Summary

Fixes the mid-session forgetfulness gap described in #2667.

The infrastructure for this feature was almost complete — compressed_buffer table with FTS5, SessionDB methods (archive_compressed_messages, search_compressed_buffer, get_compressed_buffer_count, clear_compressed_buffer), the session_buffer_search tool schema and registration, and the buffer population hook in _compress_context were all already merged. The missing piece was wiring session_buffer_search into the agent loop's special-case dispatch so it receives db and current_session_id.

Root Cause

session_buffer_search was falling through to handle_function_call → registry.dispatch(), which doesn't receive db or current_session_id as kwargs. Without those, every call returned {"success": false, "error": "Session database not available."} even when the DB was fully available.

Changes

run_agent.py — two dispatch sites, both now handle session_buffer_search:

• _execute_tool_calls_sequential (sequential path) — new elif function_name == "session_buffer_search" block, mirroring the session_search pattern
• _invoke_tool (concurrent/_execute_tool_calls_concurrent path) — same block

tests/tools/test_session_buffer_search.py (new, 36 tests):

• TestArchiveCompressedMessages — skips system/empty/pruned, counts correctly
• TestGetCompressedBufferCount — zero init, after archive, session scoping
• TestSearchCompressedBuffer — FTS5 match, no match, session scoping, role filter, limit
• TestClearCompressedBuffer — removes entries, doesn't affect other sessions
• TestSessionBufferSearchTool — no-db error, no-session error, stats mode, search, parent-chain walk, limit cap, result fields
• TestSessionBufferSearchSchema — schema name, description, required fields
• TestResolveRootSessionId — root, child, grandchild, unknown session

How it works end-to-end

1. On every context compression, _compress_context calls db.archive_compressed_messages(root_session_id, messages_being_dropped) — this was already wired
2. Archived messages are indexed in compressed_buffer_fts (FTS5) via a trigger
3. When the agent calls session_buffer_search("Tavily"), the new dispatch path injects db=self._session_db and current_session_id=self.session_id
4. The tool resolves the root session ID (walking the parent_session_id chain) and runs an FTS5 query scoped to that root — finding content from any compression round of the current conversation
5. Returns FTS5 snippets instantly, zero LLM cost

Closes #2667