Tag-Team LLM Coding
Put Claude Code and GitHub Copilot in structured debate inside VS Code, so they
converge on a root cause before any edit — breaking the single-agent
"guess → patch → analyze → repeat" loop on complex codebases.
How it works
You invoke @tag-team <problem> in the Copilot Chat panel. The extension acts as a
mediator:
- Proposal round (parallel) — Claude (via the
claude -p CLI, with file/terminal
tools) and Copilot (via the VS Code Language Model API) each return a structured
proposal: a committed root-cause hypothesis, a proposed change, confidence, and open
questions. They cannot hand back a patch without first naming a root cause.
- Cross-critique — each model is shown the other's proposal and must refute it with
evidence or concede. This is the loop-breaker: it forces engagement with an outside view
instead of re-patching its own guess.
- Convergence / checkpoint — when both agree (or after
turnCap rounds) the mediator
pauses and shows you the agreed diagnosis. You approve.
- Apply — on
/apply, Claude (the only side with autonomous file tools) makes the
edit. A design doc is saved to .tagteam/session.md.
| Concern |
Choice |
| Role split |
Symmetric debate; only Claude edits files |
| UI |
Copilot Chat participant (@tag-team) |
| Claude auth |
claude -p CLI on your Pro/Max subscription (no API key) |
| Control |
Selectable — checkpointed or auto-resolve (see Settings) |
Prerequisites
- VS Code ≥ 1.95
- GitHub Copilot installed and signed in (provides the second model)
- Claude Code CLI installed and logged in. Run
claude once in a terminal to log in.
If claude is not on the PATH VS Code sees, set tagTeam.claudePath to its absolute
path (e.g. ~/.claude/local/claude).
Develop / run
npm install
npm run package:vsix
Then press F5 ("Run Tag-Team Extension") to open an Extension Development Host.
Marketplace package
Run the packaging command below to produce the distributable VSIX:
npm run package:vsix
The generated file will be written to the project root as tag-team-llm-<version>.vsix.
Settings
| Setting |
Default |
Purpose |
tagTeam.claudePath |
claude |
Path to the Claude CLI |
tagTeam.claudeModel |
(blank) |
Optional --model override |
tagTeam.copilotFamily |
gpt-4o |
Copilot model family |
tagTeam.turnCap |
3 |
Max debate rounds before checkpoint |
tagTeam.autoResolve |
false |
When true, Claude applies the change automatically once both models agree on a root cause. If they never agree, you still decide. |
tagTeam.claudeTimeoutMs |
180000 |
Per-turn Claude timeout |
Per-invocation flag
Even if tagTeam.autoResolve is false, you can force auto-resolve for a single debate using the /auto-resolve slash command:
@tag-team /auto-resolve why does the login hang on refresh?
Type @tag-team / in Copilot Chat and the available commands will autocomplete. If the models agree on a root cause, Claude will apply the change immediately. If they never agree (even after turnCap rounds), you'll still get the tie-break prompt.
Limitations
- The Copilot model is text-only here; it shapes the plan but does not edit files.
- Copilot may not return strict JSON — parsing is tolerant; Claude is schema-enforced.
- Headless Claude turns add latency; keep
turnCap low for fast iteration.
