feat(loop-detection): add escape hatches for legitimate batch tool calls

[mvanhorn]

Closes #2511

Scientific workflows (RNA-seq differential expression, parameter sweeps, multi-file analysis) legitimately call the same tool many times in succession. The current LoopDetectionMiddleware hash + frequency layers fire on these workflows and force-stop the run with no way to opt out.

Change

Three orthogonal escape hatches, plus a clearer error message:

Mechanism	Scope	Set by
batch_friendly_tools={\"bash\", ...}	Per-tool, all calls	Middleware constructor
_loop_detection_skip: true in tool args	Single call	Tool author
loop_detection_disabled: true in runtime context	Whole run	CLI flag / agent config

Calls excluded by any mechanism are not appended to _history or counted in _tool_freq — they are effectively invisible to detection. Other calls in the same response keep being tracked normally.

When the middleware does fire a hard stop, the message now mentions all three escape hatches so the user has a discoverable path out:

[FORCED STOP] Repeated tool calls exceeded the safety limit. Producing
final answer with results collected so far. If this is an intentional
batch workflow, use runtime context loop_detection_disabled, configure
batch_friendly_tools, or set _loop_detection_skip=true in a tool call's args.

Tests

backend/tests/test_loop_detection_middleware.py adds a TestEscapeHatches class with 6 cases:

• test_batch_friendly_tools_skip_hash_layer — 6 identical bash calls with bash batch-friendly: no hard stop.
• test_batch_friendly_tools_skip_frequency_layer — 60 varying-args bash calls: no warning, no hard stop.
• test_loop_detection_skip_marker_in_args — 6 identical calls with _loop_detection_skip=true: no hard stop.
• test_runtime_loop_detection_disabled — 6 identical calls with runtime override: no hard stop, no state recorded.
• test_hard_stop_message_includes_escape_hatch_hint — default config triggers hard stop, asserts the hint text contains all three mechanism names.
• test_no_regression_when_no_escape_hatches_configured — default behavior unchanged.

============================== 52 passed in 0.15s ==============================

Verification

• cd backend && PYTHONPATH=. uv run pytest tests/test_loop_detection_middleware.py -v — 52 passed
• cd backend && uvx ruff check packages/harness/deerflow/agents/middlewares/loop_detection_middleware.py tests/test_loop_detection_middleware.py — clean
• cd backend && uvx ruff format --check ... — clean
• backend/CLAUDE.md middleware-chain bullet updated to mention the new escape hatches.

Coordination

Coordinates with #2569's alternating-pattern detector (separate PR): that one tightens detection for cases the current logic misses, this one loosens it for cases the current logic over-fires on. Together they make detection more accurate without regressing existing behavior.

🤖 Built with assistance from Claude Code and Codex.