults to its own Honcho workspace — dsh here, claude_code for claude-honcho — and a workspace is the isolation boundary, so by default they do not see each other's memory. Point them at the same workspace to merge them:

"hosts": {
  "dsh":         { "workspace": "shared" },
  "claude_code": { "workspace": "shared" }
}

Keep peerName identical across them too, since conclusions are stored per peer.

Sessions

Default: one long-lived session per project directory, named <peerName>-<dir>, matching claude-honcho. Pin a different name for any path with the root sessions map — an override always wins.

Prefer the wider scopes. Honcho's guidance is not to scope sessions too thin: the background Deriver needs a single session to accumulate enough material before it can reason well. git-branch splits a project's memory per branch, and per-session discards it on every restart.

Commands

Requirements

• Node ^22.19.0 || >=24.0.0
• A running dsh (developed and typechecked against 0.1.2-alpha.3)
• A Honcho API key, or a self-hosted Honcho at baseUrl

Development

bun install
bun test
bun run typecheck
bun run build

See ARCHITECTURE.md for the design and the reasoning behind each extension point, and RUNBOOK.md for a throwaway-VM test pass.

Credit

dsh-honcho-sync by @nanpaidashi (MIT) inspired the initial design of this plugin

The ~/.honcho/config.json contract, the session-naming convention, and src/redact.ts come from the sibling Honcho integrations — claude-honcho, codex-honcho, and their relatives.

DeepSeek Harness is MIT-licensed, and its Cordis plugin model is what made a native integration worth writing instead of a hook bridge.

License

MIT