Release Ledger

Changelog

Every Arcana release, in order. Features, fixes, and improvements — linear, dated, signed by the line.

Latestv0.3.67 Entries07 SinceJun 2026

v0.3.67 Latest

Jul 26, 2026
  • featClean run exit with finish detection. arcana run emits a done signal on completion — foreground use with full log capture via | tee. No background hacks.
  • feat--timeout flag. Bound session duration on arcana run. Defaults to 120s — stops runaway sessions for CI, benchmarks, one-shot automation.
  • featBare model-name lookup. Names without a provider prefix (deepseek-v4-flash vs provider/model) resolve across all configured providers.

v0.3.66

Jul 26, 2026
  • featDaemon respects OS signals. SIGTERM/SIGINT handled gracefully — clean shutdown, no orphaned processes.
  • featCross-directory daemon management. arcana daemon status / stop work from any working directory.
  • fixSilent TUI auto-spawn. Daemon auto-spawn no longer leaks stdout/stderr into the parent terminal.

v0.3.65

Jul 26, 2026
  • featTUI auto-connects to daemon. Detects a running daemon, auto-spawns one if missing, falls back to Worker mode. No manual management.
  • featDaemon lifecycle. Lock file, isolated error handling, detached spawn, /health endpoint, arcana daemon status|stop.
  • feat/loop standalone. Research loop works without /goal, in TUI and CLI. Native /loop hub in run.ts for headless use.
  • fixNo more double body rendering in spine entries. Overlapping Show conditions resolved.
  • fixClean slash-command input. Plain-text lines stripped; multi-slash (/foo /bar) rejected.

v0.3.64

Jul 22, 2026
  • fixPackage distribution restored. Fixed npm wrapper package for arcana-ai.

v0.3.48–v0.3.63

Jul 22–26, 2026
  • featBrowser automation via MCP. agent-browser for CDP-based control — 29 tools for navigation, screenshots, forms, JS exec. Docs at /docs/browser-automation.
  • featOllama local models. Works with locally running Ollama instances. Models auto-discovered in the picker.
  • featSubagent spine entries — clickable. Toggle open to reveal child session messages inline; live progress while running.
  • featSmart JSON formatting. Generic tool output auto-rendered as tables or key-value pairs. Todo JSON gets interactive checkboxes + braille spinner.
  • featViewport culling + scroll-to-bottom. Large spines use virtual scrolling; a scroll-to-bottom button appears when scrolled up.
  • featSession metrics bar. Real-time token usage and cost below the input prompt.
  • fix/new creates sessions properly instead of just clearing the screen.
  • fixNo more "Tool did not finish" false errors. Pending tools no longer show as failed when the agent moves on.
  • fixWindows db crash fixed. Graceful message instead of sqlite3 crash in interactive mode.

v0.3.1–v0.3.47

Jun–Jul 2026
  • featOAuth device-code login. /connect lets free users sign in via device-code flow — visit a URL, enter a code, account bound.
  • featArcana error taxonomy. Structured, readable codes replace generic failures: ARC_AUTH_INVALID, ARC_CREDITS_EXHAUSTED, ARC_RATE_LIMITED — each with next steps.
  • featAI image generation. Generate images directly from chat, backed by configured providers.
  • featFree-tier foundations. Session management, quota tracking, turn budgets, free-tier model discovery.
  • featResearch workflows engine. Autonomous research loops with CLI commands, store integration, agent prompt wiring.
  • featTUI session performance. Sessions no longer remount on switch; spine entries cached. Dramatic render-time reduction for large histories.
  • featInstant Enter. Enter feels instant — all slash commands fire on single press.
  • fixProxy TLS stability. Primary domain proxy-arcana.otnelhq.com reliable; workers.dev fallback.
  • fixOAuth error handling. Device-code URL opened server-side; console login tolerates nullish payloads.

benchmark

Jul 26, 2026
  • featMulti-Agent CLI Benchmark published — Arcana v0.3.66 vs OpenCode 1.18.5 vs Hermes v0.15.1, same model (deepseek-v4-flash).
  • resCoding suite: async semaphore w/ 10 tests — Arcana 10/10, OpenCode 10/10, Hermes 9/10.
  • resDomain suite: math/science/biology/history — all 4/4. Arcana fastest on every task: 28.4s vs 48.1s vs 86.9s.
  • res9 metrics, full reproducibility — SHA256-locked test file, reproduce.sh, archived implementations, raw logs.