Output chunks

[morganchristiansson]

Large decompiles are saved to file which need to be retrieved by additional tool calls.

The thought is nice to reduce tokens, and worth getting right.
But mostly it just results in repeat tool calls to retrieve the decompile, and having to make extra tool calls is worse for tokens (due to cached input tokens on repeat requests).
Also repeat tool call fails when requesting more than 14k chars often requiring another extra tool call.

Not sure this is useful in it's current state, should at least be configurable limit for initial and follow up chunk retrieval calls.

[themixednuts]

Yea, its a hard balance. My personal use case it helps because the binaries I'm working with produce large functions where chunking does help even with the multiple tool calls, but I know its not where I want it to be either.

hopefully this is null once the java sdk supports the 2025-11-25 spec for tasks

[morganchristiansson]

Ah.
But I would just like to increase limit and not have chunking most of the time.
Or increase max_chars 14,0000 max limit when retrieving chunks.

Maybe if I was still in broad exploration phase I would think differently. Using OOAnalyzer to recover classes/vtables/members/fields and AI with ghidra MCP to rename, retype, and generally improve has made a huge difference and most of the time I just look at single / few decompilations.

OTOH repeat requests do not have large cached input token penalty and i mostly decompile early in context window.
I can think about this a bit more.

[morganchristiansson]

OTOH the help / tool discovery can output like 5 pages of documentation & examples and is clearly not limited by chunks or max_chars. Maybe this is just one-off at start of session tho?
One of the things I appreciate with this project is that it doesn't have 200+ tools and seems more organized. Also doesn't need python proxy.

Anyway I work around that by having most commands in AGENTS.md, so most sessions it doesn't use the help system. https://github.com/morganchristiansson/mxo-launcher/blob/c439456dd1caf36ef12b3a627111b9db001398b1/code/matrix_launcher/AGENTS.md#ghidra-usage

[themixednuts]

Fixed on main by b48e9ab (fix: make MCP output limits configurable). This adds Ghidra options for the inline response budget, default read_tool_output chunk size, and maximum read_tool_output max_chars, updates read_tool_output schemas from those options, and keeps null next_cursor fields out of flattened output. Validated with targeted E2E tests for read output/listing/mutation paths plus local ci.