[xyi74976-del]

Description

Explicitly configuring LoopDetector(max_repeats=12) never detects a repeated tool call: history is capped at ten entries, so the requested window cannot be reached. Retain at least the configured window while keeping history bounded and preserving the ten-entry floor. The workflow's default threshold is unchanged.

Related Issues

No linked issue. This is separate from the empty status-probe history pollution fixed in #177.

Changes Made

• Honor explicitly configured repeat thresholds above ten.
• Test thresholds 5, 10, 12 and 20, status probes, and sequences interrupted by another tool.
• Add English and Chinese changelog entries.

Validation

Python 3.12 with the official CI lock and editable package:

• Focused tests: 10 passed; the same tests against the original source: 4 failed, 6 passed.
• Full suite: 3139 passed, 12 skipped, 1 failed. Remaining failure: tests/test_tui_input_safety.py::test_ctrl_c_clears_a_draft_before_it_interrupts. It also fails after restoring all four changed files to the clean upstream baseline.
• Added alternating-call histories to verify bounded history and the latest-five summary at default and larger thresholds. A deliberate test-only clipping-removal mutation causes both new cases to fail; the published implementation passes them.
• Native pre-commit run --all-files --show-diff-on-failure: all six automatic hooks passed; a Go module mirror was needed to install actionlint.
• uv pip check: installed dependencies compatible.
• Desktop on Node 24.21.0: 435 tests passed, ESLint, TypeScript, protocol check and Web build passed. An initial Node 26 run failed on undefined localStorage; no frontend changes were made.

Windows/native Desktop and browser end-to-end checks were not run. No live model calls were needed.

Checklist

• Changes tested locally (full-suite baseline failure disclosed above)
• Code reviewed (independent read-only AI review)
• Documentation updated
• Unit tests added

Additional Notes

Implementation and tests are AI-assisted. The independent review found no blocking code issues; it does not represent human review or upstream CI approval.