Copilot Mode

Copilot Mode uses the GitHub Copilot CLI as an AI backend, sitting alongside Claude Code Chat in the same Web UI. If you already have a GitHub Copilot subscription, it's the easiest way to switch between GPT-5 / Claude Sonnet 4.x / Gemini 2.5 Pro and other vendor models from one chat surface.

Prerequisites

1. GitHub Copilot CLI installed and logged in on the Agent machine:

   # Install (see GitHub's official docs)
   gh extension install github/gh-copilot
   # Or the standalone CLI (recommended): npm install -g @github/copilot
   
   # Log in
   copilot login

   After login, copilot --version must return >= 1.0.59 (ACP protocol dependency).
2. Optional environment variables:
   • COPILOT_BIN — explicit path to the copilot executable (default: search $PATH)
   • COPILOT_YOLO=1 — auto-approve all tool calls globally (not recommended; prefer per-session "Allow all tools")

The Yeaft Web tier does not install or download the Copilot CLI for you — it's a user-tier tool, follow GitHub's official install flow.

Create a Copilot session

1. Sidebar → + new session
2. In the session modal:
   • Agent — which machine to run on
   • Provider — drop-down → Copilot
   • Working directory — project path
   • Model — pick from the list Copilot provides (see below)
   • Allow all tools (optional) — auto-approve every tool call, no popups

Model picker

The Copilot model list is fetched dynamically — on first copilot --acp startup, Yeaft scrapes the real available list (with vendor, pricing class and preview flags) from the ACP session/new response. Subject to your Copilot subscription, common models include:

Vendor	Typical models
Anthropic	Claude Sonnet 4.6 / 4.5, Claude Haiku 4.5, Claude Opus 4.5–4.8
OpenAI	GPT-5.5 / 5.4 / 5.4-mini, GPT-5.3 Codex, GPT-5 Mini
Google	Gemini 2.5 Pro

Model can't change mid-session. The ACP protocol currently doesn't allow hot model switching — open a new session to change model.

Differences vs Claude Code Chat

Capability	Claude Code	Copilot
/compact auto-compress	✓	—
/clear reset	✓	✓
Model picker	✓	✓
MCP tools	✓	✓
Image / file attachments	✓	✓
AskUser permission popup	✓	✓
Subagent nested monitor	✓	—
Expert Panel	✓	—
Session history resume	✓	✓ (via ~/.copilot/session-store.db)

The UI hides unsupported buttons based on the provider's capability flags — buttons you don't see simply aren't supported, not a bug.

Tool permissions: the ask-user popup

By default (Allow all tools not checked), Copilot pops up a confirmation every time it wants to run bash, write files, etc.:

Copilot wants to run bash. Allow?

Buttons are usually "Allow once" / "Allow always" / "Reject", with the exact options coming from the Copilot backend. Your choice goes back to the Copilot subprocess via ACP's session/request_permission flow.

Best practice:

• Exploratory sessions — leave default, confirm each call
• Long tasks (batch refactor, running a test suite) — toggle Allow all tools when creating the session to skip popups

Session history

Copilot stores sessions in ~/.copilot/session-store.db (SQLite). Yeaft will:

• List prior sessions for the working directory in the new-session modal — click to resume
• Use ACP's session/load for resume (when the Copilot CLI supports it); fall back to a fresh session otherwise
• Convert forge_trajectory_events rows into the claude_output protocol envelope so the front-end reuses the same renderer pipeline

Environment variable: COPILOT_DB_PATH overrides the default DB path (handy for multi-account or custom storage).

Troubleshooting

"copilot ACP init failed: ... Run copilot login and ensure CLI >= 1.0.59"

• Run copilot --version on the Agent machine; require >= 1.0.59
• Run copilot login to complete the GitHub OAuth flow
• If copilot isn't on $PATH, set COPILOT_BIN=/full/path/to/copilot

The model picker is empty / only shows static fallback list

• Create one Copilot session — first connection caches the real list
• Static fallback lives in agent/providers/copilot-models.js as FALLBACK_COPILOT_MODELS

"Copilot CLI does not advertise loadSession capability — starting a new session instead of resuming"

• Copilot CLI is too old to resume sessions; upgrade to the latest

Ask-user popup appears and the page hangs

• The Copilot subprocess may have crashed; Yeaft auto-drains pending popups, refresh and you should be fine
• Check Agent logs: yeaft-agent logs

Advanced: don't confuse this with the Yeaft engine's github-copilot credential

There are two distinct "Copilots" — keep them separate:

• Copilot Mode (this chapter) — Agent spawns a copilot --acp subprocess; the subprocess is the AI backend, model selection lives inside the Copilot CLI
• Yeaft engine's github-copilot credential provider — used by Yeaft Sessions; Yeaft grabs the GitHub OAuth token itself and calls the Copilot API directly, no copilot --acp subprocess

Both authenticate against the same GitHub account, but they take different code paths. The first is for "I want Copilot CLI as a Claude replacement"; the second is for "I want a VP inside Yeaft Sessions to use Copilot-provided models".