scroll
tovarna.dev
tovarna.dev · AI issue factory

Issue in.
Merge out.

An autonomous factory for processing GitHub issues. Every issue rides the production line in an isolated git worktree — from triage through development and tests to an independent review and a fast‑forward merge into main. You hold the oversight, not the keyboard.

A coding agent writes code. Tovarna runs the whole process that decides whether that code may reach main. Claude Code can happily run inside — Tovarna is the layer above it, not its replacement.

Production line

One line. No shortcuts.

Every issue travels the belt through fixed stations — from the brief to delivery. Nothing that fails the tests or the independent review gets out. The gates are fail‑closed.

#616
01

Brief

an issue with a clear goal

02

Analysis

what changes and why

03

Architecture

how it fits the whole

04

Development

implementation in isolation

05

Test

verified against a clean state

06

Delivery

independent review · merge

automatic station may escalate to a human
Principles

Built on distrust. On purpose.

Autonomy without brakes is a gamble. Every mechanism in the factory assumes the agent can be wrong — and puts a gate in its way.

Isolated worktree

Every issue runs on its own branch in its own worktree. There is exactly one road into main: an ff-only merge of a change whose full test suite passed inside that isolation.

Independent review

The diff is reviewed by a model that never saw the development — no anchoring bias. It judges the result purely against the brief.

Human in the loop

A review conflict or a stuck station isn't swept under the rug — it escalates. You drop your decision straight into the running loop and the line moves on. No restart.

What doesn't pass the gate doesn't exist.

atomic changes · merge gate · traceability to sources

Every change has a traceable reason and a gate it had to pass.

Every failure becomes a gate.

An incident goes through RCA and turns into a permanent rule — the factory that erred today writes its own check tomorrow.

Oversight, not toil

You step in with one sentence. The line finishes the rest.

When the review finds a conflict between the brief and a test, the factory stops and asks. You write the decision — and the pipeline resumes exactly where it stood. Average intervention: one sentence, twenty seconds.

12
merges / day
steady autonomous throughput
83 %
first-pass clean
the rest via repair or escalation
0.52 USD
AI model cost / issue
Haiku triage · Sonnet dev · Opus review
3.7 min
median run / issue
p90 10.5 min · before 7.4 min
How we measure the numbers →
  • 12 merges / day — average over active days (median 7.5, maximum 50); autonomous passes with no human touching the code.
  • 83 % first-pass — the issue went from triage to merge without a repair iteration and with the critic's first verdict APPROVE; independent review ran on 100 % of merged issues (79 % in 03–07).
  • 0.52 USD — average AI-model cost per issue, including all iterations and the independent review (median 0.39; range 0.07–2.78).
  • 3.7 min — median pipeline run time per issue (p90 10.5 min); in 03–07 it was 7.4 min.

Measured on the development of Tovarna itself — the factory builds itself (dogfooding). n = 397 merged issues, 08–09/2026 (32 active days). Source: git log for pace, per-issue run metrics for first-pass, cost and duration; cost is missing for 1 of 397 issues. The previous release reported 16 / 64 % / 0.82 USD for 03–07/2026 — pace has dropped since then (independent review now runs on every issue, not 79 %), while first-pass, cost and run time improved. Numbers from external projects will follow after the pilot deployment.

Updates

Build your own factory.

The factory is just spinning up. Leave us your e-mail and we'll let you know when there's something to try.

✓ We'll write when we open beta access — no spam.
Questions: hi@tovarna.dev