[svetanis]

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• Closes: Runner drops callbackContextData written in onUserMessageCallback when it resumes an invocation #1695
• Related: feat(plugins): add SaveFilesAsArtifactsPlugin, porting adk-python's file-upload offload #1412 (where it surfaced)

2. Or, if no issue exists, describe the change:

Problem:

On a resumed invocation, resumeCore runs onUserMessageCallback on initialContext, but runResumedAgent
builds the agent's context from a second newInvocationContextBuilder, with a new callbackContextData. Data a
plugin stores in onUserMessageCallback is gone by beforeAgentCallback. A new invocation does not have this
problem: runAgentWithUpdatedSession derives the agent's context with initialContext.toBuilder(), which shares
the map. The second builder also resets endInvocation, so a plugin that ends the invocation in
onUserMessageCallback does not stop a resumed run. On every resume, onRunErrorCallback also runs on
initialContext, so it does not see what beforeRunCallback stored, even when no new message is sent. It surfaced in #1412: SaveFilesAsArtifactsPlugin saves a
file in onUserMessageCallback and reports its version from beforeAgentCallback.

Solution:

runResumedAgent takes initialContext instead of session, resolvedInvocationId and runConfig, which
initialContext already carries, and
builds rootContext with initialContext.toBuilder().agent(rootAgent).userContent(...). Its three callers in
resumeCore pass initialContext. The resume path now derives the agent's context from initialContext, as
the new-invocation path does.

Besides callbackContextData, the copy carries endInvocation, the cost manager and the checkpoint maps from
initialContext, as on the new-invocation path. initialContext makes no model call, so the cost manager's
call count is unchanged; a plugin that ends the invocation in onUserMessageCallback now ends the resumed run too, as it
already did a new one. agent and userContent are set as before, and the invocation id, run config, session,
node and branch have the same values the second builder gave them.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Test	Resumes	On main
runAsync_resume_keepsCallbackContextDataFromOnUserMessage	the root agent	❌ beforeRunCallback reads [message-1, null], beforeAgentCallback [agent:message-1, agent:null]
runAsync_resumeRoutedToSubAgent_keepsCallbackContextDataFromOnUserMessage	a sub-agent, routed past the root	❌ […, sub_agent_1:message-2, sub_agent_1:null]
runAsync_resume_endInvocationFromOnUserMessage_skipsTheModel	the root agent, after a plugin ends the invocation	❌ the model is asked again (2 requests, expected 1)
runAsync_resumeWithoutMessage_onRunErrorSeesBeforeRunData	the root agent, with no new message	❌ onRunErrorCallback reads null, expected run-2
runAsync_resumeWithStateDeltaOnly_onRunErrorSeesBeforeRunData	the root agent, with only a state delta	❌ onRunErrorCallback reads null, expected run-2

All five fail when Runner.java alone is reverted, and no other test in RunnerResumabilityTest does. The
first two cover the two branches of runResumedAgent: giving only the sub-agent's context a new map fails only
the second test. The last two cover the two resumeCore calls without a new message: giving only the
state-delta call a new map fails only the fifth test.

RunnerResumabilityTest: Tests run: 48, Failures: 0, Errors: 0.

Manual End-to-End (E2E) Tests:

SaveFilesAsArtifactsPlugin from #1412 with a stock resumable Runner: turn 1 pauses on a long-running tool;
turn 2 sends the function response plus report.pdf.

without this change   artifactDelta, turn 2 (resumed): {}               artifacts stored: [report.pdf]
with this change      artifactDelta, turn 2 (resumed): {report.pdf=0}   artifacts stored: [report.pdf]

Checklist

• I have read the CONTRIBUTING.md document.
• My pull request contains a single commit.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

[utafrali]

The fix is correct and well-scoped: switching runResumedAgent to derive its root context via initialContext.toBuilder() instead of newInvocationContextBuilder() closes the data-loss gap and aligns the resume path with the new-invocation path. The five new tests directly reproduce the reported bug and cover all described side effects including endInvocation propagation and onRunErrorCallback visibility.

[utafrali]

After this refactor, resolvedInvocationId inside runResumedAgent is used only to call userContentFor(session, resolvedInvocationId, userContent). Its value is already carried by initialContext.invocationId(), so they are always equal. As an optional follow-up, replacing that argument with initialContext.invocationId() and dropping the parameter would make the method fully self-contained through initialContext, eliminating the redundant parameter. Not blocking for this PR.

[hirematha]

Hi @svetanis, thank you for your contribution. We appreciate you taking the time to submit this pull request. The PR is currently under review by our team, and we will keep you posted if any additional information is required. Thank you.

[damianmomotgoogle]

LOW: +1 to the earlier suggestion, and the same applies to session: both values come from initialContext (initialContext.session() and initialContext.invocationId()), so passing them separately duplicates their source. Removing both parameters also trims the new @param block to the single line that adds context.

[damianmomotgoogle]

LOW: pausingAgent("agent", pauseThenSay("lro_call_id", "resumed and summarized")) from ResumabilityTestUtils, already imported here, builds the same model script and agent. The other single-agent tests can use these helpers too to keep the new tests shorter.

[svetanis]

Thanks for the review. Both comments are addressed in a single commit rebased on main:

1. Duplicated parameters: runResumedAgent now takes only initialContext and userContent, and reads the session and the invocation id from initialContext; the @param block is the initialContext line alone. A refactor only, covered by the five new tests.
2. Test helpers: runAsync_resume_keepsCallbackContextDataFromOnUserMessage, runAsync_resume_endInvocationFromOnUserMessage_skipsTheModel and the shared setup of the two onRunErrorSeesBeforeRunData tests now build their model script and agent with pausingAgent and pauseThenSay. The sub-agent test keeps its own setup for the transfer.