bishx

bishx

Full dev lifecycle plugin for Claude Code. Takes you from a raw idea to shipped code through a structured pipeline of AI agents.

idea → prompt → plan → tasks → code → test → site audit

Installation

git clone https://github.com/Bishoni/claude-bishx.git ~/.claude/plugins/bishx

Then launch Claude Code with the plugin:

claude --plugin-dir ~/.claude/plugins/bishx

Type /bishx: to verify commands appear in autocomplete.

Prerequisites

• Claude Code CLI
• bd (beads) — local task tracker used for task decomposition and execution

Commands

Command	Description
/bishx:prompt <idea>	Generate a structured planning prompt from a raw idea
/bishx:plan <prompt>	Run 10-actor verification pipeline (up to 10 iterations) to produce a bulletproof implementation plan
/bishx:plan-to-bd-tasks	Decompose the approved plan into bd tasks (Epic → Feature → Task hierarchy)
/bishx:run	Execute tasks with multi-agent orchestration (Lead → Dev → 3 Reviewers → QA)
/bishx:run <epic>	Select a specific epic by name (partial match, e.g. /bishx:run auth)
/bishx:test	Deep system testing: auto-detects stack, discovers components, runs all test types, reports bugs to bd
/bishx:site <url>	Full website audit — crawls up to 100 pages via cmux browser, analyzes UX/UI, SEO, accessibility, performance, conversion
/bishx:init	Create CLAUDE.md and AGENTS.md templates in the project root
/bishx:init-sync	Scan codebase and fill/update CLAUDE.md and AGENTS.md with real project data

/bishx:status, /bishx:cancel, /bishx:polish, /bishx:idea are defined but not yet functional.

Architecture

The plugin uses Agent Teams — independent Claude Code sessions that communicate peer-to-peer via SendMessage and coordinate through a shared TaskList.

Planning pipeline (10 actors)

                                  ┌─ Skeptic (opus)
                                  ├─ TDD Reviewer (sonnet)
Interview → Research → Planner →  ├─ Completeness (sonnet)  → Critic → Dry-Run? ×N
                                  ├─ Integration (sonnet)
                                  ├─ Security* (sonnet)
                                  └─ Performance* (sonnet)
                                    * = conditional

Iterates up to 10 times until the Critic scores ≥75% with zero blocking issues (APPROVED) and the Dry-Run Simulator passes. Complexity gate adapts the pipeline: TRIVIAL skips review, SMALL runs lite review, MEDIUM+ runs full parallel review. Each session is stored in a timestamped directory (.bishx-plan/YYYY-MM-DD_HH-MM/) with all iterations preserved for history. The approved plan is saved as APPROVED_PLAN.md inside the session directory.

Skill-library integration

After research completes, the orchestrator performs a skill-library lookup from ~/.claude/skill-library/:

1. Reads INDEX.md hierarchy to identify relevant categories by tech stack
2. Selects skills per-role: implementation patterns for Planner, verification rules for Skeptic
3. Writes PLANNER-SKILLS.md and SKEPTIC-SKILLS.md with full skill paths
4. Agents read FULL SKILL.md files themselves (no truncation, ≤2500 lines budget per agent)

Execution pipeline

Lead → Dev (opus) → 3 Reviewers (sonnet, parallel) → Validate (sonnet) → QA (opus)

Lead assigns bd tasks from the board and orchestrates the full cycle:

1. Dev (opus) implements the task
2. Three parallel reviewers (sonnet) analyze the diff independently:
   • Bug Reviewer — correctness, logic errors, syntax, tests
   • Security Reviewer — OWASP vulnerabilities, injection, XSS, SSRF, secrets
   • Compliance Reviewer — CLAUDE.md/AGENTS.md project rules
3. Per-issue validation (sonnet) — each CRITICAL/MAJOR finding is independently confirmed or rejected by a sonnet subagent to suppress false positives
4. QA (opus) runs acceptance testing

Each reviewer has formal severity definitions (CRITICAL/MAJOR/MINOR/INFO), HIGH SIGNAL filters ("what NOT to flag"), scope discipline (diff-only), and self-validation. Pass/fail is deterministic: zero CRITICAL + zero MAJOR + automated checks pass.

Review approach inspired by the Anthropic code-review plugin.

Lead performs centralized skill lookup from the skill library before each task, passing relevant skill paths to each agent (≤2500 lines budget per agent).

All teammates (Dev, Reviewers, QA) reason and communicate in English for better analytical quality. Lead communicates with the user in the user's language.

Site audit

/bishx:site https://example.com