Idea

Allow teams of coding agents (main agent + sub-agents) to share a read-only view of each other's compacted context. Currently each session has its own conversation in the same lcm.db, but sub-agents cannot query other sessions' summaries or understand what their teammates have done.

Use Case: Coding Team Orchestration

A main "chief of staff" agent spawns sub-agents for different tasks (testing, refactoring, documentation). Each sub-agent works independently but would benefit from knowing:

• What files other agents have modified (from their session summaries)
• What decisions were made in other sessions
• What the main agent's current priorities are

Currently this requires the main agent to explicitly relay context to each sub-agent, which wastes tokens and is lossy.

Possible Approaches

A. Cross-session lcm_grep

Extend lcm_grep to search across all conversations in the DB (not just the current session). Sub-agents could query lcm_grep("what has the testing agent done?") to pull summaries from other sessions.

B. Shared context assembly

Allow a sub-agent to include summaries from a designated "parent" session in its assembled context. The parent's high-level condensed summaries (depth 2+) provide team awareness without duplicating raw messages.

C. Separate team coordination DB

Create a lightweight secondary database that stores:

• Per-session activity summaries (auto-generated after each compaction)
• File modification logs
• Decision/outcome records

Sub-agents read from this DB for team context. Writes are append-only from each session's compaction lifecycle.

Considerations

• SQLite concurrent read support via WAL mode makes shared reading feasible
• The existing statelessSessionPatterns config could gate which sessions participate
• The ignoreSessionPatterns config already excludes cron sessions — similar filtering for team DBs
• Read-only access prevents sub-agents from corrupting each other's state
• Team context should use condensed summaries (depth 1+) to minimize token overhead

Related

• perf: cache DB reads during compaction to reduce SQLite lock contention #291 (DB read optimization — concurrent access patterns)
• Leaf compaction trigger destabilizes Anthropic prompt cache on high-traffic conversations #282 (cache-aware compaction — token efficiency)