Changelog

All notable changes will be documented here.

This project follows Semantic Versioning.

Unreleased

Added

• A machine-readable Perplexity, Firecrawl, Playwright, Claude Code, Codex, and Agent Plugins MCP compatibility matrix plus a read-only doctor, provenance snapshots, clean-client test plan, negative fixtures, and fail-closed portable-package gate.
• An additive Agent Plugins 1.0 portable, skills-only package with a closed root manifest, contained fixed-location skill discovery, an offline doctor, deterministic tests, and an explicit separation from Codex-native metadata and project-local MCP configuration.
• A strict, read-only capability decision engine that combines active profiles and durable task evidence into built-in, recommended, optional, missing, or blocked states with provenance, authority limits, risks, and safe next steps; it validates but never executes plan-only adapters.
• A deterministic specialist capability broker with reviewed Agency Agents provenance, profile-aware task routing, reversible manifest activation, required risk reviews, local-role fallback, independent-evaluator evidence, and no automatic external installation or runtime-authority expansion.
• Profile-aware Emil Kowalski design-engineering integration with a reviewed, pinned ten-skill manifest; a local anti-slop/motion router; explicit-only prototyping, library-selection, and animation-review gates; Claude Code and Codex installation; and deterministic coverage tests.
• Resumable PR finalization that recovers only an exact interrupted task-ledger transition, re-verifies it before commit, and waits a bounded interval for GitHub to register required checks before reporting a safe retry.
• A registry-backed ./agentic command interface that groups 24 supported workflows, classifies every shell file, hides internal policy helpers and security hooks from public discovery, and preserves existing script paths as compatibility adapters during a measured migration.
• A human-approved PR finalizer with dry-run, clean-worktree and PR identity checks, ledger-only staging, recovery from an already-ready PR, required- check waiting, and explicit guarantees that it never approves or merges.
• A machine-readable Claude Code and Codex runtime policy plus a read-only advisory/strict/JSON doctor that separates compatibility from authority and keeps optional self-hosted, cross-session, plugin, MCP, and automatic- approval capabilities human-gated.
• A read-only post-merge closeout report that verifies default-branch task truth, merged PRs, closing issue state, volatile handoff claims, and safe local cleanup targets without performing any mutation.
• A deterministic GitHub Issue ↔ task ↔ PR contract with explicit issue-free exceptions, multi-task closure safety, task-planner guidance, required offline policy validation, and optional read-only live drift inspection.
• Seven project-scoped, read-only Codex specialist roles for product, architecture, research, design, security, QA, and integration review, with deterministic authority and schema validation.
• A native Codex adapter with shared repository skills, bounded project configuration, reviewed safety-hook wiring, a skills-only plugin manifest, runtime doctor, multi-terminal worktree guidance, and automated drift tests.
• A Playwright visual-regression gate for the Showcase's normal, loading, empty, and error states across desktop and mobile, with reviewed Linux baselines and failure artifacts.
• An adaptive product-design intake that asks only relevant questions for the active platform, creates comparable visual directions, and requires explicit human approval before canonical design-system or token changes.
• Mode-aware semantic color tokens, light/dark parity validation, required WCAG contrast checks, and a generated token specimen for review evidence.
• A ready-for-review PR policy gate that requires the linked task to be done, preventing merged work from leaving stale execution state behind.

Security

• Portable MCP packaging remains blocked until credential references, deterministic package resolution, protocol negotiation, client trust and rollback, and independent security evidence all pass; no MCP was installed, authenticated, contacted, or executed during the review.
• Claude Code 2.1.224–2.1.225 version guidance for filesystem-deny, cross-session, self-hosted, workspace-trust, and OAuth hardening, with optional runtime surfaces disabled by default.
• Codex 0.147.0 capability gates for portable plugins, MCP 2026-07-28, and automatically reviewed approvals without installing or enabling them.
• Version-qualified Claude Code guidance for credential-file masking, zsh permission hardening, and corrected nested-subagent defaults without enabling new runtime settings or exposing credential paths.
• Version-qualified Claude Code worktree/background-agent fixes and Codex cyber-model review defaults without enabling models, permissions, hooks, or remote-control capabilities.
• Version-qualified Claude Code 2.1.223 guidance for shell approval, workflow-sandbox, and managed bypass-permission fixes without changing the repository's runtime or permission configuration.

0.1.0 - 2026-07-25

Added

• A repository-centered agent harness with durable product, design, engineering, execution, evaluation, and collaboration context.
• A phase-based product-design engine with product-specific authority, specialist routing, independent critique, and Playwright evidence.
• DTCG-compatible design-token sources and generated CSS, TypeScript, and React Native outputs.
• Controlled, evidence-gated daily ecosystem research that proposes changes without modifying the harness autonomously.
• A complete product-design resource catalog with source links and routing.
• Deterministic, non-destructive project profiles and a guided initializer with explicit active and inactive capability plans.
• A confirmation-gated task launcher that turns TASKS.jsonl entries into reviewed branch or worktree plans.
• Signalroom, a responsive reference product for supervising AI-agent work, including interaction, accessibility, and visual evidence.
• GitHub issue, branch, pull-request, review, CI, security, and contribution contracts for human-agent collaboration.

Security

• Deterministic hooks for destructive-command prevention and secret detection.
• Human approval gates for releases, deployments, credentials, destructive migrations, and other irreversible actions.