feat: add GitHub Copilot CLI as a supported coding agent

AGENTS.md

@@ -607,8 +607,8 @@ computes pure functions. Run instructions: `README.md`.
 - `loopany update` hands the running daemon over to the invoking (new) CLI version:
   `down` then `runEnsure({force:true})` - force skips the still-reported-online
   short-circuit (server `ONLINE_TTL` 30s outlives the local pidfile clear).
-- `loops.agent` (`CodingAgent` enum: `claude-code|codex|grok`) records the loop's host
-  coding agent AND selects the executor (at create: measured env fingerprint >
+- `loops.agent` (`CodingAgent` enum: `claude-code|codex|grok|copilot`) records the loop's
+  host coding agent AND selects the executor (at create: measured env fingerprint >
   `--agent` > server default; detection markers in `create.ts detectAgentFromEnv`).
   Editable afterward on the edit path - in `EDITABLE_LOOP_FIELDS`, `loopany edit
   --json`/`show` roundtrip, and the web `LoopForm` agent select (the next run picks
@@ -624,11 +624,29 @@ computes pure functions. Run instructions: `README.md`.
     `-m`; resume is `codex exec resume <sessionId> …`; `execEnv("codex")` forwards
     `OPENAI_API_KEY`/`CODEX_API_KEY`/`CODEX_HOME` (session/config under `~/.codex`
     free via `HOME`)
+  - `copilot` → `copilot` (`LOOPANY_COPILOT_BIN`), flags verified against a live
+    `copilot --help`: `-p`, `--allow-all` (unattended BYOA, same intent as claude/grok
+    `bypassPermissions`; also satisfies the `--allow-all-tools` non-interactive
+    requirement), `--deny-tool SELF_SCHEDULING_TOOLS` (parity with claude/grok's
+    `--disallowed-tools` — an unattended copilot run must not self-schedule either),
+    `--no-ask-user` (so an unattended run never blocks on Copilot's `ask_user` tool),
+    `--output-format json`, optional `-r/--resume <id>`/`--model`; `detectAgentFromEnv`
+    fingerprints it via `COPILOT_CLI` (verified live in-session); `execEnv("copilot")`
+    forwards `COPILOT_GITHUB_TOKEN`/`GH_TOKEN`/`GITHUB_TOKEN` (auth precedence per
+    `copilot help environment`) + `COPILOT_HOME`/`COPILOT_MODEL`/`GH_HOST`/
+    `COPILOT_GH_HOST` (GitHub Enterprise host targeting, same doc) + the
+    `COPILOT_PROVIDER_*` BYOK family (OAuth session under `~/.copilot` is free
+    via `HOME`). Its skill-install target id is `github-copilot` (verified against the
+    `skills` npm package source, NOT `copilot`), reading `~/.copilot/skills`. It has
+    NO `HOOK_TARGET_AGENTS` entry yet — Copilot CLI's `hooks` config exists but its
+    event-name schema (SessionStart equivalent) isn't confirmed in verified docs, so
+    that's a deliberate gap, not an oversight (`setup.ts`).
   **Non-Claude telemetry is DEGRADED**: grok's headless stream is grok-native
-  (`thought`/`text`/`end`, no cost/usage) and codex `--json` is not Claude
-  stream-json, so the Claude-shaped `makeStreamConsumer` parses nothing — a run
-  still marks ok on exit 0 and the agent's own `loopany report` persists the
-  result; daemon-side live-progress/cost/transcript await per-agent stream adapters.
+  (`thought`/`text`/`end`, no cost/usage), codex `--json` is not Claude stream-json,
+  and copilot's `--output-format json` is its own JSONL shape too — the Claude-shaped
+  `makeStreamConsumer` parses nothing from any of them; a run still marks ok on exit 0
+  and the agent's own `loopany report` persists the result; daemon-side live-progress/
+  cost/transcript await per-agent stream adapters.
   Grok's SessionStart hook (`setup.ts`, `~/.grok/hooks/loopany.json`, always-trusted)
   rides `HOOK_TARGET_AGENTS` (a superset of `SKILL_TARGET_AGENTS` — grok reads
   Claude's skills dir so it is NOT a skill-install target). The enum's single source

packages/daemon/src/create.test.ts

@@ -70,6 +70,10 @@ describe("detectAgentFromEnv", () => {
     expect(detectAgentFromEnv({ GROK_AGENT: "1" })).toBe("grok");
   });
 
+  test("fingerprints GitHub Copilot from its COPILOT_CLI marker (verified live)", () => {
+    expect(detectAgentFromEnv({ COPILOT_CLI: "1" })).toBe("copilot");
+  });
+
   test("returns null when no host marker is present (undetectable → caller falls back)", () => {
     expect(detectAgentFromEnv({ PATH: "/usr/bin" })).toBeNull();
   });
@@ -84,6 +88,7 @@ describe("coerceAgent", () => {
     expect(coerceAgent("claude-code")).toBe("claude-code");
     expect(coerceAgent("codex")).toBe("codex");
     expect(coerceAgent("grok")).toBe("grok");
+    expect(coerceAgent("copilot")).toBe("copilot");
     expect(coerceAgent("unknown")).toBeNull();
     expect(coerceAgent("")).toBeNull();
     expect(coerceAgent(undefined)).toBeNull();
@@ -96,6 +101,10 @@ describe("resolveAgent (precedence: measured > declared > undefined)", () => {
     expect(resolveAgent({ CLAUDECODE: "1" }, "codex")).toBe("claude-code");
   });
 
+  test("a measured Copilot host overrides a conflicting declaration", () => {
+    expect(resolveAgent({ COPILOT_CLI: "1" }, "codex")).toBe("copilot");
+  });
+
   test("falls back to the declared value when the env is undetectable", () => {
     expect(resolveAgent({ PATH: "/usr/bin" }, "codex")).toBe("codex");
     expect(resolveAgent({ PATH: "/usr/bin" }, "claude-code")).toBe("claude-code");

packages/daemon/src/create.ts

@@ -97,13 +97,13 @@ export function idempotencyKey(token: string, resolvedBody: Record<string, unkno
 }
 
 /** The coding agents Loopany can record a loop against (TS-only; cheap to widen). */
-export type CodingAgent = "claude-code" | "codex" | "grok";
+export type CodingAgent = "claude-code" | "codex" | "grok" | "copilot";
 
 /** Coerce an arbitrary declared value (--agent flag / config.agent) to a known
  *  agent, or null when it's absent/unrecognized (so it can't override a measurement
  *  and the server falls back to its own default). */
 export function coerceAgent(v: unknown): CodingAgent | null {
-  return v === "claude-code" || v === "codex" || v === "grok" ? v : null;
+  return v === "claude-code" || v === "codex" || v === "grok" || v === "copilot" ? v : null;
 }
 
 /**
@@ -120,11 +120,14 @@ export function coerceAgent(v: unknown): CodingAgent | null {
  *     shell tool under the sandbox). We deliberately ignore `CODEX_COMPANION_*`,
  *     which a Claude Code session can also export and would misattribute.
  *   - Grok Build: `GROK_AGENT=1` (grok exports it into child processes).
+ *   - GitHub Copilot CLI: `COPILOT_CLI=1` (verified against a live copilot session's
+ *     exported env; it also sets `COPILOT_AGENT_SESSION_ID`/`COPILOT_CLI_BINARY_VERSION`).
  */
 export function detectAgentFromEnv(env: NodeJS.ProcessEnv): CodingAgent | null {
   if (env.CLAUDECODE || env.CLAUDE_CODE_ENTRYPOINT || env.CLAUDE_CODE_SESSION_ID) return "claude-code";
   if (env.CODEX_SANDBOX || env.CODEX_SANDBOX_NETWORK_DISABLED) return "codex";
   if (env.GROK_AGENT) return "grok";
+  if (env.COPILOT_CLI) return "copilot";
   return null;
 }
 
@@ -155,7 +158,7 @@ export async function runCreate(args: string[], deps: CreateDeps = {}): Promise<
   const jsonArg = flag(args, "json");
   const dryRun = args.includes("--dry-run");
   if (jsonArg === undefined) {
-    process.stderr.write("loopany: usage: loopany new --json '<config>' [--dry-run] [--connect-key dk_…] [--tz <IANA>] [--agent claude-code|codex|grok]\n");
+    process.stderr.write("loopany: usage: loopany new --json '<config>' [--dry-run] [--connect-key dk_…] [--tz <IANA>] [--agent claude-code|codex|grok|copilot]\n");
     return 2;
   }
 

packages/daemon/src/runner.test.ts

@@ -274,6 +274,29 @@ describe("buildAgentSpawn", () => {
     expect(args.slice(-2)).toEqual(["--model", "grok-4"]);
   });
 
+  test("copilot: copilot bin + copilot arg vector — allow-all, no-ask-user, json, NO sys-file flag", () => {
+    // A sysFile is passed but copilot has no Claude sys-prompt-file flag — it must be dropped.
+    const { bin, args } = buildAgentSpawn({ agent: "copilot", prompt: "do it", sysFile: "/tmp/sys.md" });
+    expect(bin).toBe("copilot");
+    expect(args).toEqual([
+      "-p", "do it",
+      "--allow-all",
+      "--deny-tool", "ScheduleWakeup,CronCreate,CronList,CronDelete",
+      "--no-ask-user",
+      "--output-format", "json",
+    ]);
+    expect(args).not.toContain("--append-system-prompt-file");
+    expect(args).not.toContain("--verbose");
+  });
+
+  test("copilot: LOOPANY_COPILOT_BIN escape hatch + resume + model", () => {
+    process.env.LOOPANY_COPILOT_BIN = "/opt/copilot";
+    const { bin, args } = buildAgentSpawn({ agent: "copilot", prompt: "p", resumeSessionId: "c-1", model: "gpt-5" });
+    expect(bin).toBe("/opt/copilot");
+    expect(args.slice(0, 4)).toEqual(["-p", "p", "--resume", "c-1"]);
+    expect(args.slice(-2)).toEqual(["--model", "gpt-5"]);
+  });
+
   test("codex: codex exec arm — not claude flags; unattended + json + skip-git", () => {
     // A sysFile is passed but codex has no Claude sys-prompt-file flag — drop it.
     const { bin, args } = buildAgentSpawn({ agent: "codex", prompt: "do it", sysFile: "/tmp/sys.md" });

packages/daemon/src/runner.ts

@@ -97,7 +97,7 @@ export interface AgentSpawn {
 /**
  * Build the coding-agent spawn command (bin + argv) for one run pass.
  *
- * Three arms (BYOA — each agent's real CLI surface):
+ * Four arms (BYOA — each agent's real CLI surface):
  *   - `claude-code`: `claude -p … --output-format stream-json --verbose …`
  *   - `grok`: mirrors Claude's shape but uses `streaming-json`, drops `--verbose`
  *     (exit 2) and `--append-system-prompt-file` (no file form).
@@ -106,15 +106,23 @@ export interface AgentSpawn {
  *     `--dangerously-bypass-approvals-and-sandbox` (unattended BYOA, same intent
  *     as claude/grok `bypassPermissions`), optional `-m` / `--model`, and
  *     `--skip-git-repo-check` so non-git loop workdirs are not rejected.
+ *   - `copilot`: GitHub Copilot CLI's own `-p`/`--prompt` non-interactive surface
+ *     (flags verified against a live `copilot --help`). `--allow-all` is the
+ *     unattended-BYOA equivalent of claude/grok's `bypassPermissions` (it also
+ *     satisfies `--allow-all-tools`, required for non-interactive mode);
+ *     `--no-ask-user` disables the `ask_user` tool so an unattended run can never
+ *     block waiting on a question; `-r/--resume <id>` resumes a session; no Claude
+ *     sys-prompt-file flag, so `sysFile` is dropped like grok/codex.
  *
- * Escape hatches: `LOOPANY_CLAUDE_BIN` / `LOOPANY_GROK_BIN` / `LOOPANY_CODEX_BIN`.
+ * Escape hatches: `LOOPANY_CLAUDE_BIN` / `LOOPANY_GROK_BIN` / `LOOPANY_CODEX_BIN` /
+ * `LOOPANY_COPILOT_BIN`.
  *
- * Telemetry note: grok's headless stream is grok-native (`thought`/`text`/`end`)
- * and codex `--json` is not Claude stream-json either — the Claude-shaped
- * `makeStreamConsumer` parses nothing from either. Both still mark OK on exit 0;
- * the agent's own `loopany report` persists the result. Daemon-side live-
- * progress/cost/transcript for non-Claude agents is degraded until a per-agent
- * stream adapter lands.
+ * Telemetry note: grok's headless stream is grok-native (`thought`/`text`/`end`),
+ * codex `--json` is not Claude stream-json, and copilot's `--output-format json`
+ * is its own JSONL shape too — the Claude-shaped `makeStreamConsumer` parses
+ * nothing from any of them. All three still mark OK on exit 0; the agent's own
+ * `loopany report` persists the result. Daemon-side live-progress/cost/transcript
+ * for non-Claude agents is degraded until a per-agent stream adapter lands.
  */
 export function buildAgentSpawn(opts: {
   agent: CodingAgent;
@@ -161,6 +169,20 @@ export function buildAgentSpawn(opts: {
       ],
     };
   }
+  if (agent === "copilot") {
+    return {
+      bin: process.env.LOOPANY_COPILOT_BIN || "copilot",
+      args: [
+        "-p", prompt,
+        ...resume,
+        "--allow-all",
+        "--deny-tool", SELF_SCHEDULING_TOOLS,
+        "--no-ask-user",
+        "--output-format", "json",
+        ...modelArgs,
+      ],
+    };
+  }
   return {
     bin: process.env.LOOPANY_CLAUDE_BIN || "claude",
     args: [

packages/daemon/src/setup.ts

@@ -181,7 +181,12 @@ const installGrokHook: HookInstaller = (s, remove) => {
 };
 
 /** Installers keyed by `CodingAgent` id. An agent in `HOOK_TARGET_AGENTS` with no
- *  entry is reported `skipped (no session-hook integration yet)`. */
+ *  entry is reported `skipped (no session-hook integration yet)` — this currently
+ *  covers `github-copilot`: Copilot CLI's `~/.copilot/config.json` does have a
+ *  `hooks` key, but the verified `copilot help config`/`--help` output never
+ *  enumerates its event names (no confirmed SessionStart equivalent), so wiring
+ *  an installer here would risk shipping a malformed hook config. Deliberately
+ *  left as a documented gap until that's verified, not an oversight. */
 const HOOK_INSTALLERS: Record<string, HookInstaller> = {
   "claude-code": installClaudeCodeHook,
   codex: installCodexHook,

packages/daemon/src/skill-cli.test.ts

@@ -1,8 +1,9 @@
 /**
  * `loopany skill status` — honest, per-agent install reporting. The install now
- * targets every agent in `SKILL_TARGET_AGENTS` (Claude Code + Codex today), so
- * status must report each one's location for BOTH scopes (user + project), derived
- * from the same target list as the installer so the two surfaces cannot drift.
+ * targets every agent in `SKILL_TARGET_AGENTS` (Claude Code, Codex, GitHub Copilot
+ * today), so status must report each one's location for BOTH scopes (user +
+ * project), derived from the same target list as the installer so the two
+ * surfaces cannot drift.
  * Nothing here spawns npx or hits the network — status is pure filesystem reads.
  */
 import fs from "node:fs";
@@ -32,13 +33,13 @@ async function captureStatus(): Promise<string> {
 afterEach(() => vi.restoreAllMocks());
 
 describe("loopany skill status — multi-agent", () => {
-  test("reports every targeted agent (Claude Code + Codex) by label", async () => {
+  test("reports every targeted agent (Claude Code + Codex + GitHub Copilot) by label", async () => {
     const out = await captureStatus();
     for (const t of SKILL_TARGET_AGENTS) {
       expect(out).toContain(t.label);
     }
-    // The two CodingAgent values are exactly what we target today.
-    expect(SKILL_TARGET_AGENTS.map((t) => t.id)).toEqual(["claude-code", "codex"]);
+    // The CodingAgent values we target today.
+    expect(SKILL_TARGET_AGENTS.map((t) => t.id)).toEqual(["claude-code", "codex", "github-copilot"]);
   });
 
   test("reports each agent × scope (user + project) with a real installed/not-installed verdict", async () => {

packages/daemon/src/skill-install.test.ts

@@ -32,19 +32,19 @@ describe("installArgs", () => {
     // Repeated `-a <id>` flags, one per known agent (the comma form `-a a,b` is
     // rejected by the `skills` CLI as a single bogus agent name).
     expect(installArgs("/b/skill")).toEqual([
-      "--yes", "skills", "add", "/b/skill", "-a", "claude-code", "-a", "codex", "-y", "--copy",
+      "--yes", "skills", "add", "/b/skill", "-a", "claude-code", "-a", "codex", "-a", "github-copilot", "-y", "--copy",
     ]);
   });
 
   test("global appends -g", () => {
     expect(installArgs("/b/skill", true)).toEqual([
-      "--yes", "skills", "add", "/b/skill", "-a", "claude-code", "-a", "codex", "-y", "--copy", "-g",
+      "--yes", "skills", "add", "/b/skill", "-a", "claude-code", "-a", "codex", "-a", "github-copilot", "-y", "--copy", "-g",
     ]);
   });
 
-  test("targets exactly the two CodingAgent values, never `-a '*'`", () => {
+  test("targets exactly the known CodingAgent skill-CLI ids, never `-a '*'`", () => {
     const args = installArgs("/b/skill");
-    expect(SKILL_TARGET_AGENTS.map((t) => t.id)).toEqual(["claude-code", "codex"]);
+    expect(SKILL_TARGET_AGENTS.map((t) => t.id)).toEqual(["claude-code", "codex", "github-copilot"]);
     // one `-a` per agent, and the litter-everything wildcard never appears
     expect(args.filter((a) => a === "-a")).toHaveLength(SKILL_TARGET_AGENTS.length);
     expect(args).not.toContain("*");
@@ -56,20 +56,23 @@ describe("targetSkillDirs", () => {
     expect(targetSkillDirs({ global: true })).toEqual([
       "~/.claude/skills/loopany",
       "~/.agents/skills/loopany",
+      "~/.copilot/skills/loopany",
     ]);
   });
 
   test("project cwd → each agent's dir under the cwd", () => {
     expect(targetSkillDirs({ cwd: "/loops/cookie" })).toEqual([
       path.join("/loops/cookie", ".claude/skills/loopany"),
       path.join("/loops/cookie", ".agents/skills/loopany"),
+      path.join("/loops/cookie", ".copilot/skills/loopany"),
     ]);
   });
 
   test("default (no cwd) → cwd-relative, keeps the ./ prefix", () => {
     expect(targetSkillDirs()).toEqual([
       "./.claude/skills/loopany",
       "./.agents/skills/loopany",
+      "./.copilot/skills/loopany",
     ]);
   });
 });

packages/daemon/src/skill-install.ts

@@ -11,8 +11,8 @@
  * is the manual escape hatch (`--project` installs into the cwd instead).
  *
  * MULTI-AGENT: the skill is installed for every coding agent loopany knows about
- * (`SKILL_TARGET_AGENTS` — currently Claude Code and Codex, the two `CodingAgent`
- * values), so a Codex user (or any of the two) gets the skill, not just Claude Code.
+ * (`SKILL_TARGET_AGENTS` — currently Claude Code, Codex, and GitHub Copilot),
+ * so any of them gets the skill, not just Claude Code.
  * We deliberately DON'T pass `-a '*'`: empirically the `skills` CLI treats `'*'` as
  * "install to ALL ~72 supported agents regardless of presence", littering dozens of
  * home dirs (`~/.aider-desk`, `~/.astrbot`, …). Targeting the known agents explicitly
@@ -130,13 +130,17 @@ export interface InstallOpts {
 }
 
 /**
- * The coding agents the loopany skill install targets — the two `CodingAgent`
- * values. Each carries the `skills` CLI agent id (for `-a <id>`), a human label,
- * and the skill dir layout that agent reads (relative to the scope root: `~` for a
- * global/user install, the cwd for a project install). Verified empirically against
- * the current `skills` CLI: Claude Code reads `.claude/skills`, Codex reads the
- * universal `.agents/skills`. Extend this list to cover a new agent (installArgs and
- * `loopany skill status` both derive from it — they cannot drift).
+ * The coding agents the loopany skill install targets. Each carries the `skills`
+ * CLI agent id (for `-a <id>`), a human label, and the skill dir layout that
+ * agent reads (relative to the scope root: `~` for a global/user install, the
+ * cwd for a project install). Verified empirically against the current `skills`
+ * CLI: Claude Code reads `.claude/skills`, Codex reads the universal
+ * `.agents/skills`. GitHub Copilot's `skills` CLI agent id is `github-copilot`
+ * (NOT `copilot` — verified by unpacking the `skills` npm package source), and
+ * `copilot skill --help` confirms it reads personal skills from `~/.copilot/skills`
+ * (its `skillsRoot` here), also falling back to the universal `~/.agents/skills`.
+ * Extend this list to cover a new agent (installArgs and `loopany skill status`
+ * both derive from it — they cannot drift).
  */
 export const SKILL_TARGET_AGENTS: ReadonlyArray<{
   id: string;
@@ -146,6 +150,7 @@ export const SKILL_TARGET_AGENTS: ReadonlyArray<{
 }> = [
   { id: "claude-code", label: "Claude Code", skillsRoot: [".claude", "skills"] },
   { id: "codex", label: "Codex", skillsRoot: [".agents", "skills"] },
+  { id: "github-copilot", label: "GitHub Copilot", skillsRoot: [".copilot", "skills"] },
 ];
 
 /** The argv (after `npx`) for the install — pure, so tests can assert it. Targets