Skip to content

Add multi-agent support for named agents per channel#19

Open
ZoeFridler wants to merge 20 commits intompociot:mainfrom
ZoeFridler:dev
Open

Add multi-agent support for named agents per channel#19
ZoeFridler wants to merge 20 commits intompociot:mainfrom
ZoeFridler:dev

Conversation

@ZoeFridler
Copy link

@ZoeFridler ZoeFridler commented Feb 18, 2026

Summary

  • Adds named agent support so multiple agents can work in the same Slack channel, each with its own working directory and conversation session
  • Agents are addressed via @name <message> or broadcast to all with @all <message>
  • Agents can communicate with each other via claude-flow MCP memory tools (memory_store, memory_retrieve)

Changes

  • src/types.ts — Added AgentConfig and ParsedCommand interfaces
  • src/agent-manager.ts (new) — Agent CRUD, command parsing, and directory resolution
  • src/claude-handler.ts — Added agent session key/get/create methods
  • src/slack-handler.ts — Added agent command routing, handler methods, extracted executeQuery() helper from the streaming loop for reuse by both regular messages and agent messages

Commands

Command Action
create agent <name> on <path> Creates a named agent with its own working directory
remove agent <name> Removes an agent
list agents Lists all agents in the channel
@<name> <message> Sends message to a specific agent
@all <message> Broadcasts message to all agents concurrently

Test plan

  • Run npm run dev — bot starts without errors
  • Send create agent linux-agent on linux-metrics — agent created with resolved path
  • Send list agents — shows the agent
  • Send @linux-agent what files are in src/ — responds with [linux-agent] prefix
  • Send create agent slack-agent on claude-code-slack-bot — second agent created
  • Send @all describe your project — both agents respond concurrently
  • Send remove agent linux-agent — agent removed
  • Regular (non-agent) messages still work as before

🤖 Generated with Claude Code

zoe-fridler and others added 20 commits February 18, 2026 14:24
@zoe-fridler @claude
Add multi-agent support for named agents per channel
7e4406b 
Allow multiple named agents in the same Slack channel, each with its
own working directory and conversation session. Agents can be addressed
individually via @name or all at once via @ALL broadcast.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Persist agents to disk, bypass permissions for agents, fix MCP server…
6e82bf1 
… path

- Persist agent configs to agents.json so they survive restarts
- Bypass permission prompts for agent sessions (autonomous mode)
- Fix hardcoded permission MCP server path to use relative path
- Upgrade SDK to 1.0.128 for Node 22 compatibility
- Add agent announcement on creation in main conversation
- Add agents.json to .gitignore

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Remove debug logging from agent command parsing
b068f68 
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add agent rules, status, rename, quiet mode, templates, schedules, an…
9d0e4ce 
…d agent-to-agent messaging

New commands: set/clear rules, rename agent, quiet mode on/off, agent status,
schedule/unschedule tasks, list templates, create from template, @agent1 ask @agent2.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add help command that lists all available commands
a43ddad 
Type "help" or "commands" in DMs to see the full command reference.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Persist conversation sessions to survive restarts
11826d1 
Sessions with a sessionId are saved to sessions.json and restored on
startup, so agents remember their conversation history across restarts.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add live status heartbeat with elapsed time and current tool
7cc05af 
Status message now updates every 5s showing elapsed time and what
the agent is doing (Thinking/Working/Responding + tool name), so
you can tell if it's active or stuck.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add per-agent model selection (opus vs sonnet)
19b3a35 
Use `model <agent> sonnet` for fast responses or `model <agent> opus`
for smarter reasoning. Shown in agent list and status.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Default agent model to Opus (smart)
c240141 
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Show help reply in same context as the request
e4621ef 
If asked in the main conversation, reply there. If asked in a
thread, reply in the thread.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add multi-turn agent collaboration mode
2894f2b 
Agents take turns working on a shared task, each responding to
the other until one signals DONE or the 10-turn limit is reached.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add global rules that apply to all agents
55c726e 
Global rules are prepended to every agent's identity prompt, before
their individual rules. Persisted to global-rules.json.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add natural language command parsing fallback
746f856 
When exact command syntax doesn't match, the bot now understands
natural language like "add a global rule to not delete files",
"make lina fast", "show me my agents", "how is lina doing", etc.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Allow addressing agents by name without @ prefix
5e93749 
"lina do that", "tell lina to check the logs", "ask lina about
the auth module" all work now without needing @lina.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add colored agent messages using Slack attachments
6111de6 
Each agent is auto-assigned a unique color from a 10-color palette.
Messages from agents display with a colored sidebar via Slack
attachments API. Colors are applied to greetings, responses, tool
output, collaboration, and agent-to-agent queries. Existing agents
without colors get auto-assigned on load.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Add rich colored formatting for tool output in Slack
6dd943e 
Edit diffs now show removed code with a red sidebar and added code
with a green sidebar (like GitHub). Bash commands use a dark sidebar.
New file creation uses a blue sidebar. Search tools (Glob/Grep) get
a search icon. All via Slack attachments API for visual distinction.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Fix NLP parser deleting agents when message contains "remove"
9ebb1dd 
The natural language parser matched "remove" anywhere in text as a
remove_agent command, so "lina ... remove comments" deleted the agent.

Fixed by:
1. Moving "agent name at start" check to the top of NLP parser —
   messages starting with an agent name are always sent to that agent
2. Requiring "agent" keyword near "remove/delete" for destructive cmds
3. Adding confirmation step for agent removal (30s timeout)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Prevent image data from corrupting session history
e4c1800 
Two fixes for the "Could not process image" error:

1. Images are now embedded as base64 directly in the prompt instead
   of telling Claude to use the Read tool. This prevents image data
   from being stored in the session history, which would corrupt the
   session on resume when the API can't reprocess stale images.

2. Added auto-retry in streamQuery — if a resumed session fails with
   an image/invalid_request error, automatically clears the corrupted
   session and retries fresh instead of crashing.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Revert base64 image embedding — keep Read tool for images
2006ad6 
The base64 approach just dumped unreadable text into the prompt.
The Read tool is the correct way for Claude to view images.
The auto-retry in claude-handler.ts handles corrupted sessions
from stale images on resume.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@zoe-fridler @claude
Route messages with just agent name to that agent
8211aba 
Handles case where user sends "lina" with a file attachment but
no additional text. Previously fell through to the regular handler
which requires a working directory.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@ZoeFridler @zoe-fridler