/talk-to - Agent Messaging

Send messages to agents via Oracle threads. Each agent has a persistent channel thread.

Messaging consolidation (2026-07-06): /hey, /contacts, /mailbox, and /inbox were archived to the zombie tier — this skill is the one messaging verb now. For real-time tmux pings use maw hey directly (or --maw here). Archived skills remain installable via arra install -s <name>.

Usage

/talk-to arthur "What's your status?"          # one-shot message
/talk-to arthur --new "Hey, starting fresh"    # skip lookup, create new thread
/talk-to arthur loop ask about their work      # autonomous conversation
/talk-to #42 "follow up on this"               # post to thread by ID
/talk-to --list                                # show channels
/talk-to arthur --maw "quick ping"             # force maw transport (real-time tmux)
/talk-to arthur --thread "async question"      # force MCP thread transport
/talk-to arthur --inbox "offline message"      # force inbox transport (file write)

Mode 0: No arguments

If ARGUMENTS is empty, show usage help then run --list.

Step 0: Contacts Lookup

Always read contacts first. This is the source of truth for agent routing.

date "+🕐 %H:%M %Z (%A %d %B %Y)"
CONTACTS="$(pwd)/ψ/contacts.json"
if [ ! -f "$CONTACTS" ]; then
  # Fallback path
  CONTACTS="$(pwd)/.oracle/contacts.json"
fi

if [ -f "$CONTACTS" ]; then
  MAW=$(jq -r ".contacts.\"$AGENT\".maw // empty" "$CONTACTS")
  INBOX=$(jq -r ".contacts.\"$AGENT\".inbox // empty" "$CONTACTS")
  THREAD=$(jq -r ".contacts.\"$AGENT\".thread // empty" "$CONTACTS")
  REPO=$(jq -r ".contacts.\"$AGENT\".repo // empty" "$CONTACTS")
  NOTES=$(jq -r ".contacts.\"$AGENT\".notes // empty" "$CONTACTS")
  FOUND_IN_CONTACTS=true
else
  FOUND_IN_CONTACTS=false
fi

If agent not found in contacts AND not found in maw ls:

I don't know "{agent}".

  /contacts add {agent}    — register transport info
  /talk-to {agent} --thread "message"  — try MCP thread anyway

Offer /contacts add first. If user insists, fall through to MCP thread.

---

Transport Selection

Flag	Transport	Best For
(none)	auto — detect best	Default
--maw	maw hey (tmux sendkeys)	Real-time, local fleet, low latency
--thread	MCP oracle_thread	Async, persistent, cross-machine
--inbox	File write to ψ/inbox/	Offline, no maw/MCP needed

Auto-detect logic (when no flag):

1. Check ψ/contacts.json → has maw field? → use maw hey
2. No contacts? → maw ls 2>/dev/null | grep -q "{agent}" fallback → use maw hey
3. Neither? → use MCP thread (async, persistent)

# Auto-detect: contacts-first, maw ls fallback
if [ -n "$MAW" ]; then
  echo "USE_MAW (from contacts: $MAW)"
elif maw ls 2>/dev/null | grep -q "{agent}"; then
  echo "USE_MAW (from maw ls)"
else
  echo "USE_THREAD"
fi

Replying to a message another oracle sent you (#434)

A ping arrives in your context looking like [m5:digger] …. The reflex is to answer in your own transcript — but that text goes to Nat, not to digger, who is sitting in a different pane waiting. Reply with maw hey or the conversation is one-sided and nobody notices for hours.

maw hey <their-oracle-name> 'your reply'     # name, not a pane id

Never reply to a pane id. maw peek %3001 works, so it is tempting to paste that same %3001 into maw hey — it hard-errors (bare target '%3001' not found locally). Pane ids are a peek-only handle. maw ls -v is the source of truth for targets; a bare maw ls | grep finds the row but not the addressable name, which is why the auto-detect above should be treated as a hint, not an answer:

maw ls -v | grep -i <name>      # resolves the real target
maw hey <target> ping --dry-run # confirm it resolves BEFORE sending

--dry-run is the cheap check. It costs nothing and catches a wrong target before the message goes somewhere confusing.

When using --maw:

1. Compose message from intent
2. Use contacts maw name if available: maw hey {MAW or agent-oracle} '{message}'
3. Optionally maw peek {agent} to check response
4. Confirm: Sent via maw to {agent}

When using --inbox:

1. Compose message from intent
2. Check contacts for inbox path — if empty, error: No inbox path for {agent}. Run /contacts show {agent}
3. Write message file:

   SELF="$(basename $(pwd) | sed 's/-oracle$//')"
   echo "$MESSAGE" > "$INBOX/$(date +%Y%m%d_%H%M)_from_${SELF}.md"
   

4. Confirm: Dropped to {agent}'s inbox

When using --thread (or auto-detected thread): Fall through to Mode 3 (one-shot) below.

Routing

Pattern	Use
channel:{agent}	Persistent per-agent channel
topic:{agent}:{slug}	Topic-specific thread (with --topic)
#{id}	Direct thread reference by ID

Mode 1: --list

1. arra_threads() (no status filter)
2. Filter titles starting with channel: or topic:, exclude closed
3. Display: channel:arthur (#42) pending — 12 msgs

Mode 2: --new (fast create)

Skip lookup. One MCP call.

1. Compose message from intent
2. arra_thread({ title: "channel:{agent}", message, role: "human" })
3. Notify: Bash maw hey {MAW or agent-oracle} 'Thread #{id} from {self}: {preview}'
   • If maw hey fails → warn only, don't error (thread already sent)
4. Confirm: Created channel:{agent} (thread #{id})

Mode 3: One-shot (default)

1. Compose message from intent
2. If first arg is #{id} → post directly to that thread ID
3. Otherwise: arra_threads() → find channel:{agent}, create if missing
4. Post message to thread
5. Notify: Bash maw hey {MAW or agent-oracle} 'Thread #{id} from {self}: {preview}'
   • If maw hey fails → warn only, don't error (thread already sent)
6. arra_thread_read({ threadId }) → show any agent responses
7. Confirm: Posted to channel:{agent} (thread #{id})

Mode 4: loop (autonomous conversation)

Like Ralph loop — AI drives the conversation autonomously. No user prompts between turns.

1. Find or create thread (channel:{agent}, or --new to skip lookup)
2. Compose opening message from user's intent and post it
3. Autonomous loop (max 10 iterations): a. arra_thread_read({ threadId }) — check for new messages b. If agent responded: read their response, compose a thoughtful follow-up, post it c. If no new response: compose a follow-up question or probe deeper, post it d. After each exchange, briefly note what you learned e. Stop when: enough insight gathered, conversation circling, or 10 iterations hit
4. Notify (once, after opening message): Bash maw hey {MAW or agent-oracle} 'Thread #{id} from {self}: {preview}'
   • If maw hey fails → warn only, don't error
5. Show summary:

   Conversation with {agent} (thread #{id}) — {n} messages, {iterations} turns
   
   Key insights:
   - [insight 1]
   - [insight 2]
   

6. Leave thread open for future use

The goal is insight extraction. You are having a conversation on behalf of the human to learn something useful.

Parsing Rules

• First arg = agent name (lowercase), #id (thread ref), or --list
• --new = skip lookup, create fresh
• loop = autonomous conversation (AI drives, no user prompts)
• --topic "slug" = use topic:{agent}:{slug} instead of channel:{agent}
• Everything else = the message/intent

Message Composition

CRITICAL: You are the composer. The user gives intent, you write the message.

• Compose a clear, natural message from the user's intent
• Post immediately — do NOT ask the user what to say
• Do NOT use AskUserQuestion for message content
• Show what you posted after sending

If the message already reads like a direct message (e.g. "What's your status?"), post as-is.

Auto Notification (maw hey)

After posting to a thread, notify the target agent via maw hey:

maw hey {MAW or agent-oracle} 'Thread #{id} from {self}: {first 60 chars of message}'

• {MAW} = contacts maw field (e.g. "mawjs-oracle") — preferred over {agent}-oracle
• {self} = current Oracle's name (e.g. "Mother Oracle")
• {preview} = first ~60 chars of the posted message
• Runs once per /talk-to invocation (not per loop iteration)
• Fail-safe: if maw hey errors, log warning and continue — the thread is the source of truth

Important Notes

• Agent names are always lowercase
• Thread titles are the routing key — never modify existing thread titles
• One channel thread per agent (reuse, don't recreate)
• #{id} lets users reference any thread directly — no lookup needed
• All messages attributed with role: "human"

ARGUMENTS: $ARGUMENTS