Conversation Management Directions¶

When and how to use agent-user conversations — decision guide, response types, behavioral notes. The threads namespace is reserved for the universal-entity primitive.

When to use¶

user wants to start a conversation, or agent needs to explain a plan, ask multi-step questions, or send progress updates

Slash command: /wb-conversation

conversations/conversation_create
conversations/conversation_send
conversations/conversation_ask
conversations/conversation_poll
conversations/conversation_receive
conversations/conversation_ack
conversations/conversation_close
conversations/conversation_list

Directions¶

Start or manage an agent-user conversation via the dashboard sidebar.

Argument: $ARGUMENTS (optional -- conversation_id to resume, or blank to create new)

Conversation operations¶

mcp__work-buddy__wb_run("conversation_create", {"title": "...", "message": "..."}) mcp__work-buddy__wb_run("conversation_send", {"conversation_id": "...", "message": "..."}) mcp__work-buddy__wb_run("conversation_ask", {"conversation_id": "...", "question": "...", "response_type": "boolean"}) mcp__work-buddy__wb_run("conversation_poll", {"conversation_id": "..."}) mcp__work-buddy__wb_run("conversation_receive", {"conversation_id": "...", "consumer": "...", "generation": "...", "timeout_seconds": 110}) mcp__work-buddy__wb_run("conversation_ack", {"conversation_id": "...", "consumer": "...", "generation": "...", "message_id": "..."}) mcp__work-buddy__wb_run("conversation_close", {"conversation_id": "..."}) mcp__work-buddy__wb_run("conversation_list")

When to use conversations¶

Explaining plans before executing: create a conversation, describe the plan, ask approval
Multi-step decisions: ask a sequence of questions in one back-and-forth
Progress updates: send status messages during long tasks
Follow-up dialogue: when a notification needs back-and-forth

Response types for conversation_ask¶

Type	User sees	Response value
freeform (default)	Text input	User's text
boolean	Yes/No buttons	"true" or "false"
choice	Labeled buttons	The choice key

Behavior¶

conversation_create opens the chat sidebar on the dashboard automatically
conversation_ask with timeout_seconds blocks until response (max 110s)
conversation_poll accepts an exact message_id; omission discovers the current question. Once a wait starts, only that question can complete it, even when a newer question appears. Content-bearing pending/answer results are disclosure-accounted for scoped hosted workers before return
conversation_receive is for a leased long-running driver: it returns the oldest unacknowledged user turn without advancing the durable cursor
conversation_ack advances that cursor only over the exact delivered message; call it after the turn's reply and side effects succeed
a leased driver must pass the same consumer and generation to send, ask, poll, receive and acknowledge tools; this fences late access after a restart or close
when the lease carries a validated execution snapshot, assistant messages receive their provider/model producer provenance from that exact lease; a caller cannot assert or override it
a lease_lost result from receive, acknowledge, send, or ask means a newer driver owns the conversation; stop immediately
canonical question answers carry an exact in_reply_to; scoped disclosure stages the question and that canonical user answer, then checks the current lease and binding again after accounting. Historical unlinked answers fail closed rather than being guessed into an association
Sidebar auto-polls every 3s for new messages

Hosted form conversations¶

Form agents receive a pre-bound conversation and may not create, list or close arbitrary conversations. Consume the bound initial context first, then receive each authored turn and consume its exact form snapshot. Only then propose allowlisted edits and send a durable reply/question before acknowledgement. Use plain messages for open-ended questions; finite choices use stable question IDs, and ordinary composer text does not consume them. Stop on scope, policy, expiry or disclosure failure. See services/dashboard/react/assisted-drafts.

Naming note¶

In v5 the term Thread is reserved for work-buddy's universal entity for 'context that may need an action'. The agent-user dialogue subsystem you're using here is called a Conversation to free the name.