Tune Every Story

New

• Per-story execution settings — Every story now carries three knobs you can see and edit on its page: how many audit passes the Auditor runs, which model tier implements it, and how much reasoning effort that model spends. Trinity's calibrator picks sensible values for each story during planning; you can override any of them when you know a story needs more (or less) scrutiny.
• Audit passes, set per story — A story's audit passes are an explicit count. Set it to 0 to skip the audit loop entirely on throwaway work, leave it at 1 for the typical story, or raise it up to 10 for multi-concern or security-critical changes. On quality checkpoints the same control sets how many audit-and-fix rounds run, where 1 is an audit-only gate (it reports but doesn't auto-fix) and 2–3 fixes and re-checks until the audit comes back clean.
• Effective-effort indicator — When you pick a reasoning effort, the story page shows the effort that will actually run on the model your chosen tier resolves to — so you can tell at a glance when a model caps below the level you asked for, or takes no effort setting at all.
• Architect calibrates the stories it adds — When you use Architect to add a story, epic, or phase to an existing plan, the new stories now come back fully calibrated — their audit passes, model tier, and reasoning effort are judged against the rest of that plan, the same way a story planned from scratch is, instead of all landing on the same defaults. Edit a story's substance with Architect — its title, description, or acceptance criteria — and it re-calibrates that story too, so its settings keep pace with the change. Architect shows a progress ribbon while it writes and calibrates, and you can refresh mid-run without losing it.

Improved

• Quality checkpoints always cover all five review lenses — Every checkpoint audit now runs the full set of review angles — security, performance, error handling, maintainability, and feature completeness — every time. The audit thoroughness is fixed; the only thing you tune is how many audit-and-fix rounds it runs, set per checkpoint on its story page.
• Stories you add later still get audited — When you use Architect to add a story to an existing plan, it's now automatically wired into its phase's quality checkpoint, so it goes through the same audit as the stories planned from the start instead of slipping past every gate. (If the phase has no checkpoint, the story stays ungated.)
• Difficulty and surface area are clearer about their role — A story's difficulty and surface-area ratings now describe the story for ordering and reporting only; they no longer secretly decide how the story runs. What controls execution lives in the story's Execution Settings, where you can see and change it.

Fixed

• A failed agent turn tells you what happened — When an agent turn hits an error or you cancel it mid-run, the conversation now shows the failure reason and clears the "thinking" indicator the moment the turn ends. The error persists with the message, so reopening the thread still shows what went wrong instead of an empty gap.
• Agents reliably find your Claude Code CLI — The installed app now locates your system-installed Claude Code CLI when starting an agent, so runs no longer fail to launch because the command couldn't be found.