Codex CLI: oversized tool outputs are ingested into context and can drain 5h quota disproportionately

## Summary
Codex CLI can consume extremely large tool outputs (for example broad rg -n) and feed them directly into model context, causing disproportionate 5-hour quota burn for otherwise simple tasks.

This feels like a harness inefficiency: tool output ingestion is effectively unbounded, and quota impact is amplified across subsequent turns as context grows.

Environment

• Codex CLI: 0.118.0
• OS: Windows 11
• Model: gpt-5.3-codex
• Reasoning effort: medium

Reproduction

1. Start a normal coding session in a large repository.
2. Run a broad search command through the CLI tool path, for example:

• rg -n --hidden -S "IP Scan|scan ip|scanIP|IpScan|IPScan|Aux|Unknown|IV"

1. Continue normal follow-up turns in the same session.

Observed behavior

• A single tool output can be massive and appears to be forwarded directly into context.
• In my logs, one function_call_output shows:
• Original token count: 948372
• Total output lines: 23781
• Same day had additional very large outputs (300k+ tokens, 100k+ tokens).
• After this, later turns become expensive due to bloated context, and the 5-hour usage limit is exhausted much faster than expected.

Expected behavior

• Tool output ingestion should be token-budgeted and safe by default.
• Very large tool outputs should be summarized/truncated before entering model context.
• CLI should warn and require opt-in when predicted output is above a threshold.
• Quota UX should clearly attribute high burn to oversized tool outputs.

Additional telemetry concerns

I also observed occasional inconsistent quota telemetry (for example used_percent = 0 with reset time near Unix epoch in some events), which may be a separate bug, but worsens user trust when quota drains quickly.

Suggested fixes

• Hard cap on tool-output-to-context tokens.
• Automatic extractive summarization for oversized outputs.
• Safer defaults for broad commands (rg, git diff, large file dumps).
• Per-turn token budget warning before executing high-risk commands.
• More explicit meter diagnostics in UI/logs.