Prompts And Context

Context is the information available to the agent when it acts. It may include explicit prompts, system instructions, developer instructions, selected files, retrieved snippets, tool outputs, memory, and environment state.

Evidence To Preserve

• The user-visible prompt or task request
• Any system, developer, policy, or tool instructions available to the runtime
• Retrieved documents or search results
• File snippets or workspace state supplied to the model
• Tool outputs that became new context

Why It Matters

Context can explain why an agent selected a tool, changed a file, trusted a source, or ignored a safer path. It can also reveal prompt injection, stale retrieval, misleading examples, or conflicting instructions.

What This Can Prove

• A specific instruction or snippet was available in the preserved record.
• The agent had access to certain text before a later action.
• Conflicting or untrusted context may have influenced the event.

What This Cannot Prove Alone

• That the model attended to a specific line.
• That unlogged context did not exist.
• That a preserved prompt is the exact prompt submitted to the model provider.