agent-conf-sync

name

agent-conf-sync

description

Runs the ai-dev-agent-config-sync batch script (`sync-all`) to regenerate Cursor, Claude Code, GitHub Copilot, VS Code, and JetBrains Junie configuration from a generic source tree. Use when the user asks to "sync agent config", "regenerate rules/skills/agents/mcp", "update .cursor / .claude / .github / .junie / .vscode from source", "run sync-all", or similar — so they don't have to assemble CLI flags or pick Windows vs Unix scripts themselves. The skill detects the OS, finds the script, and infers `-i` / `-o` / `--tools` / `--items` / `--clean` from the natural-language request.

agent-conf-sync

This skill invokes scripts/sync-all.sh (macOS/Linux/Git Bash) or scripts/sync-all.ps1 (Windows PowerShell) from the ai-dev-agent-config-sync repo. The user describes what they want in plain language; you pick the right script, resolve paths, infer flags, run it, and summarize the result.

Full format reference is in the repo's README.md. This skill is only about invocation and reporting.

When to apply

Apply when the user wants to (re)generate tool-specific AI-assistant config from a single generic source tree. Trigger phrases include:

• "sync all", "run sync-all", "sync agent config", "sync the config"
• "regenerate / update the rules / skills / agents / guidelines / mcp"
• "rebuild .cursor / .claude / .github / .junie"
• "sync MCP servers", "regenerate .vscode/mcp.json", "update Claude .mcp.json"
• "sync for Cursor and Claude", "only Copilot", "skip Junie"
• "clean sync", "prune stale files", "mirror deletions"

Do not apply when the user is asking about the format of rules/skills/agents, or editing source files. For that, point them to README.md.

Step 1: Detect the OS and shell

Pick the entrypoint based on the shell you'll actually execute in:

Environment	Entrypoint	Flag style
Windows native PowerShell (pwsh, powershell.exe)	scripts\sync-all.ps1	-InputRoot, -OutputRoot, -Tools, -Items, -Clean
macOS, Linux, BSD	scripts/sync-all.sh	-i, -o, --tools, --items, --clean
Git Bash / WSL on Windows	scripts/sync-all.sh (prefer Bash)	same as Unix

Detect with: $env:OS -eq "Windows_NT" in PowerShell; uname in Bash. Do not assume Unix just because the repo is cross-platform — use what the current session runs.

Step 2: Locate the script root (CONFIG_SYNC_ROOT)

Resolve the directory that contains scripts/sync-all.sh and scripts/sync-all.ps1:

1. If the workspace root has scripts/sync-all.sh, use the workspace root.
2. Else if the user named a path (e.g. vendor/ai-dev-agent-config-sync), use that.
3. Else search the workspace for **/scripts/sync-all.sh and use its grandparent.
4. If still ambiguous or missing, ask once for the path.

Use absolute paths when invoking the script.

Step 3: Infer the required parameters

Input root — -i / -InputRoot (required)

The directory that contains AGENTS.md, agents/, skills/, rules/ (the generic authoring tree). May optionally contain mcp-servers/ — when present, sync-all also generates the per-tool MCP config (Cursor .cursor/mcp.json, Claude .mcp.json, VS Code .vscode/mcp.json; Junie prints the snippet to stdout). .vscode/mcp.json is written by the vscode tool, not by copilot (it is a VS Code file, not a Copilot file).

Mapping from user phrasing:

• "from _internal", "internal tree", "authoring root", "this repo's own source" → <CONFIG_SYNC_ROOT>/_internal
• "from examples", "the examples tree" → <CONFIG_SYNC_ROOT>/examples
• "from my source", "from <path>" → that path
• "in place", "same as output" → same path as output root

If nothing is stated, look at the workspace: presence of AGENTS.md + agents/ + skills/ + rules/ at some root indicates it's the source. Ask if unclear.

Output root — -o / -OutputRoot (required)

The project root where .cursor/, .claude/, .github/, .junie/, and the root AGENTS.md copy are written. Usually the consumer project root, or this repo's root when regenerating its own outputs.

--tools / -Tools (optional; default: all four)

Comma-separated subset of cursor,claude,copilot,vscode,junie (case-insensitive, no spaces).

Phrase	Value
"all tools", "everything", no tool mentioned	omit flag (default = all five)
"only Cursor"	cursor
"Cursor and Claude", "not Copilot, not Junie"	cursor,claude
"skip Junie"	cursor,claude,copilot,vscode
"only VS Code", "just MCP for VS Code"	vscode
"Copilot with its MCP"	copilot,vscode (Copilot Chat in VS Code reads .vscode/mcp.json)

Synonyms: cursor ← "Cursor"; claude ← "Claude", "Claude Code"; copilot ← "Copilot", "GitHub Copilot"; vscode ← "VS Code", "VSCode", "Visual Studio Code"; junie ← "Junie", "JetBrains".

--items / -Items (optional)

Single comma-separated list of basenames without extension (e.g. delegate-to-aside, commit-style). Applied to agents, skills, and rules together.

Phrase	Value
"only the aside skill", "just delegate-to-aside"	delegate-to-aside
"just foo and bar"	foo,bar
"full sync", "all of them", no item mentioned	omit flag

Caveat: --items does not trim the guidelines merge. sync-agent-guidelines always emits the full AGENTS.md + merged rules for Claude/Junie. Mention this in the report if the user expected partial guidelines.

--clean / -Clean (optional; default: off)

Set when the user asks to remove stale outputs, prune, mirror source deletions, empty generated dirs first, or says "clean sync".

Warning — data loss risk: --clean empties the entire output directory for every sync step on every selected tool before writing: .cursor/{agents,rules,skills}, .claude/{agents,skills}, .github/{agents,instructions,skills}, .junie/{agents,skills}. Any hand-authored files in those dirs are deleted. This is especially risky for .github/ since users often keep unrelated content there (workflows live in .github/workflows/ — not wiped — but a user might have put manual instruction files in .github/instructions/). Flag this risk in your pre-run note whenever -Clean is inferred, and name the specific dirs that will be emptied based on the --tools selection.

Step 4: Validate and run

Before running:

1. Check that <input_root>/AGENTS.md exists. If not, stop and tell the user.
2. Use absolute paths in the command.
3. Capture full stdout and stderr.

Example — macOS / Linux:

bash "/abs/path/to/ai-dev-agent-config-sync/scripts/sync-all.sh" \
  -i "/abs/path/to/source-root" \
  -o "/abs/path/to/output-root" \
  --tools cursor,claude \
  --items delegate-to-aside

With clean:

bash "/abs/path/to/ai-dev-agent-config-sync/scripts/sync-all.sh" \
  -i "/abs/path/to/source-root" \
  -o "/abs/path/to/output-root" \
  --clean

Example — Windows PowerShell:

& 'C:\abs\path\to\ai-dev-agent-config-sync\scripts\sync-all.ps1' `
  -InputRoot 'C:\abs\path\to\source-root' `
  -OutputRoot 'C:\abs\path\to\output-root' `
  -Tools 'cursor,claude' `
  -Items 'delegate-to-aside'

On non-zero exit, surface the error and suggest fixes (missing dirs, unknown tool name, missing AGENTS.md).

Step 5: Post-run report (required)

Reply with a compact summary, not the full log:

1. Command — one line: OS/shell, script path, effective -i/-o/--tools/--items/--clean.
2. Deletions / clean — list paths from log lines matching cleaned / removed / Clean. If --clean was not used, state that no clean step ran (overwrite in place only).
3. Generated / updated — from lines with -> (e.g. cursor agent ->, copilot skill ->) and == tool == headers. Group by tool. Call out AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, .junie/AGENTS.md when they appear.
4. Basis — one short sentence explaining what you inferred from the user's request.

Edge cases

• MCP-only sync: likewise, call scripts/<tool>/sync-mcp.{sh,ps1} -i <src>/mcp-servers -o <out> directly. The Bash variants require jq.
• Rules-only sync: sync-all has no such flag. Call the per-tool script directly (e.g. scripts/cursor/sync-rules.sh -i <src>/rules -o <out>).
• Submodule / monorepo: always resolve real filesystem paths before running.
• Input equals output: allowed; scripts skip the redundant AGENTS.md copy.
• Git Bash on Windows: prefer the .sh script with bash; only use .ps1 when the session is clearly native PowerShell.