Stories Sized for Agents, Not Engineers

Improved

• Stories sized for agents, not engineers — The Story Writer's merge test asked "would an engineer naturally do these in the same PR?" — calibrated to human cognition and 2024 LLM throughput. It now asks "can Trinity's per-story pipeline ship this within the audit-pass budget?" Two stories only get split when the split unlocks parallelism (independent lanes that run in different worktrees) or isolates a risk hot-spot (security, real-time, perf-critical) that warrants its own audit attention. Splitting along a sequential chain is explicitly flagged as added cost without speedup. The headline effect: PRDs come out shorter in story count, but each story does more end-to-end work — closer to how Trinity actually executes.
• Difficulty and surface area on every story — Stories carry first-pass difficulty (1–5) and surface-area (small / medium / large) hints right out of planning. The Calibrator phase verifies them comparatively across the plan, overrides outliers, and enforces a healthy distribution — so the values you see in the Stories page actually predict how each story will run.
• Phasing decisions are respected — When you tell Trinity "build the API before the Mobile app" during onboarding, the Architect and Dependency Mapper now treat that as authoritative — phase ordering and cross-epic dependencies match what you said instead of being inferred from scratch.
• Onboarding setup, split in two — The single Setup step becomes API Keys (paste-only services like Stripe, Resend, Anthropic) and Tools & Auth (CLI-driven services like gcloud, firebase, supabase, vercel). Less context-switching, every field still optional, and you can leave anything for later.
• Hybrid CLI flows reuse pasted keys — When a CLI step needs a value you already pasted in the Keys step, Trinity injects it into the CLI's environment automatically. No retyping the same Cloudflare or Supabase token twice.
• Worker pre-flight stops gating on paste-only services — Stories no longer pause with service_setup_required because you skipped pasting an API key. That gate now fires only for CLI-driven services that the agent genuinely can't proceed without; missing paste-only keys surface as a soft warning in Project Settings instead.
• Sturdier roadmap generation — Onboarding refines roadmap sections in parallel waves, so generation finishes faster on multi-target projects. If a single wave fails midway, the partial state is rolled back cleanly rather than leaving the project in a half-roadmap state.
• Story-writer retry on overflow — When a planning batch produces too many stories for the hard ceiling, Trinity retries once automatically. If it still overflows, you get a clear toast with a one-click retry instead of a silent failure.
• Sharper "complex story" flag in reports — The Executive Report's complex-story callout now flags any story at difficulty ≥ 4, matching the reasoning-tier routing used at execution time. The same threshold drives "hard story" callouts across recaps and audits — one definition end-to-end.

Removed

• Milestones page — The standalone Milestones page is gone. Releases are how you group and ship work now; the Releases page covers everything Milestones used to.

Notes

• Auto-update from 0.2.0-beta installs this build in place; no reinstall needed.