ModeHandlersTest.shouldInvokeExitPlanModeHandlerWhenModelUsesTool fails after CLI 1.0.57 upgrade

[edburns]

Summary

ModeHandlersTest.shouldInvokeExitPlanModeHandlerWhenModelUsesTool fails on both JDK 17 and JDK 25 after upgrading the Java SDK to depend on @github/copilot ^1.0.57.

Failing run: https://github.com/github/copilot-sdk/actions/runs/26802258445
Last passing run (CLI 1.0.55-5): https://github.com/github/copilot-sdk/actions/runs/26787985570
PR that introduced the upgrade: #1546

Root Cause

The test uses the replaying CAPI proxy with the snapshot at:

test/snapshots/mode_handlers/should_invoke_exit_plan_mode_handler_when_model_uses_tool.yaml

This snapshot was recorded against CLI 1.0.55-5 (commit 671b50a3, PR #1228). The .lastmerge file now points to commit 753d4729 which installs CLI 1.0.57 into the test harness. CLI 1.0.57 sends requests to the CAPI proxy that no longer match the recorded snapshot patterns, so the proxy returns 500 Proxy error. The CLI retries 5 times (~30s) then surfaces the error as a test failure.

The snapshot YAML itself is unchanged between the old and new .lastmerge commits — nobody has re-recorded it for the 1.0.57 request format.

Error Output

ModeHandlersTest.shouldInvokeExitPlanModeHandlerWhenModelUsesTool:100 » Execution
java.lang.RuntimeException: Session error: Execution failed: Error: Failed to get
response from the AI model; retried 5 times (total retry wait time: 30.64 seconds)
(Request-ID proxy-error) Last error: CAPIError: 500 500 Proxy error

All 3 retry attempts produce the same error.

What Changed in 1.0.57

The 1.0.57 schema introduced several new fields that likely affect the request shape sent to CAPI:

• contextTier added to SessionModelSwitchToParams, SessionStartEvent, SessionResumeEvent, SessionModelGetCurrentResult
• model field added to ToolExecutionStartEvent
• mcpToolsTokens added to context info
• suppressCustomAgentPrompt and other new fields in SessionOptionsUpdateParams
• New RPC namespaces (user, runtime)

Any of these could change how the CLI constructs the chat completion request that the replay proxy tries to match.

Fix Required

The snapshot needs to be re-recorded against CLI 1.0.57 so the replay proxy can match the updated request patterns. The snapshot is located at:

test/snapshots/mode_handlers/should_invoke_exit_plan_mode_handler_when_model_uses_tool.yaml

Additional Context

• The other mode_handlers test (shouldInvokeAutoModeSwitchHandlerWhenRateLimited) passes.
• All other 1692 tests pass cleanly.
• The snapshot format itself is simple (single conversation with an exit_plan_mode tool call), so re-recording should be straightforward.

[edburns]

The proxy has a built-in capture mechanism. When running locally (not in CI), if no snapshot matches, it falls through to the real CAPI upstream and automatically captures the exchange. When the test finishes, it writes the captured conversation back to the snapshot YAML file (writeCapturesToDisk).

So to "record" an updated snapshot for CLI 1.0.57:

1. Run the test locally (not in CI — i.e., GITHUB_ACTIONS must not be set to "true")
2. Have a valid COPILOT_GITHUB_TOKEN (or equivalent auth) so the proxy can forward to real CAPI
3. The proxy will fail to match the snapshot, forward to real CAPI, get a live response, and overwrite the YAML with the new exchange

In practice: run mvn verify locally (without -Dskip.test.harness=true) with a real GitHub token set, and the ModeHandlersTest will auto-record the updated snapshot. The result is a git diff in test/snapshots/mode_handlers/should_invoke_exit_plan_mode_handler_when_model_uses_tool.yaml that you commit.

The reason it 500s in CI is that CI sets GITHUB_ACTIONS=true, which disables the pass-through-and-capture path and instead errors immediately on mismatch.