Use plous-debug in Codex or Claude Code to investigate a Kubernetes failure, connect runtime evidence to the code in your repository, and verify a deployed fix. The skill uses Plous's cluster tools through a local MCP server while the agent uses its usual tools for code and tests.

Install

Install Plous, then run the setup command for your agent.

Agent

Setup

Invoke in your agent

Codex

  plous setup codex

  $plous-debug

Claude Code

  plous setup claude

  /plous-debug

Setup runs without prompts. Restart or reconnect your agent afterward. The server follows the current context in the agent's kubeconfig. If you switch contexts, the next tool call uses the new one.

Investigate a checkout error

Suppose checkout starts returning 502 after a deployment. Invoke the skill from your repository and give the agent the symptom and the intended target.

Checkout started returning 502 in staging/shop after the last deploy. Confirm the Kubernetes context, then check the checkout Pods, events, and logs around a failed request. Tell me what you observed and what you think caused it.

An illustrative result would separate the evidence from the suspected cause.

Example

Observed

Checkout logged a timeout calling   payments.shop.svc:8081  . The payments Service exposes port   8080  .

Suspected cause

Checkout may be using an old port in its runtime configuration.

Next check

Trace where checkout gets its payments URL. Compare the deployed setting and the code's default with the payments Service port.

Ask the agent to follow that lead into the repository.

Trace how checkout builds the payments URL. Check whether the deployment sets the port or the code falls back to a default. If you find a mismatch, fix it in the repository and add a test that would have caught it.

After you deploy the fix, ask the agent to check the original failure.

I deployed the fix to staging/shop. Confirm the intended image is running. Check fresh checkout logs and, if active diagnostics are enabled, repeat the failing request. Tell me which checks passed and which still need evidence.

When logs aren’t enough

Choose where to run an active test based on what you need to learn.

If you need to know…

Run the test from…

Whether the checkout application can reach payments

The existing checkout container, if it has the required executable.

Whether an internal endpoint responds from a separate workload

A temporary diagnostic Job with the required tools. Its Pod can have different network access from checkout.

For the checkout example, a useful follow-up is “From the checkout container, check whether the payments health endpoint responds on port 8080. Report the Pod and the result.”

Both options require explicit enablement and Kubernetes permissions. Add   --allow-exec   for commands in a container or   --diagnostic-image IMAGE   for temporary Jobs that run HTTP GET, DNS, or TCP checks. Reconnect the agent after changing setup options. Active diagnostics, permissions, and cleanup

Configuration reference

Pin a context

Use explicit options when you want the server to stay on one context. This example also sets the default namespace for diagnostics.

Shell command

plous setup codex --kubeconfig /path/to/kubeconfig --context staging --namespace shop

Use   claude   for Claude Code. If an older setup pinned your context, rerun setup without   --context   and reconnect once to start following your current context. Context and saved evidence

Installed files

Client

Skill

MCP configuration

Codex

  ~/.agents/skills/plous-debug/SKILL.md

  ~/.codex/config.toml  , or   $CODEX_HOME/config.toml

Claude Code

  ~/.claude/skills/plous-debug/SKILL.md

  ~/.claude.json

Rerun setup to update a managed installation. Setup preserves other MCP entries and stops if an unrelated skill or unmanaged   plous_debug   entry occupies the destination. Use   --home PATH   to choose a separate installation home.

Remove the integration

Delete the installed   plous-debug   skill directory. For Codex, remove the block between   BEGIN PLOUS DEBUG MCP   and   END PLOUS DEBUG MCP  . For Claude Code, remove the   mcpServers.plous_debug   entry. Reconnect your agent. Saved evidence stays on disk until you remove its standalone profiles. Evidence storage locations

Next steps

Practice with Investigate a restarting Pod, configure MCP without the skill, or read the skill instructions.

View source

On this page

Plous · Open source · Early development

Search Plous docs

Search the documentation.

Documentation versions