Bug: Team agents send messages to agentType instead of member name, creating orphan inboxes

Bug Report: Team agents send messages to agentType instead of member name, creating orphan inboxes

Summary

When agents are spawned as teammates via Task with team_name and name parameters, they use the agent type (e.g., hawking-theorist) instead of the team member name (e.g., hawking) as the recipient in SendMessage. The system silently creates new inbox files for any recipient string, so messages land in orphan inbox files that no agent reads.

Environment

• Claude Code version: latest (Feb 2026)
• OS: Windows 11 (MINGW64/Git Bash)
• Model: claude-opus-4-6

Steps to Reproduce

1. Create a team:

TeamCreate({ "team_name": "my-team" })

1. Spawn two agents with explicit name parameters:

Task({
  "name": "alice",
  "subagent_type": "hawking-theorist",
  "team_name": "my-team",
  "prompt": "Send a message to your teammate bob"
})

Task({
  "name": "bob",
  "subagent_type": "einstein-theorist",
  "team_name": "my-team",
  "prompt": "Send a message to your teammate alice"
})

1. Agent "alice" sends a message to recipient: "einstein-theorist" instead of recipient: "bob".

1. The system creates ~/.claude/teams/my-team/inboxes/einstein-theorist.json (orphan) alongside ~/.claude/teams/my-team/inboxes/bob.json (where bob actually reads).

1. Bob never receives the message.

Evidence

From a real team session with 3 agents (schwarzschild-penrose, dirac-antimatter, hawking):

Inbox files created:

inboxes/
  schwarzschild-penrose.json    # Correct (team member name)
  dirac-antimatter.json         # Correct
  dirac-antimatter-theorist.json  # ORPHAN (agent type name)
  hawking.json                  # Correct
  hawking-theorist.json         # ORPHAN (agent type name)
  team-lead.json                # Correct

hawking-theorist.json (orphan) contained 12 unread messages:

{
  "from": "schwarzschild-penrose",
  "text": "[substantive 500-word message about curvature invariants]",
  "read": false
},
{
  "from": "dirac-antimatter",
  "text": "[substantive 400-word message with algebraic proofs]",
  "read": false
}
// ... 10 more unread messages

hawking.json (actual inbox) contained only 2 messages:

Both from team-lead (the orchestrator), which used the correct name.

Result:

• Hawking never received any messages from his teammates during the first ~10 minutes
• SP and Dirac never received Hawking's early messages
• Agents reported "checking inbox" but found nothing (because they checked the correct inbox, which was empty)
• Eventually the team lead manually told Hawking to resend, partially resolving the issue
• Even after that, SP continued sending to dirac-antimatter-theorist (wrong) instead of dirac-antimatter (correct)

Root Cause Analysis

Agents construct the recipient field using information from their system prompt. The agent type names (e.g., hawking-theorist, dirac-antimatter-theorist) are more prominent in their context than the team member names from config.json. The SendMessage tool does not validate that the recipient matches a registered team member name — it silently creates a new inbox file for any string.

Two issues compound:

1. Agents default to agentType over name: Even when told to use team member names, agents gravitate toward the more descriptive type names.
2. No validation on recipient: SendMessage accepts any recipient string and creates an inbox file. There is no error, warning, or fallback when the recipient doesn't match a registered team member.

Suggested Fix

Option A: Validate recipient in SendMessage (recommended)

When SendMessage receives a recipient that doesn't match any name in the team's config.json, either:

• Error: Reject the message and return an error telling the agent to use the correct name
• Fuzzy match: If the recipient matches an agentType in config, auto-resolve to the corresponding name and deliver to the correct inbox

Option B: Merge inbox files

Use a single inbox file per agent, keyed by agentId (which is {name}@{team}). Any message sent to either the name or agentType should route to the same inbox.

Option C: Inject team roster into agent context

When spawning a teammate, include the full team roster with member names prominently in the system context, so agents know the correct names to use.

Current Workaround

Add an explicit instruction to CLAUDE.md (project instructions inherited by all agents):

### Message Discipline
- **CRITICAL: Use teammate NAMES, not agent types.** Read `~/.claude/teams/{team-name}/config.json`
  to find the `name` field for each member. Use THAT as the `recipient` in SendMessage.
  Example: if a teammate's name is `hawking`, send to `hawking` — NOT `hawking-theorist`.
  Messages sent to the wrong name create orphan inbox files and are never delivered.

This partially mitigates the issue but does not fully prevent it — agents still sometimes use the type name despite the instruction.

Impact

• Severity: High — silent message loss in team collaboration
• Frequency: Consistent — observed in every multi-agent team session
• User experience: Agents appear to be ignoring each other; the user sees "checking inbox" behavior that produces no results; debugging requires reading raw inbox JSON files to discover the orphan inboxes