Ouroboros library (`clawql-ouroboros`)

ClawQL ships a TypeScript workspace package clawql-ouroboros for specification-first multi-generation workflows: validated Seeds, Wonder / Reflect hooks, pluggable Executor and Evaluator, ontology convergence (similarity, stagnation, oscillation, regression gates), an EvolutionaryLoop, and helpers for MCP-style tool registration and a background seed poller.

It is a library you import from Node ≥ 22. In clawql-mcp, set CLAWQL_ENABLE_OUROBOROS=1 to register the three optional ouroboros_* MCP tools backed by this package. For durable lineages, configure Postgres with CLAWQL_OUROBOROS_DATABASE_URL (or split CLAWQL_OUROBOROS_DB_* env vars).

Learn walkthrough (three tools together, benefits, full MCP example): Ouroboros tools (seed, loop, lineage).

How it relates to upstream “Ouroboros”

The widely known Python project Q00/ouroboros is a full agentic stack (interview, CLI, MCP plugin, execution routing, persistence, and more). clawql-ouroboros is a ClawQL-scoped subset: overlapping concepts (seed, wonder/reflect, convergence), not API or feature parity. The npm name clawql-ouroboros avoids implying an official TypeScript port of Q00.

Install and build

From the ClawQL monorepo, depend on the workspace:

{ "dependencies": { "clawql-ouroboros": "workspace:*" } }

From npm (after you publish clawql-ouroboros):

npm install clawql-ouroboros

From a repo checkout, build the package:

npm run build -w clawql-ouroboros

Imports

Path	Contents
`clawql-ouroboros`	`SeedSchema`, `EvolutionaryLoop`, `ConvergenceCriteria`, `InMemoryEventStore`, types
`clawql-ouroboros/mcp-hooks`	`ouroborosMcpTools` — schemas + handlers for crystallize / run / lineage (wire to your MCP server)
`clawql-ouroboros/poller`	`startSeedsPoller` — interval loop with your DB `fetchPending` / `markFailed`

Minimal usage (conceptual)

Implement or inject WonderEngine, ReflectEngine, Executor, Evaluator (or use stubs for tests).
Choose an EventStore — InMemoryEventStore for dev/tests; Postgres (or similar) for production. clawql-mcp already includes a Postgres-backed store for ouroboros_*.
Construct EvolutionaryLoop, then await loop.run(seed) or run(seed, { maxGenerations, convergenceThreshold }) for per-run caps.

Using Ouroboros from `clawql-mcp` (production guide)

When CLAWQL_ENABLE_OUROBOROS=1, ClawQL registers:

ouroboros_create_seed_from_document
ouroboros_run_evolutionary_loop
ouroboros_get_lineage_status

These are standard MCP tools exposed over stdio, Streamable HTTP, and optional gRPC, so your agent can use the same route regardless of transport.

1) Turn it on

Set:

CLAWQL_ENABLE_OUROBOROS=1 to register tools
Optional: CLAWQL_OUROBOROS_DATABASE_URL (or split CLAWQL_OUROBOROS_DB_*) for durable events in Postgres

Without database config, lineages are in-memory (process-local, cleared on restart).

2) Understand the default loop behavior

The loop runs with default Wonder/Reflect/Executor/Evaluator engines unless you replace them in your embedding:

Wonder/Reflect: conservative defaults
Evaluator: lightweight pass/fail based on output signal
Executor: can now route to internal ClawQL search/execute if Seed route hints are present

This means you can run meaningful workflows immediately, then incrementally harden custom engines later.

3) Route hints: the contract

For schema-safe routing with MCP inputs, put hints in:

seed.brownfield_context.context_references[]

Supported hint objects:

Execute hint (run a specific operation):

{
  "clawql_execute": {
    "operationId": "listPets",
    "args": {},
    "fields": ["name"]
  }
}

Search hint (discover operations first):

{
  "clawql_search": {
    "query": "find pet listing operations",
    "limit": 5
  }
}

Why this field: context_references accepts flexible object payloads and is validated by SeedSchema, so hints survive tool input parsing and transport boundaries.

4) Full `ouroboros_run_evolutionary_loop` payload (copy/paste)

{
  "seed": {
    "goal": "List pets through internal execute",
    "task_type": "analysis",
    "brownfield_context": {
      "project_type": "brownfield",
      "context_references": [
        {
          "clawql_execute": {
            "operationId": "listPets",
            "args": {},
            "fields": ["name"]
          }
        }
      ],
      "existing_patterns": [],
      "existing_dependencies": []
    },
    "constraints": [],
    "acceptance_criteria": ["Return non-empty output"],
    "ontology_schema": {
      "name": "PetOntology",
      "description": "Pet listing ontology",
      "fields": []
    },
    "evaluation_principles": [],
    "exit_conditions": [],
    "metadata": {
      "seed_id": "pet-run-0001",
      "version": "1.0.0",
      "created_at": "2026-01-01T00:00:00.000Z",
      "ambiguity_score": 0.1,
      "interview_id": null,
      "parent_seed_id": null
    }
  },
  "maxGenerations": 2,
  "convergenceThreshold": 0.8
}

Expected response includes:

converged
generations
lineageId
status
summary

5) Inspect lineage output

Call ouroboros_get_lineage_status with seedId = lineageId from the run response.

Generation records include:

generation_number
seed
execution_output
evaluation_summary
phase

For route-hinted runs, execution_output contains route metadata (for example route: "execute" or route: "search"), which is useful for observability and tests.

6) Practical patterns

Deterministic execute workflow
- Put clawql_execute hint in context_references
- Keep acceptance criteria narrow and objective
- Use low maxGenerations (2-4) for predictable behavior
Discovery-first workflow
- Put clawql_search hint first
- Add explicit acceptance criteria about relevance and operation quality
- Later replace defaults with custom Reflect to convert search output into a new execute-focused seed
Document bootstrap workflow
- Use ouroboros_create_seed_from_document to crystallize raw extracted text
- Run loop
- Optionally pair results with memory_ingest and notify for durable notes + completion signals

7) Troubleshooting

Tools missing from ListTools
- Check CLAWQL_ENABLE_OUROBOROS=1
Lineage lost after restart
- Configure CLAWQL_OUROBOROS_DATABASE_URL (or split DB vars)
Custom route hint ignored
- Verify hint object is inside brownfield_context.context_references[]
- Verify key names exactly: clawql_execute or clawql_search
- Confirm operationId and/or query are non-empty strings
Validation errors on Seed
- Ensure required Seed fields are present
- Keep metadata shape compatible with SeedSchema base fields
Unexpected operation result shape
- Remember execute may project fields or fallback to REST depending on mode
- Test operation directly with search/execute before embedding in loop

8) Safety and rollout recommendations

Start with in-memory lineages in dev, then move to Postgres in staging/prod.
Keep initial maxGenerations low to bound runtime and output volume.
Log and monitor lineage status transitions (active, converged, exhausted, aborted).
Treat route hints as execution intent; gate who can author seeds in multi-tenant contexts.

npm registry

On npm, the package listing shows README.md from packages/clawql-ouroboros/README.md — install, quick start, export table, and honest scope for any Node ≥22 project (not only ClawQL). The tarball includes LICENSE (Apache-2.0).

Full guide and copy-paste examples

Canonical documentation with examples (minimal loop, convergence-only, MCP hooks, poller) lives in the repo:

clawql-ouroboros.md

Package README (short pointer): packages/clawql-ouroboros/README.md.

Ouroboros library (clawql-ouroboros)