Summary

Fixes #25516 — GPT pool entries crash with TypeError on rehydration.

Root Cause

Pool state files serialise last_status_at as an ISO-8601 string (via to_dict() / JSON round-trip). When the entry is reloaded from disk and _exhausted_until() is called, it performs

return entry.last_status_at + _exhausted_ttl(entry.last_error_code)

which raises TypeError: can only concatenate str (not "int") to str — crashing every outbound request to a GPT-family provider until the pool state file is manually deleted.

Fix

Apply the existing _parse_absolute_timestamp() helper to the last_status_at field inside PooledCredential.from_dict(). This helper already handles ISO-8601 strings, epoch-millisecond integers, and float-epoch values and is used for all other timestamp fields in this module. Adding it here keeps last_status_at a float at all times, matching its declared type annotation and the expectations of _exhausted_until().

Changes

• agent/credential_pool.py — 5-line guard in from_dict() after the dataclass is assembled; normalises last_status_at through _parse_absolute_timestamp() when the field is present and non-None.
• tests/agent/test_credential_pool.py — two regression tests:
  • ISO-8601 string round-trip no longer raises TypeError; _exhausted_until() returns a float
  • epoch-millisecond integer is correctly divided to epoch-seconds

Before / After

# Before — rehydration from pool state file with ISO timestamp
entry.last_status_at  → "2026-05-11T08:23:20.891066+00:00"
_exhausted_until(entry) → TypeError: can only concatenate str (not "int") to str

# After
entry.last_status_at  → 1747050200.891066  (float)
_exhausted_until(entry) → 1747053800.891066  (float, valid future deadline)

Testing

python3 -m pytest tests/agent/test_credential_pool.py::test_from_dict_normalises_iso_last_status_at tests/agent/test_credential_pool.py::test_from_dict_normalises_epoch_ms_last_status_at -v
# 2 passed in 1.93s