Changelog

Every layer is additive over the same drive() loop, with zero kernel change. Full commit history on GitHub · RSS feed.

v0.3.8

  • Per-round token economy: agent template bodies become provider-cacheable system prompts; a cheap repo scout runs before planning in the simple flow; the implementer escalates its model tier after two failed rounds; pathological diffs are capped with a complete per-file stat list
  • Dashboard UX overhaul: deep-linkable routes, responsive shell, a needs-you fleet view, a state-driven project page with an open-blockers gate panel and deterministic-checks chips

v0.3.7

  • Rework-loop stability: open blockers are delivered into the next spawn's context, an unparseable review blocks instead of silently passing, an empty implementer diff fails fast, and only blocking reviewers re-run on a rework round
  • Deterministic pre-review checks: the project's typecheck / lint / test run before the review panel and walk the task back on failure — with the full compiler output delivered to the implementer
  • Control-plane hardening (CSRF / rebinding / allowlist) and per-drive cost accounting

v0.3.6

  • Dashboard: tabbed archive browser, live log under the agent chain, Docker-on default
  • Hard total-spawn cap per drive — bounds runaway spend

v0.3.5

  • Reviewer blocker-park flow and guided occupied-slot resolution (loom run --replace, /proceed)
  • Per-task complexity pinning (--complexity) and a planner model tier
  • Self-diff hardening for non-Claude file-editing harnesses

v0.3.4

  • Telegram remote-control bot — submit, approve gates, ship from a chat
  • Per-agent model fallback chains; real cost from every backend
  • Per-spawn transcripts readable at the gate; push / squash-merge on demand
  • Rebuilt dashboard: tabbed settings, in-app folder picker, provider-key management, horizontal agent chain
  • Per-project state moved to <project>/.loom/ (auto-migrated)

v0.3.3

  • Pipeline hardening: clean finalize + slot release, permanent provider errors park instead of retry-looping, clean per-task sandboxes

v0.3.2

  • Observability: per-task agent-chain view (model, tokens, duration, findings / verdicts) and an archived-task browser

v0.3.1

  • Operator UX: fast-task path, per-task Docker, live model dropdowns, pause / resume / cancel, human-readable log

v0.3.0

  • Configure-once control layer (loom config / secrets / models)
  • Per-spawn multi-backend resolution: Claude Code login, OpenRouter, Ollama, Anthropic
  • Non-Claude file-editing harnesses (Aider / opencode)
  • The web dashboard — loom up

v0.2.x

  • HTTP control plane, container isolation, unattended hardening (0.2.1)
  • Headless loom run + the autonomous loom daemon (0.2.0)

v0.1.x

  • The interactive kernel, the code bundle, MCP + CLI transports