v0.5.2

selftune 0.5.2

Curated docs release notes were not found for this version, so this release falls back to the package changelog's Unreleased section.

Changelog

Added

• Fail-closed use-once helper — a separately built and signed ephemeral helper now uses the pinned https://cloud.selftune.dev authority for token-bound preview, exact sealed-object delivery, bounded package validation, interactive rights disclosure, one-winner consume, fixed agent invocation, and owned temporary-workspace cleanup; redirects, cookies, caller origins, persistent installs, and trusted telemetry remain unavailable.
• Optional Desktop-to-Cloud linking — connect or create a SelfTune Cloud account from Sync & Backup or the state-aware sidebar Cloud profile through short-lived browser approval, return the native window to the foreground for account-backed backup, and open the hosted dashboard only through a separately labeled external action—without token copy/paste or an app-wide account gate.
• Native desktop lifecycle — installed macOS builds can run the authenticated local sidecar as a crash-restarting, login-started LaunchAgent independent of the Electron window, and signed desktop releases now download GitHub Release updates in the background with explicit restart controls.
• Project Skill Sets — cache immutable package revisions in a local Library, preview and apply conflict-free Codex or Claude Code project links, and roll back only receipt-owned paths through the CLI or native Projects screen.
• Real-time improvement signal detection — prompt-log hook detects user corrections ("why didn't you use X?") and explicit skill requests via pure regex patterns. Signals are logged to ~/.claude/improvement_signals.jsonl with skill name extraction from installed skills.
• Signal-reactive orchestration — session-stop hook checks for pending improvement signals and spawns a focused selftune orchestrate --max-skills 2 run in the background. Respects a 30-minute lockfile to prevent concurrent runs.
• Signal-aware candidate selection — Orchestrator reads pending signals and boosts priority for mentioned skills (+150 per signal, capped at +450). Signaled skills bypass the minimum evidence gate and the "UNGRADED with 0 missed queries" gate.
• Orchestrate lockfile — acquireLock()/releaseLock() with PID+timestamp in ~/.claude/.orchestrate.lock. 30-minute stale threshold prevents deadlocks from crashed runs.
• Signal consumption — After an orchestrate run completes, consumed signals are marked with consumed: true, consumed_at, and consumed_by_run so they don't affect subsequent runs.

Changed

• Cross-project skill visibility and reliable source sync — the native Library discovers project-scoped installs across known workspaces, current Codex custom tool calls and user-attached skill paths count as usage, append-aware file fingerprints revisit resumed rollout logs without reprocessing unchanged history, and replayed sessions atomically replace their own prior canonical snapshot instead of inflating usage counts.
• Verified macOS automation — recurring desktop jobs use launchctl bootout, enable, and bootstrap, then require loaded-state proof instead of trusting the deprecated load command's misleading exit status.
• Context-backed workflow sets — Projects separates exploratory portfolio workflows from observed co-usage suggestions and can recommend broader Cloudflare Full-Stack and Mobile Engineering sets with reusable Matt Pocock skills and exact installed or catalog-backed identities.
• Explainable overlapping Skill Sets — larger recommendations are validated by recurrence of their discovery relationships in unseen newer sessions instead of requiring every member to recur at once, while the dashboard shows each reusable skill's source, contextual role, membership strength, and held-out relationship coverage.
• Evidence-backed Skill Set suggestions — suggestion mining now reads local source identity, stops presenting same-source pairs as finished sets, combines statistically supported co-usage edges into overlapping multi-skill communities, lets reusable core skills appear in multiple sets, and requires larger sets to meet the held-out quality floor.
• Scoped desktop supervision — Electron now composes an Effect-owned runtime with queued lifecycle transitions, connection-generation-safe health recovery, staged authenticated window rebinding, schema-decoded tray responses, removable IPC handlers, deterministic shutdown that preserves CLI and OS-service ownership, and a packaged JavaScript boundary for internal workspace modules.
• Scoped local dashboard runtime — dashboard workflows now run through a typed Effect service and managed Layer; auth, SSE resources, SPA transport, request validation, and SQLite read routes have independent owners while the Bun server remains the process composition root.
• Lean npm distribution — release packaging now flattens shared external runtimes into the root manifest and normalizes bundled workspace links, preventing npm from recursively embedding duplicate Effect and Zod development trees in the SelfTune tarball.
• Faster native Library loading — the desktop shell uses a compact summary instead of downloading the full analytics history, inventory no longer waits on GitHub update checks, repeated harness links share one package hash per scan, and portfolio combination analysis skips query-log work when it only needs positive examples.
• Compact harness inventory — the native Library shows harness logos without repeated labels in each row, preserves names in accessible tooltips, and scales the Pi mark to match the other harness icons.
• Publish version ceiling now respects live npm — export no longer reuses a stale repository package.json version when npm already has a newer release, so shipped OSS changes publish the next real patch version instead of colliding with the latest package on npm.
• GitHub releases use curated changelog entries — OSS publish now renders release bodies from the Mintlify changelog when a matching entry exists, with CHANGELOG.md used as the fallback body for unmatched versions.
• Dedicated creator-loop workflow — the shipped selftune skill now exposes "create, test, and deploy" as a first-class routed workflow while keeping Evals, UnitTest, Baseline, Evolve, and Watch as atomic workflow docs instead of demoting them to references.

Full docs changelog: https://docs.selftune.dev/changelog
Full Changelog: v0.5.1...v0.5.2