Problem
Scope-level Prompt customization currently exposes Prompt keys for Memory, Experience, Skill, and Handoff, but not Topic Memory. The Topic Memory pipeline has six generation stages whose instructions remain code-defined and cannot be overridden through the Prompt API:
- probe
- global
- planner
- evolve
- temporary
- reconcile
This makes it impossible to apply a scope-level language policy consistently. For example, a Chinese Source can produce English Topic titles/summaries depending on the generation model, while other configurable artifacts can receive a custom instruction to follow the Source language.
Expected behavior
Expose Topic Memory prompt definitions through the same scope-level custom Prompt mechanism, or provide an equivalent supported configuration. The effective instruction should be able to include a rule such as:
The output language should follow the dominant language of the source content. If the source is primarily Chinese, write the natural-language output in Chinese; preserve code, identifiers, paths, commands, and proper nouns as supplied.
The six stages should retain their existing evidence, schema, citation, and safety constraints when customized.
Acceptance criteria
- Topic Memory prompt keys are discoverable through the Prompt configuration API.
- A scope can create/replace custom instructions without replacing the built-in safety/schema rules unintentionally.
- The effective prompt for every Topic Memory stage is observable and versioned.
- A Chinese Source produces Chinese Topic title, summary, and detail in a new processing run.
- Existing auto-mode behavior remains unchanged when no custom Prompt is configured.
Problem
Scope-level Prompt customization currently exposes Prompt keys for Memory, Experience, Skill, and Handoff, but not Topic Memory. The Topic Memory pipeline has six generation stages whose instructions remain code-defined and cannot be overridden through the Prompt API:
This makes it impossible to apply a scope-level language policy consistently. For example, a Chinese Source can produce English Topic titles/summaries depending on the generation model, while other configurable artifacts can receive a custom instruction to follow the Source language.
Expected behavior
Expose Topic Memory prompt definitions through the same scope-level custom Prompt mechanism, or provide an equivalent supported configuration. The effective instruction should be able to include a rule such as:
The six stages should retain their existing evidence, schema, citation, and safety constraints when customized.
Acceptance criteria