prime-sync

Ultrathink.

Mode Detection

Detect operating mode before anything else:

1. VERSION file exists in CWD → push mode (CWD is the prime repo)
2. .claude/.prime-version exists in CWD → pull mode (CWD is a target project)
3. Neither detected → ask user which mode and for the required path

Push Mode — Resolve Target

Source = CWD (prime repo). Target = resolved project path.

State file: .claude/prime-projects.json

Tracks known target paths, last sync time, and version. Shape:

{ "projects": [{ "path": "/absolute/path", "lastSynced": "2025-12-30T10:30:00Z", "version": "1.4.2" }] }

Create as { "projects": [] } if missing.

Resolution flow:

1. Argument provided → Use that path, then record it after a successful sync
2. No argument, known projects exist → Ask the user which project(s) to sync, including an All projects option
3. No argument, no known projects → Ask the user for a path

State file management:

• Create .claude/ only if you need to persist the state file
• Update the synced project's time/version only after a successful sync
• Keep one entry per absolute path

Pull Mode — Clone Prime

Source = cloned prime repo. Target = CWD.

Flow:

1. Read current version from .claude/.prime-version in CWD
2. Generate timestamp: date +%Y%m%d%H%M%S
3. Clone prime repo: git clone https://github.com/avibebuilder/claude-prime.git /tmp/claude-prime-sync-<timestamp>/
4. Set source = /tmp/claude-prime-sync-<timestamp>/, target = CWD
5. Continue to shared process below
6. Clean up /tmp/claude-prime-sync-<timestamp>/ after sync completes (success or failure)

No state file in pull mode — the target project is self-contained.

Process

1. Validate Target

• Check that the target path exists, is a directory, and is not the same location as the source repo
• Treat a valid target as a project root where writing .claude/ is appropriate
• Read versions from target (.claude/.prime-version) and prime (VERSION)

If the path fails those checks, abort and explain why.

2. Parallel Analysis

Change Detection:

• Detect relevant changes in prime's .claude/ based on the target's recorded version state
• If the target has no recorded version, or already matches the current prime version, sync only uncommitted prime changes
• If the target is on an older prime version, diff from that version tag to HEAD and include uncommitted changes; warn if the tag is missing
• Categorize changes into commands, agents, hooks, settings, skills, and starter-skills

Stack Detection (only if skills or starter-skills changed): check target for stack indicators relevant to each changed skill/starter.

File Comparison: compare each changed file with target. Detect: NEW, UPDATE, IDENTICAL, CONFLICT.

3. Build Sync Plan

Aggregate results and present:

CHANGED (will sync): ...
IDENTICAL (skip): ...
CONFLICTS (will ask): ...
IRRELEVANT (skip, wrong stack): ...

GATE: User approves sync plan.

4. Handle Conflicts

For each conflict, show the relevant diff and ask the user how to proceed:

• Overwrite — Replace the target copy with the prime version
• Skip — Leave the target copy unchanged
• Merge — Surface the full diff and let the user choose which parts to keep before writing anything

5. Execute Sync

1. Create .claude/ in target if needed
2. Copy approved files (skills as entire folders)
3. Sync approved starter-skills to .claude/starter-skills/ in target
4. Update .prime-version with prime's VERSION

6. Report

Sync complete! (prime vX.X.X)
Updated: ...
Skipped: ...
Version: X.X.X → written to .prime-version

Push mode only: update state file (lastSynced timestamp + version). Pull mode only: clean up /tmp/claude-prime-sync-* clone directory.

Gotchas

• prime-sync stays in prime — do not copy this skill into target projects.
• Push and pull keep different state — state file is push-only; pull mode treats the target as self-contained.
• Version history can be incomplete — if the expected tag is missing, warn and continue with that uncertainty explicit.

Constraints

• NEVER modify the target project's source code
• NEVER sync without user approval at the plan gate
• In push mode, preserve one state entry per target path and update it only after success
• In pull mode, always clean up the temporary clone, even on failure

Target Path

<target-path>$ARGUMENTS</target-path>