compound-sync

name	compound-sync
description	Sync GaleHarnessCLI skills with latest changes from the reference upstream repo. Use when you need to pull upstream updates, review what changed, apply gh: renames, and re-inject HKTMemory patches. Triggers: 'sync upstream', 'update skills', 'pull latest changes', 'upstream diff'.

Compound Sync

Sync GaleHarnessCLI with upstream compound-engineering-plugin changes through per-commit patch batches.

Upstream Reference

Current baseline: Read from .upstream-ref in repo root.

cat .upstream-ref

GaleHarnessCLI was forked from EveryInc/compound-engineering-plugin at the commit recorded in .upstream-ref. All subsequent changes are independent developments.

When to Use

• Pull latest upstream compound-engineering changes
• Review upstream diff before applying
• Re-apply gh: prefix renames after upstream sync
• Re-inject HKTMemory patches after sync

Workflow

Use the upstream sync CLI as the source of truth for batch generation, per-commit worktrees, PR creation, state transitions, cleanup, and .upstream-ref updates.

Step 1: Initialize The Batch

From the repository root, run:

python3 scripts/upstream-sync/sync-cli.py init
python3 scripts/upstream-sync/sync-cli.py status

The CLI reads .upstream-ref and .upstream-repo, generates a dated batch under .context/galeharness-cli/upstream-sync/, and writes workflow state under .context/galeharness-cli/upstream-sync/state.json.

If init reports no new upstream commits, stop. Do not mutate .upstream-ref manually.

Step 2: Process One Commit

Start exactly one pending upstream commit:

python3 scripts/upstream-sync/sync-cli.py next

The CLI creates an isolated worktree, applies one adapted patch, runs verification, commits the result, pushes the branch, and opens a PR. Treat the generated PR as the review boundary for that upstream commit.

Default operating model:

• one upstream commit -> one adapted patch
• one adapted patch -> one isolated worktree
• one adapted patch -> one PR

Step 3: Re-apply GaleHarnessCLI-Specific Logic

Mechanical adaptation only handles naming and path rewrites. In the generated worktree or PR:

1. Re-apply gh: naming expectations not covered by raw text replacement
2. Re-inject HKTMemory workflow patches where upstream does not know about them
3. Update or add tests for the changed behavior
4. Run focused verification before the PR lands

Step 4: Resume Or Skip

After the PR is merged, return to the original worktree and run:

python3 scripts/upstream-sync/sync-cli.py resume
python3 scripts/upstream-sync/sync-cli.py status

resume verifies the PR target, reconciles upstream ancestry, updates .upstream-ref for the merged upstream commit, cleans up the worktree, and returns the workflow to idle for the next explicit next command. It does not automatically start the next commit.

If a commit must be abandoned, run:

python3 scripts/upstream-sync/sync-cli.py skip --reason "short reason"

For an open or closed PR that needs cleanup, use:

python3 scripts/upstream-sync/sync-cli.py skip --force-cleanup --reason "short reason"

The CLI validates branch ownership before deleting remote branches.

Step 5: Complete The Batch

Repeat next -> PR review/merge -> resume until status reports complete.

Do not hand-edit .upstream-ref. The CLI advances it only after each corresponding PR is confirmed merged and upstream ancestry checks pass.

HKTMemory Patches to Re-apply

After each upstream sync, these patches must be re-applied:

gh:compound skill

• Phase 0.4: HKTMemory retrieve before research
• Phase 2.3: HKTMemory store after solution documented

learnings-researcher agent

• Step 0: HKTMemory vector search before grep-first filtering

gh:brainstorm, gh:plan, gh:work, gh:review

• Per-phase HKTMemory read/write (see each skill's HKTMemory sections)

Notes

• GaleHarnessCLI is now fully independent (upstream remote removed)
• This skill provides an Agent-driven batch workflow instead of a direct git merge
• Always review upstream changes before applying; not every upstream commit will be mechanically clean
• The batch directory is local workflow state under .context/, not a Git-tracked artifact