Agent plugins and bug workflows

Agent plugins and bug workflows

The StatePort debugging skill guides a coding agent through reuse of an existing, exact State Card while investigating a browser bug: inspect the card, open it against current local code through StatePort MCP, and compare safe run evidence after a change. StatePort Desktop owns the cards and the local MCP server; the plugin supplies the agent workflow. Installing a plugin does not connect Desktop automatically.

Development preview: These plugins have been installed and discovered from public Git source in isolated Linux CLI profiles. In September 2026, the founder reported repeatedly testing installation, local agent connection and the complete cycle on Linux and macOS: capture a State Card from a supplied bug task, change the code, reopen the same Card and compare the result. The exact agent host and Desktop/skill versions and shareable before/after receipts are not yet recorded in the public compatibility evidence. Provider-specific compatibility and automatic skill selection are not yet documented. This is not yet a supported end-user plugin release. Review the source and use a test project with synthetic data.

Choose a focused workflow

Fix & Verify captures or reuses a browser case, records the failure, fixes current frontend code and checks a fresh run of the same case. Check Changes checks a small relevant set of saved cases without editing code by default. Harden Case checks independent valid response-data variants of one case; fixing a finding is a separate request. These are separate tasks, not an automatic three-stage run. Capture, Journey, Replay Experiments and Transfer remain available.

New checks and ordinary-Card experiments depend on the installed runtime’s advertised capabilities and its separate replay permission. Discover support before starting. The integration source version and exact host evidence live in the compatibility record. No time or token savings are established by successful functional checks.

Connect before your first State Card

1. Install StatePort Desktop and open Settings → Coding agents. With an empty Library, Set up coding agent leads to the same connection section. No Card, Developer trial or agent Capture permission is needed for setup.
2. Choose Connect an agent, review the preview, then Copy setup and paste it into your local coding agent. The short message points to the single setup entrypoint and includes Desktop’s generated configuration. The agent follows the setup for its own environment; Desktop does not restrict setup to a vendor list. A compatible local MCP client is required.
3. Restart the host/session if needed and confirm it discovers public StatePort tools. Saved config, an installed plugin and Local MCP: ready are separate from real host discovery. Setup does not list Cards or start Capture; zero Cards is a valid completed connection, with case verification still unperformed.
4. Record a case using Capture → Review → Save, or expand Capture with an agent in Capture to prepare an agent task. Agent Capture requires explicit permission and a capable runtime and host.
5. On the saved Card, expand Use with agent and review Copy task without leaving the Card. Optionally include the saved Notes / Observed / Expected, and paste it into the connected agent. No repeated installation is needed. The selected AI host may process the prompt remotely even though StatePort stores the case locally.

The public setup entrypoint indexes the existing helpers. Manual Copy MCP config remains available under Settings → Coding agents → Connection details. Existing server entries are preserved; conflicting configuration requires an explicit decision. Do not keep both legacy scenariodeck and stateport entries for the same runtime. Offline registry access does not block ordinary Desktop use.

CLI host	Install the source plugin	Host setup guide
Codex CLI	codex plugin marketplace add StatePort-Dev/stateport-agent-skills --ref main codex plugin add stateport@stateport-dev	Codex CLI
Claude Code	claude plugin marketplace add StatePort-Dev/stateport-agent-skills claude plugin install stateport@stateport-dev	Claude Code
GitHub Copilot CLI	copilot plugin marketplace add StatePort-Dev/stateport-agent-skills copilot plugin install stateport@stateport-dev	Copilot CLI

The maintained source repository records the exact package version, host checks, installation details, update paths, and current compatibility status. These commands install the agent plugin; install Desktop only through its download page.

For example, tell the agent:

Use my StatePort Card <CARD_ID> at revision <REVISION> to investigate this browser bug against my current local code. Inspect that exact Card first, reopen it after the fix, compare the safe evidence, and tell me what remains unverified.

If you use a standalone copy of the skill or another StatePort plugin in the same host, disable the duplicate before starting. Keep raw credentials, protected local values, and captured private material out of agent prompts. Authentication or interactive Journey steps still require the supported human handoff.

Current capability boundary

The public MCP contract supports inspection and reopening of a supplied Card. Agent Capture is a separate capability and permission: use it only when the connected runtime and actual host workflow support it. Setup does not grant that permission or prove Capture works. Use Capture with an agent for a new case, or capture it yourself in Desktop; an unrelated agent browser is not a StatePort Capture.

Before a code change, the agent must confirm the project’s current-code target and reproduce the symptom. Afterward it reopens the same Card/revision and compares separate run evidence. A captured frontend shell or successful Open alone cannot confirm the fix. Missing capabilities, replay misses, timeouts and inconclusive evidence must be reported without hidden live fallback.

VS Code/Cursor VSIX, Cursor Agent plugin, and IntelliJ IDEA/WebStorm plugin sources are separate development candidates, with different installation and verification status. Their local package or binary checks do not establish a working agent connection or workflow.