let-them-talk 3.7.0 → 3.9.0

package/CHANGELOG.md

@@ -1,5 +1,64 @@
 # Changelog
 
+## [3.9.0] - 2026-03-17
+
+### Added — Channels & Split Cooldown
+
+- **`join_channel(name, description?)`** — create or join a channel for sub-team communication
+- **`leave_channel(name)`** — leave a channel (can't leave #general, empty channels auto-delete)
+- **`list_channels()`** — list all channels with members, message counts, membership status
+- **`send_message` channel parameter** — send to specific channel (`channel-{name}-messages.jsonl`)
+- **`listen_group` reads all subscribed channels** — merges messages from general + channel files, sorted by timestamp
+- **Channel validation** — sending to nonexistent channel returns error with hint to create it
+- **Ghost member cleanup** — heartbeat auto-removes dead agents from channel membership
+- **#general auto-created** — `members: ["*"]` (everyone), uses existing messages.jsonl for backward compat
+- **Split cooldown (reply_to-based)** — fast lane (500ms) for addressed agents, slow lane (max 2000, N*1000) for unaddressed, incentivizes threading
+
+### Fixed
+- Task race condition — `update_task` now rejects claiming a task already in_progress by another agent, auto-assigns on claim
+
+## [3.8.0] - 2026-03-16
+
+### Changed — Group Conversation Overhaul
+
+Redesigned from the ground up based on 3-agent collaborative testing and design session.
+
+**Single-write group messages (O(1) instead of O(N)):**
+- `send_message` in group mode now writes ONE message with `to: "__group__"` instead of N copies per agent
+- `broadcast` in group mode also uses single `__group__` write
+- Old O(N) auto-broadcast loop completely removed
+- Result: with 6 agents, a message now creates 1 write instead of 6. A broadcast round that previously created 30 writes now creates 6.
+
+**`addressed_to` field + `should_respond` hints:**
+- `send_message(to="AgentName")` in group mode stores `addressed_to: ["AgentName"]` on the `__group__` message
+- `listen_group` response includes `addressed_to_you: true/false` and `should_respond: true/false` per message
+- Hint-based, not enforced — agents can still respond when they have valuable input
+- No `addressed_to` = everyone should respond (backwards compatible)
+
+**Adaptive cooldown:**
+- Cooldown now scales with team size: `max(500ms, N * 500ms)` where N = alive agent count
+- 2 agents = 1s, 3 agents = 1.5s, 6 agents = 3s, 10 agents = 5s
+- Explicit `group_cooldown` config still respected if set
+
+**Shorter stagger:**
+- Deterministic stagger reduced from 0-3000ms to 500-1500ms
+- Same agent always gets the same delay (hash-based)
+
+**Alive-only garbage collection:**
+- `autoCompact` for `__group__` messages only checks alive agents for consumed tracking
+- Dead agents no longer block message compaction forever
+- Dead agents catch up via `get_compressed_history()` which reads history.jsonl (never compacted)
+
+**Own-message filtering:**
+- Agents no longer see their own `__group__` messages in `listen_group` batches
+- Own messages are auto-consumed on sight
+- Own messages still visible in `context` array for reference
+
+### Added — 3D World: Player Mode & Improvements
+- **Player character** — users can spawn as a controllable character in the 3D Hub
+- **Spectator camera improvements** — refined controls and speed
+- **Environment updates** — campus environment refinements
+
 ## [3.7.0] - 2026-03-16
 
 ### Added — Agent Ecosystem (20 new tools, 52 total)

package/README.md

@@ -86,9 +86,9 @@ Each terminal spawns its own MCP server process. All processes share a `.agent-b
 
 - **3D virtual office** — chibi characters at desks, spectator camera (WASD+mouse), 11 hairstyles, 6 outfits, gestures, furniture, TV dashboard
 - **Managed conversation mode** — structured turn-taking with floor control for 3+ agents, prevents broadcast storms
-- **52 MCP tools** — messaging, tasks, workflows, profiles, workspaces, branching, managed mode, briefing, file locking, decisions, KB, voting, reviews, dependencies, reputation
+- **53 MCP tools** — messaging, tasks, workflows, profiles, workspaces, branching, managed mode, briefing, file locking, decisions, KB, voting, reviews, dependencies, reputation
 - **8-tab dashboard** — 3D Hub (default), messages, tasks, workspaces, workflows, launch, stats, docs
-- **Group conversation mode** — free multi-agent collaboration with auto-broadcast and cooldown
+- **Group conversation mode** — single-write `__group__` messages, adaptive cooldown, `addressed_to` hints, alive-only GC
 - **5 agent templates** — pair, team, review, debate, managed — with ready-to-paste prompts
 - **5 conversation templates** — Code Review, Debug Squad, Feature Dev, Research & Write, Managed Team
 - **Stats & analytics** — per-agent scores, response times, hourly charts, conversation velocity
@@ -175,7 +175,7 @@ The dashboard's default view is a **real-time 3D virtual office** (the "3D Hub")
 
 **Animations:** walk, sit, type, raise hand, sleep (ZZZ), wave, think, point, celebrate, stretch, idle gestures. Agents turn toward speakers during conversations.
 
-## MCP Tools (52)
+## MCP Tools (53)
 
 <details>
 <summary><strong>Messaging (13 tools)</strong></summary>

package/cli.js

@@ -9,7 +9,7 @@ const command = process.argv[2];
 
 function printUsage() {
   console.log(`
-  Let Them Talk — Agent Bridge v3.7.0
+  Let Them Talk — Agent Bridge v3.9.0
   MCP message broker for inter-agent communication
   Supports: Claude Code, Gemini CLI, Codex CLI, Ollama