Build1 publisher2 min readPublished
One prompt turned 46 SWC issues into 45 parallel pull requests on one subsystem
SWC's creator fanned a single instruction out to 45 concurrent Labor0 tasks on the ES minifier, then kept review and merge for himself. He says he does not know how long the run took; he was looking at his phone.
The Engineer · Build desk
What happened
- kdy1, who created the SWC compiler, gave Codex Desktop one prompt telling GPT 6 Astra to check every branch in every ES minifier file for incorrect conditions and file an issue for each; it opened 46.
- A second prompt asked his Labor0 system to fix every es/minifier issue he had filed that day, and Labor0 turned the 46 issues into 45 tasks, one of which covered two issues whose scopes belonged together.
- All 45 tasks ran at the same time, and because a Labor0 task generally produces one pull request, the run ended with 45 open PRs against the same subsystem.
- Labor0 assigns models by task difficulty, so GPT 6 Astra found the bugs while many of the simpler fixes were implemented by GPT 5.6 Terra.
- When Labor0 hits a decision it sends kdy1 a question by something like a web push notification and resumes once he answers, and it starts the next independent task without waiting for him.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Generation cost two typed prompts, while review of 45 branches falls to the single maintainer who also owns every decision on the project.
- decision With 45 sibling branches open on one subsystem, the maintainer has to pick a landing policy: merge serially and re-test after each, or batch them and test the combined result.
- exposure SWC's downstream users get minifier condition changes whose supporting evidence is an agent-constructed example input, and many of whose fixes were written by a cheaper model than the one that found the bug.
The constraint kdy1 documents sits in the issue text. "A large number of parallel sessions is only useful when every session receives a sufficiently detailed issue," he wrote [6]. His $add-issue skill enforces that: a bug has to arrive supported by evidence, as a self-contained, implementation-ready GitHub issue [5]. The sweep prompt told the agent to construct an example input for every incorrect condition it found before filing [3].
Concurrency then moves the ordering problem. Labor0 sequences tasks that depend on each other, and it treated these 45 as independent [10]. A task generally produces one PR [9], so all 45 branches were written against a tree that did not contain the other 44 diffs [23]. Land one and up to 44 have a base that CI never tested [23]. All 46 issues came out of one sweep of the ES minifier [3][4], so those branches sit in the same code area.
The time he documents spending himself goes to decisions. He writes PRDs with a Write PRD skill on GPT 5.6 Sol, and the skill surfaces decisions for him to resolve [19]. On Watchtower, a Sentry-compatible error monitoring and observability system he is building in public [15], PRD #13 began with his preference for Connect RPC because he likes strongly typed systems, and JSON over HTTP was recommended instead [16]. He spent a few rounds of conversation checking whether Connect had a serious limitation [16]. The concern narrowed to Rust support for Protobuf and its tooling, which he decided was not fundamental for him, so the PRD settled on Protobuf over HTTP with Connect deferred until streaming is needed [17]. Labor0 turned that PRD into PR #31, which he reviewed and merged the same day [18].
"Humans are always the bottleneck," kdy1 wrote [12]. Decisions stay with him, and he reviews and merges the pull requests himself [12]. Of the SWC run he wrote: "While the work was running, I was looking at my phone and playing games." [14] He says he does not know exactly how long it took [14], and the post does not report how many of the 45 PRs were merged [21].
Two things have to hold before that 45 transfers to another repository. The findings have to be separable enough that a scheduler can call them independent, as Labor0 did here [10]. And the review path has to be one person with commit rights: kdy1 created SWC [2] and merges its PRs himself [12]. He was surprised by how few decisions the work required of him relative to the size of the projects [20].
What to watch
- Whether kdy1 publishes merge outcomes for the 45 SWC pull requests: how many landed, how many needed rebases, how many were closed.
- Whether Labor0's dependency ordering extends past authoring into merge-time re-testing of sibling branches.
- Whether the $add-issue and Write PRD skills ship in a form other maintainers can run on their own repositories.