Feature request: programmatic orchestration API for external harnesses

Context

Evaluating jcode as a backend for an external multi-agent orchestrator (managing N concurrent coding sessions across projects, with session lifecycle, telemetry, and cross-harness resume).

The cross-harness session resume + multi-provider OAuth surface is exactly what an orchestrator wants. But to drive jcode from outside, a stable programmatic API is needed beyond the interactive TUI.

Ask

Document (or expose, if not yet present) a stable non-interactive surface covering:

1. Spawn / attach / detach — start a session headlessly with a provider profile + working directory + initial prompt, return a session ID, allow re-attach later.
2. Session inventory — list active + resumable sessions with provider, model, cwd, last-active timestamp, token/cache state.
3. Resume by ID — including the cross-harness resume already advertised (codex / claude-code / opencode / pi). What's the CLI / JSON contract?
4. Streaming I/O — stdout/stderr/events as JSON lines or a websocket, so an orchestrator can render transcripts without scraping the TUI.
5. Cache-cold signal — the warning surfaced in the UI exposed as a structured event, so external schedulers can avoid waking sessions past the 5-min Anthropic TTL.
6. MCP passthrough — confirm whether ~/.jcode/mcp.json servers are available to spawned sessions in headless mode.

Why it matters

Without these, integrators have to shell-scrape, which defeats the perf wins. With them, jcode becomes a viable engine for orchestrators (Octogent-style dashboards, CI agents, IDE plugins) rather than only a human-facing TUI.

Happy to test against a draft spec.

Start by reviewing jcode's existing TUI and any non-interactive CLI surface, then map the requested session lifecycle, inventory, resume, streaming, cache-cold, and MCP behavior. Clarify whether the work is documentation or implementation, define the CLI/JSON contract and structured events, and verify how ~/.jcode/mcp.json behaves for headless sessions.