feat(cli): wire /tasks + /background on a session-scoped TaskManager

[oratis]

What

Wires the /tasks and /background slash commands (both were 🔄 in docs/BEHAVIOR_PARITY.md) on top of the existing background-task infrastructure (TaskManager + TASK_TOOLS).

The gap: the TaskManager was constructed per runAgent call (agent.ts), so background tasks evaporated after each turn and nothing outside the agent loop (the REPL, slash commands) could see them. This re-scopes the manager to the REPL session so tasks persist across turns and are visible to both the agent and the user.

This is slice 1 (visibility) — the foundation. Slice 2 (move the in-flight turn to the background, Ctrl+B parity) needs REPL concurrency and is scoped separately.

How

@deepcode/core

• TaskManager.setRunner() — lets a host own a long-lived (session-scoped) manager while the agent loop attaches its run-local sub-agent runner each turn. The rich runner resolves named sub-agents + fires SubagentStop. Tasks already started are unaffected (their handle is captured at create() time).
• RunAgentOptions.taskManager — when provided, runAgent calls setRunner(...) on it and exposes it as toolCtx.tasks instead of creating a per-run manager. When absent, the original per-run fallback is kept verbatim, so headless / desktop / library callers are unchanged.

deepcode-cli

• The REPL constructs one TaskManager for the whole session and threads it into every runAgent call (opts.taskManager) and onto the slash-command SessionContext (ctx.tasks). Its baseline runner runs a prompt as a depth-1 sub-agent (clean context, no nested tasks; reads ctx.model/ctx.mode live) — this only matters for /background started before the first turn, after which the agent loop's richer runner takes over.
• /tasks — lists this session's background tasks (id · status · description). /tasks <id> shows a single task's status + output.
• /background <prompt> (alias /bg) — creates a background task that runs <prompt> as a sub-agent via the manager.

docs — flip /background + /tasks to ✅ in BEHAVIOR_PARITY.md; /batch stays 🔄 (batch-of-prompts not wired here — use /background per prompt).

Notes on behavior

• One shared manager means tasks the agent starts (via TaskCreate) and tasks the user starts (via /background) show up in the same /tasks list.
• Background sub-agents run without an approval/askUser callback (matching the existing sub-agent path), so an ask-verdict tool is returned to the sub-agent as a blocked tool result rather than blocking on the REPL prompt — no deadlock / interleaved prompts.
• Each task carries its own AbortController, so it isn't tied to a turn's lifecycle and keeps running while the user is at the prompt (await rl.question yields to the event loop).

Testing

• core: setRunner re-targets the runner for subsequent create() calls.
• cli: new background-commands.test.ts (10 cases) drives /tasks + /background against a stub-backed TaskManager — create / list / <id> output / aliases / usage / unavailable / runner-failure.
• Full workspace green: typecheck (6 pkgs) · lint (0 errors) · format:check · core 643 · cli 155 · desktop 54 · lsp 8.
• Manual end-to-end smoke through the real CommandRegistry confirms the rendered output for empty list, create, /bg, list-with-statuses, /tasks <id>, and usage.

🤖 Generated with Claude Code