Notify on background-workflow completion (banner + agent summary)

[agentforce314]

Problem

You ran /deep-research end-to-end — all 15 agents finished, Synthesize completed — but the REPL never told you. The only way to see it was done was to keep polling /workflows, and even then there was no pointer to the results.

Root cause: on completion, complete_workflow_task → enqueue_workflow_notification dutifully builds a <task-notification> (result + output location + token/agent stats) and pushes it onto the message_queue_manager queue… and nothing in the REPL ever drains that queue. drain_pending_notifications had zero production callers — its own docstring admits the consumer side ("WI-3.3") was deferred. The notification was built and dropped on the floor.

Fix

Wire the missing consumer, matching the Claude Code reference you showed (banner, then the agent reads + summarizes):

✔ deep-research completed · 15 agents · 847.7k tok · 6m 40s
  run journal → ~/.clawcodex/workflows/wf_1d92c718.json

● The research is done. Micron's HBM advantage comes down to its
  1-beta DRAM node (best perf-per-watt)… [agent summarizes the result]

• src/repl/task_notifications.py (new, pure + fully unit-tested): deterministic completion banner from the terminal task state (✔/✗/⊘ · name · agents · tokens · duration · run journal), an envelope-only fallback, and build_notification_turn() that wraps drained envelopes in a guiding <system-reminder> for the agent.
• REPL glue (core.py): _deliver_pending_task_notifications() drains the queue at the turn boundary, prints each banner, then hands the envelopes to the agent as one turn so it summarizes and points to where the output lives. A daemon watcher wakes the idle ❯ prompt the moment a run finishes — so it surfaces without a keystroke — and is careful to never clobber a half-typed line or a permission dialog (only wakes when the prompt is running with an empty buffer).

The cross-thread wake (app.exit via loop.call_soon_threadsafe) was verified against prompt_toolkit 3.0.52 under a PTY before building on it.

Tests — 15, all green

Pure helpers; the REPL delivery method (banner + exactly one agent turn; no-op on empty queue); the wake guard (idle+empty exits; half-typed / not-running don't); the watcher (wakes when idle+pending, silent when not at prompt); and an end-to-end test driving the real production launcher (run_workflow_task) → enqueue → REPL delivery with a fake runner.

Note: 7 pre-existing tests/tui/test_workspace_search_c5.py failures are a ripgrep/env issue on this machine (they fail on a clean tree too) — unrelated to this change.

Scope note

Banner + agent-summary is the behavior you picked. p pause / s save and other surfaces are untouched. This consumer also covers background agent task notifications (same queue), not just workflows.

🤖 Generated with Claude Code