fix(query): stop tool-failure-loop guard from conflating distinct Bash failures

[yibocai-airbnb]

Summary

_normalize_error_category's generic fallback (no named pattern matched — InputValidationError, PermissionError, etc.) keyed off the head of the tool result: text.lower()[:MAX_FALLBACK_CATEGORY_LENGTH] (120 chars).

Bash tool results are assembled as stdout + stderr + exit-code sentence, in that order. Any script that prints even a short startup banner or progress output before crashing has its actual differentiator — a traceback's exception line, a compiler's final error — pushed past that 120-char window. So a script that prints an identical banner on every run collapses every distinct bug behind it into one fallback category.

Observed live, not hypothetical: benchmarking a local model on an agentic JAX/MNIST task, three genuinely different bugs behind an identical stdout banner were categorized as one recurring signature and tripped the loop guard after attempt 3 — even though each attempt had actually fixed the prior bug and hit a new one. The guard punished the shape of the script's own output, not whether the model was actually looping — the same category of defect #777 fixed elsewhere in this same guard.

Fix

Prefer the tail of the (exit-code-stripped) text, and specifically the tail of a matched Traceback (most recent call last): block when present, since that isolates the exception line from any stderr preamble.

This only changes behavior when the text exceeds the 120-char window — below that, [:120] and [-120:] are identical, so short results are untouched. A genuinely-recurring failure (identical traceback) still produces an identical tail, so real-loop detection is not weakened; see the fuzz test below for that property proven at scale rather than asserted.

Tests

• 3 unit tests: the core regression case (distinct tracebacks behind a shared banner must not collapse), a generic tail-vs-head check, and same-traceback-still-matches.
• 2 differential fuzz tests (2000 seeded trials each), in the same spirit as fix(harness): stop the loop guard and plan mode from killing headless runs #777's own fuzz methodology:
  • Reproduces the bug at scale: the old head-slice collides on >90% of trials whenever the shared banner is >=120 chars; the new scheme collides on 0%.
  • No-regression property: an identical traceback tail is still recognized as the same category regardless of how much random, unrelated stdout noise varies in the head.

Full existing suite (test_tool_failure_loop_guard.py, test_ch06_tools_round4_A.py) still green — 42 + 13 tests, 0 failures.

Test plan

• python3 -m unittest tests.test_tool_failure_loop_guard tests.test_ch06_tools_round4_A -v — 55/55 pass
• Reproduced the original bug against the pre-fix code path (differential fuzz, TestFallbackCategoryFuzz)

🤖 Generated with Claude Code

[agentforce314]

LGTM.