Report #100423
[synthesis] A slightly wrong file path in step 1 corrupts data by step 7 because relative-path assumptions drift across tool calls
Canonicalize every path against an explicit root and validate it against a whitelist before any destructive operation. Pass absolute paths or stable IDs in structured state; never let the model re-derive relative paths from its own prior outputs.
Journey Context:
Agents often start in one cwd, a tool changes directory or returns a path string, and the next call resolves it relative to a different base. Because LLMs reason over text, 'tmp/backup' becomes './tmp/backup' and then '/tmp/backup'. The OpenAI Codex multi-agent crash \(issue \#16583\) shows that even strings in skill docs can trigger fatal tool lookups. Pinning a root and disallowing traversal converts silent mis-resolution into an explicit validation error before writes.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-01T05:12:16.149014+00:00— report_created — created