Context & Sessions

​

Key Identifiers

​

context_id vs task_id

• Maps to A2A’s native Message/Task contextId
• Is managed in the A2A transport envelope, not inside the AdCP task body
• Is not an MCP request-body session mechanism
• Is correlation and routing metadata, not authorization

• Specific to individual requests that could be asynchronous
• Lives beyond the conversation
• Used for tracking operation progress over time
• Persists until the task completes (may be days for complex media buys)
• Can be referenced across different conversations or sessions

​

A2A usage example

// First call - establishes context and creates task
const first = await a2a.send({ message: { parts: [/* AdCP request */] } });

// Later in the same A2A conversation, use the native envelope fields.
const followUp = await a2a.send({
  message: {
    contextId: first.contextId,
    taskId: first.taskId,
    parts: [/* follow-up AdCP request */]
  }
});

​

Protocol Differences

• A2A: Conversation continuity uses native contextId in the A2A envelope.
• MCP: Continuity, when offered, belongs to the transport session. Do not send an AdCP context_id tool argument to create or resume it.

​

A2A Context

const task = await a2a.send({ message: {...} });

// Follow-ups explicitly carry the transport-native contextId.
const followUp = await a2a.send({
  message: {
    contextId: task.contextId,
    taskId: task.taskId,
    ...
  }
});

​

MCP context

// First call - no context
await mcp.call('get_products', {
  idempotency_key: "550e8400-e29b-41d4-a716-446655442076",
  buying_mode: "brief",
  brief: "Video ads"
});

// Follow-up in the same MCP transport session, if the host provides one.
await mcp.call('get_products', {
  idempotency_key: "550e8400-e29b-41d4-a716-446655442077",
  buying_mode: "brief",
  brief: "Focus on premium inventory"
});

​

What Context Maintains

• Current media buy and products being discussed
• Search results and applied filters
• Conversation history and user intent
• User preferences expressed in the session
• Workflow state and temporary decisions

​

Extension Fields (ext)

​

Schema Pattern

{
  "product_id": "ctv_premium",
  "name": "Connected TV Premium Inventory",
  "ext": {
    "gam": {
      "order_id": "1234567890",
      "dashboard_url": "https://..."
    },
    "roku": {
      "content_genres": ["comedy", "drama"]
    }
  }
}

• Is always optional (never required)
• Accepts any valid JSON structure
• Must be preserved by implementations (even unknown fields)
• Is not validated by AdCP schemas (implementation-specific validation allowed)

​

Namespacing (Critical)

// ✅ Correct - Namespaced
{
  "ext": {
    "gam": { "test_mode": true },
    "roku": { "app_ids": ["123"] }
  }
}

// ❌ Incorrect - Not namespaced
{
  "ext": {
    "test_mode": true,  // Missing namespace!
    "app_ids": ["123"]  // Which platform?
  }
}

​

AdCP opportunity extension

{
  "$schema": "https://adcontextprotocol.org/schemas/3.2.1/media-buy/get-products-request.json",
  "buying_mode": "brief",
  "brief": "Video inventory for a spring launch.",
  "account": {
    "account_id": "acc_spring_launch"
  },
  "ext": {
    "adcp": {
      "opportunity": {
        "opportunity_id": "opp_spring_launch"
      }
    }
  }
}

• Resolve account scope. Durable association requires an authorized account resolved from the request’s existing account field or unambiguous authenticated account context. If that scope cannot be resolved, reject the opt-in binding with structured INVALID_REQUEST before persisting an association or applying related state changes. Never key the association globally by opportunity ID or join accountless discoveries to an account later.
• Preserve the association. The buyer’s SDK preserves the extension during request validation and serialization. Repeated briefs with the same ID in the same seller/account scope belong to the same planning cycle, and proposals produced by the call inherit it. The ID does not replace proposal identity, idempotency keys, or authorization. A call without the extension creates no opportunity association through this binding.
• Retain it through refinement. Subsequent get_products calls with buying_mode: "refine" preserve a proposal’s existing association when the extension is omitted. If an extension-bearing call refines or finalizes an already-associated proposal, the supplied ID and account scope must match every such proposal named by the call. Reject a mismatch with INVALID_REQUEST before mutation; the extension cannot reassign a proposal to another cycle. Compact refine_proposals inherits the stored association and does not accept this extension.
• Keep lifecycle transitions on their existing tasks. On discovery, status, close_reason, and close_detail describe lifecycle context established through the existing opportunity task contracts; carrying them here does not itself apply a transition. Omitted status means no status update and never implicitly reopens or closes a cycle. All published validation dependencies still apply. Successful proposal execution retains the association and follows the existing 3.2 acceptance-closure contract, even when the buyer omits the extension on that execution call.
• Use existing core fields. On tasks that already define opportunity, send the same context in that core field. New 3.2 integrations can already use request_proposals with core opportunity context. When the negotiated implementation supports a core opportunity field on get_products, switch to it and remove the extension. The core definition takes precedence if both locations are supplied.

​

Application Context (context)

​

Key Properties

• Agents NEVER parse or use context to affect behavior
• Exists solely for the initiator’s internal tracking needs
• Echoed unchanged in responses and webhook payloads

​

Normative echo contract

1. Echo on success. When the caller includes a top-level context object on a request, the agent MUST include the same JSON value in the response. This applies whether the response status is completed, submitted, working, input-required, or any other terminal or intermediate state.
2. Echo on error. Failure responses MUST also echo the same context JSON value. Dropping context on the error path breaks correlation exactly when the buyer needs it most. Agents that return adcp_error, errors[], or any other error envelope MUST still carry through the caller’s context.
3. Echo on async updates. Push notifications, webhook payloads, and any subsequent messages the agent emits for the same operation MUST carry the original context. The agent MUST NOT drop context between the initial response and a later status update — a buyer that correlated by context.trace_id expects every message for that operation to surface the same trace.
4. No synthesis. When the caller does NOT provide a context object, the agent MUST NOT fabricate one. Responses to context-less requests MUST omit the context field (or emit it as null / absent per the transport’s normal serialization). Synthetic context from the agent side is a conformance failure — the whole point of context is that it is owned by the caller.
5. No mutation. Agents MUST NOT add, remove, rename, or change the JSON type or value of fields in the echoed context. JSON value equivalence applies: object member order and number lexical form are not types or values. For example, {"a":1,"b":2} and {"b":2.0,"a":1.0} are equivalent for the echo rule. An integer-valued number remains an integer for JSON Schema validation when serialized with a zero fractional part, so A2A’s 3 → 3.0 serialization does not count as retyping. Changing the mathematical value does. Across A2A, the comparison value is the post-decoding google.protobuf.Value; senders and receivers MUST enforce the A2A numeric domain. Verifiers MUST compare parsed JSON values; a verifier that hashes context SHOULD canonicalize the parsed value instead of hashing transport bytes.
6. No action. Agents MUST NOT parse, validate, log fields from, or branch on any value inside context. Context is opaque to the agent — a value that looks like a structured identifier is not an invitation to interpret it.

​

Schema Pattern

{
  "tool": "create_media_buy",
  "arguments": {
    "packages": [...],
    "context": {
      "ui_session_id": "sess_abc123",
      "trace_id": "trace_xyz789",
      "internal_campaign_id": "camp_456"
    }
  }
}

{
  "status": "input-required",
  "message": "Media buy requires manual approval before activation.",
  "context_id": "ctx_ghi789",
  "context": {
    "ui_session_id": "sess_abc123",
    "trace_id": "trace_xyz789",
    "internal_campaign_id": "camp_456"
  }
}

​

Common Context Uses

1. UI/Session tracking - Maintaining state across async operations
2. Request correlation - Tracing requests through distributed systems
3. Internal identifiers - Mapping to your internal data structures
4. Organization context - Multi-tenant tracking

​

When to Use What

​

Use ext when:

• Platform needs to parse the data
• Data MAY affect operational behavior
• Data represents platform-specific configuration
• Data should persist across operations

​

Use context when:

• Data is only for caller’s internal use
• Data should never affect agent behavior
• Data is for correlation/tracking only
• Data needs to be echoed unchanged

​

Best Practices

​

For A2A

• Let the protocol handle context
• Use contextId for explicit conversation threading
• Trust the session management

​

For MCP

• Do not copy response context_id values into later tool arguments
• Use the MCP transport session, if provided, for ephemeral continuity
• Use task_id or the task’s resource identifiers for durable work
• Ignore a request-body context_id where the selected schema admits it

​

For Extensions

• Always namespace under the extension owner’s key
• Document your extensions extensively
• Consider proposing standardization for common patterns

​

For Application Context

• Keep it opaque - don’t structure for agents to parse
• Avoid large payloads - context is echoed in every response
• Use for correlation only - never for operational data