Keep the Copilot run alive on rate-limit errors eligible for auto-switch

[PratikDhanave (PratikDhanave)]

What

responseUpdateForSessionEvent in provider/copilotprovider/copilot.go mapped every *copilot.SessionErrorData to a terminal (done=true, error), so the run loop yielded the error and returned, terminating the session.

The Copilot SDK sets SessionErrorData.EligibleForAutoSwitch only on errorType: "rate_limit", documented as: when true, the runtime will follow this error with an auto_mode_switch.requested event and keep the session alive (or silently switch). The follow-up auto_mode_switch.requested/completed events are already handled by the provider's default (non-terminal) case, and the run naturally completes on the subsequent session.idle.

This change dereferences EligibleForAutoSwitch in the SessionErrorData case: when non-nil and true, the error is surfaced as a non-terminal raw notification (done=false, err=nil) so the loop keeps pumping the auto-switch events and the eventual idle completion. Every other session error (nil/false flag, or any non-rate-limit type) keeps the existing terminal behavior. Recovery is confined to exactly the SDK-documented rate-limit auto-switch case.

Why

This aligns the Go provider with the runtime contract the SDK exposes via EligibleForAutoSwitch: a rate-limit that the runtime is about to auto-recover from should not surface to the caller as a fatal run error. Treating these as terminal caused spurious run failures on transient, auto-recoverable rate limits.

Tests

Added to the canonical copilot_test.go, driving the full run loop through the existing fake-runtime harness:

• TestRun_WithRateLimitEligibleForAutoSwitch_KeepsSessionAlive — a session.error with errorType: "rate_limit" and eligibleForAutoSwitch: true, followed by auto_mode_switch.requested, an assistant.message, and session.idle. Asserts the run does not fail and emits the assistant message.
• TestRun_WithRateLimitNotEligibleForAutoSwitch_ReturnsError — a rate-limit error without the flag still terminates with the session error.

Both tests fail before the fix and pass after. go build ./..., go vet ./provider/copilotprovider/..., and go test ./provider/copilotprovider/... all pass.

[Copilot]

Pull request overview

This pull request updates the Copilot Go provider’s session event handling so that session.error events representing rate-limit errors that the runtime will auto-recover from (via EligibleForAutoSwitch) do not terminate the run loop. This aligns provider behavior with the Copilot SDK/runtime contract by allowing the session to continue through subsequent auto-switch events and complete normally on session.idle.

Changes:

• Treat SessionErrorData.EligibleForAutoSwitch == true as a non-terminal session notification (no error returned, run loop continues).
• Add an end-to-end fake-runtime test asserting the run survives an eligible rate-limit error and still emits the assistant message.
• Add a test asserting a rate-limit error without the eligibility flag remains terminal and returns the session error.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
provider/copilotprovider/copilot.go	Makes eligible rate-limit session.error non-terminal so the run loop can process auto-switch and reach session.idle.
provider/copilotprovider/copilot_test.go	Adds regression tests covering eligible vs non-eligible rate-limit errors through the full run loop.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Parity Review: No Issues Found

This PR modifies only an unexported internal function (responseUpdateForSessionEvent) inside provider/copilotprovider/copilot.go. No exported Go API surface is added, removed, or changed.

The behavioral change is specific to the GitHub Copilot SDK (EligibleForAutoSwitch on SessionErrorData), which is a Go-provider-specific integration contract. The upstream microsoft/agent-framework .NET and Python implementations do not have a Copilot SDK provider equivalent, so cross-repo parity does not apply here.

Verdict: This PR stays within Go-internal implementation details. Cross-repo parity is preserved. The parity-approved label remains correct. The public-api-change label is not warranted.

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• awmgmcpg

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "awmgmcpg"

See Network Configuration for more information.

Generated by Go API Consistency Review Agent · 22.2 AIC · ⌖ 5.53 AIC · ⊞ 5.9K · ◷