copilot-second-opinion

OpenCode plugin + MCP server + agent skill that runs the full GitHub Copilot PR review loop — request review, wait deterministically, triage every comment, push fixes, reply, resolve, repeat, and merge with safety gates.

What it does

• request_copilot_review — POSTs Copilot as a reviewer and verifies the request took effect by re-reading requested_reviewers. The built-in github_request_copilot_review returns HTTP 200 even when Copilot isn't actually added; this one catches the silent-fail.
• wait_for_copilot_review — blocks via gh run watch on the Copilot Actions workflow run for the head SHA. No blind polling. Falls back to REST polling for older repos.
• get_copilot_threads / reply_to_review_comment / resolve_review_thread — full thread lifecycle with both the GraphQL thread_id and the REST comment_id returned in one shot.
• enable_copilot_auto_review — idempotently creates the repo ruleset that auto-requests Copilot on new PRs.
• safe_merge_pr — gated merge: blocks unless Copilot has reviewed the current HEAD, all threads are resolved, and all checks are green. Replaces github_merge_pull_request, which does none of that.
• copilot-second-opinion skill — the playbook the agent loads to drive the loop. Auto-discoverable on phrases like "address the Copilot comments" or "Copilot didn't re-review after my push".

Why

OpenCode's built-in GitHub MCP tools have two footguns this package fixes:

1. github_request_copilot_review silently no-ops when Copilot code review isn't enabled on the repo. GitHub returns HTTP 200, so the agent thinks the request worked and waits forever for a review that's never coming. Across one user's session history (~66k tool calls), the prefixed MCP tools from this package were called zero times because the agent kept finding the broken built-in first — the skill exists in part to shadow it.
2. github_merge_pull_request does zero gating. It will happily merge a PR with failed CI, unresolved Copilot threads, or a Copilot review that's only for the previous HEAD.

Install

opencode plugin copilot-second-opinion -g

This installs the package globally and, on next session start, the plugin:

1. Symlinks the skill into ~/.config/opencode/skills/copilot-second-opinion/SKILL.md.
2. Registers the MCP server in-memory via the config() hook (no opencode.json mutation — uninstalling the plugin un-registers it automatically).
3. Disables the built-in github_merge_pull_request so safe_merge_pr is the only merge path.

Or manually:

npm install -g copilot-second-opinion

Then add "copilot-second-opinion" to the plugin array in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["copilot-second-opinion"]
}

Opt out of the in-memory config injection

Some users want full control over their opencode.json. Set:

export OPENCODE_COPILOT_REVIEW_NO_AUTOCONFIG=1

The plugin will still install the skill file, but will skip MCP registration and the tools filter. Add them manually to opencode.json instead (snippet in the manual section).

Manual install (Claude Code or custom MCP hosts)

git clone https://github.com/Adamkadaban/copilot-second-opinion
cd copilot-second-opinion
./install.sh

The bash installer copies the skill into the first existing of ~/.config/opencode/skills/, ~/.claude/skills/, or ~/.opencode/skill/, runs npm install for the MCP server, and prints the exact config snippet to paste into your MCP host.

Requirements

• Node 18+
• gh CLI authenticated (gh auth status returns a green check)
• Copilot code review enabled on the target repo. If it isn't, the first request_copilot_review call will return {requested: false, hint: "..."} and the skill will offer to enable it via repository ruleset.

Usage

Inside an OpenCode session, the skill is auto-discoverable. Any of these will trigger it:

get a second opinion from Copilot on PR 42 address the Copilot comments on this PR Copilot didn't re-review after my last push merge PR 42 if Copilot is happy and checks pass

The agent will load the skill, run the loop, and use safe_merge_pr for the final merge.

Tools

| Tool | Purpose | |---|---| | request_copilot_review | POST Copilot as a reviewer + verify the request took effect | | check_copilot_review_status | Non-blocking snapshot: done / pending / absent for the current HEAD | | wait_for_copilot_review | Blocking wait via gh run watch, with REST-poll fallback for older repos | | get_copilot_threads | List unresolved threads with both thread_id and root_comment_id | | reply_to_review_comment | Post a reply to a specific review comment | | resolve_review_thread | Resolve a thread via GraphQL resolveReviewThread | | enable_copilot_auto_review | Create/update the repo ruleset with copilot_code_review (idempotent) | | safe_merge_pr | Gated merge — blocks unless review, threads, and checks all pass |

Each tool has its own server-side timeout budget (30s–120s) so a hung gh call can't block the full session.

Development

git clone https://github.com/Adamkadaban/copilot-second-opinion
cd copilot-second-opinion
npm install
npm run check         # smoke-test plugin import
npm run mcp:smoke     # smoke-test MCP server initialize

For live development against your OpenCode session:

cd ~/.config/opencode
npm link /path/to/copilot-second-opinion

Then add "copilot-second-opinion" to your plugin array in opencode.json.

Releasing

Releases are automated via GitHub Actions. To cut a new release:

npm version patch   # or minor / major
git push --follow-tags

The publish workflow sanity-checks the plugin + MCP server, publishes to npm with provenance via Trusted Publishing, and creates a GitHub Release with auto-generated notes.

License

MIT