Merge reference implementation SDK changes (2026-05-18)

[Copilot]

Resolves #207

---

Before the change?

• CustomAgentConfig had no way to specify a model for an agent
• Hook inputs (PreToolUseHookInput, PostToolUseHookInput, SessionStartHookInput, SessionEndHookInput, UserPromptSubmittedHookInput) did not expose the session ID, breaking sub-agent hook propagation
• No support for cloud session creation with repository metadata

After the change?

• model field on CustomAgentConfig — optional model identifier (e.g. "claude-haiku-4.5") that the runtime uses for the agent, falling back to the parent session model if unavailable
• sessionId on all hook inputs — enables hooks to identify which session (including sub-agent sessions) triggered them
• Cloud session config — new CloudSessionOptions and CloudSessionRepository classes, wired through SessionConfig.setCloud() and CreateSessionRequest
• .lastmerge updated to f6c1adf8, @github/copilot synced to ^1.0.49-1

// New: specify model per agent
var agent = new CustomAgentConfig()
    .setName("fast-helper")
    .setModel("claude-haiku-4.5");

// New: cloud session with repo metadata
var session = client.createSession(
    new SessionConfig()
        .setOnPermissionRequest(PermissionHandler.APPROVE_ALL)
        .setCloud(new CloudSessionOptions()
            .setRepository(new CloudSessionRepository()
                .setOwner("my-org").setName("my-repo").setBranch("main")))
).get();

// Hook inputs now expose sessionId
hooks.setOnPreToolUse((input, invocation) -> {
    log.info("Session {} calling {}", input.getSessionId(), input.getToolName());
    return CompletableFuture.completedFuture(PreToolUseHookOutput.allow());
});

Not ported (intentionally): .NET-specific changes (CopilotTool helper, MSBuild targets, netstandard targets, C# argument validation), test snapshot reorganization, Python/Go-only doc fixes. See PR body checklist for full breakdown.

Pull request checklist

• Tests for the changes have been added (for bug fixes / features)
• Docs have been reviewed and added / updated if needed (for bug fixes / features)
• mvn spotless:apply has been run to format the code
• mvn clean verify passes locally

Does this introduce a breaking change?

• Yes
• No

---

[github-actions]

Commit pushed: f219ada

Generated by Codegen Agentic Fix

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.

[edburns]

Code Review: PR #212 (Reference Impl Sync — 14 new commits)

APPROVAL — CopilotSession.java

The two SessionModelSwitchToParams call sites correctly add null for the new reasoningSummary parameter (4th position), keeping modelCapabilities in 5th position. Matches the regenerated record:

// Before: (sessionId, model, reasoningEffort, capabilities)
// After:  (sessionId, model, reasoningEffort, reasoningSummary, capabilities)

APPROVAL — CloudSessionOptions + CloudSessionRepository (new files)

Clean DTOs following established patterns: @JsonInclude(NON_NULL), @JsonProperty, fluent setters, Javadoc with @since 1.5.0. No issues.

APPROVAL — CreateSessionRequest + SessionConfig + SessionRequestBuilder

Standard wiring of the new cloud field:

• SessionConfig adds field + getter/setter + copy() inclusion
• SessionRequestBuilder passes config.getCloud() to the request
• All follow existing patterns

APPROVAL — CustomAgentConfig (model field)

Clean addition with good Javadoc explaining fallback behavior.

WARNING — Hook Input Records (breaking positional API)

SessionStartHookInput, SessionEndHookInput, UserPromptSubmittedHookInput add sessionId as the first positional record parameter. This is source-incompatible for any consumer using positional construction:

// Previously valid, now compile error:
new SessionStartHookInput(timestamp, cwd, source, initialPrompt)
// Must become:
new SessionStartHookInput(sessionId, timestamp, cwd, source, initialPrompt)

Mitigating factors: These records are deserialized from CLI JSON internally — users shouldn't construct them directly. The existing test (SessionHandlerTest:287) uses MAPPER.valueToTree(Map.of(...)) which deserializes by property name (via @JsonProperty), so it's unaffected — sessionId will just be null.

Verdict: Acceptable for a minor version bump but should be noted in CHANGELOG.

APPROVAL — PostToolUseHookInput + PreToolUseHookInput (sessionId)

These are classes (not records), so adding sessionId is fully backward-compatible.

APPROVAL — Documentation (advanced.md, hooks.md)

• Cloud sessions documented with correct example
• Hook tables updated with sessionId column
• Fixed getTimestamp() → timestamp() (correct for record accessors)
• Agent config table adds skills and model

SUGGESTION — Missing Tests

No new test coverage for:

• Cloud session options wiring through SessionRequestBuilder
• sessionId on hook inputs
• model on CustomAgentConfig

These would ideally have unit tests verifying serialization round-trip. However, E2E coverage depends on reference-impl snapshots, so this is acceptable to defer.

---

Summary

Category	Count	Items
APPROVAL	7	All changes are correct and follow patterns
WARNING	1	Hook record positional API break (note in CHANGELOG)
SUGGESTION	1	Missing unit tests for new features
CRITICAL	0	—

Recommendation: Merge-ready. The one warning is a documented acceptable trade-off for reference-impl parity. Consider adding a CHANGELOG note about the hook input record signature change.

[edburns]

@copilot I added some tests, and there are test failures.

Look at the job run https://github.com/github/copilot-sdk-java/actions/runs/26061898666 to discern the failures and fix them, pushing additional commits to the topic branch for this work copilot/reference-impl-sync-14-new-commits.

[Copilot]

@copilot I added some tests, and there are test failures.

Look at the job run https://github.com/github/copilot-sdk-java/actions/runs/26061898666 to discern the failures and fix them, pushing additi...

Fixed in commit ba6e981 → new commit pushed. The three test failures were caused by the codegen producing enum types where tests expected strings:

1. GeneratedEventTypesCoverageTest — used "approve"/"edit"/"reject" as ExitPlanModeAction values (valid values are exit_only/interactive/autopilot/autopilot_fleet); used McpServersLoadedServerStatus where the generated record now uses McpServerStatus; compared "project" string against SkillSource.PROJECT enum.
2. SessionEventDeserializationTest — same ExitPlanModeAction issue.
3. ModeHandlersTest — compared "interactive" string against ExitPlanModeAction.INTERACTIVE enum for selectedAction(), causing the event listener to never match and the test to timeout.