As a Nablarch developer, I want to use nabledge in long sessions without hitting context limits so that I can work sustainably

[kiyotis]

Situation

When nabledge executes any workflow (QA, code analysis, keyword search, semantic search), search results and knowledge file contents are loaded entirely within the main agent context and remain there for the rest of the session. As knowledge files grow, each workflow execution adds a significant and permanent token load to the main context.

This behavior affects all workflows across all five versions (v6 / v5 / v1.4 / v1.3 / v1.2) and both CC (Claude Code) and GHC (GitHub Copilot) modes.

Pain

Nablarch developers running multiple nabledge queries in one session may hit context limits mid-session, breaking their workflow. All workflow processing except user interaction (question intake and answer delivery) should run in a separate context — but currently it does not.

Benefit

• Developers can run multiple nabledge queries in a single session without hitting context limits
• Developers can use nabledge in long coding sessions alongside other tools without the context being consumed by prior nabledge queries

Success Criteria

• Investigate: Determine whether GHC supports running workflow steps in a separate context, and identify the mechanism or fallback if not
• QA workflow (v6, CC and GHC): executing a query produces measurably less main context growth than before the change
• Code analysis workflow (v6, CC and GHC): executing an analysis produces measurably less main context growth than before the change
• Keyword search workflow (v6, CC and GHC): executing a search produces measurably less main context growth than before the change
• Semantic search workflow (v6, CC and GHC): executing a search produces measurably less main context growth than before the change
• The same changes are applied to all remaining versions (v5 / v1.4 / v1.3 / v1.2)
• Benchmark scores for all five versions do not decrease vs. baseline after the change

[kiyotis]

Related: nablarch/nabledge#16