Fix empty tool-use response handling

[Bill-Billion]

Summary

• use concrete tool_use blocks as the execution signal across the current Agent Loops
• prevent empty user/tool-result turns in parent, subagent, and teammate runtimes
• sync the trilingual course text, diagrams, generated web data, and regression coverage

Context

The current course still inherited the boundary reported in #224: a compatible provider can return stop_reason="tool_use" without an actual tool call. The loop would append an empty user turn and request the model again. This PR fixes the current S01-S15 course structure rather than changing the frozen agents/s04_subagent.py.

Reported by @miketam in #224.

Verification

• pytest -q tests/test_agent_loop_boundaries.py tests/test_agent_teams_runtime.py tests/test_compaction_tool_pairs.py tests/test_background_tasks.py -> 116 passed, 75 subtests passed
• pytest -q -k "not test_s16_scenario_matches_the_deterministic_runtime" -> 206 passed, 1 deselected, 84 subtests passed
• full suite -> 206 passed, 84 subtests passed, with the pre-existing S16 concurrent progress-order fixture failure also reproducible on main
• Web extraction is idempotent
• 48 modified SVG files pass xmllint and match their Web asset mirrors
• git diff --check passes

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
❌ Deployment failed View logs	learn-claude-code	168fff8	Aug 14 2026, 04:04 PM