# Agentic Runtime Debug Loop

> Use when debugging a deployed service/profile, investigating CI or Kubernetes runtime failure, explaining deployment drift, or producing a runtime debug session summary

- Skill: `dolphinai2026/agentic-runtime-debug-loop` (Agent Skill)
- Install (CLI): `npx skillmds@latest add dolphinai2026/agentic-runtime-debug-loop`
- Raw SKILL.md: https://api.skillmd.com/api/skills/dolphinai2026/agentic-runtime-debug-loop/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: dolphinai2026 (https://skillmd.com/u/dolphinai2026)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/dolphinai2026/agentic-runtime-debug-loop

---


# Agentic Runtime Debug Loop

Use this skill when debugging a deployed service/profile and a bounded evidence loop is needed across code, CI/CD, Kubernetes, artifacts, logs, Grafana observability, and Phase 4 runtime resources.

## Phase 6.8A Delegation

For product-level diagnostics, construct a `HarnessEvidenceRequest` and use `agentic-harness-evidence-loop` / `agentic-evidence` to collect the local bundle and produce `RepairRoute`.

This skill retains service/profile collector knowledge for runtime evidence input, but it is not the product-level evidence loop contract. Do not execute code-changing repair, Git mutation, CI trigger, deploy, rollback, Control Plane writes, or storage upload from this skill.

## Inputs

Required:
- `service`
- `profile`
- `plan`

If `service`, `profile`, or `plan` is missing, ask for it before starting the session or running remote commands. If a controlled write action may be needed, the plan must contain an `agentic_debug_loop` fenced YAML block.

## Plan Policy

Read the implementation plan before remote commands. The policy shape is:

```yaml
agentic_debug_loop:
  max_iterations: 3
  success_criteria:
    - "Kubernetes pods for app-api are ready"
    - "latest CI pipeline succeeds"
  allowed_actions:
    - jenkins.build
    - gitlab.run_pipeline
    - kubernetes.rollout_restart
  allowed_resources:
    - main-postgres
    - main-redis
```

No `agentic_debug_loop` block means one read-only iteration. Do not execute write actions without explicit `allowed_actions`.

## Session Commands

Start a session:

```bash
bin/agentic-runtime --json loop start --service <service> --profile <profile> --plan <plan>
```

Record each material command:

```bash
bin/agentic-runtime --json loop record-step --session <session-id> --kind status --status ok --summary "<short summary>" --evidence <path>
```

Finish after writing the runtime summary:

```bash
bin/agentic-runtime --json loop finish --session <session-id> --final-status <fixed|likely_fixed|not_fixed|blocked|error> --summary-path <summary-path>
```

Session records are written under `.generated/runtime/debug-sessions/`.

## Workflow

1. Parse the plan policy and start a debug session.
2. Run local-only checks:
   ```bash
   bin/agentic-runtime --json status --service <service> --profile <profile> --local-only
   bin/agentic-runtime --json correlate --service <service> --profile <profile> --local-only
   ```
3. Run remote read-only evidence commands:
   ```bash
   bin/agentic-runtime --json status --service <service> --profile <profile>
   bin/agentic-runtime --json correlate --service <service> --profile <profile>
   bin/agentic-runtime --json observe doctor --service <service> --profile <profile>
   bin/agentic-runtime --json observe metrics --service <service> --profile <profile>
   bin/agentic-runtime --json observe logs --service <service> --profile <profile>
   bin/agentic-runtime --json observe trace --service <service> --profile <profile> --trace-id <trace-id>
   ```
4. Inspect evidence in this order:
   - config/parameter resolution
   - CI status and logs
   - artifact existence and digest/tag
   - deployed Kubernetes image
   - Kubernetes pod state, events, and logs
   - Grafana observability metrics, logs, dashboard links, and trace hints
   - configured debug_loop resource hints
5. If status, logs, or correlate evidence includes a trace id or request id, run trace lookup before Phase 4 resource checks:
   ```bash
   bin/agentic-runtime --json observe trace --service <service> --profile <profile> --trace-id <trace-id>
   ```
   Record the result with `--kind observability`.
6. Record observability commands with `--kind observability` and evidence paths under `.generated/runtime/observability/`.
7. When evidence points to a data or middleware issue, run resource doctor before query or inspect:
   ```bash
   bin/agentic-runtime --json resource doctor --target <resource-id>
   ```
8. Query or inspect only resources allowed by the plan and service/profile config.
9. Search code/config only after evidence points to a likely area.
10. If a controlled action is authorized, run dry-run before confirm:
   ```bash
   bin/agentic-runtime --json act --service <service> --profile <profile> --action <action> --plan <plan> --dry-run
   ```
   Inspect the dry-run target, parameters, risk, and confirmation id.
   Run `--confirm` only when the action is listed in `allowed_actions` and the plan policy unambiguously authorizes automated confirmation for the current loop.
   If that authorization is absent or ambiguous, stop and ask for a plan update or human direction.
   ```bash
   bin/agentic-runtime --json act --service <service> --profile <profile> --action <action> --plan <plan> --confirm <confirmation-id>
   ```
11. Re-observe status if iterations remain.
12. Write the runtime summary and finish the session.

## Stop Conditions

- Stop when success criteria are met.
- Stop when `max_iterations` is exhausted.
- Stop when required secrets, kubeconfig, or runtime config are missing.
- Stop when Grafana observability is configured but its required secret env vars are missing.
- Stop when an unlisted action or resource is needed.
- Stop when Phase 4 resource safety blocks the operation.
- Stop if evidence is contradictory; summarize the contradiction and ask for human direction.