Codeman/docs/copilot-integration.md

GitHub Copilot CLI sessions

Codeman can drive GitHub Copilot CLI (copilot, npm @github/copilot) as a session backend, alongside Claude Code, OpenCode, Codex, Gemini, Antigravity, Pi, Grok, DeepSeek Harness and OMP. copilot is a run mode: its own PTY, its own tmux session, its own tab identity (purple, a drawn mark rather than GitHub's logo). It is not a location overlay like Docker or remote-SSH cases, and it is not a web tab.

Everything here was measured against Copilot CLI 1.0.94 and 1.0.95 inside tmux. The entry itself is registry data (src/config/cli-registry/stock.ts); see cli-registry.md for the field-by-field reference.

Install

npm install -g @github/copilot

CLI management (Settings → Agents & CLIs) can run exactly that for you. Codeman resolves the binary from the server's PATH and the usual npm/nvm prefixes, then runs copilot --version and accepts it only if the output says GitHub Copilot CLI <x.y.z>. A bare copilot on PATH that answers anything else is ignored. Check what it resolved:

curl -s localhost:3000/api/copilot/status | jq
# { "available": true, "path": "/home/you/.local/bin", "version": "1.0.95" }

Authenticate

Codeman never logs Copilot in or stores its credentials.

• Interactive. Run copilot once and use /login (or copilot login), or do it inside a Codeman tab. The token goes to the OS keychain when one is available.
• Headless server. There is no keychain on a service, so Copilot offers to keep the token as plain text in ~/.copilot/config.json (authTokens). That is where it ends up when you accept. To keep a token out of a file, set COPILOT_GITHUB_TOKEN in the service environment; Copilot reads it first, then GH_TOKEN, then GITHUB_TOKEN.
• Per-session override. COPILOT_* is the only prefix the env allowlist admits for Copilot. GH_TOKEN and GITHUB_TOKEN are deliberately not accepted as per-session overrides: the allowlist is one global list with no mode context, so admitting them would make them settable on every session. Use COPILOT_GITHUB_TOKEN.
• No GitHub sign-in is needed when a session points at your own endpoint (below).

What Codeman wires up

Thing	Behaviour
Launch	copilot, in Copilot's own Manual Approval mode.
Run button	copilot --yolo (allow every tool, path and URL), like the other CLIs' bypass switches, so an unattended session never stops on an approval.
Permission clamp	allowAll is a privilegedParam. A multi-user owner without the bypass grant has it forced back to false; an absent config is never materialized (a bare copilot is already safe). COPILOT_HOME, COPILOT_ALLOW_ALL, COPILOT_PROVIDER_* and COPILOT_MODEL are privileged env keys, and so is every other COPILOT_* variable the CLI reads that switches approvals (COPILOT_PLAN_THEN_AUTOPILOT), runs code with no approval row (COPILOT_CLI_DIST_DIR, the package caches, COPILOT_EDITOR, COPILOT_PROVIDER_API_KEY_COMMAND, skills, hooks) or sends the sign-in token to another host (COPILOT_API_URL, COPILOT_GH_HOST, ...): the list is in stock.ts, read from the 1.0.98 bundle. A variable a later Copilot release adds is not covered until it is listed there.
Model	--model <id>, sent only by an API caller that sets copilotConfig.model (the value must start with a letter or digit, like resumeSessionId). Nothing in the UI sets it: a saved custom endpoint sets COPILOT_MODEL instead (see below).
Resume	--resume <value>, sent only by an API caller that sets copilotConfig.resumeSessionId (Past Sessions does not). The value (an id, an id prefix or a session name) must be one plain word starting with a letter or digit: --resume's value is optional, so --resume --yolo would otherwise make --yolo its own flag and bypass the clamp.
Continue	--continue, from Past Sessions: it resumes the most recent conversation Copilot has for that directory. Nothing sets it on a respawn. Skipped when a resume value is set.
Name	A tab name you set yourself is passed as --name <tab name>, so Copilot's own session list reads like that tab. An automatic name (w1-myapp, or an auto-named title) is not pinned. Copilot refuses --name with --resume/--continue, so it is dropped there and a continued session keeps the name it had.
Working / idle	The footer. A turn shows ◉ Working esc edit prompt (the dot alternates ◉/◎); the composer is the ❯ between two rules. A tool waiting for approval replaces the composer and reads as idle.
Model detection	The right-hand field of the footer's last row, at rest and mid-turn.
Status probe	GET /api/copilot/status, plus an entry in the CLI-installed probes and codeman doctor.
MCP	The entry declares mcpConfig (~/.copilot/mcp-config.json, dialect copilot-json, relocated by COPILOT_HOME), so MCP server sync treats it as a target and a source. copilot mcp disable is honoured (read from settings.json, never written).
Docker	The agent image installs @github/copilot. No credential directory is seeded (see below).

Your own endpoint (BYOK)

With custom model endpoints on, a saved endpoint is offered in the Run menu for Copilot and launches straight onto it (no restart; the env is applied before the process starts):

Codeman field	Variable
base URL (+ /v1)	COPILOT_PROVIDER_BASE_URL
key	COPILOT_PROVIDER_API_KEY
model	COPILOT_MODEL

COPILOT_PROVIDER_TYPE is not set: it defaults to openai, which is what a Codeman custom endpoint is (llama.cpp, vLLM, Ollama). All of these are privileged env keys.

Checked at the CLI level against a local OpenAI-compatible endpoint (1.0.95): it posted to /v1/chat/completions with the key and the model name, needed no GitHub sign-in, and --resume <id> --model <m> / --continue --model <m> resumed earlier sessions. Tool calls through a custom endpoint were not exercised. Only keys are persisted (never values).

Terminal behavior

Copilot is an alternate-screen TUI with mouse tracking on, so it needed one non-default decision, the same one OpenCode needed:

• Selection. Left in, xterm reports a plain drag to the TUI instead of selecting, so Auto Copy and copy-on-select silently did nothing (only Shift+drag selected). The entry is altScreen: 'strip-mux-and-mouse': the tracking DECSETs are stripped under tmux, a drag is a local selection, and clicks still reach Copilot through the browser's hand-encoded tap (cliMouseTracking is published as the server strips).
• Wheel and touch. Stripping the DECSETs also stops xterm encoding the wheel, so _shouldForwardWheelToApp forwards it by hand as SGR reports while Copilot is tracking the mouse; a phone swipe goes through the same gate. Shift+wheel scrolls Codeman's local history. If the tracking flag is stale after a server restart, the gesture falls back to PageUp/PageDown rather than going dead (Copilot pages its own transcript on those keys: three PageUp moved a 140-line answer from lines 109-140 to 010-042).
• Keys. Keystrokes from tmux send-keys and POST /api/sessions/:id/input reach the composer; Shift+Enter inserts a newline; shell tool calls complete inside tmux (github/copilot-cli#4180 and #4223 report both failing on 1.0.70 to 1.0.74; neither reproduces on 1.0.94).
• Scrollbar. Copilot draws a ┃ gutter at the right edge of every row; the response reader strips it.

codeman agent verbs

The three CODEMAN_* variables are exported into every pane, so the session-to-session verbs work from inside a Copilot tab and against Copilot workers:

• spawn waits for the composer. The ready mark is the footer hint / commands · ? help, not ❯: the folder-trust dialog draws ❯ 1. Yes, so the glyph would call a worker ready while it is still asking whether to trust the folder. An untrusted folder therefore makes spawn warn that the composer was not seen instead of claiming ready.
• send --wait resolves on idle. --until stop is refused (400, "no Claude Code hooks") like every hook-less mode; use idle, exit or a --match marker.
• read returns the last answer. Copilot's output stream is absolute cursor moves rather than lines, so the route reads the pane's visible screen (tmux capture-pane) and cuts it with a per-CLI pane dialect (src/web/response-viewer-transcript.ts): ❯ <prompt> HH:MM, ● <answer> with indented continuation, and the cwd / Session: N AIC used row, composer, footer and scrollbar as chrome. Everything before the first prompt (banner, trust dialog) is dropped, so a session that has not been asked anything answers nothing.
• interrupt sends a bare ESC; rm refuses itself.

Docker cases

The agent image installs @github/copilot, and a docker case runs Copilot like any mode. No credential directory is seeded: without a keyring the sign-in token is plain text in config.json, and CRED_STORES (src/docker-hosts.ts) deliberately has no .copilot row. Sign in inside the container (copilot login). Setting COPILOT_GITHUB_TOKEN for the session does not reach it: a docker quick-start refuses envOverrides, and Copilot's dockerExecEnvNames is empty, so a token in the server environment is not forwarded either.

Remote SSH cases

A remote or Docker pane runs a bare copilot (defaultRemoteCommandForMode, pinned in test/location-overlay-commands.test.ts), in Copilot's own Manual Approval mode. The copilotConfig flags (--yolo, --model, --resume, --continue) are a local-session feature and are not sent there, so there is nothing for the privileged-param clamp to do. A remote host needs its own copilot on its own PATH and its own sign-in. This path has unit coverage for the built command but was not exercised against a real remote host.

What Codeman deliberately does NOT wire up

• No login or token storage. See Authenticate.
• No bypass inferred. --yolo is sent only by the Run button's allowAll, never guessed from anything else, and is clamped for non-granted owners.
• No --api-key-style flag on the command line. Secrets travel as environment variables through tmux setenv.
• No Claude-only features. Auto-resume on a usage limit, the Approvals Inbox, Read My Mind, the Ralph tracker, subagent windows and the bundled agent skill are Claude-shaped.

Known gaps

• agent read does not model tool-call rows: they open with the same ● bullet and read as part of the answer.
• agent read sees the visible screen, so an answer longer than one screen is its tail.
• Custom endpoints: tool calls through a BYOK endpoint were not exercised.
• Remote SSH was not exercised against a real remote host.