Build1 distinct publisher3 min readPublished
The leverage in this rebase sits in routing rather than prompting, and the three regexes that decide which hunks skip the model entirely are tuned to one repository's five months of noise. Copying the method means re-deriving them.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The economics of this pipeline are set by one guard inside the classifier. `classify()` calls a hunk mechanical only when every line on both sides is blank or matches one of three regexes [16]. Everything else falls through to the model or to the human queue. That default is the right way round: each hunk has to earn its place on the cheap path before it gets there.
The three patterns tell you what repository this was. One matches lines beginning with `import` or `from`, one matches a quoted key followed by a quoted version string, one matches Markdown-style bullets [17]. Between them they cover import block collisions, lockfile churn, formatter disagreements and two branches appending to the same list [15]. That is 37.5 percent of the hunks routed around the model [19]. For that share to transfer, your drift has to be made of the same noise. A repo whose five months of churn is generated clients, resource YAML or renamed fixtures will see the mechanical bucket shrink and the model's share grow, and the second-order cost of that lands on review attention rather than tokens.
Now the gate. A passing batch is committed and recorded in `resolutions.md`; a failing batch is reverted and re-run with the failure output attached [14]. Worth being precise about what that buys, because the author's own statement of the danger is a resolution that compiles and passes the tests that happen to exist while shipping a bug that looks like it was always there [11]. A green build and a green suite cannot detect that, by construction. What actually stands against picking a side to make the markers go away [7] is the human-only queue and the written record. `resolutions.md` is doing the job the chat session failed at: conflict #12 in `adapter.ts` settled the new retry signature, conflict #88 in a caller reintroduced the old one, and nothing in the session tied the two together [8].
Batch size sets the clock. 171 semantic hunks at ten per batch is 18 batches at minimum [20], so the build-and-test cycle runs at least 18 times before any revert-and-retry loop [14]. Against 11 hours of wall clock that averages about 37 minutes per batch for context assembly, model time, gate, review and commit [21]. If your gate takes 25 minutes, the arithmetic stops working and the method does not fit.
The piece I would want before copying this is the context bundle. The published excerpt breaks off inside the classifier and never says what goes into one [25], and that is exactly where intent is supplied, which the author identifies as the thing the agent was never shown [10]. Everything around it is scheduling, and scheduling is the easy half. Two weeks was the hand estimate [3]; on a branch kept alive by "we'll land it next sprint" [2], estimates had a track record.
Ranked by verification strength, evidence, and original report placement.
The rebase took 11 hours instead of the two weeks the author had budgeted.
Running `git rebase main` on the branch produced 312 conflict hunks across 137 files, including content conflicts in src/storage/adapter.ts and src/storage/index.ts and an add/add conflict in src/storage/cache/lru.ts; `git diff --name-only --diff-filter=U | wc -l` returned 137.
The branch was a storage-layer rewrite started in February and kept alive by "we'll land it next sprint"; by July it was five months and 1,900 commits behind main.
The author estimated three days of solid work at best and two weeks realistically to resolve the conflicts by hand, plus a meaningful chance of silently breaking something.
The first attempt was to open the repo in Claude Code and say "resolve the merge conflicts"; the agent resolved about 40 hunks before things went sideways.
Failure mode one: on genuinely ambiguous hunks the agent took HEAD or theirs to make the markers go away, which was syntactically clean and semantically a silent revert of somebody's bug fix.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 30, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
Agent memory rots by accumulation, and the missing primitive is a supersession key1 distinct publisher
build
Five coding agents on one repo is a staffing problem, and worktrees only fix the filesystem1 distinct publisher
build
Your inference bill is an architecture defect: declare the task before you call the model1 distinct publisher
build
Claude Code walks the whole process table to inherit one shell's environment1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Specific, checkable arithmetic, one witness
The artifacts are unusually concrete for a post of this kind — real rebase output, a classifier you could paste into a repository tomorrow, named conflict numbers — and the figures close: 117 plus 171 plus 24 is exactly 312, and 171 hunks at ten a batch does land near eighteen gated passes across eleven hours. What no part of it survives is outside scrutiny. There is no repository to inspect, no second run, no test suite to look at, and the context-bundle script cuts off before it finishes.
One branch, one engineer, one sitting
Everything observed happened inside a single storage-layer rewrite: one successful run and one abandoned attempt at the naive version, both from the same person. Nobody has picked the method up, and the part that does the routing is explicitly local — three regexes tuned to the import collisions, lockfile churn and bullet-list appends of one codebase over five months.
The baseline flatters the result
Eleven hours is probably true. Two weeks is a number the author invented before starting, and the whole seven-to-one framing hangs on it — as does the omission of however long the 24 hand-resolved conflicts took. Pulling the other way, the post is genuinely more interested in how the agent failed than in how it succeeded, and it keeps the model away from the conflicts where both sides changed a business rule. The headline says Claude Code resolved 312 merge conflicts; the body says a script decided which 117 it was not allowed to see.
Attention-paid write-up, no disclosed stake
A "5 Lessons" post is paid in readers, and the number in the title is what buys them. No vendor relationship appears anywhere and none is implied — the pressure here is subtler: the same person set the two-week baseline, ran the clock, and chose which figures to publish. The absences fit that pattern. Token spend, hours on the human-only queue, revert counts and any post-landing bug report would all complicate a clean result, and none of them appear.
Mechanics believable, outcome unverified
We'd bet on the failure modes: conflict #12 keeping a retry signature that conflict #88 quietly undid is too specific to be a rhetorical invention, and the classifier is real code doing what the text says it does. Our hesitation is about everything downstream of that — one voice, a truncated script, no visibility into the test suite the gate depends on, and no way to know whether the merged branch is still correct.