[gulldan]

Stack: #406 → #405 → #407 → #411 → #395. Stacked on #411: review only the last commit.

Fixes #393.

Problem

subscribe_to_task resubscribes only to a task whose execution is still registered. An execution that ends in an interrupted state (INPUT_REQUIRED, AUTH_REQUIRED) is unregistered as soon as the executor returns, and the task-store fallback from #277 turns only a terminal stored task into UnsupportedOperationError. So a stored task that waits for input falls through to TaskNotFoundError (-32001), although GetTask returns it.

§3.1.6: SubscribeToTask "can be used with any task that is not in a terminal state", and the stream MUST open with a Task carrying the current state. The official TCK builds on that: STREAM-SUB-001 (the first event is a Task), STREAM-SUB-002 (subscribe to an input-required task, send the follow-up message, the stream closes at the terminal state) and STREAM-ORDER-002..004 all create their task with tck-input-required. Against a2a-rs all five are skipped because the subscription errors.

Fix

ExecutionManager keeps, per task id, the subscriptions that found no registered execution (waiting). subscribe_to_task registers one under the registry lock before it reads the store, so an execution that registers from then on finds it: start subscribes every waiter to the new broadcast channel before the executor runs, so none of the turn's events is missed, and from there the stream is an ordinary resubscription (the task as the turn found it, the turn's events, closing where the execution ends). After the store read the errors are unchanged — a missing task is TaskNotFoundError, a terminal one UnsupportedOperationError — and an execution that attached meanwhile is served as is. Otherwise the stream opens with the stored Task and continues with the next execution's events once a follow-up message starts it.

A waiter is a Drop guard: a client that disconnects before the next turn removes its own entry, so nothing stays in the registry and no task is spawned for it. Lock order is the existing one (executions, then the new std::sync::Mutex around waiting, never held across an await). A subscription that is not polled while the next turn publishes more than EXECUTION_BUFFER_CAPACITY events lags and ends with the existing sanitized error, as a resubscription does.

§11.7 closes a turn's stream at an interrupted state. A subscription opened between turns has no turn to close with; following the next one is what STREAM-SUB-002 expects, and that stream still ends when the execution it attached to ends, interrupted or terminal, exactly as a resubscription to it would.

Tests

• test_subscribe_to_a_task_awaiting_input_opens_with_the_stored_task — the regression; on main it fails with -32001.
• test_subscriptions_to_a_task_awaiting_input_follow_its_next_turn — two subscribers, and the whole next turn runs before either stream is polled: both see Task INPUT_REQUIRED (stored), Task INPUT_REQUIRED (as the turn found it, with the follow-up in its history), COMPLETED, then close; a new subscription is then UnsupportedOperationError.
• test_a_dropped_subscription_to_a_task_awaiting_input_unregisters; the unknown-task and terminal-task tests now also assert the registry is left empty.
• just ci on Linux (stable 1.97.1): lockfile --locked, cargo hack check --all-targets --each-feature, cargo fmt --check, cargo hack clippy --all-targets --each-feature -- -D warnings, cargo test --workspace (doc tests included), copyright headers — all green.
• Official TCK (a2aproject/a2a-tck d4d6fae, JSON-RPC + HTTP+JSON) against an agent built on the patched crates ([patch.crates-io] → this branch): MUST 69 passed / 4 failed / 20 skipped, before 64 / 4 / 25 — the same four failures as before (TCK-side, fix(a2a-server): SSE event ids, keep-alive, and a request body limit #202/fix(ci): exempt bot commits in bot-authored pull requests from DCO #231/chore: release #240), and STREAM-SUB-001, STREAM-SUB-002, STREAM-ORDER-002, STREAM-ORDER-003, STREAM-ORDER-004 pass instead of being skipped; STREAM-SUB-003/004 still pass.

Overlap with the A2A 1.1 stack

This touches ExecutionManager and subscribe_to_task, which #385–#392 rewrite (handler/live.rs, subscribe_stream). The stack tip (refs/pull/392/head) answers this case with the same -32001 — the regression test above fails on it the same way — so the fix is needed on either base. Happy to rebase onto the stack once it lands; the same idea fits its per-task change notifier and the subscribe_stream fallback.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.00885% with 9 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
a2a-server/src/handler.rs	95.52%	9 Missing ⚠️

📢 Thoughts on this report? Let us know!