Ouroboros for DeepSeek Harness

36 Tools, Five Stages: Inside Ouroboros's MCP Bundle

Plugin quick facts

Setup and first check

Install the repository subdirectory into a named profile and inspect the resulting layer:

dsh plugin --profile web add "github:Q00/ouroboros#main&path:integrations/dsh-plugin"
dsh --profile web --dump-config

Keep uv on the system path. If the workflow uses ooo auto, configure an executable agent runtime such as Codex, Claude Code, or OpenCode. DSH removes credential-shaped environment variables from child processes, so pass only the provider credentials explicitly required by the selected Ouroboros backend. Test an interview-only request before enabling execution.

Because DeepSeek Harness is in developer preview, a new release may require changes to the bundle or its profile configuration.

The DSH bundle exposes 36 tools under the ouroboros namespace. They cover more than one command, but the primary workflow can be understood as five connected stages:

1. Interview uses structured questioning to surface assumptions and missing decisions.
2. Seed turns the answers into an immutable specification with constraints and acceptance criteria.
3. Execute hands the specification to the selected agent runtime.
4. Evaluate applies mechanical, semantic, and multi-model gates according to the selected workflow.
5. Evolve feeds evaluation results into another generation when important uncertainty remains.

Inside DSH, a user can start with a concise phrase such as ooo interview <goal> or ooo auto <goal>. The harness model can discover the matching MCP tool from its description. Supporting tools expose status, cancellation, resumption, QA, and recovery-oriented paths rather than forcing the entire process through a single opaque call.

The bundle is config-only. It adds no custom plugin code. One configuration row mounts DSH's existing MCP client against ouroboros mcp serve. That makes the integration easier to inspect, but the Ouroboros process and its own requirements remain separate from DeepSeek Harness.

Configure uv, Credentials, and the Execution Runtime

The plugin starts the Ouroboros MCP server through uvx, so uv must be available on the system path. The interview and planning tools can run through that server, while ooo auto also needs an executable agent runtime for its execution stage. Configure the runtime explicitly and keep its credentials separate from the DSH process.

Install the repository subdirectory into a named DSH profile:

dsh plugin --profile web add "github:Q00/ouroboros#main&path:integrations/dsh-plugin"

Then start that profile and inspect its composed configuration. The profile flag matters because DSH needs to know where the layer belongs.

If interview, Seed, or QA work should route through a different backend, set the corresponding backend and credential variables explicitly. DSH scrubs credential-shaped environment names before starting child processes. The bundle forwards a short allowlist, and any additional provider key must be intentionally added to the plugin configuration. This is a useful security boundary, not an invitation to pass the whole host environment.

Keep the first pilot simple: one disposable repository, one bounded goal, one explicitly selected runtime, and no production credentials. Confirm the interview path before allowing the execution stage to edit files.

How the Evaluation and Evolution Loop Actually Runs

Evaluation should be tied to the Seed, not to a general impression that the output looks plausible. Mechanical checks can cover commands, files, tests, or other explicit acceptance evidence. Semantic checks ask whether the result solves the defined problem. A consensus step can add another model view when the workflow calls for it, but it also adds latency and usage.

Evolution is a feedback loop rather than a full restart. The evaluation output becomes input to the next generation of the specification and plan. This can be useful when a task reveals a missing concept, conflicting constraint, or architecture decision during implementation. It can also waste time when the acceptance criteria are already clear and the remaining work is a small local fix.

Before running the loop, define a stopping rule:

• the required tests and static checks pass;
• the stated acceptance criteria have evidence;
• unresolved items are named rather than hidden;
• the change stays within the approved repository scope;
• a human reviews any security, migration, or destructive action.

For a solo builder, the main value is an audit trail between the original goal and the delivered change. That trail is useful only if the Seed remains specific enough to test.

What Happens When a Step Fails Mid-Loop

A failed MCP connection does not have to prevent DSH from starting. The Ouroboros tools may be unavailable while other plugins continue to work. After fixing a missing executable, configuration path, credential, or runtime, reload the plugin or restart the profile. Do not assume a disconnected client will always recover on its own, because reconnect behavior depends on the DSH build.

For an in-flight execution, use status information before repeating a stage. Ouroboros provides session tracking, a resume path for in-flight sessions, and cancellation for stuck or orphaned executions. Its persistence layer records events and checkpoints, so reattaching is preferable to launching a duplicate run when recoverable state exists.

A safe failure sequence is:

1. Identify whether the failure is transport, configuration, execution, or evaluation.
2. Preserve the Seed, execution identifier, logs, and last accepted artifact.
3. Fix only the failing boundary.
4. Resume or rerun the smallest applicable stage.
5. Reapply the evaluation gate before calling the task complete.

Related Guides

• dsh-memory-plugin for DeepSeek Harness