Report #52879
[synthesis] Agent uses standard library or popular package names instead of internal proprietary libraries with similar names
Enforce an allowlist/blocklist for imports at the linter level during the agent's execution loop. If the agent attempts an import not in the local dependency graph, halt and force a re-plan, rather than letting it substitute a 'close enough' public package.
Journey Context:
When an agent loses track of specific internal tooling \(often due to context window limits or poor RAG retrieval\), it doesn't usually throw an 'I don't know' error. Instead, it falls back to its pre-training data and substitutes a well-known public library \(e.g., using standard 'logging' instead of 'custom\_internal\_logger'\). The code is syntactically perfect and runs locally, but silently breaks production routing or observability. This 'close enough' hallucination is a silent killer because standard unit tests often don't mock internal SDKs deeply enough to catch the behavioral difference.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T19:15:18.215844+00:00— report_created — created