dsh-ouroboros
Mount Ouroboros — a spec-first AI dev workflow engine (Socratic interview → Seed spec → execute → evaluate → evolve) — into DeepSeek Harness as native tools. Once installed, type ooo interview <goal> or ooo auto <goal> directly in the dsh chat — the model finds and calls the matching mcp__ouroboros__* tool on its own.
This is a config-only bundle: it contains no custom plugin code, just one row that mounts dsh's existing @deepseek-ai/dsh-mcp-client against ouroboros mcp serve.
Requirements
uvonPATH. Nothing else —uvxfetches and
runs Ouroboros in an isolated environment on first launch, no pip install step.
- Python >= 3.12 (whatever
uvresolves). - Nothing else for execution: this bundle defaults
OUROBOROS_AGENT_RUNTIME
to host, so dsh's own model does ooo run/ooo auto's execution step — no CLI to install. Set OUROBOROS_AGENT_RUNTIME (see below) to an executable runtime (claude-cli, codex, opencode, ...) if you'd rather have that CLI do the work instead.
Install
This bundle lives inside the main Ouroboros repository as a subdirectory (it has no independent release cadence or code of its own), so install it straight from GitHub with pnpm's subdirectory syntax. --profile is required by dsh plugin and names the profile to install into:
dsh plugin --profile <your-profile> add "github:Q00/ouroboros#main&path:integrations/dsh-plugin"Then boot as usual (dsh --profile <your-profile>, or dsh web if your profile is named web). dsh --profile <your-profile> --dump-config shows the # == dsh-ouroboros layer once it's composed.
Configuration
Set these as environment variables before launching dsh. They are read from the dsh host environment and passed through to the spawned ouroboros mcp serve process:
| Variable | Purpose |
|---|---|
OUROBOROS_AGENT_RUNTIME | Agent runtime for ooo run/ooo auto's execution step. Defaults to host (dsh's own model does the work — see below); set to claude-cli, codex, opencode, ... to use that CLI instead. |
OUROBOROS_LLM_BACKEND | LLM backend for interview/Seed/QA. Leave unset to keep your ouroboros setup default. dsh routes those calls through DeepSeek Harness — read the next section first, it needs two more variables. |
OUROBOROS_DSH_CONFIG_PATH | Absolute path to the trusted Cordis composition the dsh LLM backend loads. Required whenever OUROBOROS_LLM_BACKEND=dsh. |
OUROBOROS_DSH_CLI_PATH | Path to the dsh-acp-demo bin, when it isn't on PATH. |
Override any field — timeout, args, a pinned Ouroboros version — from your own profile's cordis.patch.yml by targeting the mcp-ouroboros row id; see "Package and install a plugin" for the override mechanics.
Routing interview/Seed/QA back through DeepSeek Harness
OUROBOROS_LLM_BACKEND=dsh is not a one-variable switch, and it does not reuse the dsh process you're already running. Ouroboros spawns its own dsh-acp-demo ACP child, and that child fails closed with invalid_config unless it is given a composition to load. To use it:
- Build DeepSeek Harness from source (
pnpm install && pnpm run build,
Node.js >= 22) and make its dsh-acp-demo bin reachable — on PATH, or named by OUROBOROS_DSH_CLI_PATH. A source build is the working path today: installing the published @deepseek-ai/dsh-acp-demo still fails on a peer-dependency conflict inside its own dsh-tool-bash chain.
- Point
OUROBOROS_DSH_CONFIG_PATHat an absolute composition file.
Relative paths are rejected on purpose (they would resolve against the untrusted project cwd). The file must sit where dsh's node_modules (or workspace) is reachable: plugin package names in a composition resolve relative to the composition file's own directory.
- Provide whatever credential that composition names —
DEEPSEEK_API_KEYfor
DeepSeek's own models. See the credentials note below: it must be present in the dsh host environment, and this bundle forwards it explicitly.
Without steps 1–2 the Ouroboros tools still register and list fine; the first interview/Seed/QA call is what fails. Everything except the dsh backend works without them.
Credentials
dsh does not hand child processes the harness environment wholesale. Its subprocess seam scrubs every credential-shaped name — anything matching /KEY|PASSWORD|SECRET|TOKEN/i — plus every DSH_* name, so harness credentials never leak into a spawned process implicitly. A plugin's explicit env layer merges after that scrub.
So this bundle names a short allowlist rather than passing everything through:
ANTHROPIC_API_KEY— Ouroboros' default LLM backend.DEEPSEEK_API_KEY— thedshbackend loopback above.
To forward another (OPENAI_API_KEY, OPENROUTER_API_KEY, GOOGLE_API_KEY, ...), override the mcp-ouroboros row in your own profile's cordis.patch.yml with that one extra name in env. A later layer replaces a row's entire config rather than deep-merging it, so copy this bundle's config block and add your line to it. Everything non-credential-shaped — PATH, HOME, the OUROBOROS_* selectors — passes through untouched and needs no row at all.
Startup failures and recovery
Connection failures are non-fatal (failOnStartupError: false): a machine without uv on PATH yet still boots dsh normally, just without the Ouroboros tools, and every other plugin keeps working.
Recovery is not automatic in general. Whether mcp-client retries at all depends on your dsh build — the published 0.0.1-rc.1 has no reconnect loop, current main has one with bounded exponential backoff that gives up after a capped number of consecutive failures. Either way, after fixing the cause (installing uv, setting a runtime), reload the plugin or restart dsh rather than waiting for a reconnect.
What you get
36 tools under the ouroboros namespace — mcp__ouroboros__ouroboros_interview, mcp__ouroboros__ouroboros_auto, mcp__ouroboros__ouroboros_evaluate, mcp__ouroboros__ouroboros_ralph, and more. Each carries its own description, so a plain ooo interview: <vague idea> or ooo auto: <goal> in chat is enough — no extra prompting required.
License
MIT