fix(nous): ignore non-finite rate limit resets

[lin-hongkuan]

Summary

• reject non-finite Nous rate-limit reset values from headers and error context
• avoid persisting Infinity breaker state and clean up non-finite state files when read
• ignore non-finite bucket values when deciding whether a Nous 429 is a genuine account limit

Why

Malformed or non-finite rate-limit headers such as retry-after: inf could be recorded as an infinite cross-session cooldown. The stored Infinity state then made nous_rate_limit_remaining() return inf, and format_remaining() or bucket parsing could raise OverflowError instead of falling back safely.

Tests

• python -m pytest tests/agent/test_nous_rate_guard.py -q
• python -m pytest tests/agent/test_nous_rate_guard.py tests/agent/test_rate_limit_tracker.py tests/run_agent/test_nous_429_fallback_reentry.py -q
• python -m py_compile agent/nous_rate_guard.py
• python scripts/check-windows-footguns.py --all

[tonydwb]

Hermes Agent Review

Verdict: Approved — Clean, well-scoped change with appropriate tests.

Reviewed as part of batch review session 2026-06-24d.

Reviewed by Hermes Agent

[teknium1]

Thanks for addressing the malformed-header cooldown path; the current-main premise is real (agent/nous_rate_guard.py:62-64 accepts positive infinity, which reaches persistence at lines 111-114).

Problems

• agent/nous_rate_guard.py:160 converts persisted reset_at with float(...), but the enclosing handler does not catch ValueError. Valid JSON containing a nonnumeric string will now escape instead of failing closed. Please catch it and clean up/ignore that invalid state; add a regression test.
• The same header class remains in agent/rate_limit_tracker.py:78-89: _safe_int("inf") raises OverflowError, while _safe_float("inf") can feed _fmt_seconds() at line 148. Please reject non-finite values there too and add parser/display coverage.

Suggested changes

• Extend the guard-state test matrix with a nonnumeric persisted reset_at.
• Add non-finite rate-limit-tracker tests for both integer and reset fields.

Automated hermes-sweeper review.

[teknium1]

float() raises ValueError for a valid JSON value such as "reset_at": "bad", but this function's except clause still excludes ValueError. Please fail closed here (and remove or ignore the invalid state) so a corrupted-but-parseable state file cannot escape this guard.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs touch the Nous rate-limit guard but address distinct failure classes: #18782 makes text I/O explicitly UTF-8 across several modules, while #51951 rejects non-finite rate-limit values that could create an infinite cooldown or formatting failures.

Related pull requests

• #18782 [closed] related — (+197/-12) — merged via #71078: This broader encoding fix adds explicit UTF-8 reads and writes across Copilot, model metadata, the Nous rate guard, cron, and delivery paths; it remains relevant because its commits were cherry-picked into the class-wide encoding cleanup with authorship preserved.
• #51951 related — (+118/-9) — keep open pending fixes: The diff correctly rejects non-finite reset and bucket values, removes persisted infinite cooldowns, and guards duration formatting. The contributor keep_open review identifies two remaining gaps: a nonnumeric persisted reset_at can still raise ValueError, and the equivalent non-finite parsing/display paths in agent/rate_limit_tracker.py are not covered.

Suggested consolidation

Merge #51951 after addressing the contributor keep_open review with invalid-string persisted-state cleanup and corresponding rate_limit_tracker.py guards and tests; #18782 is not a duplicate and needs no reopening because its relevant changes were merged through #71078.

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 23 kB of PR diffs, 5 kB of issue/PR text, 4 kB of discussion (5 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.