Bug Report: Rollout Log Size Explosion from Unbounded Base64 Image Data #32128

Bug Report: Rollout Log Size Explosion from Unbounded Base64 Image Data

Environment

• Codex CLI Version: 0.142.5 (desktop build 26.707.3748.0)
• Platform: Windows 11
• Storage: ~/.codex/sessions/ and ~/.codex/logs_2.sqlite

Problem

Codex's rollout JSONL logging system stores full base64-encoded image data directly in function_call_output and message payload fields. This causes exponential log growth:

• A single session with image operations can grow to 15-16 MB (normally a session should be < 1 MB)
• 94.6% of the inflated session size is base64 image data embedded in data:image/png;base64,... strings
• The largest single JSON line observed: 4.8 MB (one base64 image)
• Cumulative base64 across all sessions: 11.3 MB in image data alone

Root Cause Analysis

1. function_call_output payloads store complete base64 image strings from image_url.url fields (format: data:image/png;base64,iVBORw0...)
2. These are written verbatim to JSONL files without any size limit or truncation
3. session_meta duplicates base_instructions text (~20 KB) in every session file
4. SQLite WAL files are not periodically checkpointed, causing logs_2.sqlite-wal and state_5.sqlite-wal to grow

Impact

Evidence

• Largest session: 2118 lines, 515 function_call_output entries, 2 base64 occurrences totaling 4.67 MB
• All sessions combined: 31,770 response_items, 12,947 function_call_outputs, 11.3 MB total base64
• Base64 accounts for 94.6% of function output size in affected sessions

Proposed Solutions

P0: Filter/Truncate base64 in function_call_output

Detect data:image/*;base64,<string> patterns and replace with [BASE64_IMAGE_DATA_REMOVED]

P1: Add configurable log size limits

[logging]
max_response_item_size = 10000
max_base64_per_output = 1024

P2: Auto-checkpoint SQLite WAL on idle/close

Execute PRAGMA wal_checkpoint(PASSIVE) when a session ends or app goes idle

P3: Deduplicate base_instructions in session_meta

Store in a shared file, reference by hash/path instead of duplicating ~20 KB per session

Severity

High - Rapid disk consumption (hundreds of MB in days of normal use), can fill SSDs on limited-storage devices.

Reported by: user diagnosing their own Codex installation