Add upstream-sync agent skill (cherry-pick from microsoft/terminal)#218
Open
yeelam-gordon wants to merge 81 commits into
Open
Add upstream-sync agent skill (cherry-pick from microsoft/terminal)#218yeelam-gordon wants to merge 81 commits into
yeelam-gordon wants to merge 81 commits into
Conversation
Cherry-picks new commits from microsoft/terminal into this fork one at a
time, preserving per-commit author/committer/dates (verified by smoke
test: 5 picks from upstream/main~5..upstream/main matched character-for-
character including timezone offsets).
Pipeline (scheduler-friendly, resumable):
fetch upstream -> compute pending (drop revert pairs + empties)
-> branch upstream-sync/YYYY-MM-DD -> cherry-pick loop
-> always write report -> open PR (or push direct to main)
Safety:
- state.json checkpoint (last_synced_upstream_sha + stuck-lock)
- Tier-0 known-take-upstream auto-resolve (seed: spelling2.yml)
- Stuck -> abort, open GitHub issue, set lock, exit 10 (not an error)
- 'DO NOT squash' banner on PR body; -AutoMergeStrategy rebase arms
gh auto-merge with the correct strategy
Bootstrap: baseline = a325a2f ('Fix settings error text legibility in
High Contrast mode' #20098), the merge-base between fork main and
upstream/main on 2026-06-04. 16 upstream commits currently pending; the
first real sync run will be a follow-up PR after this one merges.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Pull request overview
Adds an upstream-sync Agent Skill to periodically cherry-pick commits from microsoft/terminal into this repo with stateful tracking (state.json) and automated reporting.
Changes:
- Introduces
.github/skills/upstream-sync/with PowerShell scripts to fetch upstream, compute pending commits, cherry-pick with basic auto-resolution, and open PRs/issues. - Adds reference documentation describing bootstrap, workflow, conflict triage tiers, reporting, known conflict rules, and the
state.jsonschema. - Bootstraps
.github/upstream-sync/state.jsonplus a reports folder placeholder.
Reviewed changes
Copilot reviewed 19 out of 20 changed files in this pull request and generated 11 comments.
Show a summary per file
| File | Description |
|---|---|
| .github/upstream-sync/state.json | Initial sync state (baseline SHA, lock fields, history). |
| .github/upstream-sync/reports/.gitkeep | Ensures reports directory is present in git. |
| .github/skills/upstream-sync/SKILL.md | Skill definition + operator-facing workflow, gotchas, and usage. |
| .github/skills/upstream-sync/scripts/Common.ps1 | Shared helpers for repo root/state/report paths and JSON read/write. |
| .github/skills/upstream-sync/scripts/00-bootstrap.ps1 | One-time state initialization (baseline SHA) on a dedicated branch. |
| .github/skills/upstream-sync/scripts/01-fetch-upstream.ps1 | Creates/fetches the upstream remote and resolves upstream HEAD SHA. |
| .github/skills/upstream-sync/scripts/02-compute-pending.ps1 | Computes pending commit list and drops revert pairs / empty commits. |
| .github/skills/upstream-sync/scripts/03-cherry-pick-one.ps1 | Performs one cherry-pick with Tier-0/Tier-1 conflict handling and JSON status output. |
| .github/skills/upstream-sync/scripts/04-run-batch.ps1 | Orchestrates a full run: gate on lock, pick loop, report, finalize. |
| .github/skills/upstream-sync/scripts/05-write-report.ps1 | Produces Markdown run reports used as PR/issue bodies. |
| .github/skills/upstream-sync/scripts/06-finalize-pr.ps1 | Commits state+report, pushes branch, opens PR, optionally arms auto-merge. |
| .github/skills/upstream-sync/scripts/06b-finalize-direct.ps1 | Optional direct fast-forward of main to the sync branch. |
| .github/skills/upstream-sync/scripts/07-open-stuck-issue.ps1 | Pushes stuck branch, opens an issue, sets stuck-lock on main, copies report. |
| .github/skills/upstream-sync/scripts/clear-stuck.ps1 | Clears stuck-lock after manual resolution has been merged. |
| .github/skills/upstream-sync/references/workflow.md | Authoritative step-by-step workflow and exit code semantics. |
| .github/skills/upstream-sync/references/state-schema.md | state.json schema and operational/concurrency notes. |
| .github/skills/upstream-sync/references/reporting.md | Report + stuck-issue templates and retention rationale. |
| .github/skills/upstream-sync/references/known-conflicts.md | Tier-0 “known resolution” list format and guidance. |
| .github/skills/upstream-sync/references/conflict-triage.md | Tiered conflict-resolution rubric, including line-ending cautions. |
| .github/skills/upstream-sync/references/bootstrap.md | Baseline SHA discovery approaches and bootstrap procedure. |
This comment has been minimized.
This comment has been minimized.
…kill Closes the PR #220 audit gap: clean cherry-picks were producing broken PRs because git-level conflict detection missed content-level issues (duplicate .resw keys, dropped fork-specific warning suppressions). New gates run AFTER the cherry-pick loop, BEFORE push/PR: 1. Toolchain preflight -> detect missing PlatformToolset (infra) 2. Static breakage scan -> resw dup baseline-diff + fork invariants 3. Try-build -> razzle + bz no_clean (45min timeout) Failures become Tier-4 stuck (post-pick validation), distinct from Tier-3 cherry-pick conflicts. State schema extended with a new stuck_validation field alongside the existing stuck_on_sha; either being set causes the scheduler to skip. Smoke-tested 08-static-scan against the known-bad PR #220 worktree: returns 26 critical resw-duplicate findings + 1 high C4459 invariant finding, blocking=true (matches the PR #220 audit exactly). Toolchain preflight verified to detect missing v143 on a v145-only host. 10-try-build is wired but not end-to-end tested (would take 30+ min). Also fixed: - gh now uses -R microsoft/intelligent-terminal explicitly on issue/pr calls (the upstream remote was making gh default to microsoft/terminal). - workflow.md: --tags=false -> --no-tags (invalid flag). v1 scope: resw duplicate keys (baseline-aware) + fork invariants from references/fork-invariants.json (seed: C4459 suppression). v2 deferred: missing-include / vcxproj drift (try-build catches these with zero false positives, so not worth a separate static check). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This comment has been minimized.
This comment has been minimized.
yeelam-gordon
added a commit
that referenced
this pull request
Jun 4, 2026
Two content-level issues slipped past the clean git-level cherry-pick (now caught by the new static-scan in PR #218): 1. Duplicate <data name> entries in 20 .resw files (10 locales x 2 projects, 210 blocks total). Root cause: a fork-local commit (ddf97ab) appended ConfirmCloseDialog_* / NotificationMessage_TabActivity* keys; upstream commit b753e3d then renamed the old QuitDialog_* keys to the same names in-place. Resolution: drop the fork-appended duplicates, keep upstream's in-place positions (which preserves alphabetical ordering). 2. C4459 warning suppression dropped from src/common.build.pre.props. Root cause: upstream removed unrelated C4467 in the same DisableSpecificWarnings list during the ATL/MFC cleanup, and the take-upstream resolution wholesale lost the fork-specific C4459 too. C4459 (declaration of 'X' hides class member) is required because the fork builds with TreatWarningsAsError=true and fork-specific code triggers it. Restored. Verified with .github/skills/upstream-sync/scripts/08-static-scan.ps1: blocking=false, 0 critical, 0 high. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
… audit miss) Get-FileTextOnDisk used [System.IO.File]::ReadAllText with a relative path. .NET resolves relative paths against [System.Environment]::CurrentDirectory, NOT PowerShell's $PWD, so the scan was silently reading from `Q:\official\intelligent-terminal` (the unrelated main worktree) while the orchestrator was cd'd into `Q:\official\it-sync` (the sync worktree with the duplicates). PR #220 symptom: the scan returned �locking=false / 0 critical even though the cherry-pick batch had introduced 63 new duplicate `<data name>` entries across 6 pseudo-locale files, and the build subsequently failed with PRI175 / PRI277. Fix: - Get-FileTextOnDisk: resolve to absolute via Join-Path (Get-RepoRoot) when the input is relative. - Get-FileTextAtRef: capture `git show` output through a temp file via `cmd /c "git show ref:path > tmp 2>nul"` instead of PowerShell's pipeline, which truncates/mangles high-Unicode pseudo-locale bytes when the subprocess's stdout binds to PSObject string formatting. Verified against PR #220 commits: - HEAD = fd8332b (the buggy state): 6 critical findings, blocking=true ✅ - HEAD = d6b4dc3 (the fixed state): 0 critical, blocking=false ✅ Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This comment has been minimized.
This comment has been minimized.
yeelam-gordon
added a commit
that referenced
this pull request
Jun 4, 2026
Two content-level issues slipped past the clean git-level cherry-pick (now caught by the new static-scan in PR #218): 1. Duplicate <data name> entries in 20 .resw files (10 locales x 2 projects, 210 blocks total). Root cause: a fork-local commit (ddf97ab) appended ConfirmCloseDialog_* / NotificationMessage_TabActivity* keys; upstream commit b753e3d then renamed the old QuitDialog_* keys to the same names in-place. Resolution: drop the fork-appended duplicates, keep upstream's in-place positions (which preserves alphabetical ordering). 2. C4459 warning suppression dropped from src/common.build.pre.props. Root cause: upstream removed unrelated C4467 in the same DisableSpecificWarnings list during the ATL/MFC cleanup, and the take-upstream resolution wholesale lost the fork-specific C4459 too. C4459 (declaration of 'X' hides class member) is required because the fork builds with TreatWarningsAsError=true and fork-specific code triggers it. Restored. Verified with .github/skills/upstream-sync/scripts/08-static-scan.ps1: blocking=false, 0 critical, 0 high. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Update skill docs to match same-repo PR creation, direct-push mode, and pinned upstream author/committer dates. Harden native git error checks and empty cherry-pick skip validation, keep state writes newline-terminated, seed the Tier-4 state field, and route stable spelling identifiers through patterns. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This comment has been minimized.
This comment has been minimized.
Extend the upstream-sync spelling pattern to cover underscore-prefixed toolsets identifiers emitted by the report script. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Validate native git branch switches before stuck-lock writes, seed Tier-4 state during bootstrap, and wait for asynchronous build output callbacks before closing the build log writer. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
02-compute-pending.ps1:
- `git log --format='%H'` lines could carry trailing CR or whitespace
on some PS hosts/terminals; the SHA regex would then drop them and
silently truncate the pending list. Each line is now `.Trim()`'d
before the `^[0-9a-f]{40}\$` check.
03-cherry-pick-one.ps1:
- `Get-ConflictPaths` split `git diff --name-only` on LF but didn't
strip a trailing CR. A CRLF-normalized stream would yield paths ending
in `\r`, missing the exact-string lookup in 03-known-conflicts.md
and unnecessarily escalating Tier-0-eligible commits to Tier-3.
Added `.TrimEnd(\"\
\")` per line.
04-try-build.ps1:
- `-LogDir` overrides now get rooted to the repo when supplied as a
relative path. Previously the path stayed relative through Join-Path,
then `ConvertTo-RepoRelativePath` threw because the input wasn't
rooted under the repo (since round 16 removed the silent absolute-path
fallback). Absolute paths outside the repo still abort, as intended.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
SKILL.md prerequisites: - Added explicit prereq that `origin/main` must have full git history. Watermark discovery scans up to 5000 commits on origin/main for cherry-pick trailers and the pending walk uses `merge-base --is-ancestor`. A shallow clone (e.g. GitHub Actions default `fetch-depth=1`) will produce empty/wrong results. CI should use `actions/checkout@v4` with `fetch-depth: 0` or `git fetch --unshallow`. SKILL.md step 8 (gh pr create retry): - The retry loop used `Select-Object -Last 1` to capture gh's output, which drops every line before the last — usually losing the actual error context (auth failure, missing head ref, etc.). Now captures full output via `Out-String`, surfaces it in the `Write-Warning`, and extracts the URL line by filtering for `^https://github.com/` so the success path still works. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
04-try-build.ps1: - `\` used nested `cmd /c \"cd /d \"\\" && …\"` to ensure the build ran in the repo root, but `ProcessStartInfo.WorkingDirectory` is already set to `\` two lines below — the `cd /d` prefix was redundant *and* the double-nested quoting breaks when the repo path contains spaces or quotes (cmd.exe quote parsing is famously fragile). Dropped to a plain `/c \`; the working directory is enforced by the ProcessStartInfo, not by an in-line `cd`. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
02-compute-pending.ps1: - Trailer regex and SHA validator now accept uppercase A–F as well as lowercase a–f. Hand-edited or seed-commit trailers using uppercase hex no longer slip past watermark discovery. - `git merge-base --is-ancestor` previously treated *any* non-zero exit as `not an ancestor`. That conflated the documented exit 1 (semantic `not an ancestor`, keep the SHA) with exit 128 (bad object / missing ref / shallow-clone error). Now switches on the exit code: 0 = ancestor (drop), 1 = keep, anything else throws so a broken repo state can't silently produce a wrong pending list. 03-cherry-pick-one.ps1: - `Get-ConflictPaths` now captures stderr via `2>&1` and throws on non-zero `git diff --diff-filter=U` exit, so a real repo error can't be misreported as `no conflict paths`. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…UpstreamShas 03-cherry-pick-one.ps1: - `git cherry-pick` output was streamed straight to stderr, so when the exit was non-zero AND there were no unmerged paths (e.g. `cherry- pick` rejected a merge commit without `-m`, or a git version rejected `--keep-redundant-commits`), the stuck result said only `exited N with no conflict paths` and the actionable git message was nowhere in the JSON / issue body. - Now captures the stream into `\` first, forwards it to stderr as before for live visibility, AND grabs the last 20 lines as `\`. The Write-Warning and the `result.error` field both include that tail so the operator can diagnose the failure without rerunning to reproduce. 02-compute-pending.ps1 (real bug — Copilot was right): - `Get-PendingUpstreamShas` ended with `return ,\`. The comma wraps the string[] as a single pipeline object, so the caller's `\ = @(Get-PendingUpstreamShas …)` ended up with one element that was itself the SHA[], and the subsequent `foreach (\ in \)` iterated once with `\` set to a System.Object[]. Verified the bug interactively (`count=1, type[0]=Object[]`) and the fix (`return \`, `count=N, type[0]=String`). PowerShell's `@(...)` already flattens the stream — the comma was actively harmful here. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…eflects failure class
03-cherry-pick-one.ps1:
- When git cherry-pick fails *before* starting (bad args, unsupported
option on older git, hook reject), the in-progress state is empty
and the subsequent `git cherry-pick --abort` itself fails with
`no cherry-pick or revert in progress`. Previously we threw on
that, which discarded the carefully-built stuck JSON and aborted
the orchestration with a misleading error.
- Now capture --abort's exit and output, classify the "nothing to
abort" case as benign, and fold any abort diagnostic into
`result.error` so the operator still gets the JSON-stuck payload
with full context.
SKILL.md step 5a (stuck-issue body):
- Body said "stopped at a conflict" unconditionally, which misled
operators when 03 stuck on a non-conflict failure (empty conflict
list, error like "merge commit without -m"). Body now branches on
whether conflict_paths is non-empty:
- conflict path > 0 → "stopped at a conflict that needs human
judgment"
- empty → "non-conflict cherry-pick failure (e.g. merge commit
without `-m`, unsupported git option, hook failure)"
and the "Conflicting paths" section renders `_(none …)_` instead
of an empty list when there are none.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…it log SKILL.md Get-OwnerRepo: - Earlier rewrite (round 15) split out HTTPS + scp-style only. A remote like `ssh://git@github.com/microsoft/terminal.git` (the fully-qualified SSH-protocol form, including `ssh://…:22/…`) matched neither branch and was treated as "wrong fork" — the identity guard then aborted the sync. - Now three explicit branches: https, ssh://[user@]host[:port]/, and scp-style `user@host:owner/repo`. Verified against eight URL variants (with/without `.git`, with/without `user@`, with trailing slash, with port) — all resolve to `microsoft/terminal`. 03-cherry-pick-one.ps1: - Tier-1 empty-commit detection ran `git diff-tree` and `git log` without checking `\0`. If either failed (corrupt ref, plumbing change, hook issue), an empty stdout would `-not \` true → silently reset --hard \; non-empty stdout from a failure would proceed as if the pick had real changes. - Now captures stderr, fails fast on non-zero exit with the actual git diagnostic, and only then evaluates the change list / cherry-pick footer match. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…n catch under Stop; exit-check git diff --cached 04-try-build.ps1: - Was `Process::Start(psi)` then attached OutputDataReceived / ErrorDataReceived afterwards. Anything the cmd shell, razzle.cmd preamble, or MSBuild banner emitted before .NET wired up the handlers was dropped, so `log_tail` could miss the only line that explained the failure (e.g. "razzle: bad arg"). - Now construct a `New-Object Process`, assign StartInfo, attach both handlers, `[void]\.Start()`, then BeginOutputReadLine / BeginErrorReadLine. No window between Start and the handler subscription. - Outer `catch` did `Write-Error \.Exception.Message` while `\Continue='Stop'` was set at the top of the script. That turns each Write-Error into its own terminating error, so the `throw` after it never sees the original exception — operators got the second-order stop record instead of the actual git/msbuild failure. Switched to `[Console]::Error.WriteLine` for the diagnostic and kept the bare `throw` to rethrow `\` unmodified. 03-cherry-pick-one.ps1: - `git diff --cached --name-only` after a failed `cherry-pick --continue` ran without an exit-code check. If `git diff` itself failed (corrupted index, missing object) the stderr text would be assigned to `\`, become truthy, and the script would proceed to `git cherry-pick --abort` instead of taking the `--skip` branch — or worse, masked the underlying repo problem with a misleading stuck issue. - Now captures stderr (`2>&1`), throws with the actual git message on non-zero exit, and filters whitespace-only lines before deciding the staged set is empty. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The comment said "ancestor (already on origin/main, drop it)" but the ancestor check is `\` vs `\` (the watermark from `Get-LastSyncedUpstreamSha`), not `origin/main` directly. The distinction matters: the watermark is the *newest* upstream SHA we've already picked, recorded in the trailer; the filter drops any SHA at or before the watermark on upstream/main's history. Rewrote to call that out — "reachable from \ (the watermark)" — and explained exit-1 as "newer than the watermark, keep it". No behavior change. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…ete duplicated reference After convergence on Copilot review, a fresh structural review caught 3 concrete issues against `.github/instructions/agent-skills.instructions.md`: 1. SKILL.md was 568 lines, exceeding the hard 500-line cap (`agent-skills.instructions.md:364` — "500 is the hard maximum"). 2. `references/04-build-verification.md` duplicated SKILL.md step 7 without adding unique content. 3. (Considered: absolute URLs in SKILL.md gh-issue/PR-body heredocs. Kept absolute — GitHub-rendered issue/PR bodies do not resolve relative repo paths, so absolute is correct in those contexts.) Changes: - Extracted ``Direct-to-main``, ``First-time sync``, and ``Squash-merge recovery`` (55 lines) from SKILL.md into a new `references/recovery-procedures.md`. Replaced with a 6-line pointer section in SKILL.md. - Compacted the ``After-PR review handling`` section (40 lines) into a 10-line summary plus a pointer to the existing `references/follow-up-pr.md` (which already carries the full rubric, so the in-SKILL.md version was duplicate prose). - Deleted `references/04-build-verification.md`. Its JSON output contract lives in `04-try-build.ps1`'s docstring; its Generated-Files artifact note is already in SKILL.md step 7. - Updated SKILL.md References list to drop 04-build-verification and add recovery-procedures. - Fixed two stale ``#first-time-sync`` anchor links (in the State Model bullet at line 52 and in Troubleshooting) to point at the extracted section in `references/recovery-procedures.md`. Result: SKILL.md is now 495 lines (down from 568), under the 500-line cap. No information is lost — all extracted content is preserved in the references, and the SKILL.md body stays focused on the runbook the agent actually executes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
SKILL.md: 495 → 156 lines (well under the 200-line soft target from agent-skills.instructions.md L312; matches sibling skill copilot-pr-review-loop/SKILL.md which is 117 lines). The full 8-step procedure with commands, JSON contracts, and failure-handling moved verbatim into references/run-a-sync.md (336 lines). SKILL.md now keeps only the ASCII flow diagram + four key invariants and points at the runbook — same progressive-disclosure pattern the sibling skill uses. Gotchas compressed from 13 to 9 entries; kept only items Copilot would not know from training (the trailer-as-watermark trick, squash-merge hazard, gh pr create retry pattern, take-upstream tier-0 file, single-host scheduler caveat, CRLF/LF on manifests). Generic git advice dropped. No information lost — everything is reachable via the References list. The 8-step contract, all scripts, and all reference docs are unchanged. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Three findings on Squash-merge recovery section, all correct: 1. `-AutoMergeStrategy rebase` is not a real flag in this skill — the step-8 recipe uses `gh pr merge --rebase --auto`. Wording updated. 2. The watermark resolver in 02-compute-pending.ps1 walks trailers bottom-up (newest-first, lines 60-66), so a squash commit with the trailers preserved picks the NEWEST trailer, not the first/oldest. Watermark does NOT move backward in that case. Doc was wrong. 3. Therefore recovery is conditional: needed only when the squash body was hand-edited to strip the newest trailer (or all trailers). Re-seeding unconditionally would be unnecessary work most of the time. Section rewritten to make the condition explicit. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…case revert SHAs
Two findings from Copilot:
1. SKILL.md troubleshooting table listed `union` as a Tier-0 strategy.
03-cherry-pick-one.ps1:191 emits "union strategy not implemented
yet" and escalates to stuck, and 03-known-conflicts.md:12 says so
explicitly. Dropped `union` from the parenthetical and added the
"reserved, currently escalates" note so operators don't add entries
that won't auto-resolve.
2. Revert-pair regex `[0-9a-f]{40}` is case-sensitive. While `git
revert` itself emits lowercase, the body could carry an uppercase
SHA (hand-edited message, scripted commits, third-party tools).
Switched to `[0-9a-fA-F]{40}` so we don't silently miss a
revert-pair and re-apply both halves on the fork.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot review: synopsis described the default as the full `cmd /c "tools\razzle.cmd && bz no_clean"` wrapper, but `-BuildCommand` and the `command` field in the JSON output are just the inner cmd string (the `/c` wrapper is added at line 106). Clarified the synopsis to match what callers actually pass. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot review claimed `$u -split "`n"` produces a bogus single
`System.String[]` entry when `$u` is a string[]. Verified
interactively that PowerShell's binary -split operator iterates over
arrays element-wise, so the original code worked correctly:
$u = @('a/b.txt','c/d.txt')
@($u -split "`n") -> @('a/b.txt','c/d.txt') (count=2, String)
So the bug Copilot described doesn't exist — but the -split was
redundant. Removed it; iteration is now obvious to any reader. Also
quoted `$_` in the TrimEnd call to defend against the (theoretical)
case where 2>&1 pulls an ErrorRecord into the pipeline.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds the
upstream-syncagent skill: a cherry-pick-per-commit pipeline that periodically pulls new commits frommicrosoft/terminalinto this fork, preserving the original author, committer, and dates on every commit.What's in this PR
.github/skills/upstream-sync/— full skill (SKILL.md + references + scripts)..github/upstream-sync/directory and nostate.json. The skill derives all state from authoritative sources:(cherry picked from commit <sha>)trailer onorigin/mainthat is an ancestor ofupstream/main.upstream-sync-stucklabel. Closing the issue clears the lock; the issue body is plain markdown (stuck SHA, upstream URL, author, branch, conflicting paths) — no machine-parseable metadata.git log --cherry-pick --right-only origin/main...upstream/main, then drop ancestors of the watermark.Generated Files/upstream-sync/<date>/, which is gitignored at the repo root and never committed..github/actions/spelling/excludes.txtnow excludes the whole.github/skills/tree from spell-check (per reviewer feedback that skill docs naturally contain technical strings the project-wide allow/expect lists shouldn't have to track).Why cherry-pick (not rebase / not merge)
agentic-terminalrepo and blew up because the fork's ownMerge upstreamcommits get replayed and conflict against themselves.Verified
Merge method
First-time sync
Because the watermark comes from a
cherry picked from committrailer, a freshly bootstrapped fork needs one seed commit. SKILL.md → "First-time sync" walks through it: onegit commit --allow-emptywith the trailer pointing at the upstream SHA the fork was branched from. No scripted bootstrap, no state file.Concurrency
The stuck-lock (open labeled issue) is a read-then-check gate. The skill assumes a single scheduler host. For multi-host fan-out, wrap the runner in a GitHub Actions
concurrency: upstream-syncgroup.Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com