Skip to content

getting-started.md

Latest commit

 

History

History
187 lines (131 loc) · 5.64 KB

getting-started.md

File metadata and controls

187 lines (131 loc) · 5.64 KB
Raw
title Getting started
description Install Bug Hunter, verify the target, run a scan-only audit, and interpret the current precision-first result.
prompt Guide a user through installing and verifying Bug Hunter, running a safe scan-only audit, understanding unresolved results, and choosing loop or plan modes without granting unintended mutation authority.

Getting started

Before you begin

You need:

  • Node.js 22 or newer;
  • Git for branch/PR scope and the complete fix pipeline;
  • a coding agent that can read skill files and run normal repository checks;
  • a repository to audit.

The default workflow is scan-only and single-pass. It does not edit source files. Use --loop when you explicitly want the queued scope worked until coverage completes.

1. Install for one agent

For the latest published package, install Bug Hunter for Codex with:

npm exec --yes --package=@codexstar/bug-hunter@latest -- bug-hunter install --agent codex

Use another target when needed, for example:

npm exec --yes --package=@codexstar/bug-hunter@latest -- bug-hunter install --agent claude-code
npm exec --yes --package=@codexstar/bug-hunter@latest -- bug-hunter install --agent cursor
npm exec --yes --package=@codexstar/bug-hunter@latest -- bug-hunter install --agent copilot
npm exec --yes --package=@codexstar/bug-hunter@latest -- bug-hunter install --agent droid

If you intentionally want the current GitHub main source instead of the latest published package:

npx --yes https://github.com/codexstar69/bug-hunter/archive/refs/heads/main.tar.gz install --agent codex

See agent installation for every supported target, source installs, custom paths, and updates.

2. Verify the installed copy

Use the same package source and target you installed.

Published package:

npm exec --yes --package=@codexstar/bug-hunter@latest -- bug-hunter doctor --agent codex

Current GitHub source:

npx --yes https://github.com/codexstar69/bug-hunter/archive/refs/heads/main.tar.gz doctor --agent codex

A complete target check reports that the managed runtime and install manifest are current for the CLI package performing the check. Do not mix an old global CLI with a newer source install when diagnosing version mismatches.

Restart the coding agent if it was open during installation.

3. Open the repository

Start your coding agent in the repository root. The agent should have permission to read source and run the project's normal validation commands.

4. Request the safe first scan

Send this portable natural-language prompt:

Use the bug-hunter skill to scan this repository. Do not edit files.
Return the final report and call out every manual-review or unreviewed item.

If the agent exposes slash skill commands, the short form is:

/bug-hunter

Natural language is preferred in shared instructions because it works across agent interfaces.

5. Understand what the scan is doing

The current pipeline is precision-first:

triage -> optional adaptive plan -> Recon -> retrieval -> Hunter
       -> Skeptic -> Referee -> optional hybrid verification -> report

Important guarantees include risk-ordered scope, token-bounded chunks, source-hash integrity, assigned-file enforcement, truthful per-file coverage, and fail-closed required verification.

The public skill defaults to one pass. For large targets where you require the entire queued scope to reach a terminal coverage state, request:

/bug-hunter --loop

6. Review the result

Start with:

  • .bug-hunter/report.md — human-readable report;
  • .bug-hunter/scan-report.json — joined machine-readable result;
  • .bug-hunter/referee.json — final verdict evidence;
  • .bug-hunter/coverage.json — per-file coverage when persisted/loop work is used;
  • .bug-hunter/verification-report.json — hybrid verification evidence when requested;
  • .bug-hunter/adaptive-plan.json and .bug-hunter/retrieval-plan.json — bounded execution/retrieval policy when those layers are active.

Treat manual-review, unreviewed, failed coverage, and required-verification failure as unresolved. They are not clean results.

7. Plan fixes without editing

After reviewing the report:

Use the existing Bug Hunter findings to build a fix plan. Do not edit files.

Slash form:

/bug-hunter --plan

Review .bug-hunter/fix-strategy.json, .bug-hunter/fix-plan.json, and the resulting immutable scope before any mutation run.

8. Apply reviewed fixes

When the plan is acceptable:

Use the bug-hunter skill to apply the approved fix plan.
Ask before every edit. Do not commit.

Slash form:

/bug-hunter --fix --approve

--approve requests the host's reviewed/default permission mode. Approval prompt behavior depends on the coding agent. --auto-commit is separate and must not be assumed from edit permission.

Do not grant autonomous fixing or commit permission unless that behavior is intended.

Quality note

The bundled benchmark and pnpm quality:world-class validate Bug Hunter's measurement and regression contracts. They do not prove universal superiority across every repository or model. See world-class protocol.

Next steps