prism-mcp-orchestrator

Files (4)

• SKILL.md
• smoke.yaml · evals/
• REFERENCE.md · references/
• collect.sh · scripts/

SKILL.md (first lines)

# Prism MCP orchestrator skill

Use this skill for orchestration and bounded typed decisions. Delegate evidence-heavy specialist work via Prism MCP; use the decision workflow below for configured Laya/Jev tools.

## Activation rule

Activate when prompts mention things like: CI failures, PR blockers, Kubernetes incidents, Argo drift/workflow failures, release notes/docs lookup, large logs/JSON, bounded Laya/Jev decisions, or "delegate/run_agent/MCP".

Do **not** activate for tiny single-step tasks where delegation overhead is not worth it.

## Fail-closed evidence gate

Before delegation, check for minimum identifiers:

- GitHub: `repo` + (`PR`, `run URL`, or `run ID`)
- Kubernetes: `context` + `namespace` + workload (`deployment`/`statefulset`/`pod`)
- Release notes/docs: product/library + version range or target release

If required identifiers are missing, **do not run specialists yet**. Ask the parent LLM to request only missing fields.

## First-party skill precedence

When a subtask targets a specific platform/tool/vendor, prefer that vendor's
first-party skills/docs before community or generic sources.

Examples:

- LaunchDarkly tasks -> prefer LaunchDarkly first-party skills (for example from
  the LaunchDarkly `ai-tooling/skills` source) and official docs.
- Cloud/provider SDK tasks -> prefer official provider skills/docs first.

If a first-party source exists but is not available in the current environment,
callback to parent with `insufficient_evidence` and request permission/inputs to
access the first-party source.

## Required sequence for specialist delegation (must follow)

1. Call `list_agents` first.
2. Call `list_resources`, then `get_resource` for:
   - `prism://resource/tooling/run_agent`
   - `prism://resource/tooling/orchestration-guide`
   - `prism://resource/agent/<agent_id>/constitution` (for selected agents)
3. Optional: `list_prompts` + `get_prompt` for payload templates.
4. Call `run_agent` with bounded task + valid `skill_names`.
5. Synthesize specialist outputs in parent model.

If evidence is insufficient at any step, callback to parent with:

```json
{
  "status": "insufficient_evidence",
  "needs_from_parent": [
    "repo owner/name",
    "PR URL or workflow run URL",
    "namespace/deployment"
  ],
  "reason": "Identifiers required for evidence-backed delegation are missing."
}
```

## Typed decisions through downstream MCP

Use a configured Laya or Jev MCP tool for a bounded classification, a choice
among explicit alternatives, or a rubric score when that judgment helps the
task. The parent model supplies the question and options, interprets the result,
and produces any subsequent general tool-call arguments. Prism executes calls.

1. Call `list_mcp_servers` to discover configured decision servers. If none is
   configured, continue the task without a decision call. Registration and agent
   access setup are described in `docs/decision-services.md` in the Prism repo.
2. Call `list_mcp_server_tools` with `include_schema: true` for the selected
   server. Use its actual tool schema; Laya and Jev wrappers differ.
3. Call `call_mcp_tool` with the discovered server and tool. For Laya's
   `laya_predict`, an example is:

   ```json
   {
     "server": "laya",
     "tool": "laya_predict",
     "arguments": {
       "model": "english",
       "state": {"body": "I was charged twice."},
       "questions": {
         "department": {
           "type": "choice",
           "instructions": "Which team should handle this request?",
           "criteria": {
             "billing": "Invoices and refunds",
             "technical": "Bugs and outages"
           }
         }
       }
     }
   }
   ```

4. Check `is_error` and `truncated` before interpreting the result. Laya's
   decision JSON may be encoded in `content` or in `structured_content.result`;
   decode that JSON to read `answers`. Report uncertainty or continue with parent
   reasoning if the call fails. A decision does not grant permission for an action.

Laya custom questions include choices on each call; `laya_preset` supplies only
its built-in question sets. The remote Laya wrapper forwards inference to its
HTTP server and does not support `laya_shortlist` or `laya_tournament`. Keep
finite choices small (at most 20), narrowing them first when needed.

This downstream MCP path is separate from `recommend_tools`, whose built-in
reranker uses `PRISM_DECISION_*` over HTTP. Reconnect a running Prism MCP host
after changing registrations. A CLI probe tests transport; a specialist also
needs access to the server through its configured MCP allowlist.

## Adapted routing checkpoint

When `laya-routing` is registered, use it only for English next-tool choice
questions. Keep general classification, rubric scoring, and built-in presets on
`laya` or the configured Jev service. Inspect the actual MCP schema first.

For a routing call:

From bryanbarton525/prism, MIT.

More from bryanbarton525/prism