bug: copilot-sdk executor fails first message with context canceled on Windows

[RascoApps]

Summary

On Windows, Waza 0.38.7's copilot-sdk executor creates a session but every evaluated task fails when sending the first message with context canceled. The standalone Copilot CLI works with the same authenticated account and model.

Environment

• Waza: 0.38.7
• GitHub Copilot CLI: 1.0.80
• OS: Windows 11 Enterprise (10.0.26200.0)
• Shell: PowerShell
• Waza executor: copilot-sdk

Steps to reproduce

1. Authenticate the Copilot CLI normally.
2. Confirm the requested model appears in waza models --json.
3. Run a single Waza task without cache, for example:

waza cache clear
waza run eval.yaml --executor copilot-sdk --model <model-from-waza-models> --trials 1 --task <task-id>

Actual behavior

The session starts, but sending the first task message fails:

session.start
failed to send message: context canceled

The task receives no agent response and fails. Clearing .waza-cache and rerunning produces the same result.

Control cases

• A direct non-interactive Copilot CLI prompt succeeds under the same account/environment.
• Repeating the Waza run after clearing the cache fails identically.
• Pointing COPILOT_CLI_PATH at the independently installed Copilot CLI 1.0.80 changes the failure to:

Time.UnmarshalJSON: input is not a JSON string

This suggests a possible Waza/Copilot SDK/CLI protocol or timestamp-schema compatibility issue, although the exact root cause is not yet confirmed.

Expected behavior

The copilot-sdk executor should send the task prompt and return the model response, or report a specific actionable SDK/CLI compatibility error rather than context canceled.

Additional context

Related upstream issue: github/copilot-sdk#1165 tracks SDK protocol-version drift detection, but no existing Waza issue was found for this failure signature.

[spboyer]

Validation update: the report is credible and the Time.UnmarshalJSON result is consistent with a Copilot SDK/CLI protocol mismatch, but I could not reproduce the Windows-specific failure on the available macOS environment. The targeted execution, orchestration, and validation tests pass on the current checkout.

Before routing this to coding-agent, please provide:

• the complete verbose output from the failing waza run (with tokens, paths, and account data redacted);
• waza version, waza models --json, and the exact embedded/COPILOT_CLI_PATH CLI version;
• go env output and the smallest eval/task YAML that reproduces it; and
• confirmation whether the failure still occurs on current main.

The Windows reproduction plus these diagnostics will distinguish a Waza error-reporting bug from SDK/CLI wire-version drift.

[spboyer]

Validation follow-up: this is a credible Windows Copilot SDK integration report, but it is not independently reproducible in the available macOS environment. The repository currently uses Copilot SDK 1.0.2 while open PR #549 is updating that dependency, so this issue should not be routed to coding-agent while the related PR is in progress. Please retest current main and the PR #549 dependency level on Windows, then attach redacted verbose output, waza version, waza models --json, the embedded/override CLI version, go env, and the smallest eval/task YAML.

[github-actions]

⚠️ No squad member found matching label squad:richard park. Check .squad/team.md (or .ai-team/team.md) for valid member names.