Initial Beta Release

New

• Multi-Agent Execution Pipeline — Stories flow through Analyst → Implementer → Auditor (with simplify loop) → Documenter, each stage run by a dedicated AI agent via the Claude CLI. Automated PR creation, merge, and conflict handling
• Parallel Execution — Run many stories concurrently with a configurable worker pool, each isolated in its own git worktree per repo. Workers know which release they belong to, so release context flows all the way into agent prompts
• Checkpoint & Release Pipeline — Analyst → Audit/Fix loop → Refactor → Audit/Fix loop → Documenter → Knowledge Base consolidation → human gate approval. Multi-perspective quality audits scale with story difficulty
• First-Class Releases — Group PRDs into shippable units with a lifecycle (created → ready → releasing → released), a dependency DAG, per-repo semver via configurable tag_template + bump_strategy, AI-generated release notes per target, per-release automation overrides, and retry tagging on partial failure. Releases (not PRDs) are Trinity's primary unit — dashboard, stories list, story detail, and Architect all scope by active release, with a release selector that syncs across devices
• Architect — Conversational entry point for changing your plan. Classifies requests into story | epic | phase | prd | roadmap scope, returns a mode (add / modify / restructure / new-prd), and surfaces smart signals: release hint, new-target detection, stack implications, PRD size warnings, actionable risk callouts. Every PRD is assigned to a specific release — Architect suggests existing vs. fresh releases or auto-creates the first one. Sessions persist across refreshes and sync across devices
• Architect Restructures Mid-Project — When a request implies a structural change, Architect proposes one of add_target_same_repo | add_target_new_repo | consolidate_to_mono | consolidate_to_turbo. New MCP tools (add_project_repo, update_project_target, merge_project_repos) let it re-home targets, spawn new repos, or collapse a polyrepo — gated by a project-quiet check so it can't race in-flight work
• PRD Planning — AI-powered roadmap generation via a 5-phase pipeline (Architect → Story Writer → Dependency Mapper → Target Mapper → Calibrator), producing phases, epics, and stories with dependency graphs and per-story target assignments
• Onboarding Wizard — Multi-step guided project setup: discovery chat, tech-stack refinement, phasing, design preferences, project structure picker, per-repo draft review, section-by-section roadmap review, service API-key setup, and Claude skills configuration. The Repos step auto-suggests each repo's GitHub description and shows a per-repo "Contains" preview of which targets will land where, so placement is visible before you click Create. Roadmap generation runs sections in parallel and streams them in as they finish, with a Back-button guard that prevents falling back into Repos mid-generation. The discuss-this-question chat can propose fresh options when none of the listed answers cover what you're describing — the chat panel renders the proposed options inline with an "Apply these options" button so you can steer the question itself, not just pick from preset answers
• Project Import — Bring an existing codebase under Trinity. 8-step flow (scan → files → repos → secrets → docs → skills → save → done) with deep codebase analysis, quality audit, and PRD cross-referencing so already-planned work doesn't clutter the audit checklist. A dedicated Files step lets you exclude any folders or individual files from the workspace snapshot before capture runs. Polyrepo imports show explicit target→repo mapping you can override. Per-repo "Read-only dep" toggle marks upstream forks or vendored libraries — Trinity clones them so agents can read but never branches, commits, PRs, merges, or tags them. Ambient (non-repo) files at the import root are snapshotted into the workspace so trunk is fully reconstructible from server data on any machine
• Monorepo / Turborepo / Polyrepo — The structure picker answers one question: how is the project organized in git? Plain-folder monorepo, declared-workspace turborepo (detected via pnpm / turbo / nx / workspaces config), or polyrepo (separate repos at a workspace root). Roadmap and first-PRD prompts shape themselves around the answer
• Polyrepo Execution Gates — Three gate types specific to multi-repo work: merge_conflict (retry / skip / abort), external_pr_closed (someone closed the PR on GitHub outside Trinity — recreate / skip / abort), and partial_merge_failed (release tagged some repos but not others — retry the failures or accept a partial release)
• Placeholder Audit at Checkpoint + Release — Full-worktree scan with two independent regex families: images (placeholder.svg, picsum.photos, Lorem ipsum) and contacts (example.com, 555-*, John Doe, stale copyright years). Per-family skip toggles, optional agent triage that classifies each hit as leak / fixture / docs / false_positive with suggested replacements. On-demand scrub available from project settings
• Agents Can Raise Their Hand — Any pipeline phase can call the signal_blocked MCP tool when it genuinely cannot make progress. The story_blocked gate carries the escalating phase, a prose explanation, and the list of approaches already tried — so you know exactly what not to suggest
• Agents Own Their Repo Docs — AGENTS.md is the canonical per-repo agent doc; CLAUDE.md and GEMINI.md are thin pointers. Trinity seeds it once, then agents enrich a ## Project-specific context section as they learn durable facts. Polyrepo projects also get a trunk-root AGENTS.md for cross-repo context — stored as a workspace asset that round-trips through the merge step, so edits persist across worktree rebuilds and sync across devices
• Design Guide System — Per-target design references (Web App, Website, Mobile, Desktop, CLI, Library, Extension) loaded automatically by the UX skill based on your project targets
• Per-Target Accessibility — WCAG 2.1 conformance level (A / AA / AAA / off) per UI-bearing target, so a polyrepo can hold a marketing site at AA next to an internal admin tool with no a11y requirement. Story prompts take the strictest level across the targets the story is mapped to; releases use the strictest level across the project. Onboarding's a11y question fires after targets when at least one UI target exists
• Knowledge Vault — Database-backed knowledge base (books, chapters, pages) with markdown rendering, tagging, search, and rich zip + PDF exports. Agents query it mid-execution to compound project knowledge over time
• Gotchas Library — Stack-aware common issues and solutions for your languages, frameworks, and tools
• Codebase Map — Living file index with AI-generated descriptions, synced across team members. The Project Settings → Codebase Rescan button enqueues a first-class task that spans both scan and describe phases, so the Tasks pill stays accurate end-to-end and a completion toast lands you back on the codebase tab
• Code Viewer — Read-only code browser with syntax highlighting (CodeMirror 6), branch switching, file search (Cmd+P), and virtualized file tree. Uses an isolated bare clone so browsing never interferes with execution
• Dependency Graph — Visual DAG of your stories with saved layouts, highlighting, and fullscreen mode
• Align — Phase-adaptive project health check that evaluates roadmap quality pre-PRD and story/execution health post-PRD. Results auto-save to the vault and stay interactive on reload
• On-Demand Codebase Audit — Scan any codebase for quality gaps, security risks, and improvement opportunities. Audit pages feed directly into Architect via an "Address in Architect" button
• Docker Integration — Per-story Docker Compose isolation with automatic port allocation
• Model Provider Cascade — Settings → AI Models has three selects per tier: harness → provider → model. Switch providers without losing your model choice. Supported: Anthropic (Claude Opus 4.7 default for reasoning, Sonnet + Haiku for lighter tiers), DeepSeek, Ollama, Moonshot (Kimi K2/K2.5/K2.6/K2 Turbo), Z.ai (GLM 4.5/4.6/4.7), Qwen (Coder Plus/Next/Flash, Qwen3 Max, Qwen3.5 Plus/Flash), and automatic fallback between them
• 5-Level Effort Ladder — Every Anthropic call uses the canonical low | medium | high | xhigh | max ladder. The harness auto-clamps when a model doesn't support the requested level, and ai_events.effort records the clamped value for accurate cost and latency attribution
• Per-Entity Automation Overrides — Override automation and model settings at the story or release level. Cascade is global → team → project → entity → job
• Platform Quality Checks — Preflight checks for TypeScript, iOS, Android, Web, API, CLI, Desktop, Website, Library, and Extension targets
• PDF Reports with Charts — Executive and technical reports with cover pages, visual charts, and polished formatting. Uses any installed Chromium-based browser (Chrome, Edge, Brave, Chromium) for PDF export — and if none is installed, Trinity offers to install one for you in-place via the same setup-runner panel that handles GitHub CLI install (brew install --cask google-chrome on macOS, per-distro chromium package on Linux)
• Teams — Team scopes with Turso-replicated databases, roles (owner | manager | member), permission cascade, real-time presence, GitHub repo permission gating, collaborator management (invite / auto-invite / revoke), transfer policies with approval workflows, and business email enforcement
• Activity Feed — Global audit trail with filters (project, category, actor), field-level entity history, before/after diffs, provenance badges, and comments
• Seat Sponsorship — Gift subscription seats to other users via token-based accept/decline. Unified /requests page surfaces pending invites, sponsorship offers, and project transfer approvals
• Multi-Account Sign-In — Stay signed into multiple Trinity accounts at once. Account picker lets you add, switch, and sign out of individual accounts with parallel device-code flows; each account keeps its own warm sync pool and workspace state
• Free Trial — New accounts automatically start a 15-day free trial with full feature access. Trial-expired users see a dedicated screen with a subscribe CTA
• Encrypted Secrets, Cascade & Materializer — AES-256-GCM encrypted store with a three-axis schema (purpose × target × sensitive). Cascade resolution at consume time picks target-specific → project-wide → global for the same key — one project-wide value works for every target unless you override. Top-level Vercel-style Secrets page in the sidebar with table view, target filters, purpose filters, paginated grouping by key name. Add and edit dialogs let one key fan out to multiple targets at once. .env paste import with per-key conflict policy (keep / overwrite / skip) committed in a single transactional write. File-purpose secrets carry destinations (target, repo, or trunk base) and get materialized to disk under a Trinity-managed .gitignore block — removed secrets unlink the file and prune the line on the next pass. Team secrets share a team key. Decryption happens server-side — keys never reach the desktop
• Setup Runner & Service Setup Gate — pty-driven CLI orchestrator that walks declarative pipelines (regex match, send-to-pty, OTC + URL capture for device-code flows, paste prompts for post-completion values, and a surface_action step kind for "do this yourself, then click I'm done" gates without a pty). Pipelines have a cwd_scope so service CLIs run inside the right target dir, and an optional outputs[] declaration lets the runner capture config files the CLI emits — a server-side classifier sniffs filename + content and decides whether each output becomes an encrypted env var, an encrypted on-disk file, a public asset, or a .gitignore-only entry. Triage on exit (Haiku) classifies the run as success / failure / cancelled. Worker pre-flight raises a service_setup_required gate before the analyst phase if any stack row that needs keys isn't configured — the gate dialog runs the same setup panel inline, pre-bound to the missing target. Agents can also raise the gate mid-execution via the signal_service_required MCP tool when they discover a service the pre-flight check didn't catch. Reference playbooks: gh auth login (requests delete_repo scope upfront and chains gh auth setup-git so HTTPS push/fetch flips green afterward), plus a full homebrew-install → gh-install → gh-auth prerequisite chain that walks the user from a clean machine to a working gh (per-distro Linux dispatch for apt / dnf / pacman / zypper / apk; no-spawn surface_action gate to brew.sh since Trinity won't curl … | bash on the user's behalf)
• Pluggable Cloud Storage — Managed Trinity Cloud (R2) by default, or bring your own S3-compatible backend (AWS S3, Cloudflare R2, Backblaze B2, DigitalOcean Spaces, MinIO). Credentials cascade from team / global defaults; usage bars show consumption against quota with threshold-based warnings. The reset / delete dialogs surface a Delete uploaded assets toggle that follows storage type — locked on for managed (Trinity-hosted assets can't outlive the project), off by default for BYO S3 / Local Only so you keep the files you've uploaded
• Pre-Flight Tool Checks — Contextual verification of git, gh, and claude CLI before project creation, project deletion, and execution. The gh check verifies that gh auth setup-git has routed git's credential helper through gh (so HTTPS push/fetch can't hang on a stdin prompt) and that the token carries the delete_repo scope before reset/delete dialogs run; tokens missing the scope get a one-click gh auth refresh flow embedded directly in the alert. Missing or misconfigured tools surface the exact fix command inline
• Run Page — Splits into Active (running + gated stories with a live pipeline caption showing the current agent) and Queue (pending / ready / awaiting-merge / upcoming). Gate review is a focused modal with per-gate-type bodies — checkpoint gates tab between Overview, Findings, Fixes, Preflight, Notes, and Consolidation
• macOS + Linux Desktop App — Native Tauri v2 app for macOS (Apple Silicon) and Linux (amd64 + arm64) with OS-keyring auth, a custom draggable title bar on macOS, and an execution-aware auto-updater that never interrupts running stories
• Theme System — Six built-in themes: Dark, Light, Trinity Dark (neon-green Matrix aesthetic with monospace font), Trinity Light, Cyber Dark (cyberpunk purple/magenta), Cyber Light — plus system-follow mode
• Help Assistant — Floating chat with context-aware answers from the built-in user guide
• Bug Reports — In-app bug reporting with screenshot attachments, submitted to the admin dashboard
• Built-in User Guide — 27-page guide shipped as markdown, auto-seeded into the knowledge base on startup. On first sign-in the Knowledge and Gotchas pages auto-refresh while the background seed batch runs, so content lands without a manual reload