Harness Shaping — Operator Guide

How the Federal Frontier SRE Investigator is shaped by the SME catalog: Remuda namespace, first tools, local scout, and quality gate.

The Federal Frontier SRE Investigator is shaped by the SME catalog: alert-class playbooks select Remuda namespace, first tools, and scout brief. Architecture: ADR-033. Catalog inventory: /docs/sre/sme-catalog/.

This is the operator enablement path for harness shaping already defined in FFP (posse roles + Remuda slices) - not a second runner fleet.

Enable local scout (lab)

On sre-dispatch (lab, e.g. texas-dell-04):

- name: RUNNER_RUNTIME
  value: opencode
- name: OPENCODE_LOCAL_SCOUT
  value: "1"
- name: OPENCODE_LOCAL_ONLY
  value: "0"   # do NOT default local-forever
- name: REMUDA_MCP_BASE
  value: http://internal-wanaku-router.remuda.svc.cluster.local:8080
- name: OPENCODE_RUNNER_IMAGE
  value: harbor.vitro.lan/ffp/opencode-runner:v1.4.3

Pinned images for this docs revision: sre-dispatch:v1.3.6, opencode-runner:v1.4.3.

Playbooks / SMEs

Source of truth: dispatch-controller/app/playbooks.py (living inventory).

Export Command
Markdown table python -m app.playbooks md (from dispatch-controller/)
YAML python -m app.playbooks yaml
Stats python -m app.playbooks stats
Wrapper scripts/export-sme-catalog.md
HTTP GET /sme-catalog (?format=md|yaml|json|stats)

The Investigator routes only ga / lab SMEs today; planned entries remain in the catalog as roadmap depth until Remuda tools and classifiers are wired.

Job env when a playbook matches:

  • HARNESS_PLAYBOOK - SME id (e.g. k8s-pod-crash)
  • HARNESS_SME_DISPLAY_NAME - specialist display name
  • HARNESS_FIRST_TOOLS - Remuda MCP tool order (invoke as MCP tools, never bash)
  • HARNESS_SCOUT_BRIEF - one-line brief
  • LOCAL_FORCE_OUTCOME_AFTER_SECONDS - playbook override
  • Remuda MCP URL = REMUDA_MCP_BASE + playbook remuda_path

Quality gate (before Bedrock)

  1. Local Gemma scout runs with a short Remuda-only prompt (<=~2500 chars).
  2. Scout must call Remuda MCP tools (k8s_get / k8s_logs / …) - never bash or kubectl shell wrappers.
  3. Weak or shell-failed local outcomes escalate to Opus with the local transcript.
  4. Do not set OPENCODE_LOCAL_ONLY=1 as the lab/prod default.

Verify

  1. Fire CrashLoop vs PVC vs unknown alerts.
  2. Confirm Job label harness-playbook and env HARNESS_PLAYBOOK.
  3. Confirm session mcp.json path matches the SME (/ns-9/mcp vs /public/mcp).
  4. Confirm scout transcript shows MCP tool_use for Remuda tools (not bash).

Eupraxia Labs · Federal Frontier · Platform docs