Feature: Human-readable live view of background agent activity

Problem

When background agents run (via Agent tool with run_in_background: true or scheduled tasks), there is no way to see what they are doing in real-time. The raw output file at /tmp/claude-*/tasks/*.output is NDJSON with deeply nested tool calls — completely unreadable for humans.

Users have to either:

• Wait blindly until the agent finishes
• Try to tail -f the output file and parse raw JSON mentally
• Write custom parsers (which is what we did)

Proposed Solution

Add a built-in command to watch agent activity in a human-readable format:

claude agent watch <agent-id>
# or
claude agent list   # show running agents with picker

The output should show:

• Tool calls formatted as: $ git status, Read file.swift, Edit file.swift, Grep /pattern/
• Agent text (thinking/reasoning) in plain text
• Results — brief summary (green for success, red for errors)
• Progress — which file is being read/edited right now

Example output:

Watching agent: Fix Swift Profile (acc994ac...)
────────────────────────────────────────────────
Read SwiftProfileAlgorithms.swift
Read BGTargets.swift
  ✓ 48 lines
Edit SwiftProfileAlgorithms.swift
  guard let bgTargetsData → added error logging
$ xcodebuild -workspace ... build
  ✓ BUILD SUCCEEDED (0 errors)
$ git commit -m "fix: ..."
  ✓ fc2358bca

Workaround

We wrote a Python parser (50 lines) that does this: parses the NDJSON, extracts tool names + inputs, colorizes output. It works but should be a first-class feature.

Context

This is especially important for:

• Scheduled tasks running overnight (user can't interact)
• Multiple parallel agents (need to see which one is stuck)
• Debugging why an agent failed or made wrong decisions
• Trust — seeing what the agent does builds confidence

Related: #43829, #37991, #33629