Skip to main content

AIFactory 3.4: watch the build, message the agent, and a token-saving solo mode

· 3 min read
Olaf Krasicki-Freund
Olaf Krasicki-Freund
Creator of AIFactory

The 3.4 line is about seeing and steering a build while it runs — and trimming the overhead when a job is small. Here's what landed.

See where the build is going: observability panels​

Every task now has an Observability tab with three live panels:

  • Token usage — total tokens vs the context window, total cost, and a per-category breakdown (system prompt / user messages / tool outputs / thinking). An over-large system prompt or a runaway tool-output category is now obvious at a glance, and attribution survives a context compaction instead of resetting.
  • Resource usage — live CPU and memory for the agent's own process, polled every few seconds. This is how you tell a working build from a hung one without tailing logs.
  • Send to agent — a box to drop a message straight into the agent's inbox.

See Task observability panels.

Steer it: agent messaging & mentions​

Builds are mostly sequential (plan → code → QA → fix), but some coordination has to happen mid-run. 3.4 adds a file-backed inbox with between-turn delivery: messages are picked up at the next turn boundary (atomic writes, read-back-verified), so you — or another agent — can nudge a running build without restarting the session.

Message text is parsed for @agent and #task mentions, with the boring-but-important details handled: mentions inside inline code are ignored, and someone@example.com is never mistaken for a mention. See Agent messaging & mentions.

Trim the overhead: solo mode​

For a one-line fix or a throwaway script, the full four-session pipeline is overkill. Solo mode collapses it to a single self-directed agent that writes its own plan and works it in the same loop — far fewer tokens and sessions. It's opt-in (env var, per-task flag, or global config) and off by default, because there's no independent QA pass; keep the full pipeline for anything you'd want reviewed. See Solo mode.

Providers​

  • Antigravity CLI — the Gemini provider, renamed, with portal-driven install/update (Settings → CLI Tools) so you can pull a new version without leaving the UI. The isolated worktree is trusted automatically so the CLI can edit files.
  • OpenCode CLI — a community/self-host runtime, with an automatic workaround for the sandbox model-catalogue problem. See the multi-provider page for the tier note and caveats.
  • Plus fixes: Codex builds on ChatGPT-account logins (account-default model, no forced --model), and openai-compatible Bedrock/Azure/Vertex routing through the LiteLLM gateway.

Fixes worth calling out​

  • Worktree core.bare self-heal — a stray core.bare=true on the primary checkout used to break every git operation; the worktree manager now heals it on init and guards each call.
  • Project-context reconciliation — the kanban board, sidebar, and project dropdown no longer drift apart on load (the dropdown could previously get stuck on the wrong project).
  • Honest build state — all-failed builds now surface as failed instead of being quietly labelled "human review".

Container images ship for linux/amd64; native linux/arm64 images are being restored on a dedicated runner.

— As always: provider-agnostic, self-hostable, and the code is right there to read.