knowledge-maintain

Knowledge Maintain

Apply targeted fixes to Basic Memory notes — missing sections, relations, structural tidies, archival — with the user confirming changes inline. This is the scoped, interactive sibling of the knowledge-maintainer agent: it owns the common case (repair a few named notes, where being in the main session means the user sees each edit) and delegates heavy, autonomous remediation to the agent.

Write discipline (non-negotiable):

• edit_note with find_replace ONLY. Never append with section= — it appends to end-of-file, not end-of-section, the documented BM footgun.
• write_note and delete_note are intentionally excluded. Create new notes via /intel; archive via move_note to archive/.
• Read the live note before every edit (verify-before-fix). A flagged issue may be a parse artifact, or the note may already be correct.
• One edit_note per note per message — issue same-note edits sequentially and re-read before re-anchoring after any lands (concurrent same-note edits are a precautionary hazard; see intel's note-write-mechanics.md single-writer rule).
• schema_validate after every edit. If its output appears to repeat the same value, trust the note body (read_note JSON) as the source of truth rather than inferring a duplication from the validator response alone.

1. Decide venue (scoped inline vs delegate)

Classify the request first:

Delegate to the agent when the work is heavy or autonomous — any of:

• "Fix the whole audit", "remediate the graph", "fix everything".
• The fix requires creating notes or running research: "research and document missing packages", anything that would spawn /intel.
• Brew-note refresh batches, graph-wide orphan linking, or any sweep across a whole ecosystem/type.

To delegate, launch the write agent and stop: call Agent(subagent_type="knowledge-maintainer", description="Graph remediation from audit", prompt="<the user's original request, verbatim; include the gardener report if one is in context>") — i.e. pass a description of "Graph remediation from audit" and a prompt that is the user's original request verbatim, including the gardener report if one is already in context.