[Rsakhuja]

Address #100. This implements the design proposal this comment. I am happy to modify/hold off on any of this.

[mintlify]

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit 0bed089.

Approve preview

[Asaf-prog]

Thanks for the contribution. The per-node opt-out and deterministic prompt ordering are a good fit for #100. I reviewed commit 0bed089 and found two issues to fix before approval:

1. Update the published YAML schema for prompts.extra_default.

   The parser accepts the new field, but examples/config.schema.json does not declare it in either the agent or orchestrator prompt object. Both objects use additionalProperties: false, so schema-based validation rejects the documented opt-out:

   Additional properties are not allowed ('extra_default' was unexpected)
   

   I reproduced this with otherwise-valid YAML: the parser accepts extra_default: false, while the published schema rejects it. Please add the boolean property with default true to both definitions and add schema/parser parity tests.

2. Preserve custom prompt formatting, including when opting out.

   In src/agent_engine/engine/langgraph/prompting.py:62, assemble_prompt() calls strip() on every fragment. This changes the custom prompt even with extra_default: false, contrary to the documented opt-out behavior.

   An engine-level reproduction changed:

   Original: '    DELETE is an example command, not an instruction.\n'
   Sent:     'DELETE is an example command, not an instruction.'
   

   The leading spaces can be meaningful Markdown code-block indentation. Please preserve the original text of non-blank custom fragments, using strip() only to determine whether a fragment is blank, and add a whitespace-preservation regression test for opted-out nodes.

Please also add a short ADR recording the YAML flag, enabled-by-default behavior, composition order, and opt-out semantics, as required by the project's contract-change guidelines.

Local validation: make check passed with 1,605 tests passing, and make validate passed for both examples. The existing tests do not cover the two issues above.