[0xacee]

Why this matters: A local-model agent with a long JSONL conversation causes every authenticated session-status poll to load the entire history into memory. A 12,000-turn fixture (~13 MB) allocated 13,724,595 bytes in the existing HTTP fallback. Read the file incrementally while preserving turn counts, tool counts, history characters, malformed-line handling, file size and recommendations.

Evidence and validation:

• Regression exercises GET /api/session-status with database fallback and a real local history, including a measured peak-memory budget of 3 MB after warming HTTP machinery.
• Before: 1 regression failed, 1 counting control passed. After: full Token Spy suite 24 passed, 1 skipped. Changed-file Ruff E/F/W checks and git diff --check pass.
• Separate mixed-record fixture verifies blank/malformed lines and string-encoded messages.
• Overlap check: open/closed title and changed-file searches found fix(token-spy): narrow broad excepts in local session-status helpers #2691 (narrow exception handling in these helpers); its patch does not stream the history. This uses its intended I/O exception boundary for the stream. No remote SSH behavior or reset selection is changed.

AI assistance: implemented and validated with Codex. Local HTTP tests use synthetic session files and no live agent. Review required before merge: independent review of session monitoring, which feeds automatic reset decisions.

Combined validation: c08b6695ff14 on public-beta base 9fc9f61. Token Spy: 64 passed, 1 skipped, including the public HTTP memory-budget regression.

Merge guidance: tested Token Spy main.py order is #4561, #4569, #4583, #4634, #4640. The #4583/#4634 poller conflict must retain await asyncio.to_thread for status and reset, include_session_id=True on assessment, session_id=status["_reset_session_id"] on deletion, and cooldown updates only after action == "killed".