Synara 0.3.0 release notes.

Synara 0.3.0

• Automations are a real workspace surface — Synara now has first-class Automations for scheduled agent work, with sidebar navigation, list/detail pages, run history, triage actions, and inline editing.

  This release wires automation contracts, persistence, scheduler leases, run tracking, RPC methods, sidebar badges, Current/Paused views, detail routes, editable fields, previous-run history, and result triage so scheduled work lives inside the same thread/provider/worktree pipeline as normal chat work.

• Heartbeat automations can stop when the goal is met — Heartbeat automations can store an AI-evaluated stop clause, evaluate it after successful runs, and disable themselves with a recorded reason when the condition is satisfied.

  Completion policies now support natural-language stop conditions, dedicated background evaluation, visible completion results, timeout handling, stale-result guards, legacy-row defaults, and archive/read preservation so a stop check cannot silently undo user triage state.

• Automation creation understands natural prompts — The composer can turn automation-style prompts into scheduled drafts, including intervals, daily/weekly timing, cron-like schedules, heartbeat targets, and review dialogs.

  Automation intent parsing now covers explicit and generated prompts, English and Italian stop/schedule phrasing, bounded fast-loop safety, draft review, source-thread handling, restored plan source metadata, inline editing from composer text, and stricter confirmation for LLM-generated automations.

• Scheduled runs are harder to lose or corrupt — Automation scheduling, recovery, and run reconciliation were hardened so crashes, duplicate wakes, approval waits, stale cache updates, and cleanup failures are handled more predictably.

  The automation service now has occurrence dedupe, scheduler leases, crash replay, failed-run rollback, startup recovery, bounded completion-evaluation queues, recovery/lease observability, approval ownership re-checks, standalone worktree cleanup, equal-timestamp cache merging, and DST/long-downtime schedule coverage.

• Files attach, preview, and download more reliably — Chat now supports generic file attachments alongside images, with better chips/cards, safer upload normalization, worktree-aware previews, and in-app local image downloads.

  File attachments now flow through contracts, upload storage, composer paste/drop, provider prompts, Kanban dispatch, recap/bootstrap surfaces, optimistic timeline rendering, caps/rollback, attachment-bearing plan follow-ups, explicit unsupported-file rejection, worktree-backed file preview roots, and blob-based download handling that keeps failed local image downloads inside Synara.

• Providers and models stay scoped to the right project — OpenCode and Claude startup paths are more careful about cwd, model discovery, config scope, and sticky plan mode so new threads inherit less accidental state.

  OpenCode model discovery can fall back to opencode models --verbose, managed OpenCode/Kilo paths run in the request/session cwd, warm server reuse is scoped, file config is no longer replaced with synthetic empty config, OpenCode resume preserves cwd, and fresh Claude threads avoid inheriting plan mode from the previous active thread.

• Chats and side panels stay in sync — Deleted chats disappear immediately, the Environment panel behaves better in constrained layouts, automation cards show up in the transcript, and file previews avoid extra full-thread subscriptions.

  Client projections now use delete tombstones and responsive archived bulk-delete updates, environment-panel open/close preferences survive chat switches, constrained/floating layouts stay calmer by default, thread automation summaries appear in the environment panel, created automation cards render in chat, and file preview routing avoids unnecessary full thread subscriptions.

• Profile skill counts reflect more real work — Profile stats now count repeated /skill and $skill usage more accurately, including retained history that should still contribute to your local activity picture.

  Skill aggregation now includes retention-hidden threads while still excluding manually deleted data, counts repeated slash/dollar skill tokens inside one prompt, avoids double-counting structured references, and has regression coverage for retained threads and repeated skill invocation.