Multi-Agent Pipeline

Complete automated workflow from branch to handoff.

• •Phase 0 - Init: branch, identity, worktree setup
• •Phase 1 - Analysis: codebase exploration with Opus
• •Phase 2 - Planning: task decomposition + Plan Approval Gate (clarification max 2 rounds + free-text edit loop) — normal mode only
• •Phase 3 - Dev: TDD red-green-refactor cycles (component tasks delegate to figma-to-component skill)
• •Phase 4 - Review: CLI-aware parallel review + Opus triage (Claude: 2-model · Copilot: 3-model)
• •Phase 5 - Test: optional UI testing
• •Phase 6 - Commit: conventional commits → PR → issue update
• •Phase 7 - Report: Jira comment → Wiki + Figma screenshots → Confluence → Report + log → Knowledge & memory capture

Reviewer set depends on the host CLI — GPT-5.4 is only natively reachable from Copilot CLI, so Claude Code runs a 2-model set and Copilot CLI runs a 3-model set. Opus triage runs on both.

• •Deterministic gates first: build + lint + test + secret scan
• •Claude Code (2 parallel): Opus (security + architecture) + Sonnet (quality + correctness)
• •Copilot CLI (3 parallel): GPT-5.4 (edge cases + cross-provider diversity) + Opus + Sonnet
• •Opus triage pass filters false-positives and out-of-scope items on both CLIs
• •Only actionable findings loop back to Phase 3

Phase 2 now refuses to start Phase 3 on underspecified tickets without a human read-through.

• •Clarification mode (max 2 rounds): if the Jira/issue is ambiguous — missing acceptance criteria, UI task without a Figma link, API task without an endpoint contract, vague language, parent-story scope drift — the pipeline asks structured questions, regenerates the plan with answers, and caps at 2 rounds to prevent grooming loops.
• •Approval loop: plan rendered with summary / approach / risk / scope / files / todos. User responds with 'onayla' / 'iptal' / free-text edit (e.g. "let's touch the auth service too but keep LoginView out of scope"). Edit requests loop through Opus + validator until approve or abort. No iteration cap — user controls exit.
• •Skipped for --dev, autopilot, and --dev autopilot to preserve their fast/zero-interaction contracts.

Browse and launch tasks from your terminal.

• •/multi-agent:jira - lists your open Jira issues, pick one to start
• •/multi-agent:issue - lists unassigned GitHub issues, auto-assigns on pick
• •Shared interactive flow: choose branch → mode → autopilot → launch
• •No more copy-pasting URLs

Per-phase token limits for cost control.

• •Defined in token-budget.json with warn/max thresholds
• •smoke-token-budget.sh enforces in CI
• •Total pipeline cost: 13,282 tokens (−28% from v2.6)
• •Phase docs lazy-loaded - only current phase is read

Purpose-built agents for each concern.

• •code-reviewer - CLI-aware parallel review + Opus triage (Claude: 2-model · Copilot: 3-model)
• •explorer - Phase 1 codebase scanner
• •ios-architect - iOS architecture review
• •android-architect - Android architecture
• •backend-architect - API design review
• •security-auditor - OWASP-based security audit

6 automated scripts installed to ~/.claude/scripts/.

• •pre-commit-check.sh - secret detection (API keys, AWS AKIA, private keys, .env, Firebase JSON)
• •stack-swap.sh - auto-detect project stack on session start
• •keychain-save.sh - save tokens/JSON to macOS Keychain (credentials never visible in ps)
• •github-ssh-setup.sh - SSH key generation + config
• •ui-tree-dumper.swift - iOS accessibility tree dumper
• •sync-parity-check.sh - Claude↔Copilot↔repo drift detection

Multiple modes for different workflows.

• •Normal - all 8 phases with full review and Report delivery
• •--dev - Init → Dev (Opus) → Commit → Report for fast iteration
• •--local - works directly on your branch without worktrees
• •autopilot - skips all confirmations for zero-interaction runs
• •Combine: --dev autopilot is the fastest path from issue to PR

Comprehensive CI-ready test coverage.

• •13 test suites (62 tests) covering every script and phase doc
• •10 adversarial eval fixtures for triage semantic regression
• •4 JSON schema validations - all runnable via 'npm test'
• •smoke-pre-commit.sh tests secret detection with real temp git repos
• •smoke-token-budget.sh enforces per-phase token limits

Auto-detects project stack and loads matching guides.

• •iOS/Swift, Android/Kotlin, Backend, Frontend, or mobile combined
• •Switch on the fly with /multi-agent stack [platform]
• •Platform-specific compliance rules map 1:1 to audit checks

Automated visual testing and compliance audits.

• •Boots simulators, captures screenshots across light/dark and large text
• •Accessibility audits and store compliance checks
• •Uses native CLI tools (xcrun simctl, adb, codesign, aapt2)
• •On-demand: dark mode, accessibility, store-ready, biometric, performance

Incremental learning that persists across sessions.

• •Captures patterns, decisions, and architecture notes per project
• •Phase 7 extracts user corrections, constraints, and references into auto-memory
• •Phase 1 injects memories so future runs respect prior feedback
• •Max 3 new memories per run, merge-over-duplicate strategy
• •Context auto-compacts at 65% usage (vs default ~80%)

Full component generation from Figma URL.

• •Produces Configuration, View, Modifiers, Code Connect, and docs
• •Design token mapping and accessibility IDs
• •3-layer testing: ViewInspector + Snapshot + Unit
• •14-item review checklist
• •Wiki page + issue-body update as nested SubPhases of Phase 3
• •Activated automatically when task contains a Figma URL

Full command parity between Claude Code and Copilot CLI.

• •18 slash commands: issue, jira, status, log, resume, kill, review, purge, dev, test, autopilot, help, sync, setup, enrich, refactor, dev-autopilot + main
• •Claude Code: /multi-agent:command — Copilot CLI: multi-agent-command
• •/multi-agent:sync keeps both platforms identical
• •/multi-agent:refactor scores codebase, researches competitors, plans improvements
• •/multi-agent:setup onboards Keychain tokens + Git identity

All external delivery consolidated in Phase 7 Report: Jira comment, Component Wiki + Figma screenshots, Confluence (optional), then internal report + knowledge capture.

• •Four wiki layouts supported: submodule, in-repo, github-wiki, separate-repo
• •Auto-detects mode from figma-project-config or project structure
• •Fetches Figma screenshot via REST API (auto-scaled 2x, LTR+RTL, light+dark)
• •Wiki → Jira comment triad (humanizer'd) closes the loop with QA/PMs
• •local-overlay guard prevents /multi-agent:sync from flattening v5 customizations
• •Issue progress flags auto-updated: Wiki ⚪ → 🟢

Deterministic validation of Phase 4 triage output.

• •validate-triage.mjs runs after every review
• •Exit 0: valid · 1: invalid (retry+fallback) · 2: over-rejection guard (>80%)
• •Exit 3: contradiction auto-corrected (bidirectional approved↔blocking)
• •triage-output.schema.json v3 with reviewer enum locked to {opus, sonnet}
• •'fix' mandatory on accepted items

Stateful card-stack renderer for both CLIs.

• •phase-tracker.sh renders identical ASCII cards (○ ● ✓ ✗ ↷) with sub-phases
• •State persists per-task with atomic writes for concurrent worktrees
• •Phase 0 Step −1 bootstraps the tracker as the first thing every run

Per-model cost tracking and pipeline analytics.

• •log-metric.sh appends to metrics.jsonl (atomic, never fails pipeline)
• •aggregate-metrics.mjs: review cycles avg/p95, triage rates, cost rollup
• •Per-model breakdown: calls + duration_ms + tokens_in + tokens_out
• •Phase 7 embeds last-30-day summary

10-fixture regression suite for triage prompt quality.

• •Cases: empty findings, real blocker, out-of-scope defer, false positive, mixed
• •Plus: severity mismatch, duplicate findings, stylistic misclassification, cascading, deferred with cross-reference
• •eval-triage.mjs: structural validation + coverage + no-invention assertions
• •Runs in CI on every push

One skill set, two CLIs, zero drift.

• •Previously Claude got 84 iOS-only skills, Copilot got 60 Android/web + pipeline skills — zero overlap
• •v3.5.0 merges everything into pipeline/skills/shared/ (145 total)
• •Both Claude Code and Copilot CLI install the same source
• •Covers: pipeline orchestration (multi-agent-*), iOS/SwiftUI, Android/Compose, Kotlin, Swift, web (React/Next/Vue), backend (Python/Node/FastAPI), CI/CD, HIG, and more
• •Pipeline is now truly generic — no leaked Jira keys, project directories, or domain references