feat: GitHub Copilot CLI adapter

[arielbk]

What

Add GitHub Copilot CLI as the fourth session tool (copilot), alongside Claude, Codex, and Cursor.

• Parse Copilot events.jsonl transcripts under ~/.copilot/session-state/<uuid>/ (honoring COPILOT_HOME) into the common transcript shape: title, model, message head/tail, and token totals.
• Locate live Copilot sessions by walking ancestor PIDs against session-state/<uuid>/inuse.<pid>.lock files, since Copilot exposes no session-id environment variable.
• Accept Copilot lifecycle hooks (sessionStart / agentStop / subagentStop) with their camelCase payloads, keyed on the sessionId carried in stdin.
• Ship the integration as part of the Trace plugin — manifest, version-1 hooks JSON, and a Copilot protocol skill resource — installed with copilot plugin install ./plugin (documented in the README).
• Render Copilot sessions on the board with their own icon and theme token.

Why

Trace already captures Claude, Codex, and Cursor sessions; Copilot CLI sessions were invisible. This brings task binding, transcript rendering, state-freshness checks, and re-entry to Copilot with no manual setup beyond the plugin install.

Implementation details

• Transcript adapter (packages/core/src/copilot-adapter.ts): reads the append-only events.jsonl, tolerating both data-nested and top-level event fields, and skipping malformed lines. Token totals are output-only — Copilot records no input or cache tokens (outputTokens per assistant.message).
• Session locator (apps/cli/src/copilot-session.ts): core stays fs/process-free via an injected resolveCopilotSession callback (same pattern as Cursor), wired once in the CLI identity composition root. Copilot identity takes precedence over Cursor cwd-guessing only when Claude/Codex identities are absent.
• Hooks: every payload carries sessionId + cwd. Copilot ignores command-hook stdout at sessionStart, so the binding nudge rides the prompt-type hook (auto-submitted text). agentStop feeds the payload id into state check's explicit-binding gate.
• Subagent capture: subagentStop is accepted but intentionally a no-op — no parent/child linkage was observed in the event stream. To be probed empirically during dogfood; a stable relation would enable nested subagent sessions later.
• No Drizzle migration: the tool column is unconstrained text; the SessionTool enum is TS-only.

Cut from this PR: Windows VDI support (follow-up task), session scan --copilot backfill (Cursor precedent: live capture only), and reading the session-store.db SQLite mirror.

Testing

• Copilot adapter unit suite — 384 core tests, covering transcript identity, model, message head/tail, output-token totals, and malformed-line handling.
• Live-locator fake-tree tests — matching ancestors, concurrent sessions, stale locks, and COPILOT_HOME.
• Hook/stop tests — 40 focused tests asserting registration, freshness, and exit-code behavior equivalent to the Claude hooks.
• Plugin scaffold checks — manifest, lifecycle hooks, pinned commands, skill resources, and README install command.
• Typecheck and lint pass on the changed packages.

Remaining before undraft

The end-to-end macOS dogfood (the task's done condition) is still to run: install the plugin per README, bind a task from a real Copilot session, verify board rendering and re-entry, probe subagent linkage, and record the resume invocation + prompt-hook dynamism findings. The full QA runbook lives in the task's copilot-cli-adapter.qa.md.

🤖 Generated with Claude Code