[Eduard van Valkenburg (eavanvalkenburg)]

Description

Fix request-owned steering acknowledgments for resilient multi-turn tasks. Rejected queue appends previously left unreturned futures registered, so repeated SteeringQueueFull responses could grow per-chain memory without a durable queue slot.

The fix persists opaque acknowledgment IDs alongside durable queue slots and binds append, drain, cancellation, and result delivery by ID rather than list position. Append and drain retries reconcile only the exact transition whose conditional PATCH may have committed before returning a conflict. A later drain cannot redispatch stale active state or consume the next queue head.

Turn completion now acts as a persistence barrier: return, failure, and cancellation are acknowledged, and queued successors are dispatched, only after the terminal task-state PATCH succeeds. If persistence fails, the current caller receives that persistence error, local queued handles receive TaskConflictError(current_status="in_progress"), and durable state remains available for recovery. Waiter-aware write-lock reclamation also prevents rejected remote steers from retaining per-task locks.

The branch is rebased onto the current Core 2.3.0b2 preview line. It no longer creates a stable 2.2.1 release entry; a Chronus fix change records the work for the next preview release.

Tests use the real local file provider and deterministic failure injection for queue saturation, pre/post-commit append failures, committed drain conflicts, failed Phase-3 cleanup plus failed terminal persistence, duplicate inputs/public IDs, legacy/remote backlogs, attachment drains, cancellation, return/raise paths, and waiter-safe lock cleanup. The full TaskManager suite passes (740 tests on Python 3.11), focused fault regressions pass on Python 3.10 and 3.14, and pylint, mypy, pyright, and CI-equivalent cspell pass.

Downstream dependency: microsoft/agent-framework#8794; steering review thread. The downstream dependency floor must not change until a patched AgentServer core release is officially published and verified.

All SDK Contribution checklist:

• The pull request does not introduce [breaking changes]
• CHANGELOG is updated for new features, bug fixes or other significant changes.
• I have read the contribution guidelines.

General Guidelines and Best Practices

• Title of the pull request is clear and informative.
• There are a small number of commits, each of which have an informative message. This means that previously merged commits do not appear in the history of the PR. For more information on cleaning up the commits in your PR, see this page.

Testing Guidelines

• Pull request includes test coverage for the included changes.

[github-actions]

Thank you for your contribution Eduard van Valkenburg (@eavanvalkenburg)! We will review the pull request and get back to you soon.

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
10 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Copilot review overview

🟡 Changes recommended

Ambiguous hosted PATCH outcomes can still remove a future for an input that was durably appended, corrupting acknowledgment ordering.

Review effort: Balanced
Findings: 1

Open (1)

• Ambiguous PATCH outcomes can orphan futures and misroute queued results · New

What changed in this PR

Fixes steering acknowledgment-future leaks when queue appends fail.

Changes:

• Serializes future registration and queue append.
• Cleans up futures after failed or cancelled appends.
• Adds failure, cancellation, capacity, and acknowledgment tests.

File	Description
_manager.py	Adds targeted future cleanup.
_decorator.py	Makes registration and append atomic locally.
test_steering.py	Expands steering failure coverage.
CHANGELOG.md	Documents the fix.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🔵 Needs a closer look

The concurrency-sensitive durable queue migration and cross-worker compatibility require final human validation.

Review effort: Balanced
Findings: None

Resolved since last review (1)

• Ambiguous PATCH outcomes can orphan futures and misroute queued results

[Copilot]

Copilot review overview

🟡 Changes recommended

A committed PATCH retry can still duplicate queued inputs and leave cancelled work executable.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Retrying committed PATCH duplicates ack and breaks cancellation · New
• Rejected remote steers leak cached task write locks · New

[Copilot]

Copilot review overview

🔵 Needs a closer look

Durable queue and concurrency changes require human validation of hosted behavior and coordinated worker upgrades.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Retrying committed PATCH duplicates ack and breaks cancellation
• Rejected remote steers leak cached task write locks

[Copilot]

Copilot review overview

🟡 Changes recommended

A committed drain followed by a retry-induced conflict can skip an accepted input and strand its acknowledgment future.

Review effort: Balanced
Findings: 1

Open (1)

• Drain retry skips queued input after committed PATCH returns 412 · New

Resolved since last review (2)

• Retrying committed PATCH duplicates ack and breaks cancellation
• Rejected remote steers leak cached task write locks

[Copilot]

🟡 Changes recommended

Failed turn-start cleanup can leave stale drain state and dispatch a phantom input.

1 open finding

• Atomically clear drain flag when clearing active turn fields · New

1 resolved since last review

• Drain retry skips queued input after committed PATCH returns 412

🧠 Review effort: Balanced

[Copilot]

🔵 Needs a closer look

Rejected remote steers still accumulate unbounded entries in the manager’s etag cache.

0 open findings

1 resolved since last review

• Atomically clear drain flag when clearing active turn fields

Previously missed (1)

In code that hasn't changed since last review

Clear stale etags when reclaiming inactive remote task locks

sdk/​agentserver/​azure-ai-agentserver-core/​azure/​ai/​agentserver/​core/​tasks/​_manager.py:3192

The final-user path reclaims the remote task's lock but clears its etag only when _active_tasks_pop() previously marked cleanup pending. A rejected remote steer has no local active entry, while the initial _provider_get_tracked() still caches its etag, so the new 20-task regression empties _task_write_locks but leaves 20 entries in _task_etag_cache; repeated distinct remote IDs still grow manager memory indefinitely. Clear the etag whenever this path removes a no-longer-active lock, and extend the regression to assert the cache is empty.

🧠 Review effort: Balanced