Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
repo-sync
description
Smart repository synchronization across workspace-hub ecosystem — diagnoses and fixes pull failures (detached HEAD, diverged branches, uncommitted changes)
type
reference
version
1.1.0
category
workspace-hub
last_updated
"2026-02-14T00:00:00.000Z"
source
internal
invoke
/repo-sync
capabilities
[]
requires
[]
see_also
[]
tags
["git","sync","multi-repo","automation"]
related_skills
["repo-structure","ecosystem-terminology"]
freedom
low
Repo Sync
Smart pull-all with automatic diagnosis and repair for the workspace-hub multi-repo ecosystem.
Usage
/repo-sync — Pull all repos, diagnose failures, and fix them.
Optional arguments:
/repo-sync pull — Pull all repos (default)
/repo-sync status — Status check only, no pulls
/repo-sync push — Push all repos with unpushed commits
repository_sync pull <repo-name> — Pull one configured repository (for example ./scripts/repository_sync pull mkt-a)
CLI syntax pitfall: repository_sync <repo-name> is not valid and returns Unknown command. If the user gives a bare repo name with repository_sync, infer the intended default operation as a single-repo pull and run ./scripts/repository_sync pull <repo-name> after a quick status check.
What It Does
Phase 1: Bulk Pull
Run ./scripts/repository_sync pull all to attempt pulling every repo.
Phase 2: Diagnose Failures
For each repo that failed, check:
Detached HEAD — submodules pinned at a commit, not on a branch
Diverged branches — local and remote have diverged, needs merge
Uncommitted changes — dirty working tree blocking pull
No upstream — no tracking branch configured
Phase 3: Auto-Fix
Apply the appropriate fix per failure type:
Failure
Fix
Detached HEAD (submodule)
git checkout main && git pull --no-rebase
Diverged branches
git pull --no-rebase (merge strategy)
Uncommitted changes blocking pull
git stash && git pull --no-rebase && git stash pop
Line-ending-only dirty tree
Diagnose with git diff --ignore-space-at-eol --shortstat; if empty, report as line-ending churn and do not normalize/discard without explicit user approval. See references/line-ending-only-dirty-working-tree.md
No upstream
Report only, no auto-fix
Deterministic generated-file conflict
Regenerate from the documented generator, validate with a parser, then stage; do not hand-merge arbitrary conflict hunks. See references/generated-file-conflict-regeneration.md
Large repo partial checkout / timed-out pull
Use no-untracked status probes, kill hung sync/status processes if needed, preserve recovery artifacts under .git/recovery-backups/, restore accidental tracked deletions; see references/large-repo-partial-checkout-recovery.md
Archived/read-only remote
Do not force workaround; preserve/export local-only commits and open a follow-up issue
Archived / Read-only Remote Exception Handling
If a repo is ahead locally but push fails because the remote is archived or read-only:
Confirm the state explicitly:
current branch
local HEAD vs upstream HEAD
ahead commit list (git log @{u}..HEAD)
diffstat for the ahead range
Do NOT try to bypass the archive/read-only state.
If the ahead range is non-trivial, export a preservation artifact:
Write a short decision memo with explicit options:
abandon
archive-only
selective port to a writable successor
Open a GitHub issue documenting the exception and link the exported patch artifact.
Default recommendation when there is no clear writable successor and the diff is large/destructive: archive-only.
This came up with pyproject-starter, where the remote was archived/read-only and the only ahead commit was a massive destructive sync commit. The safe action was to preserve the patch in repo artifacts and require explicit human review before any migration.
Phase 4: Encoding Health Check
After pulling, run the encoding check against all work queue and skill files
to surface any Windows-created UTF-16 / CRLF files that came in via the pull:
.claude/hooks/check-encoding.sh
This runs in warn-only mode (post-merge behaviour) — it reports bad files but
does not block. Any files flagged should be converted and committed immediately:
# Convert a UTF-16 file to UTF-8
iconv -f UTF-16 -t UTF-8 <file> | sed 's/\r//' > /tmp/fixed.md
mv /tmp/fixed.md <file>
git add <file> && git commit -m "fix(encoding): convert <file> to UTF-8"
Phase 5: Ecosystem Health + Summary
After the encoding check, spawn an ecosystem health agent in the background:
Task(
subagent_type="Bash",
description="Ecosystem health check",
prompt="Run the 14-check suite from /ecosystem-health skill. Report pass/fail/warn for each group.",
run_in_background=True
)
This does not block the pull summary. Report health results alongside repo status.
See /ecosystem-health skill for the full check suite and pass conditions.
Implementation
When this skill is invoked, execute these steps:
Step 1: Run pull
For all repos:
./scripts/repository_sync pull all
For a single configured repo:
./scripts/repository_sync status <repo-name>
./scripts/repository_sync pull <repo-name>
Capture output. Identify repos marked with ✗ Failed or any nonzero command exit.
Step 2: For each failed repo, diagnose
cd <repo_path>
# Check if on a branch
git branch --show-current # empty = detached HEAD# Check for uncommitted changes
git status --porcelain
# Check divergence (only if on a branch with upstream)
git rev-list --count @{u}.. 2>/dev/null # ahead
git rev-list --count ..@{u} 2>/dev/null # behind
Step 3: Apply fixes
Run all independent repo fixes in parallel using the Bash tool.
Detached HEAD:
cd <repo_path> && git checkout main && git pull --no-rebase
Diverged branches:
cd <repo_path> && git pull --no-rebase
Uncommitted changes blocking pull:
cd <repo_path> && git stash && git pull --no-rebase && git stash pop
If stash pop has conflicts, report to user — do NOT auto-resolve.
When an issue number is part of the sync/commit request, include a clickable GitHub issue hyperlink in the final report. Resolve it live with gh issue view <number> --json url --jq .url so the link is grounded in the current repository, not memory.
Important Notes
Never force-push or reset --hard without explicit user approval
Never rebase diverged branches — always merge (per workspace CLAUDE.md)
digitalmodel and worldenergydata are submodules — detached HEAD is normal when workspace-hub pins a specific commit
After fixing submodules, the workspace-hub git status will show them as modified (new submodule pointer) — this is expected
If stash pop fails with conflicts, stop and report to user
Use --no-rebase on all pulls to avoid rebase surprises on diverged repos
Encoding check runs after every pull — UTF-16 files from Windows editors
crash generate-index.py and other parsers silently. Fix immediately on detection.
If the user explicitly wants all dirty/untracked files tracked, run a second pass that stages with git add -A, commits on the repo's current branch, and pushes that branch before attempting broader branch-merging work.
For branch-merging across many repos, merge into each repo's actual default branch (main or master) detected from GitHub / origin/HEAD; do not assume main.
For large batch merges, use a temporary worktree checked out from the default branch (prefer origin/<default> if available) so merges are isolated from the user's current working tree and local dirty state.
Workspace-hub can mutate state during commit/push hooks (.claude/state/*, logs, generated reports). After a commit or failed push, always re-run git status and re-fetch before retrying; apparent ref-lock push failures may be stale-expectation races rather than true divergence.
Cron/report jobs often run inside a dirty, stale local main while only a narrow generated report should be committed. If the report files are ignored, use git add -f for only those artifacts; if local main is behind or unrelated dirty state makes merge risky, create a clean temporary worktree from origin/main, cherry-pick the narrow report commit, push HEAD:main, verify the remote artifact, then remove the temporary worktree. See references/cron-report-push-from-dirty-main.md.
Session-learning commits can generate a tracked skill ledger update after the primary commit (for example logs/orchestrator/hermes/skill-patches.jsonl). Treat that as intentional repo-ecosystem metadata: inspect it, commit it separately, and only then declare the worktree clean.
Before pushing after a merge/commit sequence, fetch and compare HEAD with origin/<branch>. Some workspace-hub hooks or concurrent syncs can leave the remote-tracking ref already equal to HEAD; if HEAD == origin/<branch> and ahead/behind is 0/0, report "push not needed" instead of pushing blindly.
If direct push to the default branch is rejected by branch protection (for example required status checks / "changes must be made through a pull request"), do not retry, force, or bypass. Preserve the exact committed ahead range by pushing HEAD to a topic branch, open a PR against the protected branch, and report the PR URL plus the still-ahead local main state as a known exception until the PR merges.
If a merge conflict is in a deterministic generated artifact, prefer the documented generator plus format validation over manual conflict editing. Example: config/ai_agents/ai-tools-status.yaml can be regenerated with bash scripts/maintenance/ai-tools-status.sh and YAML-parsed before staging. See references/generated-file-conflict-regeneration.md.
If a repo is fully synced but has many tracked modifications, check whether the diff is only line-ending churn with git diff --ignore-space-at-eol --shortstat. If that output is empty while normal git diff --shortstat is large, do not stash/pop, normalize, or discard automatically; report it as a non-blocking local working-tree condition and ask before changing line endings. See references/line-ending-only-dirty-working-tree.md.
New-branch pushes in workspace-hub may trigger expensive pre-push checks across tier-1 repos and can time out. If the user has approved sensible commands and the goal is repo hygiene/sync rather than validation, git push --no-verify may be necessary after verifying local/remote state.
Archived/read-only repos can still be committed locally for preservation, but push/merge to remote will fail; report them explicitly as blocked rather than retrying.
In huge repos, do not run broad git status --porcelain -uall, full du, or full checkout/pull loops as the first recovery move. Start with no-untracked status and ahead/behind probes, then follow references/large-repo-partial-checkout-recovery.md.
If interrupted sync creates .claude.partial-pull-backup-*, .codex.partial-pull-backup-*, or generated provider cache trees, preserve them inside .git/recovery-backups/ instead of committing or deleting them; they are recovery artifacts unless the user explicitly says otherwise.
Iron Law
No destructive git operation (force-push, reset --hard, auto-resolve conflicts) shall be performed without explicit user confirmation — ever.
Rationalization Defense
Excuse
Reality
"The conflict is trivial — I can auto-resolve it"
Trivial-looking conflicts have destroyed production data. Stash pop conflicts go to the user, always.
"Force-push is safe here because nobody else uses this branch"
You do not have full visibility into all machines and sessions using this repo. Assume others are active.
"I need to reset --hard to get to a clean state"
Use git stash to preserve work, then pull. Hard reset is data destruction, not cleanup.
"The encoding check is slow, I'll skip it this time"
UTF-16 files crash parsers silently. Skipping the check means shipping a time bomb.
Red Flags
These phrases signal you are about to violate the Iron Law:
"let me just force-push to fix this"
"I'll reset to a clean state"
"this merge conflict is simple enough to auto-resolve"
