feat(plugin): surface execution input and result by wangyb-A · Pull Request #616 · aws/aws-durable-execution-sdk-python

wangyb-A · 2026-08-06T22:11:55Z

Summary

Widens the plugin invocation hooks so instrumentation plugins can record execution
input/output:

InvocationInfo.execution_input: Any — the deserialized execution input, populated by
durable_execution() from the invocation's input event
InvocationEndInfo.execution_result: str | None — the serialized result JSON from
DurableExecutionInvocationOutput.result (None on failure/suspend), threaded through
from_durable_execution_invocation_output

Both fields are kw-only with None defaults, so existing plugins, hook constructors,
and positional callers are unaffected — this is a purely additive API widening.

Motivation

The JS SDK already exposes this data to plugins (InvocationInfo.executionInput /
InvocationEndInfo.executionResult), and its Workflow Insight plugin depends on it to emit
execution records carrying input/output. Porting that plugin to Python (validated
against the cross-SDK insight conformance suite,
aws-durable-execution-conformance-tests#73) is blocked without these fields — the plugin
cannot honestly emit data the SDK never hands it. The Java SDK is adding the analogous
plugin accessor in aws-durable-execution-sdk-java#596.

Changes

File	Change
`src/aws_durable_execution_sdk_python/plugin.py`	the two kw-only fields + docstrings; `from_durable_execution_invocation_output` carries them; `PluginExecutor` start-info plumbing
`src/aws_durable_execution_sdk_python/execution.py`	one line: pass `input_event` into the invocation-start plugin info
`tests/plugin_test.py`	fixtures extended; new tests: default-None backward compatibility, and the end-info factory carrying input+result

Testing

hatch run test:all packages/aws-durable-execution-sdk-python/tests/plugin_test.py — 70 passed
hatch run test:all packages/aws-durable-execution-sdk-python/tests/execution_test.py — 82 passed
hatch fmt --check clean on all touched files
End-to-end: with these fields, the preview Python Workflow Insight plugin passes the
18-requirement insight conformance suite 18/18 on both record sinks (live, us-west-2)

The handler docstring claimed the Python invocation infos have no execution-input or execution-result field. They do now; the canonical dump still omits them because the requirement puts the payload surfaces out of GA scope and asserts nothing about them. Refs #616

Addresses both AI review comments on #616. Adds two durable_execution()-driven tests asserting the SDK actually forwards the deserialized input to the invocation hooks, plus the serialized result on the end hook. Verified as real regression guards: both fail when execution_input=input_event is removed. Also corrects the execution_input field comment. It claimed None for empty input, but execution.py initializes input_event to {} and only replaces it when a non-empty payload parses, so an empty payload is surfaced as {}. None occurs only when the field is left unpopulated. Refs #616

Convert the execution_input / execution_result field comments to attribute docstrings prefixed EXPERIMENTAL, matching the wording of the existing marker on the plugins= parameter in execution.py. Unlike # comments these render in IDE hover and doc tooling, so plugin authors see the stability caveat at the point of use. Content is otherwise unchanged, including the corrected {}-vs-None note. Purely documentation: the fields keep their None defaults and kw_only=True. Refs #616

Adds InvocationInfo.execution_input (the deserialized execution input) and InvocationEndInfo.execution_result (the serialized result JSON) so instrumentation plugins can record execution I/O. Both are kw-only with None defaults, so existing plugins and callers are unaffected. Mirrors the JS SDK's InvocationInfo.executionInput / InvocationEndInfo .executionResult (the Java SDK is adding the analogous accessor in aws-durable-execution-sdk-java#596).

Addresses both AI review comments on #616. Adds two durable_execution()-driven tests asserting the SDK actually forwards the deserialized input to the invocation hooks, plus the serialized result on the end hook. Verified as real regression guards: both fail when execution_input=input_event is removed. Also corrects the execution_input field comment. It claimed None for empty input, but execution.py initializes input_event to {} and only replaces it when a non-empty payload parses, so an empty payload is surfaced as {}. None occurs only when the field is left unpopulated. Refs #616

Convert the execution_input / execution_result field comments to attribute docstrings prefixed EXPERIMENTAL, matching the wording of the existing marker on the plugins= parameter in execution.py. Unlike # comments these render in IDE hover and doc tooling, so plugin authors see the stability caveat at the point of use. Content is otherwise unchanged, including the corrected {}-vs-None note. Purely documentation: the fields keep their None defaults and kw_only=True. Refs #616

Addresses both Codex review comments on #616. durable_execution() handed the same mutable object to the user handler and to on_invocation_start, so the aliasing ran both ways: a plugin mutating info.execution_input changed the handler's event and could alter execution behaviour, and a handler mutating its event changed what the frozen start info -- and the end info derived from it -- reported afterwards. Both directions were confirmed reproducible. PluginExecutor now deep-copies the input for the plugin view. The copy is eager because the handler starts immediately after the hook, and is skipped when no plugins are registered so non-plugin executions pay nothing. On copy failure it falls back to the shared reference rather than dropping the input. Adds unit tests for both mutation directions (verified to fail without the copy) and an e2e suite under tests/e2e/ per AGENTS.md, covering the input and serialized result through complete invocations, a suspend/replay pair where only the replay carries a result, and deep isolation from the handler. Refs #616

Return None rather than the shared reference when the execution input cannot be deep-copied, so handler isolation holds even when a snapshot cannot be created. Applied from an AI review suggestion. Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Mark execution_input and execution_result with metadata={"experimental": True} and a leading EXPERIMENTAL docstring, matching the pattern #625 established for OperationInfo.result / OperationInfo.error / InvocationEndInfo.error. Unlike a docstring alone the metadata is introspectable at runtime, so the generic test #625 added -- which requires the marker on every plugin info field named for input, output, result or error -- now covers these two fields. That test fails on this branch without them. This matters more since #625 removed the plugins= FutureWarning: the per-field markers are now the only stability signal on these surfaces. Refs #616

The handler docstring claimed the Python invocation infos have no execution-input or execution-result field. They do now; the canonical dump still omits them because the requirement puts the payload surfaces out of GA scope and asserts nothing about them. Refs #616

github-actions · 2026-08-11T18:48:45Z

+    execution_result: str | None = field(
+        default=None,
+        kw_only=True,
+        metadata={"experimental": True},
+    )


Codex AI review

High: Dataclass fields appear in repr by default. The bundled OTel plugins already log invocation info objects wholesale, and the plugin example does so at INFO, so arbitrary inputs/results, including secrets or large payloads, will now be logged implicitly. Set repr=False on both payload fields and add a regression test ensuring their values are omitted from repr.

Looks concerning

github-actions · 2026-08-11T18:48:46Z

+    execution_input: Any = field(
+        default=None,
+        kw_only=True,
+        metadata={"experimental": True},
+    )


Codex AI review

Medium: This field participates in the generated dataclass equality and hash. Normal dict/list inputs make previously hashable InvocationStartInfo objects raise TypeError when hashed, while comparisons against objects constructed using the prior fields now fail. Preserve additive compatibility by setting compare=False, hash=False on both new payload fields and test those semantics.

github-actions · 2026-08-11T18:48:48Z

Codex AI review

Two regressions remain in the new public hook fields; OTel logging and dataclass compatibility need coverage.

Reviewed commit d92013f69d7b92a984d67ea98683dfd62d7785b1. Workflow run

zhongkechen

Requires a follow-up to resolve the concerning issue.

Addresses both AI review comments on #616. Adds two durable_execution()-driven tests asserting the SDK actually forwards the deserialized input to the invocation hooks, plus the serialized result on the end hook. Verified as real regression guards: both fail when execution_input=input_event is removed. Also corrects the execution_input field comment. It claimed None for empty input, but execution.py initializes input_event to {} and only replaces it when a non-empty payload parses, so an empty payload is surfaced as {}. None occurs only when the field is left unpopulated. Refs #616

Convert the execution_input / execution_result field comments to attribute docstrings prefixed EXPERIMENTAL, matching the wording of the existing marker on the plugins= parameter in execution.py. Unlike # comments these render in IDE hover and doc tooling, so plugin authors see the stability caveat at the point of use. Content is otherwise unchanged, including the corrected {}-vs-None note. Purely documentation: the fields keep their None defaults and kw_only=True. Refs #616

Addresses both Codex review comments on #616. durable_execution() handed the same mutable object to the user handler and to on_invocation_start, so the aliasing ran both ways: a plugin mutating info.execution_input changed the handler's event and could alter execution behaviour, and a handler mutating its event changed what the frozen start info -- and the end info derived from it -- reported afterwards. Both directions were confirmed reproducible. PluginExecutor now deep-copies the input for the plugin view. The copy is eager because the handler starts immediately after the hook, and is skipped when no plugins are registered so non-plugin executions pay nothing. On copy failure it falls back to the shared reference rather than dropping the input. Adds unit tests for both mutation directions (verified to fail without the copy) and an e2e suite under tests/e2e/ per AGENTS.md, covering the input and serialized result through complete invocations, a suspend/replay pair where only the replay carries a result, and deep isolation from the handler. Refs #616

github-actions · 2026-08-11T19:20:36Z

Claude AI review

Summary

This is a clean, purely additive API widening. InvocationInfo.execution_input and InvocationEndInfo.execution_result are added as kw-only fields with None defaults, durable_execution() passes input_event into the start hook, and from_durable_execution_invocation_output threads both through to the end info.

I found no correctness, regression, determinism, or compatibility defects.

Verification performed:

API compatibility: every construction of InvocationInfo/InvocationStartInfo/InvocationEndInfo across the core, otel, testing, examples, and conformance packages uses keyword arguments. The new kw-only defaulted fields leave the positional __init__ signature unchanged (the pre-existing non-kw error field and kw-only status are undisturbed), so no positional caller breaks.
execution_result semantics: map correctly on every terminal path — small success → JSON string, large-payload success → "" (checkpointed out-of-band), suspend/PENDING → None, user failure and RETRY → None — confirmed by tracing to_dict/from_dict round-tripping in PluginExecutor.handle_durable_output.
Determinism / side-effect safety: the eager copy.deepcopy snapshot is taken synchronously on the handler thread before the user function is submitted, and the handler always receives the original input_event. This correctly prevents a plugin from mutating the input and altering execution behavior, and prevents the handler from retroactively changing the frozen hook view. self._invocation_status is reset in run()'s finally, so no cross-invocation leak.
Robustness: on_invocation_end still short-circuits when on_invocation_start was never reached (early non-retryable-fetch returns), so no new crash path.

Test coverage for the new behavior is strong (forwarding to both hooks, empty payload → {}, top-level and nested isolation, suspend-no-result, replay-with-result, and the end-info factory).

Residual test risk (non-blocking): The except Exception fallback in plugin.py _snapshot_execution_input (log + return None) has no covering test. It is effectively unreachable for json.loads output (which is always deep-copyable and acyclic), so this is a defensive branch rather than a real gap; worth a note only because dev-core:cov enforces --cov-fail-under=98. Additionally, the execution_input docstring states durable_execution() "always populates it" — technically the deepcopy-failure fallback can yield None — but this is a documented edge and not a behavioral defect.

Reviewed commit d92013f69d7b92a984d67ea98683dfd62d7785b1. Workflow run

Addresses the medium-severity Codex review comment on #616. The payload fields joined the generated __eq__ and __hash__, so the widening was not additive. execution_input holds arbitrary deserialized JSON, and a dict or list value made a previously hashable InvocationStartInfo raise TypeError on hash(); both fields also made infos built from the earlier field set compare unequal to infos carrying a payload. Both effects were reproduced first. Set compare=False, hash=False on execution_input and execution_result. Identity fields still drive equality, so payload-only differences now compare equal -- payloads are incidental data on what is otherwise an event record. Adds tests for hashability across dict, list, nested and scalar payloads, for equality against the prior field set, and for the field declarations themselves. Note OperationInfo.result / OperationInfo.error from #625 remain in compare. They are hashable types so they do not break hash(), but the equality asymmetry with these fields is worth a maintainer decision. Refs #616

Addresses the high-severity Codex review comment on #616. Dataclass fields land in the generated repr, and instrumentation logs hook infos wholesale: the bundled OTel plugins at debug level, and the plugin example at info. Customer input and results -- potentially secrets, potentially megabytes -- would therefore be written to logs implicitly, just by adding these fields. Reproduced before fixing. Set repr=False on execution_input and execution_result. Identity fields still render, so the repr stays useful, and the values remain readable through the attributes for plugins that deliberately record them. Adds a regression test asserting secret-looking values never appear in either hook info's repr, plus one pinning the field declarations. Refs #616

Addresses the medium-severity Codex review comment on #616. The payload fields joined the generated __eq__ and __hash__, so the widening was not additive. execution_input holds arbitrary deserialized JSON, and a dict or list value made a previously hashable InvocationStartInfo raise TypeError on hash(); both fields also made infos built from the earlier field set compare unequal to infos carrying a payload. Both effects were reproduced first. Set compare=False, hash=False on execution_input and execution_result. Identity fields still drive equality, so payload-only differences now compare equal -- payloads are incidental data on what is otherwise an event record. Adds tests for hashability across dict, list, nested and scalar payloads, for equality against the prior field set, and for the field declarations themselves. Note OperationInfo.result / OperationInfo.error from #625 remain in compare. They are hashable types so they do not break hash(), but the equality asymmetry with these fields is worth a maintainer decision. Refs #616

This was referenced Aug 10, 2026

feat(plugin): operation maps and payloads on invocation hooks #622

Closed

feat(plugin): operation maps on invocation hooks #623

Closed

wangyb-A force-pushed the feat/plugin-invocation-input-result branch from d1f3d59 to 05c3de7 Compare August 11, 2026 05:10

wangyb-A had a problem deploying to ai-pr-review August 11, 2026 05:10 — with GitHub Actions Error

wangyb-A changed the title ~~feat(plugin): surface execution input and result on invocation hooks~~ feat(plugin): surface execution input and result Aug 11, 2026

wangyb-A closed this Aug 11, 2026

wangyb-A reopened this Aug 11, 2026

wangyb-A had a problem deploying to ai-pr-review August 11, 2026 05:13 — with GitHub Actions Error

wangyb-A force-pushed the feat/plugin-invocation-input-result branch from 05c3de7 to 1060dde Compare August 11, 2026 05:26

wangyb-A deployed to ai-pr-review August 11, 2026 05:26 — with GitHub Actions Active

wangyb-A temporarily deployed to ai-pr-review-runtime August 11, 2026 05:33 — with GitHub Actions Inactive

github-actions Bot reviewed Aug 11, 2026

View reviewed changes

Comment thread packages/aws-durable-execution-sdk-python/src/aws_durable_execution_sdk_python/execution.py

github-actions Bot reviewed Aug 11, 2026

View reviewed changes

Comment thread packages/aws-durable-execution-sdk-python/src/aws_durable_execution_sdk_python/plugin.py Outdated

This comment has been minimized.

Sign in to view

wangyb-A had a problem deploying to ai-pr-review August 11, 2026 05:50 — with GitHub Actions Error

wangyb-A deployed to ai-pr-review August 11, 2026 06:20 — with GitHub Actions Active

wangyb-A temporarily deployed to ai-pr-review-runtime August 11, 2026 06:30 — with GitHub Actions Inactive

wangyb-A marked this pull request as ready for review August 11, 2026 06:38

This comment has been minimized.

Sign in to view

wangyb-A temporarily deployed to ai-pr-review-runtime August 11, 2026 06:44 — with GitHub Actions Inactive

github-actions Bot reviewed Aug 11, 2026

View reviewed changes

Comment thread packages/aws-durable-execution-sdk-python/src/aws_durable_execution_sdk_python/execution.py

github-actions Bot reviewed Aug 11, 2026

View reviewed changes

Comment thread packages/aws-durable-execution-sdk-python/tests/execution_test.py

github-actions Bot reviewed Aug 11, 2026

View reviewed changes

Comment thread packages/aws-durable-execution-sdk-python/src/aws_durable_execution_sdk_python/plugin.py Outdated

This comment has been minimized.

Sign in to view

zhongkechen reviewed Aug 11, 2026

View reviewed changes

Comment thread packages/aws-durable-execution-sdk-python/src/aws_durable_execution_sdk_python/execution.py

Alex Wang and others added 6 commits August 11, 2026 18:34

wangyb-A force-pushed the feat/plugin-invocation-input-result branch from fce6081 to d92013f Compare August 11, 2026 18:38

wangyb-A temporarily deployed to ai-pr-review-runtime August 11, 2026 18:38 — with GitHub Actions Inactive

github-actions Bot reviewed Aug 11, 2026

View reviewed changes

zhongkechen approved these changes Aug 11, 2026

View reviewed changes

wangyb-A merged commit 6967db7 into main Aug 11, 2026
30 of 32 checks passed

wangyb-A deleted the feat/plugin-invocation-input-result branch August 11, 2026 19:09

wangyb-A mentioned this pull request Aug 11, 2026

fix(plugin): make payload fields additive on hook infos #627

Merged

This was referenced Aug 11, 2026

feat(plugin): operation maps on invocation hooks #629

Open

feat: add Workflow Insight instrumentation plugin #632

Draft

Navigation Menu

Uh oh!

feat(plugin): surface execution input and result - #616

feat(plugin): surface execution input and result#616
wangyb-A merged 6 commits into
mainfrom
feat/plugin-invocation-input-result

wangyb-A commented Aug 6, 2026

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

Uh oh!

github-actions Bot Aug 11, 2026

Uh oh!

zhongkechen Aug 11, 2026

Uh oh!

github-actions Bot Aug 11, 2026

Uh oh!

github-actions Bot commented Aug 11, 2026

Uh oh!

zhongkechen left a comment

Uh oh!

Uh oh!

github-actions Bot commented Aug 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

wangyb-A commented Aug 6, 2026

Summary

Motivation

Changes

Testing

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

Uh oh!

github-actions Bot Aug 11, 2026

Choose a reason for hiding this comment

Uh oh!

zhongkechen Aug 11, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot Aug 11, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Aug 11, 2026

Codex AI review

Uh oh!

zhongkechen left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions Bot commented Aug 11, 2026

Claude AI review

Summary

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants