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": "${HONCHO_API_KEY}" },
"hosts": {
"dsh": { "workspace": "dsh" },
},
}
Get a key at app.honcho.dev. HONCHO_API_KEY in the environment also works on its
own — the config file is only needed to change defaults.
What it does
Injects what Honcho knows at the start of a session — your profile, a summary of this project's session so
far, and the conclusions Honcho has drawn that are relevant to what you just asked. One API call, shaped to a
character budget, refreshed as you work.
Captures every turn. User and assistant messages go to Honcho in the background, debounced, and flushed at
turn boundaries, before compaction, and on shutdown. Secrets are redacted first.
Gives the model three tools:
| Tool | For |
|---|
honcho_search | Looking something up. Searches raw messages and derived conclusions. |
honcho_chat | Asking a question of judgment. Reasons over everything Honcho knows. Slow. |
honcho_remember | Saving a durable fact, preference, or decision. |
/honcho shows status and a link to the session in the Honcho dashboard — see Commands below.
Configuration
Everything behavioral lives in ~/.honcho/config.json under hosts.dsh — the same file claude-honcho,
codex-honcho, and the other integrations read. Root holds identity and connection; the host block holds
behavior.
{
"peerName": "your-name",
"workspace": "honcho",
"baseUrl": "https://api.honcho.dev", // bare host or …/v3 both fine
"timeoutMs": 30000,
"auth": { "apiKey": "${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>
"sessionPrefix": "", // literal prefix on every generated name, e.g. "vps-"
"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. sessionStrategy: "git-remote" and sessionPrefix are proposed for claude-honcho as well, so a repo worked on from both can be
pointed at one session on purpose rather than by accident.
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.
sessionStrategy | Session name | Notes |
|---|
per-directory (default) | <peer>-<dir> | Stable across restarts and branches |
per-repo | <peer>-<repo-root> | Same memory from any subdirectory |
git-remote | <peer>-<host-owner-repo> | From the origin URL, so the same repo on two machines is one session. Falls back to per-directory outside a repo or without an origin |
git-branch | <peer>-<dir>-<branch> | Falls back to per-directory outside a repo or on a detached HEAD |
per-session | <peer>-chat-<id> | A clean slate every restart |
global | <peer> | One memory for everything |
per-directory and per-repo name a session after a folder, so two different projects that both live in a
web directory silently share one memory across machines. git-remote is the fix: it reduces the origin URL
to host/owner/repo, so git@github.com:you/web.git and https://github.com/you/web agree, and you/web and
someone-else/web do not. It compares the URL as configured, so machines that clone through different ssh
Host aliases — github.com-work on one, github.com on the other — still get separate sessions.
sessionPrefix puts a literal string in front of every generated name — "vps-" gives vps-you-web — for
when the machine a session came from should be visible in it. It applies to every strategy. A name pinned in
sessions is used exactly as written and is never prefixed.
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
| Command | Does |
|---|
/honcho | Status: peer, workspace, session, strategy, pending uploads, last sync, last fetch |
/honcho config | Resolved settings, the file they came from, and any ignored injection components |
/honcho flush | Sync now |
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
