Integrating vitel: CI deploy-gate & Verel registration

vitel turns telemetry into a pass / warn / fail verdict. Two integrations make that verdict block bad releases.

1. CI deploy-gate — block a release on error-budget burn

vitel check exits non-zero on fail, so it gates a pipeline with no glue code.

Exit codes: 0 = pass (and warn), 1 = fail, 2 = usage/source error. Add --warn-as-fail to also block on warn (a stricter gate).

# .github/workflows/deploy.yml (excerpt) — see examples/ci/deploy-gate.yml
- name: Deploy
  run: ./deploy.sh

- name: Post-deploy vitals gate
  run: |
    pip install "vitel[prometheus]"
    # pull post-rollout vitals and fail the deploy if the budget is burning
    vitel check "$PROM_QUERY" \
      --backend prometheus \
      --expect "availability 99.9%" \
      --expect "must: error_budget_remaining > 20%" \
      --expect "should: p99(latency_ms) < 300ms" \
      --window 600
  env:
    VITEL_PROMETHEUS_URL: ${{ secrets.PROMETHEUS_URL }}

If post-deploy vitals violate a must: threshold (or burn the error budget), the step exits 1 and the job fails — the rollout is gated. For a file/exported-metrics source, point vitel check at a JSON/CSV file or a /metrics URL (--backend scrape) instead.

2. Register vitel as a Verel sense

Verel withholds "done" until its required senses return a passing verdict. vitel plugs into Verel's verdict bus as the vitals sense.

The afferent signal

vitel implements agentsensory.Sense (vitel.Vitals) and every grade distills to a Handoff:

import asyncio
from vitel import perceive
from vitel.slo import SLO

slo = SLO.from_inputs(expect=["availability 99.9%", "must: error_budget_remaining > 20%"])
handoff = asyncio.run(perceive(POST_DEPLOY_METRICS, slo=slo))

handoff.perceived      # Verdict.FAIL when the budget is blown
handoff.next_action    # NextAction.REVISE  (never DONE on a failing verdict)
handoff.todo           # actionable items the brain should resolve

Because a failing verdict yields next_action = REVISE (and a passing one DONE), Verel's gate cannot close a deploy task while vitals are failing.

Trust tier

• check / watch are deterministic, grounded graders → they may block (precise gate).
• analyze's LLM critique is advisory → Verel clamps it to ≤ warn (informs, never blocks), matching Verel's "precise gates, advisory informs" reducer.

Wiring (lives in the Verel repo)

Verel mounts vitel either over MCP (vitel-mcp: vitals_check / vitals_watch / vitals_status) or by importing vitel.Vitals. A ready-to-drop recipe is in examples/verel/post_deploy_vitals.recipe.json: it triggers a post-deploy vitals_check, treats a fail verdict as a blocking gate, and lists the remediation skill. Register it in Verel's sense/recipe registry (per the Verel repo's instructions).