[Reliability] Repeated tool-call errors (path/retry patterns) inflate latency; need stronger fail-fast + dedupe
Summary
Repeated tool-call failures in a single turn can cascade into extra retries/attempts and significantly increase latency.
Observed symptoms (local logs)
Top tool errors:
terminal: 41 errorsread_file: 12 errorssearch_files: 5 errorspatch: 4 errors
Common signature classes:
- path not found / bad path normalization
- repeated "tool returned error" in same turn
Expected behavior
- Error-class aware retry policy (non-retryable/path errors should fail-fast).
- Duplicate-error suppressor within a turn.
- Better remediation hints after first failure (especially path normalization on Windows).
Actual behavior
- Repeated failures can continue in same turn, adding avoidable token and wall-clock cost.
Related existing discussions (possible overlap)
Why file this anyway
Even if some sub-problems are tracked, this is still reproducible in real usage and has high user-cost impact.
Environment
- OS: Windows 11
- Hermes profile: default
[Reliability] Repeated tool-call errors (path/retry patterns) inflate latency; need stronger fail-fast + dedupe
Summary
Repeated tool-call failures in a single turn can cascade into extra retries/attempts and significantly increase latency.
Observed symptoms (local logs)
Top tool errors:
terminal: 41 errorsread_file: 12 errorssearch_files: 5 errorspatch: 4 errorsCommon signature classes:
Expected behavior
Actual behavior
Related existing discussions (possible overlap)
Why file this anyway
Even if some sub-problems are tracked, this is still reproducible in real usage and has high user-cost impact.
Environment