Final summary: - I verified your recollection: the condenser does contain prompt-level structure, but it does not enforce true structured output. - Specifically, the condenser prompt asks for labeled sections like

USER_CONTEXT

,

TASK_TRACKING

,

COMPLETED

, etc., and

LLMSummarizingCondenser

sends that prompt to the LLM and stores the result as plain text. - It does not use

response_format

, JSON Schema, a Pydantic output model, validation, or a repair/retry loop. - So my conclusion is that a condenser-style approach is not sufficient to solve issue #2566 if the goal is reliable typed output. - I think a better solution would be to build on the SDK’s existing tool schema / Pydantic validation path, likely via a structured finish/final-result tool or

conversation.run(output=MyModel)

-style API. - Per your instruction, I responded directly on the GitHub issue, introducing myself as OpenHands-GPT-5.4 and sharing these findings and opinions. - I did not create a PR or modify code.

Summary of my work: - I verified that the current SDK condenser does not enforce true structured output. It only uses a prompt that encourages a structured-looking summary, then stores the result as plain text. - I posted that analysis directly on issue #2566 as requested, introducing myself as OpenHands-GPT-5.4 and sharing my recommendation. - I then cloned the OpenHands repo from the same org and inspected the legacy V0 condenser. - In V0, there really was a more structured approach: - it defined a

StateSummary

Pydantic model - exposed it as a tool/function schema (

create_state_summary

) - forced the LLM to answer via that tool call - parsed the returned JSON arguments - validated them with

StateSummary.model_validate(...)

- However, even V0 did not keep the structured object all the way through. After validating it, it converted it back to plain text and stored that text summary in memory. ### Direct answer Yes — your memory maps much more closely to legacy V0 than to the current SDK condenser. ### Status - [x] Checked the current SDK condenser - [x] Posted findings on the GitHub issue - [x] Cloned the OpenHands repo - [x] Inspected the legacy V0 condenser - [x] Compared how V0 handled structured output vs the SDK ### Conciseness - No code changes were made. - No PR was created. - No extraneous repository modifications were introduced.