Result of running find-reset-candidates.ts across the repo and hand-
reviewing each candidate.
- mcp/oauth-provider.ts: kilocode_change markers were stale — the
branding (Kilo, kilo.ai) already lives in the translate transforms,
so the marker annotations no longer point at any Kilo-specific code.
- acp/session.ts: matches upstream's import order.
- tui/ui/dialog.tsx: matches upstream's prettier line-wrap on a long
onMouseUp prop.
No behavior change.
Add a new setup script to configure `merge.conflictStyle=zdiff3` locally
within the repository. This ensures that the common ancestor is visible
during merge conflicts, which is required for structural resolution by
`mergiraf` and simplifies manual conflict resolution.
The configuration is applied during `bun install` via the `postinstall`
hook and remains as a fallback for the upstream merge utilities.
Upstream opencode stamps its own version into shared files (the Zed extension toml + any package.json it bumps), which leaves the tree pointing at an upstream release tag that doesn't exist on Kilo's pipeline — Zed download URLs silently 404 as a result.
script/sync-versions.ts rewrites every top-level package.json version and the Zed extension toml (version + 5 download URLs) to a single target in one shot. Defaults to the root package.json version, accepts an explicit version argument, and skips packages/kilo-jetbrains/ which tracks its own cadence.
Documented in .kilo/command/upstream-manual-merge.md as a dedicated step-10 commit after 'resolve merge conflicts' so version churn stays out of the behavioural merge commit.
Fixes hang when running the finder from the repo root without a path
scope. Root cause: upstream-only packages (packages/console, packages/web,
packages/desktop, etc. in config.skipFiles) contain multi-megabyte
media assets that stall concurrent Bun `git show` subprocesses on the
pipe buffer.
Changes:
- Batch upstream blob sizes in one `git cat-file --batch-check`
subprocess before classifying anything. Files with missing upstream
land directly in `upstream-missing`; files > 256 KB land in the new
`too-large` bucket. Only sane-sized survivors get per-file `git show`.
- Filter out files matching the merge config's `keepOurs` and
`skipFiles` globs (.github workflows, translated READMEs, upstream-
only packages, etc.) — these are intentionally preserved or removed
in Kilo and would otherwise pollute the report and tempt incorrect
resets.
Full-repo dry-run now completes in ~2s.
Incorporate lessons from the v1.14.30 merge: explain reasoning before
showing diffs, investigate why Kilo-specific helpers exist before
bypassing them, prefer commenting-out over deletion with kilocode_change
markers, use TODO (not NOTE) for follow-ups, check adjacent files that
structural upstream splits force you to touch, grep for usages before
deleting apparently-unused upstream declarations, and include PR-body
guidance and the 'resolve merge conflicts' commit convention.
Users that hand us BOM-less UTF-16/32 are violating the documented
contract; trust jschardet + iconv.encodingExists and let the result
be whatever it is.
Replace the subprocess-based line diff in classifyDrift with an
in-process multiset diff (approxDiff) so concurrent classifications
don't deadlock on big text files like sprite.svg via Bun $.quiet()
pipe-buffer stalls.
Also filter non-code assets (SVG, PNG, fonts, archives, lock files,
etc.) before classification so they don't bloat the report or stress
git subprocesses. Switch kilo-only path excludes to glob pathspecs so
every packages/kilo-*/ and **/kilocode/** dir is excluded without
maintaining a hand list.
Rename the 'whitespace-only' bucket to 'cosmetic-only' — with the
multiset diff it also catches line reordering, so the old label was
misleading.
When autocompleting a file, folder, or special mention via the @ menu
in the chat textarea, insert a single trailing space (unless the next
char is already whitespace) so typing immediately after selection does
not get appended to the path and break attachment parsing.
* fix(cli): restore kilo branding in MCP auth toast and client name
The 'Run: opencode mcp auth' popup / toast and the MCP protocol client
'name' field were reverted to 'opencode' by an upstream refactor
(opencode PR #22913, pulled in via Kilo PR #9346). Re-apply the
Kilo-branded strings with narrow kilocode_change markers and add a
regression test in packages/opencode/test/kilocode/ so the next upstream
merge fails a Kilo-owned test instead of shipping an 'opencode' popup.
* fix: use kilocode_change block markers for startAuth signature
---------
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
Co-authored-by: marius-kilocode <marius@kilocode.ai>
Bulk-finds files that have drifted insignificantly from the last merged
upstream and (optionally) resets them. Extracts shared clean/changed
marker helpers into utils/markers.ts and shared reset logic into
utils/reset.ts so both the single-file reset helper and the new finder
use the same classification pipeline.
Add .github/TEAM_MEMBERS and other files Kilo has deleted to skipFiles
so future upstream merges auto-remove them via git rm instead of
surfacing delete-by-us conflicts that need manual resolution.
Also drop the stale github/.gitignore entry from keepOurs (the file
does not exist in Kilo main; keepOurs was a no-op) and move it to
skipFiles so upstream's copy is dropped on merge instead of added.
* fix(cli): serialize OAuth callback port tests
* fix(cli): close MCP OAuth browser test listener race
The mock subprocess emitted its error via setTimeout(10ms), which raced the error listener attachment on slow Windows CI and left BrowserOpenFailed unpublished. Emit on newListener('error') via a microtask so the listener is always attached first. Revert the unrelated test-runner serialization.
* fix(cli): deliver MCP OAuth mock error to listener directly
Override subprocess.on in the mock so the error is queued to the registered listener as soon as it attaches. The previous newListener + emit chain still lost events on slow Windows CI because EventEmitter dispatch can race microtask draining.
Mark IJbema