Multi-Agent Sessions

PizzaPi supports multiple agent sessions running in parallel through a single orchestration model: linked parent/child sessions. Linked children communicate upward through triggers, and parents communicate downward through respond_to_trigger and tell_child.

---

Communication model

┌─────────────────────────────────────────────────────┐
│  Parent Session                                     │
│                                                     │
│  spawn_session(prompt)                              │
│       │                                             │
│       │ ◄── ask_user_question trigger ──┐           │
│       │ ◄── plan_review trigger ────────┤           │
│       │ ◄── session_error trigger ──────┤           │
│       │ ◄── session_complete trigger ───┤           │
│       │                                 │           │
│  respond_to_trigger(id, response)       │           │
│  tell_child(sessionId, message)  ──►    │           │
│  escalate_trigger(id)  ──► human        │           │
│                                         │           │
│                              ┌──────────┴─────────┐ │
│                              │  Child Session     │ │
│                              │  (linked)          │ │
│                              └────────────────────┘ │
└─────────────────────────────────────────────────────┘

Spawned sessions are always linked. Child events arrive automatically in the parent’s conversation as triggers, and the parent responds with respond_to_trigger, tell_child, or escalate_trigger.

---

spawn_session

Spawn a new independent agent session on the runner. The new session runs in parallel and can be monitored through the PizzaPi web UI.

Parameters

Parameter	Type	Required	Default	Description
prompt	string	✅	—	Initial prompt for the new session. Must be self-contained — the child has no context from the parent.
model	{ provider, id }		Runner default	Model to use. Call list_models to discover available values.
cwd	string		Parent’s cwd	Working directory for the new session.
runnerId	string		Current runner	Target runner ID. Usually not needed.

Return value

On success, returns the session ID, runner ID, working directory, status (Pending or Ready), and a web UI share URL.

Linked child sessions

{
  "prompt": "Refactor the auth module to use JWTs",
  "model": { "provider": "anthropic", "id": "claude-sonnet-4-20250514" }
}

• Parent’s session ID is sent as parentSessionId
• Child events (questions, plans, completion, errors) arrive as triggers in the parent’s conversation
• No manual polling needed — triggers are injected automatically
• If the linked relationship cannot be restored after a reconnect, the harness surfaces that as an orchestration failure instead of silently degrading into peer messaging

Caution

Don’t stall while waiting for children. Triggers arrive automatically as injected messages — do not use sleep loops or idle waits. Simply stop responding and your session will resume when the child’s trigger arrives.

---

Trigger system

When a linked child session hits certain lifecycle events, it fires a trigger to the parent. Triggers arrive in the parent’s conversation prefixed with <!-- trigger:ID --> metadata.

Trigger types

ask_user_question

Fired when a child session calls AskUserQuestion. The trigger payload includes the question text, options, and the full structured questions array. The parent can answer on behalf of the user with respond_to_trigger.

<!-- trigger:abc123 -->
Child session "auth-refactor" is asking:
  "Should I use RS256 or HS256 for JWT signing?"
  Options: RS256, HS256

Use respond_to_trigger(triggerId, response) to answer.

plan_review

Fired when a child calls plan_mode to propose a multi-step plan. The payload includes the plan’s title, steps, and optional description.

Respond with an action:

• approve — Accept the plan. The child proceeds with execution.
• cancel — Reject the plan. The child stops.
• edit — Provide feedback in response. The child revises and resubmits.

session_complete

Fired when the child session finishes. The payload includes the child’s final output and an exitReason:

• exitReason: "completed" — Normal completion.
• exitReason: "error" — The child hit a usage limit or provider failure.

Respond with an action:

• ack — Acknowledge and terminate the child process. This sends SIGTERM to the child and tears down its relay session. Only use when the child is truly finished.
• followUp — Send more work to the child in response. The child resumes with a new turn.

Danger

ack terminates the child. It is not a passive acknowledgement — it kills the child process and cleans up relay resources. If the child’s output says it’s still dispatching workers or waiting for sub-tasks, do not call ack. Wait for the next session_complete trigger.

session_error

Fired before session_complete when a child hits a usage limit or provider error. This gives the parent an early signal to react — retry with a different model, skip the task, or escalate to the user. Check the exitReason field on the subsequent session_complete to distinguish error exits from successful completions.

Trigger lifecycle

1. Child hits a lifecycle event (asks a question, proposes a plan, completes, or errors)
2. Child emits a session_trigger event over the relay WebSocket
3. Relay routes the trigger to the parent session
4. Trigger is injected into the parent’s conversation with a <!-- trigger:ID --> prefix
5. Parent responds via respond_to_trigger, escalate_trigger, or tell_child
6. Response is routed back to the child, which unblocks and continues

Triggers have a 10-minute TTL. If the parent doesn’t respond within that window, the trigger expires and the child times out (5-minute default for ask_user_question and plan_review).

---

respond_to_trigger

Respond to a pending trigger from a child session.

Parameters

Parameter	Type	Required	Description
triggerId	string	✅	The trigger ID from the <!-- trigger:ID --> prefix
response	string	✅	Response text sent to the child
action	string		Structured intent — see below

Actions

Action	Used with	Effect
approve	plan_review	Accept the child’s plan; child proceeds with execution
cancel	plan_review	Reject the plan; child stops
edit	plan_review	Provide feedback in response; child revises the plan
ack	session_complete	Acknowledge completion and terminate the child process
followUp	session_complete	Send response as new input to resume the child

For ask_user_question triggers, action is not required — just provide the response text.

Example

// Child asks a question
respond_to_trigger("abc-123", "Use RS256 — it supports key rotation")

// Child proposes a plan
respond_to_trigger("def-456", "Approved", "approve")

// Child completes — send follow-up work
respond_to_trigger("ghi-789", "Now add unit tests for the JWT module", "followUp")

// Child completes — done, clean up
respond_to_trigger("jkl-012", "Looks good, shutting down", "ack")

---

escalate_trigger

Pass a child’s trigger to the human viewer when the parent agent can’t or shouldn’t handle it.

Parameters

Parameter	Type	Required	Description
triggerId	string	✅	The trigger ID to escalate
context	string		Additional context for the human

The trigger is re-emitted as an escalate type trigger targeting the parent’s own session, where it surfaces in the web UI for the human to respond. The original trigger remains pending — when the human responds (via the web UI), the response routes back through respond_to_trigger to the child.

---

tell_child

Proactively send a message to a linked child session. Unlike respond_to_trigger (which responds to a child’s request), tell_child initiates a new turn in the child’s conversation — the message is delivered as agent input.

Parameters

Parameter	Type	Required	Description
sessionId	string	✅	Child session ID
message	string	✅	Message to send

Use cases

• Redirect a child mid-task: tell_child(id, "Stop the refactor — focus on the auth bug instead")
• Provide additional context the child didn’t ask for
• Coordinate between multiple children by relaying information

---

Relationship durability

Linked parent/child relationships are durable across transient relay outages:

• The server preserves durable linkage metadata even when a reconnect temporarily clears parentSessionId
• Trigger and follow-up routing falls back to durable linkage checks instead of treating the child as unrelated
• If the harness cannot prove the relationship after reconnect, it surfaces an explicit broken-link/orchestration error instead of timing out silently

---

fire_trigger

Fire a trigger into any session — not just children. This enables peer-to-peer trigger communication between sessions that aren’t in a parent-child relationship.

Parameters

Parameter	Type	Required	Description
sessionId	string	✅	Target session ID
type	string	✅	Trigger type (e.g. "service", "webhook", "godmother:idea_started")
payload	object	✅	Arbitrary payload delivered to the target session
source	string		Source identifier shown in trigger history
deliverAs	string		"steer" (default, interrupts current turn) or "followUp" (queued after turn)

Delivery modes

• steer (default) — Interrupts the target session’s current turn. The trigger is injected immediately into the conversation.
• followUp — Queued and delivered after the target’s current turn completes. Use this when the trigger isn’t urgent and shouldn’t disrupt ongoing work.

fire_trigger uses the HTTP Trigger API with API key auth, with a Socket.IO fallback for offline/local mode.

---

Pattern comparison

	subagent tool	spawn_session (linked child)
Communication	Tool call → result (synchronous)	Triggers + tell_child
Blocking	Yes — parent waits for result	No — parent continues, triggers arrive async
Context	Isolated in-process session	Fully independent session + process
UI	Inline card in parent session	Separate session in web UI
Overhead	Near-zero (in-process SDK)	Higher (relay round-trip, new PTY)
Model selection	Via agent definition model field	Via model parameter
Best for	Quick tasks: research, review, refactor	Long-running parallel work, background tasks
Agent definitions	File-based (~/.pizzapi/agents/)	Prompt-based (inline)
Modes	Single, parallel, chain	One session per call

Tip

Prefer subagent for most delegation. It’s simpler, manages context isolation automatically, and returns results inline. Use spawn_session only when you need independent lifecycle, background execution, or async trigger-based coordination. See the Subagents guide for full details.

---

Real-world patterns

Orchestrator → workers

A parent spawns multiple linked children to work in parallel, then synthesizes their results:

// 1. Spawn workers
spawn_session({ prompt: "Review packages/server/ for security issues" })
// → sessionId: "worker-1"

spawn_session({ prompt: "Review packages/ui/ for accessibility issues" })
// → sessionId: "worker-2"

// 2. Triggers arrive automatically as children complete:
//    <!-- trigger:t1 --> session_complete from worker-1
//    <!-- trigger:t2 --> session_complete from worker-2

// 3. Acknowledge each worker
respond_to_trigger("t1", "Received security review", "ack")
respond_to_trigger("t2", "Received a11y review", "ack")

// 4. Synthesize results into a summary

If a worker asks a question mid-task, the parent gets an ask_user_question trigger and can answer directly or escalate to the human:

// Worker asks: "The auth module uses both bcrypt and argon2. Which should I focus on?"
respond_to_trigger("t3", "Focus on argon2 — that's the active path")

// Or if unsure:
escalate_trigger("t3", "Worker needs domain knowledge about our crypto choices")

Review ping-pong

A parent spawns a reviewer and iterates on feedback:

// 1. Spawn reviewer
spawn_session({ prompt: "Review this PR diff for bugs: ..." })
// → sessionId: "reviewer-1"

// 2. Reviewer proposes a plan (plan_review trigger)
respond_to_trigger("plan-t1", "Skip the style checks, focus on logic bugs", "edit")

// 3. Reviewer revises plan and resubmits (another plan_review trigger)
respond_to_trigger("plan-t2", "Looks good", "approve")

// 4. Reviewer completes (session_complete trigger)
// If findings need action:
respond_to_trigger("complete-t1", "Fix the null check bug you found in auth.ts", "followUp")

// 5. Reviewer fixes and completes again
respond_to_trigger("complete-t2", "Done, thanks", "ack")

Error recovery

When a child hits a usage limit, the parent gets a session_error trigger before session_complete:

// session_error arrives: "Rate limit exceeded for claude-sonnet-4"
// Option 1: Retry with a different model
spawn_session({
  prompt: "Continue this task: ...",
  model: { provider: "google", id: "gemini-2.5-pro" }
})

// Option 2: Escalate to human
escalate_trigger("err-t1", "Child hit rate limit — need human to decide next steps")

---

Push notification behavior

Linked child sessions do not trigger push notifications. Only top-level sessions — those started by a user or the runner daemon directly — send push notifications. This prevents notification spam when orchestrating multiple child sessions.

---

Related guides

• Subagents — File-based agent definitions, parallel/chain modes, model routing
• Runner Daemon — How sessions are spawned on runners
• CLI Reference — Full tool parameter reference