No way to access the resolved system prompt (from CLAUDE.md and similar files) via the Python SDK

[Nithish-R-Nair]

Context

While building agents using the Claude Agent SDK (Python SDK) the behavior of my agents is heavily driven by files like CLAUDE.md, things like tone, formatting rules, and domain constraints all live there.

From the Python SDK layer, I can track what goes in (the prompt) and what comes out (the response). But I have no visibility into what the CLI loads internally — specifically the content from CLAUDE.md and other context files that get merged into the effective system prompt before the model is called.

Question

Is there any way to access this resolved system prompt from the Python SDK? Even a read-only view of what was actually sent to the model would be really useful for logging and debugging purposes.

If this is not currently supported, is it something that could be added? Or is this intentionally hidden due to any specific reason? Would love to understand the reason either way.

SDK Version: claude-agent-sdk>=0.1.64

[Ameen-Alam]

Hi @Nithish-R-Nair — looking at the source, the resolved system prompt is not currently exposed by the SDK, and here is why.

Why it is not available today

The CLI assembles the final system prompt internally (merging CLAUDE.md, user/project settings, memory files, etc.) before making API calls. The control protocol the SDK talks over — specifically the initialize response — returns only available commands and the current output style. The assembled system prompt is never sent back to the SDK layer, so there is nothing for the SDK to expose.

You can verify this by looking at Query.initialize() in _internal/query.py and the _initialization_result stored from the response — it has no system prompt field.

Current workarounds

1. Read CLAUDE.md files directly — the CLI merges these in a predictable order: <cwd>/CLAUDE.md, ~/.claude/CLAUDE.md, and any CLAUDE.md files in parent directories up to the repo root. You can replicate the merge manually if you just need the content for logging.

2. Use get_server_info() on a connected ClaudeSDKClient — it returns the stored _initialization_result, but as noted above, this does not include the system prompt.

What a real fix would look like

The CLI would need to add a new control protocol response field (e.g. resolved_system_prompt) to the initialize response, or add a new control request subtype to fetch it on demand. That is a CLI-side change, not an SDK-side one.

Happy to help scope out what that would look like if the maintainers want to pursue it.