- pull requests merged in two days
- 15
- backend and frontend tests in the first one
- 169 + 66
- backend tests passing by Saturday night
- 466
- test suite, once run in parallel (approximate)
- 389 s → 55 s
What it is
After nearly four weeks with no commits in any of our repositories, a new one started on Friday, August 14, at 16:53: Orca-strator. Its whole first version landed in one commit at 23:48 and merged eleven minutes later, with 169 backend and 66 frontend tests.
Orca-strator is a local-first visual workflow builder for AI work. A planner node splits a request into tasks, the tasks fan out to concurrent model calls, and each call streams its tokens into its own lane inside one node on the canvas. A join node collects the results. Ten days later it would become the engine behind our hosted agents, but on August 14 it was a local tool for one machine.
The first pull request set the design rules that everything after it followed:
- Any provider, any model. Calls go through LiteLLM, and the model field is free text on purpose, because “a model released this morning must work”. Which controls a node shows (temperature, reasoning effort) comes from a lookup on the model string, so a model that rejects temperature simply has no temperature slider.
- Compare mode. A worker can run every task against several models at once, so one prompt can be diffed across them. A “passthrough” planner makes no model call and needs no API key.
- Budget on the expected cost, not the worst case. A run is refused before it starts if its expected cost exceeds the budget, and once the budget is spent no new lane starts. A worst-case check, using each model’s maximum output, would refuse almost every honest flow, and a check that refuses everything gets switched off. Lanes already streaming are never killed, because their tokens are billed either way.
Its description also said, plainly: “no API keys exist in this environment, so no real provider call has ever been made.”
Saturday: fourteen more pull requests
- Legible on first sight. The first review of the UI was blunt, and it found real bugs: React Flow applies its own default styles to node types whose names match its built-ins, so our
inputnode rendered inside a small white box, and the Workers node said “waiting” forever after a successful run. - An approval gate. A flow can pause after planning and before any worker is billed, and the user can untick individual tasks, so approving a subset genuinely narrows the fan-out. Run history and tool calling with MCP servers followed.
- CI that didn’t exist. Three places in the repository said there was CI. There wasn’t. Standing it up exposed holes in the checks themselves: the generator that keeps TypeScript types in sync with the Python schema dropped every numeric bound, so changing a limit from 900 to 600 produced no TypeScript diff at all.
- A release that shipped broken twice. Every merged pull request now cuts a version. The release job updated four version files and committed three, leaving
uv.lockbehind, so every later pull request failed in about nine seconds. It had happened in the previous release too. The fix: one list of version files shared by the script and the workflow, and a check that refuses to push if any is left modified.
Several of that day’s pull requests were about cost and token accounting that read zero when the truth was unknown. That is its own note.
The handoff document
On Saturday afternoon we wrote docs/HANDOFF.md, and its opening section is the most useful thing the project produced that week:
Every load-bearing claim in this project was wrong at least once until it was measured.
Late on Saturday two large pull requests implemented the roadmap from that document (prompt caching, saved settings, agent presets, condition nodes, durable checkpoints, fallback routing, loops and sub-flows, and an evaluation harness) and took the version to 0.3. Running the tests in parallel took the suite from about 389 seconds to about 55.
What we’d tell someone else
Write down what you measured, not what you believe, and keep a list of the claims that turned out wrong. It reads like a confession, and it’s the most useful page for whoever comes next.