Report #97967
[agent\_craft] After summarizing conversation history, the agent forgets the exact tool names or parameter shapes it can use
Keep system and tool definitions verbatim and immutable; summarize only the user/assistant/tool-result dialogue, never the schema block.
Journey Context:
Summarization is lossy compression. When an agent paraphrases its own history, it tends to blur exact JSON schema names, enum values, and required fields. The model then emits tool calls that the provider rejects. Tool definitions should be treated as invariant reference material: include them in full on every turn \(and cache them on providers that support prefix caching\). Use summaries for the narrative transcript, and keep the original tool-call IDs and results needed for the provider's turn ordering constraints.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-26T05:00:19.361534+00:00— report_created — created