[BUG]: A2A conversation history omits assistant turns — follow-up messages don't see prior answers #368

Ref: #325, #328

Summary

A follow-up A2A message on an existing contextId is given the prior user turns but never the agent's own responses. Every follow-up therefore behaves as if it has never spoken.

Observed against aura-web-server with A2A enabled:

task 1 (user): "In exactly three sentences, explain what a log pipeline is."
               -> completes, answers correctly

task 2 (user, same contextId): "How many sentences did you just use?"
               -> "I haven't responded to that prompt yet, so I used zero sentences."

Expected: the follow-up sees the previous assistant answer and replies "three".

Root cause

get_history_for_context (crates/aura-web-server/src/a2a/agent_executor.rs:549) rebuilds history by listing tasks for the context and flat-mapping each task's history field:

let chat_history: Vec<aura::Message> = task_history
    .into_iter()
    .flat_map(|t| t.history.unwrap())
    .filter_map(|m| convert_a2a_msg_to_aura(&m))
    .collect();

The context lookup itself works — list(context_id) returns every task in the conversation. The problem is what lives in task.history: the agent's response is written to artifacts, not to history. A completed task looks like this:

status: TASK_STATE_COMPLETED
history messages: 1
   role=ROLE_USER   text="In exactly three sentences, explain what a log pipeline is."
artifacts: [("response", 84 parts), ("final", 1 part)]

ROLE_AGENT present in history? False

So flat_map(|t| t.history) can only ever yield user turns. convert_a2a_msg_to_aura has a Role::Agent => Message::assistant(...) arm that is unreachable in practice.

The same call also passes include_artifacts: Some(false), so the assistant text is not even fetched from the store.

Reproduction

Pre-requisites

• aura-web-server running with --enable-a2a and any single-agent config (no HITL gating).

Steps

1. Send a first message and note task.contextId:

curl -s -X POST http://localhost:8080/a2a/v1/message:send \
  -H "Content-Type: application/json" -H "x-aura-model: <agent-alias>" \
  -d '{"message":{"messageId":"m1","role":"ROLE_USER","parts":[{"text":"In exactly three sentences, explain what a log pipeline is."}]}}'

2. Poll GET /a2a/v1/tasks/{id} until TASK_STATE_COMPLETED. Note that history holds only the ROLE_USER message while the answer sits in the response artifact.

3. Send a follow-up on the same contextId:

curl -s -X POST http://localhost:8080/a2a/v1/message:send \
  -H "Content-Type: application/json" -H "x-aura-model: <agent-alias>" \
  -d '{"message":{"messageId":"m2","contextId":"<CTX>","role":"ROLE_USER","parts":[{"text":"How many sentences did you just use?"}]}}'

4. Poll it. The answer denies having responded.

Why the existing test does not catch this

test_sequential_tasks_share_context (crates/aura-web-server/tests/a2a_test.rs:713) passes, because the only thing it asserts is that task 2 can recall what was asked in task 1:

task 1 (user): "Call mock_tool with message='context-test-{uuid}'. Do not say anything else."
task 2 (user): "Tell me what was asked in the first call for this context."
assert: response contains "context-test-{uuid}"

That UUID originates in task 1's user message — the half of history that does carry. The test validates the user-turn path and is blind to the missing assistant turns.

The test's own comment states the expectation that is currently false:

// Wait for task 1 to complete so its full history (tool call, tool result,
// assistant response) is persisted in the task store before task 2 starts.

Tool calls, tool results, and assistant responses are not in task.history.

Suggested fix

Reconstruct the assistant turn from the task's response artifact (requiring include_artifacts: Some(true) in the ListTasksRequest), or write the agent's final message into task.history as ROLE_AGENT when the task completes.

Whichever way, extend test_sequential_tasks_share_context (or add a sibling) to assert on something only the assistant turn could supply — the current phrasing passes without it.

Scope note

Pre-existing, not a regression. git diff main..feature/GH-327-session-storage-trait-and-in-memory touches only shared_task_store.rs under crates/aura-web-server/src/a2a/; get_history_for_context dates to 349014dc feat(a2a): initial implementation.

Relevant to the session-storage roadmap: #328 (Redis/Valkey A2A task store) is motivated partly by making "conversation history by context_id survive load-balancing". Making that history durable and cross-pod will not help while the assistant half of every exchange never reaches the store. Worth fixing before or alongside #328.