Debug a runtime failure across the fleet by correlating logs on an OTel trace_id / sessionId / requestId, or by reading a pasted Grafana / stack-trace screenshot, and mapping the failure back to source.
Installs into .claude/skills of the current project.
Are you the author of Debug?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/talayash-debug)
---
name: debug
description: Debug a runtime failure across the fleet by correlating logs on an OTel trace_id / sessionId / requestId, or by reading a pasted Grafana / stack-trace screenshot, and mapping the failure back to source.
argument-hint: <trace_id | sessionId | "error snippet" | (or paste a screenshot)>
allowed-tools: Bash, Read, Grep, Glob, Task
---
Debug a cross-service runtime issue in the Spring Boot fleet.
Input: `$ARGUMENTS`: one of:
- an OTel **trace_id** (32 lowercase hex chars): preferred, modern path.
- a legacy **sessionId** / **requestId**: fallback for older fleets.
- a **pasted error or stack trace**: agent extracts a trace key from
nearby log lines.
- a **pasted screenshot** of a Grafana / Tempo / Jaeger / Kibana panel
or an IDE / terminal stack; the model reads it directly (Claude is
multimodal); extract the trace_id or service + timestamp from the image
and use that as the correlation handle.
Steps:
1. Locate `spring-fleet.config.json` (cwd, else ask the user / suggest `/fleet-init`).
2. Use the **debugging-runtime-logs** skill.
3. **If the input is an image:** read it first. Look for a `trace_id` /
`traceparent` value, a sessionId, a service name + timestamp window, or an
exception class + message. Use whichever you find as the correlation
handle (prefer trace_id). Quote what you extracted so the user can sanity-check it.
4. Dispatch the **log-correlator** agent with the config path and the
resolved handle to build the cross-service timeline and isolate the failure
origin. Prefer the `correlate_by_trace` MCP tool when available.
5. If the cause is cross-service, dispatch **fleet-explorer** to confirm the
call path and map the failing hop to `file:line`.
6. Report: what happened, the timeline, failure origin (service + `file:line`),
how it surfaced upstream, the **ROOT-CAUSE HYPOTHESIS** block from the
log-correlator (WHAT / WHERE / WHY / CONFIDENCE / SUGGESTED FIX), and any
service missing a log file.
## Outside Claude Code
- `$ARGUMENTS` is the text the user supplied with the request.
- No plugin subagents (Codex): read `agents/log-correlator.md` and `agents/fleet-explorer.md` under the plugin root (two directories above this SKILL.md) and do that work yourself instead of dispatching.
- spring-fleet MCP tools: pass `config_path` (absolute path of the project's `spring-fleet.config.json`). Codex starts the server outside the project, so it cannot find the config on its own.