Commit As You, Ship As You Go

New

• Commits are authored as you — Every commit and tag Trinity makes now lands under your own author identity instead of a generic bot. For providers that don't expose a public commit email (Bitbucket, Forgejo/Gitea), you can set a commit email right in the account connect flow, or override it per account in Settings. If a commit would otherwise land under the wrong author, Trinity pauses and asks you to set one — and resumes exactly where it left off the moment you do — rather than quietly committing as someone else.
• Work is saved as it's built — The implementer now commits incrementally while it works through a story, so progress lands on the branch continuously. If a run is interrupted, completed work is already committed instead of lost, and resuming picks up from what's there rather than starting over.
• Reset a story from scratch — You can now reset a story back to a clean slate: Trinity closes its open pull requests, deletes its branch and worktree, and clears its run history so you can start it over fresh. Merged stories are protected, and a story that's currently running has to be stopped first.
• Native release promotion — A new promotion worker walks a release through each repo's staging states on its own, with a release-status rollup so you can see at a glance where every repository stands. The release queue now surfaces blockers and the PRDs that make up each release.
• Markdown and diagrams in AI conversations — Assistant messages, onboarding discussions, and handoff and failure reports now render full markdown, and authoring agents can draw Mermaid diagrams. Explanations, tables, and architecture sketches come through formatted instead of as a wall of raw text.

Improved

• Real-time updates are on by default — Live updates are now enabled everywhere, so changes from your runs and your teammates show up near-instantly; periodic refresh stays on only as a fallback.
• Smarter Start Run scope — Starting a run now cascades the scope from the PRD you pick, so you begin with the right stories selected instead of the whole release.
• Recover a failed story without starting over — When a story hits a hard failure — the agent hangs without reporting back, times out, or runs out of retries — Trinity now parks it at a recovery gate instead of quietly marking it failed. Open a diagnostic chat to investigate and re-run it with your guidance, or cut it loose; a suggested fix from the architect is usually already waiting. Failures that can't be fixed by retrying are caught after a single attempt instead of burning through every retry first.
• Per-story cost in recaps — Recaps now show the AI spend attributed to each story right on its line, so you can see what each piece of work cost at a glance.
• Clearer run controls — Gate action buttons give immediate feedback the moment you click them, and the story currently running pulses in the dependency graph so it's obvious what's in progress.
• Sign-in and subscription checks are faster — Reworked how subscription status and sign-in tokens are cached and shared, so they resolve quicker and no longer trip rate limits when several people sign in at once.
• Runs ride out transient AI hiccups — When the AI service returns a brief, temporary error mid-run, Trinity now retries with backoff instead of failing the story, so a momentary network or backend blip no longer derails a run.
• Sharper run status — The Run page now shows a "Preparing" step while a job spins up, derives the PR, Merge, and Done stages from what actually merged, and labels each queued story with its real state instead of a blanket "Queued".
• Checkpoints show inline in recaps — A checkpoint now appears as a badge right alongside the work it covers instead of in a separate section, and recaps refresh the moment a job finishes.
• Smarter checkpoint reviews — Checkpoint analysis and audits now act on warnings, not just critical issues, and take the project roadmap into account.

Fixed

• Stopping a run no longer looks like a failure — Killing or stopping a run now reads as a deliberate stop, with no misleading error toast or spinner left hanging.
• No more lost work on "already done" stories — Fixed cases where a story Trinity judged already-satisfied could discard work it had just committed. Stories that finish without needing changes now count as done instead of waiting forever to merge.
• Runs resume cleanly after an interruption — Resuming a story skips the phases that already finished and recovers its committed files correctly, instead of redoing work or wiping progress.
• Open pull requests aren't flagged as closed — An open PR is no longer incorrectly reported as closed when its status couldn't be read from the provider.
• Stuck jobs stop retrying forever — A job that has terminally failed is now stopped instead of being re-queued endlessly.
• Onboarding questions keep their options — Fixed the onboarding Discuss chat dropping the answer options it should reshape as the conversation moves.
• Story links work by display ID — Opening a story by its display ID (like 1:1.1.1) now works directly, not just by internal ID.
• Notifications are fully opaque — Error and success toasts no longer render semi-transparent and hard to read.
• Pushes recover when the remote is already ahead — A push that fails because the branch already contains your work on the remote now self-heals instead of erroring out.
• Real merge conflicts are reported right away — A genuine merge conflict now fails fast with a clear result instead of being retried as if it were a temporary glitch.
• Stopping a run cleans up promptly — After Stop or Kill, active and queued work clears within seconds, leftover background processes from a worktree are reaped, and jobs left without a live owner are cancelled.
• The code viewer shows the latest branches — The viewer now fetches the newest branches before listing them, so a branch you just pushed shows up.
• Auto-approve technology deviations saves correctly — Fixed the team- and scope-level setting for auto-approving technology deviations being rejected when you saved it.