Skip to main content

Gitmoot v0.3.0-beta.1

Gitmoot v0.3.0-beta.1 adds a full interactive dashboard cockpit and an end-to-end train-run TUI for the SkillOpt optimization loop, on top of the 0.2.0 parallel-PR foundation.

Highlights​

Interactive dashboard (gitmoot dashboard)​

  • A bubbletea cockpit with pages for Attention, Trains, Agents, Sessions, Jobs, Locks, Health, and Config. It launches automatically on a terminal and falls back to the byte-stable --plain / --json snapshot everywhere else.
  • Attention: answer interactive prompts inline; retry or cancel blocked/failed jobs; start the daemon when it is down.
  • Agents: create (pick a template or write a custom prompt in $EDITOR), delete, revert a template version, switch runtime (codex/claude), optimize, and read a version's prompt in a pager. Internal training agents are hidden.
  • Jobs: a scrollable list with timestamps, request/result detail, and retry/cancel.
  • Sessions: per-session detail and a stop action. Locks: a "what to do" guidance line.
  • Health: environment/daemon doctor checks. Config: inline scalar edits (comment-preserving) plus an $EDITOR handoff with validation.

Train-run + train-init TUI (SkillOpt)​

  • gitmoot skillopt train init interactive wizard (line and TUI) with template resolution, optional GitHub repo creation, and a confirm summary.
  • A train-run view that drives generate → review → optimize → promote with a live phase bar, streaming generation/optimizer progress, candidate preview, and promote/reject — which now also posts the decision to the candidate-review issue and closes it.
  • A lazy, self-documenting review-reply template: ranking is the only required field; everything else is optional with inline allowed-values hints.

Reliability​

  • Creating a generation repo now provisions a usable checkout (initial commit + clone + register), so the optimizer can run in it.
  • The codex optimizer and target backends default to the codex CLI's configured model instead of an unsupported gpt-4o.
  • Failed generate/continue steps surface the real runtime error (e.g. a codex usage limit) instead of a cryptic stdin line.

Validation​

  • GOTOOLCHAIN=go1.26.0 go test ./... — green except the known daemon flake TestRunQueuedJobsSerializesSameRuntimeSessionAcrossRepos.
  • go vet, gofmt, and git diff --check clean.
  • High-effort code review on every change; live exercise of the dashboard and a full train run end-to-end (generate → optimize → candidate promote) on a disposable GitHub repo.

Known Beta Limits​

  • The dashboard runs against the current checkout's home; running it from anywhere to track every repo is tracked in #267.
  • Per-agent model selection is not wired yet (#268), and the dashboard does not yet surface optimizer / blocked_config errors in the TUI (only generation errors) — check the session log if an optimizer step stalls.
  • Release assets include gitmoot_linux_amd64 only.