On October 8, 2026, Jim Clausing published two new Python scripts designed to execute an AI agent forensic review by extracting and reconstructing activity logs from AI coding agents as detailed on isc.sans.edu. As engineering teams deploy code assistants across local developer stations, incident responders frequently encounter proprietary state databases that obscure prompt histories, reasoning traces, and API transactions. These utilities bridge that visibility gap by parsing raw local storage into structured review formats without requiring live execution runtimes.
Tracing OpenCode SQLite Artifacts
The opencode-chat-replay.py script targets the default SQLite database path located at ~/.local/share/opencode/opencode.db. Modern developer tools often fragment message content across multiple relational tables, but this parser handles both legacy separate message and part tables alongside newer consolidated session_message storage schemas. When consolidated records exist for a given session, the script prioritizes them to reconstruct full conversation flows.
Execution requires Python 3.10 or higher with zero external dependencies beyond the standard library. Operators can inventory available sessions without rendering full dialogs by executing the utility with the list flag. Target selection supports exact session IDs, unique ID prefixes, or human-readable slugs. If an operator provides an ambiguous session ID prefix, the script aborts immediately rather than guessing the intended target.
Rendering Transcripts and Structured Exports
By default, the replay tool outputs markdown reports featuring session metadata such as working directories, version numbers, cost metrics, token counts, and turn volumes. To prevent voluminous model reasoning blocks from drowning out human dialog, the script wraps reasoning sequences and tool invocations inside collapsible HTML detail blocks. Investigators can alter this behavior by passing suppression flags for reasoning or tool output.
For automated ingestion pipelines, structured JSON and JSONL exports capture session objects and turn arrays. This capability allows security analysts to reproduce native tool exports on isolated forensic workstations where the original assistant software is never installed. Tool input and output rendering defaults to a two-thousand-character truncation cap to balance legibility against log bloat, though investigators can disable this limit entirely via command-line switches.
Extracting Hermes Evidence and Telemetry
The companion hermes_forensic_extract.py utility addresses a different evidence footprint by gathering comprehensive application logs, model usage metrics, and raw API request dumps. While the opencode utility focuses tightly on conversational transcript reconstruction, the Hermes extraction script pulls wider environmental context from target system images or tarballs.
Both utilities accept time boundaries through start and end parameters formatted in standard date-time syntax, allowing analysts to isolate suspicious windows during an active compromise. Analysts can point the scripts to specific database files or whole home directories recovered from disk images. This capability aligns with defensive engineering efforts that underscore the necessity of auditing automated logic behavior across complex environments.
Operational Considerations for Forensics Teams
Forensic examiners must verify whether target workstations store agent data outside standard paths before running extraction routines. Because developer agents frequently cache tokens, API keys, and sensitive source code snippets within local SQLite states or flat files, analysts handling these evidence stores must enforce strict access controls on exported json and markdown transcripts to prevent secondary credential exposure.
When conducting an AI agent forensic review, investigators often confront encrypted storage constraints or missing state records caused by aggressive log rotation policies within the host operating system. To mitigate these gaps, engineers should establish baseline file integrity monitoring on developer workstations to capture SQLite write operations before local database compaction routines execute. Furthermore, organizations deploying local coding agents must update their standard operating procedures to include memory dumps and temporary file sweeps during initial triage stages, ensuring ephemeral context window data is preserved alongside persistent database artifacts for subsequent analysis.
Related coverage
- cPanel root vulnerability exposes shared-hosting isolation flaws
- Malicious npm Package indexed-btree Executes Loader at Runtime After npm Lifecycle Script Block
- Vercel Patches Critical Next.js ImageResponse RCE Flaw (CVE-2026-94545)