Loading Forkward…
Loading Forkward…
Every upstream fix, forwarded to your fork.
Forkward merges upstream into your fork and proposes a resolution for every conflict, from deterministic rules and a team of AI agents. It applies one on its own only under a certified error bound, and the rest wait for you in a pull request.
Recorded sync · hunk 1 of 3
Reading the conflict.
…/translate/ops.rs#1
// ATen accumulates in the declared output dtype: bool and narrow // integers promote to i64 by default, and dtype= overrides that. // Casting only the scan result would lose carries or overflow first.
// ATen casts BEFORE scanning: absent dtype promotes bool/integer // inputs to int64; an explicit dtype overrides that default. PT2's // output metadata has already resolved both cases, even when the
Why forks fall behind
Every sync means conflicts, so long-lived forks put syncs off. The longer they wait, the more security and stability fixes they miss, and the bigger the next merge gets.
29 to 158
Stable releases behind
How far the vendor device kernels in our study trailed their upstream.
65
Weekly syncs, zero conflicts
Raspberry Pi merged 65 stable releases in a row into rpi-6.12.y without one conflict. Syncing often keeps each merge small.
300,000
Verified hunks for five nines
Certifying 1 wrong auto-merge in 100,000 at 95% confidence takes about that many. Every merged pull request adds to the record.
Install the GitHub app on the forks you want synced.
Seven rules and several AI agents resolve each conflict.
A hunk is applied on its own only under a certified error bound.
Everything else waits for you in the pull request, and what you merge informs the next sync.
Strategy search
A child replaces the incumbent only when it wins on validation by more than the noise.
The 95% interval, −4.4 pts to +17.8 pts, still includes zero.
Train is 70 NXP hunks and validation is 45 ST hunks. Validation picks the winner, so its line runs optimistic.
Results
Scored line for line, ignoring whitespace, against what NXP, ST and Rockchip engineers committed, with DeepSeek V4.1 Flash as the resolver model.
76.1%
Matched the engineers' resolution
54 of 71 hunks on Rockchip, a vendor never used in tuning, tested once.
71.8%
Always keeping the fork's side
51 of 71. Forkward gets 3 more right, a gap not yet shown to be significant.
2.5x
Faster with agents in parallel
1.9 s a hunk against 4.7 s, measured on 8 hunks.
8 of 18
Gate-accepted hunks that were wrong
At the frozen threshold on Rockchip. Auto-apply stays off until the gate certifies its error bound.