Skip to content

rbraga01/a-team

v1.4.1MIT

A complete engineering team in one folder: 26 specialist agents, 20 skills, orchestrated workflows and quality gates.

GitHubOpen Website

Changelog

All notable changes to A Team are documented here. Format follows Keep a Changelog. Versioning follows Semantic Versioning.

[Unreleased]

[1.4.1] — 2026-10-16

Added

  • Portable root plugin.json for Codex and Agent Plugins; Codex lifecycle hooks use the single root hooks/hooks.json instead of a duplicate under .codex-plugin/.
  • OpenCode project commands now forward $ARGUMENTS, include frontmatter for discovery, and preserve the explicit human approval stop before implementation/dispatch.
  • Offline tests validate the portable manifest, Codex hook location and 10 OpenCode commands.

Repository consistency and distribution audit (October 2026):

  • Added an automated negative-control CI test that deliberately corrupts a catalog row and requires the inventory checker to reject it.
  • Added a complete agent and skill catalog under docs/catalog.md and moved detailed operational documentation into docs/how-it-works.md.
  • Extended the CI consistency checker from numeric matching to exact catalog membership, and added Bash/PowerShell installer tests (including a Windows runner).
  • Included CLI integration directories and artwork in all installation methods, plus non-overwriting root AGENTS.md and CLAUDE.md project instructions, while keeping plugin activation as a separate, host-specific step.

Consistency shield — counts can no longer drift silently:

  • packs.json — single truth source for the builder-* domain packs (name, repo, pages URL, version, skill/agent counts).
  • check_consistency.py extended from 23 to 75 checks: agent count is now enforced from the real truth source (.claude/agents/*.md file count) across 23 references, skill-count coverage extended to CLAUDE.md, the claude plugin manifest, marketplace.json, and CITATION.cff, and packs.json is enforced against README, docs/index.html ecosystem cards, and marketplace.json (roster, versions, and advertised skill/agent counts).
  • scripts/check_packs_remote.py + packs-sync workflow — daily cron, workflow_dispatch, and repository_dispatch (type pack-updated) compare packs.json against the LIVE builder-* repos: roster, real file-tree skill/agent counts, manifest version, and GitHub description. Drift opens (or updates) a pack-drift issue automatically and closes it when resolved.

Fixed

  • Corrected missing agent/skill table entries and classified api-contract-first and data-migration as hard gates.

  • Fixed Codex and Cursor plugin-root paths and a missing Codex icon reference. Declared package hooks where supported; actual CLI activation and enforcement remain subject to end-to-end verification.

  • Removed inaccurate multi-CLI enforcement-parity claims from installation guidance.

  • Fixed the PowerShell installer on environments without an explicit TEMP variable, and made it merge missing files into existing directories without overwriting project settings.

  • Updated GitHub Pages installation examples to match actual distribution contents.

  • Condensed the root README and surfaced an illustrative orchestrator quick start.

  • builder-growth was advertised as "6 skills and 3 agents, v1.0.0" in marketplace.json and the landing-page ecosystem card; the repo actually ships 14 skills and 5 agents at v1.1.0. Corrected everywhere, including the builder-growth GitHub repo description.

Release note

  • Patch release focused on packaging, installation and documentation consistency; no new agents or skills.
  • GitHub Copilot CLI plugin functionality and real-host parity are not verified; its manifest version is synchronized solely as release metadata, without code changes.
  • Native host activation for Codex, Cursor, Claude Code and OpenCode remains tracked in issues #34–#37.
  • OpenCode v2.0.30 CLI command discovery was attempted without model/API keys, but its local command registry returned no entries even for a synthetic JSON control; host execution remains unverified.

[1.4.0] — 2026-07-19

Added

1 new skill (total: 19 → 20):

  • architecture-audit — systematic architecture review workflow. Maps the system as-built (not as-documented), identifies load-bearing decisions, dispatches parallel specialist reviews (architect, security-reviewer, performance-profiler, database-reviewer, infra-reviewer), and produces evidence-backed findings with severity verdicts (SOUND / SOUND-WITH-RISKS / INTERVENTION-REQUIRED). Every finding requires a file reference and a consequence. Registered in using-a-team under Architectural Decisions.

Changed

  • Skill count updated to 20 across README.md, CLAUDE.md, docs/index.html, docs/overview.md, CITATION.cff, and all plugin manifests.

Fixed

  • Corrected unsupported sparse-checkout arguments in both installers and added end-to-end regression coverage for Bash and PowerShell.
  • Restored runtime enforcement scripts to the installed scripts/ directory while keeping generated project state under .agent-sync/.
  • Made watcher process detection safe on Windows.

[1.3.0] — 2026-06-16

Added

Platform: GitHub Copilot CLI (4 → 5 supported platforms):

  • New .copilot-plugin/plugin.json declaring A Team's agents (../.claude/agents/) and skills (../skills/) for Copilot CLI auto-install.
  • New .copilot-plugin/hooks/hooks.json carrying SessionStart, PreToolUse, PostToolUse, and SessionEnd hooks — full enforcement parity with Claude Code's .claude/settings.json (not just SessionStart like Codex CLI today).
  • Install command: copilot plugin install RBraga01/a-team:.copilot-plugin.
  • No new agents, skills, or rules — purely additive platform plumbing.

Changed

  • README.md, CLAUDE.md, AGENTS.md, CONTRIBUTING.md, and the PR template updated to reference Copilot CLI as a supported platform.
  • Version bumped to 1.3.0 across all plugin manifests for consistency.

[1.2.0] — 2026-06-16

Added

1 new agent (total: 25 → 26):

  • typescript-reviewer — TypeScript / React / frontend review specialist. Covers TS strictness (strict, no any/@ts-ignore), React idioms (no state mutation, effect deps, focused components), design-state completeness (loading/empty/error/partial/uncertain/success/streaming), accessibility, PWA/service-worker correctness, and client-side security (XSS, token handling). Enforces the full TS gate (tsc --noEmit && eslint . && vitest run), never a subset. Tier 2. Fills a real gap: A Team shipped Go/Python/Rust/Kotlin/Swift/Flutter reviewers but no TypeScript/JS specialist — the most common web stack.

[1.1.0] — 2026-06-02

Added

2 new workflow skills (total: 16 → 18):

  • five-whys — Root cause analysis skill for structured 5-Why chains; guides agents to escalate through causal layers before proposing a fix
  • smart-init — ROADMAP-aware session initialisation; reads ROADMAP.md at session start to set context, runs an interview script when none exists, and writes INIT.md

Observability stack:

  • scripts/metrics.py — Shared module with append_metric and read_events for structured event logging
  • scripts/metrics-report.py — CLI with parse, report, and log rotation subcommands
  • scripts/status.py — Cross-platform status line script with staleness detection; wired into Claude Code statusLine and all platform SessionStart hooks

Safety guard:

  • Pre-tool-use hook with session export — prevents silent failures on tool calls (merged from imports-v1)

[1.0.0] — 2026-05-30

First public release

25 specialist agents across five categories:

  • Leadership: orchestrator, architect, planner
  • Quality: code-reviewer, security-reviewer, tdd-guide, debugger, refactor-cleaner, build-error-resolver
  • Language specialists: go, python, rust, kotlin (Android), swift (iOS), flutter (Dart), database (PostgreSQL)
  • Domain: infra-reviewer (IaC), compliance-reviewer (GDPR/COPPA/PCI/SOC2/HIPAA), ai-reviewer (LLM code), performance-profiler
  • Operations: e2e-runner, doc-updater, chief-of-staff, loop-operator, harness-optimizer

18 workflow skills — 6 hard gates + 12 workflow skills:

  • Hard gates: verification-before-completion, test-driven-development, brainstorming, systematic-debugging, api-contract-first, data-migration
  • Workflow: using-git-worktrees, subagent-driven-development, executing-plans, writing-plans, dispatching-parallel-agents, finishing-a-development-branch, incident-response, performance-audit, writing-skills, using-a-team (meta), five-whys, smart-init

Key capabilities:

  • Orchestrator init + prune protocol (INIT.md → automatic team sizing)
  • File Lock Protocol with worktree path normalisation (prevents parallel agent collisions)
  • Pipeline Auditor (harness-optimizer) with merge blocking on rule evasion
  • Surgical Changes enforcement across all agents
  • Model-agnostic tier system: Claude, OpenAI (GPT-4.1/o3/GPT-5.5), Google (Gemini 2.5)
  • Cross-platform support: Claude Code, Codex CLI, Cursor, OpenCode

Platforms supported:

  • Claude Code (full: hooks, slash commands, sub-agent dispatch)
  • Codex CLI (full: agents, skills, rules, session hook)
  • Cursor (strong: agents, skills, rules, session hook)
  • OpenCode (good: 10 command aliases)