fix: stop remote adapter turns through the shared cancellation owner by AlexanderZ-Band · Pull Request #835 · band-ai/band-sdk-python

AlexanderZ-Band · 2026-10-18T04:02:32Z

Why

Remote backend work can survive STOP or INTERRUPT, and unfinished cancellation can overlap replacement work or teardown.

What

Integrate Copilot SDK, OpenCode, Letta and A2A with BackendTurns. Cover timeout-abort ownership, bootstrap rollover, next-turn isolation and shutdown ordering through recording backend probes.

Scope

Second layer of the backend cancellation stack split from #791, built on the merged cancellation foundation (#834). Closes INT-1842; contributes to INT-1133. Parlant engine child ownership is the final layer: #791 adds Parlant and the complete inventory guard. Independent fixes: #832 and #833.

Not covered here: STOP before a new A2A task answers when the remote does not stream (INT-1862), and splitting the OpenCode test package into one-class files (INT-1863).

linear-code · 2026-10-18T04:02:38Z

INT-1842

INT-1133

AlexanderZ-Band

Full-PR review: no verified new findings across all six review groups. Converged in cycle 1. The suspected interrupted Copilot eviction leak reproduces on the base as well; the new owned stop improves abort completion. Full unit validation is still running.

Copilot

🟡 Changes recommended

Replacement tools become accessible before cancellation settles, and new A2A tasks can escape cancellation before their IDs arrive.

4 open issues

What changed in this PR

Extends the shared backend cancellation foundation to Copilot SDK, OpenCode, Letta, and A2A, aiming to stop remote work before replacement turns or teardown.

Changes:

Integrates four adapters with BackendTurns.
Adds recording probes for cancellation, timeouts, continuation, and shutdown ordering.
Documents OpenCode’s stopped-turn behavior.

File	Description
tests/integrations/a2a/kit.py	Adds a recording A2A client.
tests/framework_conformance/test_backend_stop.py	Registers four adapters’ stop probes.
tests/framework_conformance/stopprobes/opencode.py	Covers aborts, timeouts, and cleanup.
tests/framework_conformance/stopprobes/letta.py	Covers cancellation modes and failures.
tests/framework_conformance/stopprobes/copilotsdk.py	Covers aborts, session reuse, and shutdown.
tests/framework_conformance/stopprobes/a2a.py	Covers task cancellation and bootstrap continuation.
tests/adapters/opencode/test_lifecycle.py	Waits for the asynchronously owned abort.
tests/adapters/lettakit.py	Adds a recording Letta client.
tests/adapters/copilot_sdk/fakes.py	Records lifecycle calls and supports held aborts.
src/band/integrations/a2a/adapter.py	Adds task cancellation ownership and settlement waits.
src/band/adapters/opencode/adapter.py	Coordinates watchers, aborts, and timeout reporting.
src/band/adapters/letta.py	Adds owned run cancellation and cleanup waits.
src/band/adapters/copilot_sdk.py	Adds owned session aborts and shutdown waits.
docs/adapters/opencode.md	Documents cancellation ordering.

_{Review effort: Balanced}

…ation Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

AlexanderZ-Band

Blocking — would be REQUEST_CHANGES on a PR by someone else.

Full-PR review of fae4bf310, cycle 1: 6 issues, 13 nits, 5 suggestions. The issues are a STOP that cancels an already-finished A2A task, and five stop/teardown probes that stay green when the ordering they claim to guard is removed.

Not repeated here because they are already open as threads: replacement tools published before the previous turn settles (Copilot, Letta, OpenCode) and the A2A stop before the first task event.

Findings whose line is outside the diff hunks:

src/band/adapters/opencode/adapter.py:502 — nit: The comment says reading room_state after the prompt await "would wire this turn's watch task to the wrong turn's future and usage dict". Watcher creation now sits above it (lines 495-498) and takes turn directly, so no code after the await wires a watch task. Only the release_future/turn_future snapshot is still used afterwards. Trim the comment to the reason that still holds: a message racing in during prompt_async calls _begin_turn again, so room_state.turn would point at the next turn.
tests/adapters/opencode/test_lifecycle.py:463 — nit: The diff touches this 679-line file, which holds 13 module-level async def test_... functions and no class Test.... AGENTS.md says one test class per file with no module-level test functions, and 'split a file you touch that breaks it'. The edited test is test_interrupting_a_turn_stops_the_reply_and_frees_the_room.

Move the edited interrupt test into its own class TestInterruptedTurn file in tests/adapters/opencode/ (with a conftest or helpers module for shared fixtures) as part of this change; splitting the rest of the file can follow.

Copilot SDK, Letta and OpenCode published the replacement turn's tools before waiting for the stopped turn's abort or cancel, so a late tool call from the stopped turn ran against the replacement turn. Each now settles the stop first (Copilot and Letta inside running(), OpenCode at the top of on_message), as Codex and Claude already do. OpenCode also keeps its turn watcher only on the turn and drops a snapshot of the turn's futures. Its probes, and the Copilot and Letta ones, now guard the stop ordering: the next message, teardown and a stopped watcher's silence. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

A STOP before a new task's first event reached no task, and a STOP while a finished task was being delivered sent tasks/cancel for it. The first event is now read through BackendTurns.started(), the handle carries its own finished state, and a cancel releases the task through the single finalize path before it cancels the task remotely. Without streaming the first event is the final answer, so a STOP before it still cannot cancel the task; the setting's documentation says so. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

conversations.cancel first ships in 1.7.1 and the conversations resource in 1.6.5, while the floor was 1.0.0, so a shared-mode STOP raised on older releases. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

AlexanderZ-Band · 2026-10-19T16:55:51Z

Answers to the two findings in the cycle-1 review body, which have no thread:

src/band/adapters/opencode/adapter.py:502 (stale comment): confirmed. The watcher is created above it and takes turn directly, so nothing after the await wires a watch task. Fixed in 71b9852: the comment is gone, along with the two snapshot locals, which are now plain reads of turn.
tests/adapters/opencode/test_lifecycle.py:463 (file breaks the one-class-per-file rule): true, but out of scope here. A compliant split moves the whole file (679 lines; the package has 116 module-level tests) and does not belong in a change about stopping remote backend turns. Tracked in INT-1863.

AlexanderZ-Band

Blocking — would be REQUEST_CHANGES on a PR by someone else.

Full-PR review of bcb647de0, cycle 2: 6 issues, 7 nits, 2 suggestions. The issues are a Copilot message that queues behind a STOP and can return on the stopped turn's aborted idle (reproduced against the real CLI), an OpenCode STOP that does not silence a released turn's timeout report, a test that no longer tests what it claims, a probe module that breaks collection in lanes without the Copilot extra, and two A2A guards no probe covers.

No thread from earlier cycles is open.

The CLI emits the aborted turn's idle just after acknowledging the abort, and send_and_wait ends on any idle, so a message queued behind a STOP could return on it with no reply. The stop now waits for that idle while a prompt is in flight. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

… next-message ordering Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

…stop docs Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

Cancelling on_message only schedules the watcher's cancellation, so drain() could return before the abort was spawned; on Python 3.11 it did. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

…rns-throug-INT-1842

AlexanderZ-Band added this pull request to stack #836 October 18, 2026 04:02

AlexanderZ-Band force-pushed the fix/fixpython-stop-remote-adapter-backend-turns-throug-INT-1842 branch 2 times, most recently from 1d5e54a to 3a65127 Compare October 18, 2026 04:20

AlexanderZ-Band commented Oct 18, 2026

View reviewed changes

AlexanderZ-Band force-pushed the fix/fixpython-stop-remote-adapter-backend-turns-throug-INT-1842 branch from 3a65127 to c81118e Compare October 18, 2026 05:49

AlexanderZ-Band requested a balanced review from Copilot October 19, 2026 08:15

Copilot started reviewing on behalf of AlexanderZ-Band October 19, 2026 08:15

View session

Copilot AI reviewed Oct 19, 2026

View reviewed changes

Comment thread src/band/adapters/copilot_sdk.py

Comment thread src/band/adapters/letta.py

Comment thread src/band/adapters/opencode/adapter.py Outdated

Comment thread src/band/integrations/a2a/adapter.py Outdated

Base automatically changed from fix/fixpython-own-backend-turn-cancellation-across-acp-INT-1841 to main October 19, 2026 08:39

AlexanderZ-Band requested a review from amit-gazal-band as a code owner October 19, 2026 08:39

AlexanderZ-Band added 3 commits October 19, 2026 08:39

fix: stop remote adapter turns through the shared cancellation owner

96cd682

refactor: simplify OpenCode timeout completion flow

61d5470

test: control backend stop deadlines and shutdown ordering

4651e01

AlexanderZ-Band force-pushed the fix/fixpython-stop-remote-adapter-backend-turns-throug-INT-1842 branch from c81118e to 4651e01 Compare October 19, 2026 08:39

AlexanderZ-Band mentioned this pull request Oct 19, 2026

fix(tests): give the ACP pipe-holder test room to start its peer #863

Open

refactor: trim dead guards and probe boilerplate in backend-stop migr…

fae4bf3

…ation Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

AlexanderZ-Band commented Oct 19, 2026

View reviewed changes

AlexanderZ-Band and others added 3 commits October 19, 2026 19:52

AlexanderZ-Band commented Oct 19, 2026

View reviewed changes

AlexanderZ-Band and others added 3 commits October 19, 2026 22:12

test: guard A2A finished-task and cleanup waits, OpenCode timeout and…

a777fcb

… next-message ordering Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

refactor: drop a no-op aiter and refresh stale OpenCode comments and …

c76f54c

…stop docs Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YNMWPLGSHPnZpadgLr4sPA

AlexanderZ-Band and others added 2 commits October 20, 2026 05:51

Merge branch 'main' into fix/fixpython-stop-remote-adapter-backend-tu…

f67b7bf

…rns-throug-INT-1842

AlexanderZ-Band added this pull request to the merge queue Oct 20, 2026

Merged via the queue into main with commit ecd2b1b Oct 20, 2026
42 checks passed

AlexanderZ-Band deleted the fix/fixpython-stop-remote-adapter-backend-turns-throug-INT-1842 branch October 20, 2026 04:13

band-release-please-public Bot mentioned this pull request Oct 20, 2026

chore: release main #794

Open

Conversation

AlexanderZ-Band commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Why

What

Scope

Uh oh!

linear-code Bot commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AlexanderZ-Band left a comment

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AlexanderZ-Band left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AlexanderZ-Band commented Oct 19, 2026

Uh oh!

AlexanderZ-Band left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

AlexanderZ-Band commented Oct 18, 2026 •

edited

Loading

linear-code Bot commented Oct 18, 2026 •

edited

Loading