[agentserver] Enable resilient tasks in invocations LRA samples + add minimal hello samples

[Nathandrake229]

What & why

The durable-task subsystem is strictly opt-in as of azure-ai-agentserver-core 2.2.0b1: AgentServerHost only constructs the TaskManager (and runs the crash-recovery scan) when set_resilient_tasks_enabled(True) is called before host startup. With the switch off, get_task_manager() / <task>.start() raise TaskManagerNotInitialized and durable execution + crash recovery are inactive.

The resilient invocations samples declare durable tasks (@task / @multi_turn_task) but never enable the subsystem, and there is no host auto-enable for invocations. As shipped, they would raise TaskManagerNotInitialized and would not actually run as long-running/crash-recoverable agents.

Changes

Enable the subsystem in the existing resilient samples (one line, at import time, before app = InvocationAgentServerHost()):

• resilient_research/app.py
• resilient_multiturn/app.py
• resilient_langgraph/app.py

from azure.ai.agentserver.core.tasks import set_resilient_tasks_enabled
set_resilient_tasks_enabled(True)

Add two minimal LRA samples that depend on only azure-ai-agentserver-core + azure-ai-agentserver-invocations (no LLM, no azure-ai-projects, no langgraph) — a low-footprint on-ramp:

• resilient_hello_world — smallest start-and-poll LRA: a durable @task that counts N steps, checkpointing to a state store after each, and resumes on entry_mode == "recovered".
• resilient_hello_forever — minimal indefinite durable worker: a while True loop with per-iteration checkpointing, graceful-shutdown handling (ctx.shutdown → exit_for_recovery()), and a cancel path.

Verification (local, against core 2.2.0b1)

• All five resilient invocations samples import with the durable subsystem enabled.
• resilient_hello_world and resilient_hello_forever: resume from their checkpoint after a hard kill (recovery re-enters with entry_mode == "recovered" and continues past the crash point rather than restarting).
• resilient_hello_forever: stops on cancel.
• resilient_research: streams end-to-end (token/stage events) once enabled.

Notes

• No package/shipped-code changes — samples only; no CHANGELOG entry needed.
• For contrast, the responses resilient samples already call set_resilient_tasks_enabled(True); this brings the invocations samples in line.

[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]

🟡 Changes recommended

Invocation identity, polling, post-timeout cancellation, and lifecycle test coverage have unresolved moderate issues.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Enables resilient task hosting in existing invocation samples and adds minimal finite and indefinite durable-worker examples.

Changes:

• Enables resilient tasks before host startup in three samples.
• Adds checkpointed start/poll and start/poll/cancel examples.
• Documents recovery behavior and minimal dependencies.

File summaries

File	Description
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py	Enables resilient tasks.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py	Enables resilient tasks.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py	Enables resilient tasks.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/requirements.txt	Declares minimal dependencies.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md	Documents finite resilient execution.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py	Implements start and polling handlers.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/agent.py	Implements checkpointed finite work.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/__init__.py	Initializes the sample package.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/requirements.txt	Declares minimal dependencies.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md	Documents indefinite resilient execution.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py	Implements start, polling, and cancellation handlers.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py	Implements the checkpointed worker.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/__init__.py	Initializes the sample package.

Review details

Suppressed comments (12)

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:26

• This timeout description does not match the task implementation or the cooperative watchdog contract. Expiry only sets ctx.timeout_exceeded and ctx.cancel; the loop deliberately continues and therefore is not re-entered. Reserve recovery here for process crashes and graceful redeploys.

It also sets `timeout=timedelta(days=7)` (the maximum per-turn budget), because
each *turn* is watchdog-bounded (default 1 day). When any interruption occurs —
crash, redeploy, or turn-budget expiry — the task is re-entered with
`ctx.entry_mode == "recovered"` and resumes from its checkpointed iteration.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:61

• Strict opt-in was introduced in core 2.1.0b1 and is present in stable 2.1.0, not first in 2.2.0b1. Update this version reference.

As of `azure-ai-agentserver-core` **2.2.0b1** the durable-task subsystem is
**strictly opt-in**. Before host startup:

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:85

• This description of timeout behavior is incorrect. The core watchdog is cooperative-only: it sets timeout_exceeded and cancel but neither interrupts the handler nor releases its lease, so this branch simply ignores the signal and keeps executing; no recovered re-entry occurs at seven days. Update this comment and the README/module documentation to state that timeout is intentionally ignored, or implement a supported turn-boundary strategy.

            #    A cancel sets ``ctx.cancel`` without ``timeout_exceeded``; the
            #    per-turn watchdog also sets ``ctx.cancel`` but WITH
            #    ``timeout_exceeded`` — in that case we do NOT stop, so the
            #    framework re-enters (recovered) and the loop keeps going.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:50

• The core CHANGELOG places strict opt-in in 2.1.0b1, and stable 2.1.0 includes it. Correct the stated version boundary.

# Resilient tasks (durable execution + crash recovery) are strictly opt-in as of
# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` builds the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:49

• The implementation does not use task metadata or call ctx.metadata.flush(); it writes each checkpoint with FoundryStateStore.set_item(). As written, this explanation directs readers to an API that is absent from the sample.

2. The task loops, sleeping between steps and calling `ctx.metadata.flush()`
   after each — a **durable checkpoint**.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:73

• Recovery reads completed_steps from the explicit FoundryStateStore, not from ctx.metadata. Keep the recovery walkthrough aligned with agent.py so users know where their checkpoint belongs.

   reads `completed_steps` from `ctx.metadata`, and continues from the next

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:75

• The checkpoint remains after task completion, so the polling handler returns completed rather than 404. This second recovery walkthrough should not promise a response that the code cannot produce.

   `completed_steps` climb past where it crashed (then `404` when it finishes).

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:56

• This behavior dates to core 2.1.0b1 and is already in stable 2.1.0, per the core CHANGELOG. Replace 2.2.0b1 so the sample states the correct minimum behavior boundary.

As of `azure-ai-agentserver-core` **2.2.0b1** the durable-task subsystem is
**strictly opt-in**. `AgentServerHost` builds the `TaskManager` (and runs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:47

• Core's CHANGELOG records this opt-in change under 2.1.0b1, and stable 2.1.0 already contains it. Correct the version boundary here.

# Resilient tasks (durable execution + crash recovery) are strictly opt-in as of
# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` builds the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:86

• The strict opt-in behavior was introduced in core 2.1.0b1 and is included in stable 2.1.0, not first introduced in 2.2.0b1. Correct this compatibility statement.

# Resilient tasks (durable execution + crash recovery) are strictly opt-in as of
# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:49

• The strict opt-in behavior was introduced in core 2.1.0b1 and is present in stable 2.1.0, according to the core CHANGELOG. Replace 2.2.0b1 here to avoid documenting the wrong compatibility boundary.

# Resilient tasks (durable execution + crash recovery) are strictly opt-in as of
# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:104

• The core CHANGELOG places the strict opt-in breaking change in 2.1.0b1 (and it is included in stable 2.1.0), not 2.2.0b1. Correct the version so users do not infer that this behavior requires an unreleased 2.2 build.

# Resilient tasks (durable execution + crash recovery) are strictly opt-in as of
# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

• Files reviewed: 11/13 changed files
• Comments generated: 7
• Review effort level: Balanced

---

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

[Nathandrake229]

Thanks for the review — pushed d88c1b9 addressing the inline comments:

Invocation-id keying (app.py, both samples) — both now key the durable task by request.state.invocation_id and address it via GET/POST /invocations/{invocation_id}, dropping the sample-only ?task_id= query param. Docstrings, READMEs, and curl examples updated. Verified a standard client can POST, read the returned invocation_id, and poll /invocations/{invocation_id}.

hello_world 404 right after POST — handle_get now consults get_active_run() when no checkpoint exists yet and returns in_progress with completed_steps: 0 while the run is active. Verified: the immediate GET after POST returns in_progress/0, not 404.

hello_world README 404-on-completion — corrected. The durable checkpoint persists after the ephemeral @task record is cleaned up, so a finished run polls as completed. Removed the wrong 404 contract and a stale ctx.metadata reference.

hello_forever cancel after the timeout watchdog — good catch. The cancel path (get_active_run().cancel()) sets the cancel event but not cancel_requested, and after the per-turn watchdog fires timeout_exceeded stays true, so neither boolean reliably identifies an explicit stop. Replaced the boolean check with a durable stop marker written by the cancel endpoint (separate key, no ETag race with the checkpoint) and checked by the worker — robust across the per-turn timeout and across recovery. Verified: cancel transitions the worker to stopped.

Automated coverage (agent.py comments) — these two samples are currently validated manually against core 2.2.0b1 (start → checkpoint poll → hard-kill recovery → completion for hello_world; start → cancel → stop, and crash → resume for hello_forever). Adding contract/E2E coverage parallel to the existing resilient samples is a reasonable follow-up; happy to add it in this PR if you'd prefer it here rather than a follow-up.

[Copilot]

🟡 Changes recommended

One or more issues must be addressed before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (11)

Previously missed (3) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:67

• A syntactically valid JSON scalar or array reaches data.get and raises AttributeError, producing a 500 response. Reject non-object bodies with a 400 response before reading name.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:66
• A valid non-object JSON body makes data.get raise, a nonnumeric steps value makes int raise, and steps <= 0 completes without ever writing a checkpoint so subsequent polling returns 404. Validate that the body is an object and that steps is positive, returning a client error instead of a server error or an unpollable run.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:21
• This claims turn-budget expiry re-enters the task, but the watchdog is cooperative and this handler keeps looping after ctx.timeout_exceeded is set. Correct the module documentation so it does not promise recovery behavior the implementation cannot provide.

This issue also appears on line 95 of the same file.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:26

• The timeout watchdog is cooperative: it only sets ctx.timeout_exceeded and ctx.cancel; it never interrupts or re-enters the handler. This loop deliberately ignores a timeout without a stop marker, so the documented turn-budget-expiry recovery does not occur. Describe the actual behavior instead.

It also sets `timeout=timedelta(days=7)` (the maximum per-turn budget), because
each *turn* is watchdog-bounded (default 1 day). When any interruption occurs —
crash, redeploy, or turn-budget expiry — the task is re-entered with
`ctx.entry_mode == "recovered"` and resumes from its checkpointed iteration.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:98

• The durable stop marker is only read after this process-local ctx.cancel event is set. If cancellation reaches another replica (where get_active_run() returns None) or the process dies after writing the marker but before signaling the run, recovery starts with a fresh unset event and this worker ignores the persisted stop forever. Check the marker independently on every iteration.

            if ctx.cancel.is_set():
                if await store.get_item(stop_key) is not None:
                    logger.info("stop requested — stopping at iteration %d", n)
                    return {"name": name, "iterations": n, "stopped": True}

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:61

• Existing resilient invocation samples have structural and end-to-end tests, but this new infinite worker has no automated recovery or cancellation coverage. Add it to the structural gate and test both crash resume and cancellation, especially cancellation from a process that does not own the active run.

@task(name="hello_forever", timeout=timedelta(days=7))
async def hello_forever(ctx: TaskContext[dict]) -> dict[str, Any]:

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:73

• Ctrl-C initiates graceful host shutdown, so this procedure does not exercise the hard-crash recovery path described by the sample and PR. Use an actual hard kill so the verification matches the claimed scenario.

2. Poll until `completed_steps` is a few in, then **kill the process** (Ctrl-C).

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/agent.py:45

• The repository has a structural gate plus end-to-end recovery tests for every existing resilient invocation sample, but this new sample is absent from both. Add it to test_resilient_samples_structure.py and cover checkpoint continuation through the crash harness so the core behavior demonstrated here cannot regress.

@task(name="hello_world")
async def hello_world(ctx: TaskContext[dict]) -> dict[str, Any]:

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:91

• These statements are contradictory: when langgraph_session.start() raises TaskManagerNotInitialized, the agent does not continue in a non-durable mode. Clarify that execution cannot start and crash recovery is unavailable.

# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``langgraph_session.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:54

• These statements are contradictory: when session_workflow.start() raises TaskManagerNotInitialized, the agent does not continue in a non-durable mode. Clarify that execution cannot start and crash recovery is unavailable.

# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``session_workflow.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:109

• These statements are contradictory: when deep_research.start() raises TaskManagerNotInitialized, the agent does not continue in a non-durable mode. Clarify that execution cannot start and crash recovery is unavailable.

# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``deep_research.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

• Files reviewed: 11/13 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

Unresolved moderate issues affect durable cancellation, input validation, dependency compatibility, and regression coverage.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (15)

Previously missed (13) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:101

• The durable marker is only read while this process's in-memory cancel event is set. If the process crashes after the endpoint writes the marker but before run.cancel() is delivered, recovery creates a fresh context with ctx.cancel unset and the worker ignores the persisted stop forever. Check the marker independently on every iteration so it actually survives crash/recovery as advertised.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:67
• A syntactically valid JSON body can be a list, string, or null, in which case data.get(...) raises AttributeError and the sample returns a 500. Validate that the decoded body is an object and return a 400 for unsupported shapes.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/requirements.txt:8
• These unbounded requirements can silently resolve to incompatible future major versions. The package itself constrains core to >=2.1.0,<3.0.0 (pyproject.toml:24), and shipped invocation samples constrain invocations to <2.0.0 (for example samples/simple_invoke_agent/requirements.txt:1); apply the same compatibility bounds here.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:66
• steps accepts zero or negative values, so the task completes without ever writing a checkpoint. Because one-shot task records are deleted on completion, the documented poll endpoint then returns 404 for a run that successfully returned 202. Reject non-positive (and non-object/non-integer) input with 400 before starting the task.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/requirements.txt:8
• These unbounded requirements can silently resolve to incompatible future major versions. The package itself constrains core to >=2.1.0,<3.0.0 (pyproject.toml:24), and shipped invocation samples constrain invocations to <2.0.0 (for example samples/simple_invoke_agent/requirements.txt:1); apply the same compatibility bounds here.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:26
• The timeout watchdog does not interrupt or re-enter this handler. It only sets cooperative context flags, which the current loop ignores when no stop marker exists, so turn-budget expiry does not cause the recovery described here.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:60
• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). The README should identify the first affected release correctly.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:21
• Core does not interrupt or re-enter a task when this timeout expires; its watchdog only sets cooperative flags, and this loop continues when no stop marker exists. Rewrite this explanation unless the handler is changed to implement an actual rollover path.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:50
• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). Keeping the wrong version here makes the sample's compatibility guidance inaccurate.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:58
• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). The README should identify the first affected release correctly.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:48
• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). Keeping the wrong version here makes the sample's compatibility guidance inaccurate.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:49
• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). Keeping the wrong version here misleads users about which released core versions require this call.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:104
• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). Keeping the wrong version here misleads users about which released core versions require this call.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:86

• The opt-in behavior was introduced in core 2.1.0b1, not 2.2.0b1 (see azure-ai-agentserver-core/CHANGELOG.md:28-44). Keeping the wrong version here misleads users about which released core versions require this call.

# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:89

• Core's CHANGELOG records the strict opt-in under 2.1.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-44), not 2.2.0b1. Also, without the switch langgraph_session.start() raises before work begins; it does not run non-durably. Correct both claims so the sample accurately identifies affected users and failure behavior.

# Resilient tasks (durable execution + crash recovery) are strictly opt-in as of
# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the
# ``TaskManager`` ONLY when this switch is on, and it must be set before host
# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``langgraph_session.start()`` raise ``TaskManagerNotInitialized`` and the agent

• Files reviewed: 14/16 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

A critical cross-replica cancellation defect and unresolved request-validation and dependency issues remain.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (16)

Previously missed (6) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:67

• A valid non-object JSON body such as [] or null reaches data.get and becomes a 500 response. Validate that the decoded body is an object and return 400 for unsupported shapes.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/requirements.txt:8
• These unconstrained requirements can be satisfied by an older installed package that lacks the local state-store/task behavior this no-cloud sample needs. Use the package's supported release ranges, consistent with pyproject.toml:24 and the bounded requirements in samples/simple_invoke_agent/requirements.txt:1.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:66
• Invalid request bodies are either silently accepted as defaults (malformed JSON) or turn into 500s ([], non-numeric steps), and non-positive steps produce no checkpoint so later polls return 404. Validate the JSON object and positive integer before starting the task, returning 400 for invalid input.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/requirements.txt:8
• These unconstrained requirements can be satisfied by an older installed package that lacks the local state-store/task behavior this no-cloud sample needs. Use the package's supported release ranges, consistent with pyproject.toml:24 and the bounded requirements in samples/simple_invoke_agent/requirements.txt:1.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:26
• Turn-budget expiry does not cause re-entry: the watchdog is cooperative-only and this handler deliberately continues after its signal. This walkthrough therefore promises a recovery mode that users will not observe.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:21
• Core's watchdog is cooperative-only and explicitly does not end or re-enter an ignoring handler (tasks/_manager.py:1613-1616). This description incorrectly promises recovery at turn-budget expiry; the current loop simply continues in the same turn after the signal.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:60

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

As of `azure-ai-agentserver-core` **2.2.0b1** the durable-task subsystem is

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:50

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` builds the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:58

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

As of `azure-ai-agentserver-core` **2.2.0b1** the durable-task subsystem is

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:48

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` builds the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:86

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:91

• A direct @task.start() call does not degrade to non-durable execution when the manager is disabled; it raises and the task does not run, as core's contract states. Remove the claim that this agent still runs non-durably.

# ``langgraph_session.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:49

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:54

• A direct @task.start() call does not degrade to non-durable execution when the manager is disabled; it raises and the task does not run, as core's contract states. Remove the claim that this agent still runs non-durably.

# ``session_workflow.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:104

• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28-40). Stating the later version can mislead users of 2.1.x into omitting the required switch.

# azure-ai-agentserver-core 2.2.0b1: ``AgentServerHost`` constructs the

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:109

• A direct @task.start() call does not degrade to non-durable execution when the manager is disabled; it raises and the task does not run, as core's contract states. Remove the claim that this agent still runs non-durably.

# ``deep_research.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

• Files reviewed: 14/16 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

Critical checkpoint isolation and multiple correctness, dependency, and test issues remain unresolved.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (10)

Previously missed (6) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:102

• A successful POST can still be followed by a transient 404: start() schedules the worker, but the worker awaits store creation/read before writing its first checkpoint, so this branch can run before either item exists. This is especially visible when polling is routed to another replica. Persist an initial durable iterations: 0 record as part of starting the invocation so every replica can distinguish a just-started worker from an unknown ID.

This issue also appears in the following locations of the same file:

• line 107
• line 124
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/requirements.txt:8

• These unbounded requirements can resolve to a future incompatible major release, contrary to the package's sample dependency convention (samples/simple_invoke_agent/requirements.txt:1) and the runtime Core bound (pyproject.toml:24). Keep the sample dependency-light while constraining it to the compatible release lines.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:66
• steps <= 0 starts successfully but executes no loop iteration, so no checkpoint is ever created; after the ephemeral task record is removed, every poll returns 404 instead of completed. Reject non-positive step counts before starting the task (or persist a completion checkpoint for zero steps).
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/requirements.txt:8
• These unbounded requirements can resolve to a future incompatible major release, contrary to the package's sample dependency convention (samples/simple_invoke_agent/requirements.txt:1) and the runtime Core bound (pyproject.toml:24). Keep the sample dependency-light while constraining it to the compatible release lines.
  sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_forever.py:34
• This collection-time environment mutation is process-global and is never restored, so merely collecting this module can make unrelated tests run as non-hosted. The fixture already removes this variable with monkeypatch, which restores it after each test; remove the module-level mutation.
  sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_world.py:37
• This collection-time environment mutation is process-global and is never restored, so merely collecting this module can make unrelated tests run as non-hosted. The fixture already removes this variable with monkeypatch, which restores it after each test; remove the module-level mutation.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:26

• The timeout claim is incorrect for this implementation. Task timeouts are cooperative: the watchdog only sets ctx.timeout_exceeded and ctx.cancel, while this loop ignores both, so budget expiry does not interrupt or re-enter the task. Document that the worker intentionally continues after the watchdog, or add explicit timeout handling consistent with the desired lifecycle.

It also sets `timeout=timedelta(days=7)` (the maximum per-turn budget), because
each *turn* is watchdog-bounded (default 1 day). When any interruption occurs —
crash, redeploy, or turn-budget expiry — the task is re-entered with
`ctx.entry_mode == "recovered"` and resumes from its checkpointed iteration.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/agent.py:52

• FoundryStateStore is agent-scoped by default and explicitly has no built-in session isolation, but this constant shares checkpoints and stop markers across every session. A caller-supplied invocation ID can therefore collide with another session, including making one session's cancel marker stop another worker. Scope the store name by session and persist that session ID in the task input for recovery.

CHECKPOINT_STORE = "hello_forever_checkpoints"

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:107

• The stop marker records that cancellation was requested, not that the worker has terminated. handle_cancel writes it before run.cancel(), so an immediate poll reports stopped while the owner may still be sleeping or executing another iteration. Use separate durable stop_requested and terminal stopped state (the worker should write the latter immediately before returning), and report cancelling until that acknowledgement exists.

            "status": "stopped" if stop_marker is not None else "running",

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:126

• This creates a durable stop marker even when the invocation never existed. Consequently POST /unknown/cancel returns success and the subsequent GET reports a fabricated stopped invocation, whereas the invocations contract tests require unknown cancellation to return 404. Verify the durable task/checkpoint before writing, and also reject an already-present stop marker.

    store = await FoundryStateStore.get_or_create(CHECKPOINT_STORE)
    try:
        await store.set_item(f"{invocation_id}{STOP_SUFFIX}", {"stop": True})

• Files reviewed: 14/16 changed files
• Comments generated: 3
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

Malformed byte bodies can return 500, and the structural test enforces an invalid lifecycle ordering.

Review details

Suppressed comments (12)

Previously missed (10) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:81

• json.loads() raises UnicodeDecodeError for a non-UTF-8 byte body, so this handler returns 500 for malformed input despite its 400 contract. Catch that exception alongside JSONDecodeError.

This issue also appears on line 112 of the same file.
sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:68

• json.loads() raises UnicodeDecodeError for a non-UTF-8 byte body, so this handler returns 500 for malformed input despite its 400 contract. Catch that exception alongside JSONDecodeError.

This issue also appears on line 111 of the same file.
sdk/agentserver/azure-ai-agentserver-invocations/tests/test_resilient_samples_structure.py:331

• This ordering requirement does not match the host lifecycle. AgentServerHost reads the flag inside its ASGI lifespan (core/_base.py:312-350), not during InvocationAgentServerHost() construction, so app = InvocationAgentServerHost(); set_resilient_tasks_enabled(True) is valid but this test rejects it. Keep the module-level-call check, but remove the constructor-order assertion and update the surrounding explanation.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/README.md:68
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:66
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:62
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:53
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:86
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:49
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:104
• The opt-in breaking change shipped in core 2.1.0b1, not 2.2.0b1 (azure-ai-agentserver-core/CHANGELOG.md:28,36-45). Using the later version misleads users of 2.1.0b1 into thinking the switch is unnecessary.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:114

• After cancellation, hello_forever returns normally, so its one-shot TaskManager record is deleted, but this non-expiring checkpoint and stop marker remain. A retry with the same invocation ID then swallows this conflict, starts a fresh task, returns 202, and that task immediately exits on the stale marker. Detect the existing stop marker here and return the already-stopped invocation instead of starting a new worker.

        await store.create_item(task_id, {"name": name, "iterations": 0})
    except FoundryStorageConflictError:
        pass  # already seeded by a concurrent request or a previous attempt

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:115

• A completed one-shot task is deleted by the framework (core/tasks/_manager.py:2656-2660), while this checkpoint remains. A retry of the same invocation ID after completion therefore reaches this conflict branch, then successfully creates a new task instead of hitting TaskConflictError; it can return 202 started while doing no work, or continue the old checkpoint with different input. Treat an already-completed checkpoint as the existing invocation rather than starting another task.

        await store.create_item(
            task_id, {"name": name, "steps": steps, "completed_steps": 0}
        )
    except FoundryStorageConflictError:
        pass  # already seeded by a concurrent request or a previous attempt

• Files reviewed: 15/17 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

Moderate correctness and test-isolation issues must be resolved before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (6)

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/requirements.txt:8

• These unbounded requirements can resolve incompatible future major releases. This package explicitly constrains sample dependencies (CHANGELOG.md:40-43), and neighboring published-package samples use bounded ranges (for example samples/basic_voice_agent/requirements.txt:1); add compatible minimum and upper bounds here as well.

azure-ai-agentserver-core
azure-ai-agentserver-invocations

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:76

• Ctrl-C performs graceful host shutdown, not a hard crash. This task never checks ctx.shutdown or calls exit_for_recovery(), so after the shutdown grace period TaskManager cancels it and deletes the ephemeral one-shot task record (core/tasks/_manager.py:2076-2101); a restart then has nothing to recover. Use an actual hard kill in this crash-recovery walkthrough, or add graceful-shutdown handling to the task.

2. Poll until `completed_steps` is a few in, then **kill the process** (Ctrl-C).

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/requirements.txt:8

• These unbounded requirements can resolve incompatible future major releases. This package explicitly constrains sample dependencies (CHANGELOG.md:40-43), and neighboring published-package samples use bounded ranges (for example samples/basic_voice_agent/requirements.txt:1); add compatible minimum and upper bounds here as well.

azure-ai-agentserver-core
azure-ai-agentserver-invocations

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:91

• This fallback description is incorrect for invocations: get_task_manager() raises and this .start() call is not caught, so the request fails rather than running the agent non-durably (core/tasks/_manager.py:289-306). Reserve the non-durable fallback wording for protocols that explicitly swallow this exception.

# ``langgraph_session.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:54

• This fallback description is incorrect for invocations: get_task_manager() raises and this .start() call is not caught, so the request fails rather than running the agent non-durably (core/tasks/_manager.py:289-306). Reserve the non-durable fallback wording for protocols that explicitly swallow this exception.

# ``session_workflow.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:109

• This fallback description is incorrect for invocations: get_task_manager() raises and this .start() call is not caught, so the request fails rather than running the agent non-durably (core/tasks/_manager.py:289-306). Reserve the non-durable fallback wording for protocols that explicitly swallow this exception.

# ``deep_research.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

• Files reviewed: 15/17 changed files
• Comments generated: 17
• Review effort level: Balanced

[Copilot]

Note

Copilot was unable to run its full agentic suite in this review.

Pull request overview

Copilot reviewed 15 out of 17 changed files in this pull request and generated 5 comments.

[Copilot]

🟡 Changes recommended

One or more issues must be addressed before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (7)

Previously missed (2) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/requirements.txt:8

• Bound both dependencies to compatible release lines. Unbounded requirements make this sample resolve future incompatible major versions, contrary to the package convention demonstrated by samples/basic_voice_agent/requirements.txt:1 and recorded in CHANGELOG.md:42-43.
  sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/requirements.txt:8
• Bound both dependencies to compatible release lines. Unbounded requirements make this sample resolve future incompatible major versions, contrary to the package convention demonstrated by samples/basic_voice_agent/requirements.txt:1 and recorded in CHANGELOG.md:42-43.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:234

• This treats every existing checkpoint as an active worker. If the checkpoint already has status == "failed", cancel writes a stop marker; subsequent GETs prioritize that marker and incorrectly change the terminal result from failed to stopped. An already stopped worker is also reported as cancelling. Check the durable status and existing marker first, and return the existing terminal state without writing a new marker.

        if await store.get_item(task_id) is None:

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:76

• Ctrl-C initiates the host's graceful lifespan shutdown, but this task never checks ctx.shutdown or calls exit_for_recovery(). After the shutdown grace period, TaskManager cancels the execution and deletes the ephemeral task record, so the restart described here has nothing to recover. Either add graceful-shutdown handling to the task or instruct users to perform an actual hard kill.

2. Poll until `completed_steps` is a few in, then **kill the process** (Ctrl-C).

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:91

• This says the agent runs non-durably, but langgraph_session.start() raises TaskManagerNotInitialized, so no task execution occurs. Reserve the non-durable fallback wording for protocol paths that explicitly catch this exception.

# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``langgraph_session.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:54

• This says the agent runs non-durably, but session_workflow.start() raises TaskManagerNotInitialized, so no task execution occurs. Reserve the non-durable fallback wording for protocol paths that explicitly catch this exception.

# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``session_workflow.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:109

• This says the agent runs non-durably, but the preceding call raises TaskManagerNotInitialized, so the durable task does not run at all. Reserve the non-durable fallback wording for protocol paths that explicitly swallow this exception.

# ``deep_research.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

• Files reviewed: 15/17 changed files
• Comments generated: 3
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

Moderate issues remain in start-failure reconciliation, checkpoint cursor handling, and dependency constraints.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (5)

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:140

• The checkpoint is created before start(), but only TaskConflictError is handled. If provider access or scheduling raises any other exception, the request returns 500 while the durable record remains running; retries are permanently rejected and polling reports a worker that does not exist. Reconcile the seed on every start failure, while handling the possibility that a remote start was accepted before a transport error.

    try:
        await hello_forever.start(
            task_id=task_id,

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:76

• The documented Ctrl-C path does not demonstrate crash recovery for this task. TaskManager.shutdown() signals graceful shutdown, waits up to its grace period, then cancels handlers that do not call exit_for_recovery(); cancellation deletes the ephemeral task record, leaving nothing for the startup recovery scan. Use an abrupt process/container kill in these instructions, or make the task shutdown-aware.

2. Poll until `completed_steps` is a few in, then **kill the process** (Ctrl-C).

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/agent.py:181

• completed is captured only at task entry and is never advanced by the loop. If a later checkpoint or terminal write raises after earlier steps succeeded, this handler overwrites the durable cursor with that stale entry value, so polling regresses and the recorded resume point is lost. Use the just-fetched current item as the failure cursor.

                        "completed_steps": completed,

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/app.py:144

• The checkpoint is created before start(), but only TaskConflictError is handled. If provider access or scheduling raises any other exception, the request returns 500 while this durable item remains in_progress; retries then hit the existence gate with 409 and polling reports a task that never started. Reconcile the seed on every start failure (for example, persist an explicit failed-to-start state or safely roll it back while accounting for an ambiguous remote accept).

    try:
        await hello_world.start(
            task_id=task_id,

sdk/agentserver/azure-ai-agentserver-invocations/tests/test_resilient_samples_structure.py:56

• This makes resilient_cancellable part of the required shipping surface, but the PR description says that exactly two minimal samples are being added and does not mention or verify this third sample. Please either update the PR scope/verification to include the cancellable sample or remove it from this change.

    "resilient_cancellable",

• Files reviewed: 20/23 changed files
• Comments generated: 6
• Review effort level: Balanced

[github-actions]

[Pilot] PR Pipeline Failure Analysis

What failed

Azure Pipeline build 6788933 (python - pullrequest / "Build Analyze" job) failed at the CSpell spell-check step. Tests passed for the new resilient_hello_world / resilient_hello_forever samples and their e2e tests (all 20 collected tests in this artifact passed), but CSpell flagged the word fstore as unknown in 5 locations within sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_handlers.py (lines 157, 158, 159, 235, 236), causing the job to exit non-zero.

Relevant pipeline output

CSpell: Files checked: 5, skipped: 15, Issues found: 5 in 1 file.
##[error]sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_handlers.py:157:5 - Unknown word (fstore)
##[error]sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_handlers.py:158:16 - Unknown word (fstore)
##[error]sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_handlers.py:159:15 - Unknown word (fstore)
##[error]sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_handlers.py:235:5 - Unknown word (fstore)
##[error]sdk/agentserver/azure-ai-agentserver-invocations/tests/e2e/test_resilient_hello_handlers.py:236:16 - Unknown word (fstore)
##[error]Spelling errors detected. To correct false positives or learn about spell checking see: https://aka.ms/azsdk/engsys/spellcheck
##[error]PowerShell exited with code '1'.

Recommended next steps

• Rename the identifier fstore to a dictionary-recognized word (e.g. store or file_store) in test_resilient_hello_handlers.py, or add fstore to the repository's CSpell custom word list (e.g. .vscode/cspell.json or a root word-list file) if the abbreviation should be preserved.
• This is a pure spell-check/lint failure — no test logic, build, or compilation errors were found; all 20 tests in the artifact passed before CSpell ran.
• See https://aka.ms/ci-fix

Automated fix: Requested

Generated by Pipeline Analysis Next Steps · auto · 42.6 AIC · ⌖ 2.84 AIC · ⊞ 9.6K · ◷

[Copilot]

🟡 Changes recommended

Six critical and four moderate findings remain in persistence concurrency, local storage, recovery, and cancellation behavior.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (3)

Previously missed (2) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:262

• This writes a stop marker even when the durable record is already failed or already stopped. Because handle_get gives the marker precedence, cancelling a failed worker permanently changes its visible status from failed to stopped and hides its error; repeated cancellation also reports cancelling for a terminal worker. Check the persisted status and existing marker before writing, and return the existing terminal state instead.
  sdk/agentserver/azure-ai-agentserver-invocations/tests/test_resilient_samples_structure.py:56
• This introduces resilient_cancellable as a third complete sample, but the PR description and verification plan say that only resilient_hello_world and resilient_hello_forever are being added. Please either document and verify this additional scope or move it to a separate change.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:83

• Ctrl-C initiates graceful host shutdown rather than a crash. This task does not observe ctx.shutdown; it either finishes during the grace period or is cancelled afterward, and one-shot cancellation deletes its TaskManager record, so the documented restart will have nothing to recover. Tell users to use a hard kill here, consistent with the cancellable sample's recovery instructions.

2. Poll until `completed_steps` is a few in, then **kill the process** (Ctrl-C).

• Files reviewed: 20/23 changed files
• Comments generated: 9
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

The stop-state reporting and resilient-task opt-in regression guard have unresolved moderate-severity correctness issues.

Review details

Suppressed comments (8)

Previously missed (2) — in code that hasn't changed since the last review.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_forever/app.py:228

• The stop marker records a stop request, not that the worker has exited. handle_cancel can write it while hello_forever is still sleeping (and it may checkpoint another iteration), so this immediately reports terminal stopped; it also masks a persisted failed status because the marker wins. Persist status: stopped in the worker before returning, report cancelling while only the marker exists, and preserve failed precedence.
  sdk/agentserver/azure-ai-agentserver-invocations/tests/test_resilient_samples_structure.py:290
• This accepts any literal True anywhere in the call, so invalid shapes such as set_resilient_tasks_enabled(value=False, typo=True) satisfy the regression guard; the documented enabled=True keyword would itself raise because the public parameter is named value. Match exactly one positional True or value=True.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_cancellable/README.md:1

• The PR description says this change adds two minimal samples (resilient_hello_world and resilient_hello_forever), but this introduces a third substantial cancellable sample and associated tests. Please update the title/description and change list to reflect this scope, or split this sample into a separate PR.

# Minimal resilient long-running agent — **cancellation**

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_hello_world/README.md:83

• Ctrl-C initiates graceful host shutdown rather than the hard crash this procedure intends to demonstrate. Because hello_world does not handle ctx.shutdown with exit_for_recovery, it can continue during the shutdown grace period, making this recovery walkthrough nondeterministic; use SIGKILL as in the verified scenario.

2. Poll until `completed_steps` is a few in, then **kill the process** (Ctrl-C).

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_langgraph/app.py:91

• These statements contradict each other: when get_task_manager()/.start() raises TaskManagerNotInitialized, this invocation agent does not run non-durably—it does not start. Describe the fail-fast behavior so users do not expect an in-process fallback that this API does not provide.

# ``TaskManager`` ONLY when this switch is on, and it must be set before host
# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``langgraph_session.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_multiturn/app.py:54

• These statements contradict each other: when get_task_manager()/.start() raises TaskManagerNotInitialized, this invocation agent does not run non-durably—it does not start. Describe the fail-fast behavior so users do not expect an in-process fallback that this API does not provide.

# ``TaskManager`` ONLY when this switch is on, and it must be set before host
# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``session_workflow.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/samples/resilient_research/app.py:109

• These statements contradict each other: when get_task_manager()/.start() raises TaskManagerNotInitialized, this invocation agent does not run non-durably—it does not start. Describe the fail-fast behavior so users do not expect an in-process fallback that this API does not provide.

# startup (i.e. at module-import time). Without it, ``get_task_manager()`` /
# ``deep_research.start()`` raise ``TaskManagerNotInitialized`` and the agent
# runs non-durably with no crash recovery — defeating the purpose of a resilient
# long-running agent.

sdk/agentserver/azure-ai-agentserver-invocations/tests/test_resilient_samples_structure.py:56

• The PR description says this adds two minimal samples and verifies five resilient samples, but this third sample plus its implementation/tests makes six. Update the title, description, and verification scope to include resilient_cancellable, or remove it from this PR, so the reviewed scope matches what ships.

    "resilient_cancellable",

• Files reviewed: 20/23 changed files
• Comments generated: 0 new
• Review effort level: Balanced