Fix recurring merge queue failure causes

[stephentoub]

Repeated merge-queue attempts for #2507 exposed several independent SDK test and CI defects: replay mismatches after aborts, uninitialized proxy traffic, fragile Windows shell execution, truncated Python downloads, and accidental Maven Central publication during local validation.

Changes

• Normalize the runtime's bare Session aborted result to the replay harness's stable interruption representation.
• Initialize each Go replay proxy before runtime startup traffic, and run the Windows shell marker command through cmd.exe in an isolated working directory instead of nested PowerShell.
• Retry http.client.IncompleteRead failures while continuing to verify complete downloads with checksums or package integrity metadata.
• Explicitly set skipPublishing during Java native local publication validation so release-version commits do not require Maven Central credentials.

The MCP OAuth startup race is intentionally excluded because github/copilot-agent-runtime#18931 and #2511 already own that work. This also complements #2484, which improved teardown, diagnostics, artifact reruns, and registry retries but did not cover these failure modes.

Validation

• Replay harness: 34 tests passed
• Harness TypeScript check passed
• Python Ruff checks and 9 downloader tests passed
• Go E2E package compilation passed
• Windows Go shell E2E passed
• Java workflow YAML parsed successfully
• Opus review found no significant issues

Generated by Copilot

[github-actions]

SDK Consistency Review

Reviewed the full PR diff (7 files) against all six SDK implementations.

Summary of changes:

• .github/workflows/java-sdk-tests.yml — CI-only flag addition (-DskipPublishing=true), no SDK API surface.
• go/internal/e2e/rpc_shell_and_fleet_e2e_test.go + go/internal/e2e/testharness/context.go — Go E2E test/harness fixes (writing marker file in a subdirectory, initializing the replay proxy with a dummy snapshot before model traffic). Internal test infra only, no public Go SDK API changed.
• python/copilot/_cli_download.py (+ its test) — expands the retriable exception set for CLI/runtime downloads to include http.client.IncompleteRead alongside HTTPError/URLError.
• test/harness/replayingCapiProxy.ts + .test.ts — shared test-harness normalization logic (used by all SDKs' E2E suites), tightened to also match a bare "Session aborted" tool result.

Cross-SDK consistency assessment:

The _cli_download.py retry-hardening only applies to Python because Python is the only SDK that implements its own bespoke CLI/runtime download-with-retry logic in this repo. The other SDKs acquire the CLI/native runtime through their respective language package managers (npm for Node.js, Go modules, NuGet for .NET, Maven for Java, crates.io conventions for Rust) rather than a custom HTTP downloader with manual retry loops — I found no equivalent _fetch_url_bytes/retry-loop pattern in nodejs/, go/, dotnet/src/, java/sdk/src/main/java/, or rust/src/. Since there's no parallel implementation to update, this isn't a parity gap — it's a Python-specific implementation-detail fix.

The Go E2E test change and the shared test/harness/ normalization update are test-infrastructure fixes that benefit every SDK's E2E suite equally (the harness is shared), and the CI workflow tweak is Java-build-only with no code-level API implication.

No action needed. This PR does not introduce any new public API in one language without a counterpart elsewhere, nor does it change SDK behavior inconsistently across languages. No inline comments added.

Generated by SDK Consistency Review Agent for #2515 · copilot · sonnet50 · 23.1 AIC · ⌖ 12 AIC · ⊞ 9.7K · ◷

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused reliability fixes are consistent with their surrounding implementations and include appropriate validation coverage.

Review tier: Balanced
Findings: None

What changed in this PR

Fixes recurring merge-queue failures across replay handling, Python downloads, Go E2E setup, and Java validation.

Changes:

• Normalizes aborted tool results and retries truncated downloads.
• Stabilizes Go proxy initialization and Windows shell execution.
• Prevents Java local validation from publishing to Maven Central.

File	Description
test/​harness/​replayingCapiProxy.ts	Normalizes bare session-abort results.
test/​harness/​replayingCapiProxy.test.ts	Tests abort-result normalization.
python/​copilot/​_cli_download.py	Retries incomplete HTTP reads.
python/​test_cli_download.py	Tests truncated-response recovery.
go/​internal/​e2e/​testharness/​context.go	Initializes the replay proxy before clients start.
go/​internal/​e2e/​rpc_shell_and_fleet_e2e_test.go	Stabilizes Windows shell marker execution.
.github/​workflows/​java-sdk-tests.yml	Explicitly disables Maven Central publication.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.