[zpzjzj]

Summary

Run the evaluated Agent once, then execute named judges sequentially in the same case runtime (none, Docker or OpenSandbox). Each Agent judge starts a fresh session; timeouts, verdicts and archived materials remain separate by judge ID. Remove judge Snapshot/fork support and retain normal case cleanup.

Filesystem mutations, installed tools, services, volumes, environment and network policy remain shared. Later judges still run after an earlier judge FAIL/ERROR; case cancellation skips pending members. Separate sessions do not provide filesystem or external-service isolation.

Changes and scope

• Keep shared sequential judge orchestration, all_required aggregation and per-judge artifacts together with their runtime tests. Independent engine provisioning and provider routing are supplied by main via feat(judge): support independent judge engine configuration #314.
• Include the three OpenSandbox E2E files and bilingual verification documentation previously reviewed in test(runtime): verify sequential judges in a shared OpenSandbox #316. That PR is superseded by this one; no separate OpenSandbox implementation PR remains.
• Cover shared filesystem/dependencies, fresh session IDs, material readback, script FAIL / later judges PASS / aggregate FAIL, independent OpenCode provider configuration, and default-provider routing readback.

Dependencies

Based on main at 6d121cb02970e355fe5ea00c55a215f09a8542b1, including #314's independent-engine/provider support and #323's bounded GoReleaser build concurrency. The PR contains only shared-runtime orchestration, Snapshot removal, consolidated regressions/documentation, and the OpenCode judge connection-order fix. #316 is closed and its tests/docs are included here.

Part of #308. Filesystem-isolation-specific acceptance remains outside this shared-runtime design, so #308 stays open.

Validation

• After rebasing onto main with fix(ci): bound GoReleaser build concurrency on self-hosted runners #323, local make fmt verify test build e2e test-action passed on 83583ba4dffd0c138d727b5e8336f22e2ab83095, including race tests, plugin tests and Skill regression scripts. E2E-tag lint passed with zero findings. The real OpenCode 1.14.24 loopback HTTP order regression also passed in both directions. All standard CI checks passed for this exact SHA: CI, Extended CI, CodeQL, docs, and release tooling.
• Standard CI covers build/race tests, lint, Linux E2E, Windows E2E, Docker integration, CodeQL, documentation and release-tool checks. Service/model-gated cases are compiled but skip without explicit flags and credentials; these skips are not live verification.
• Historical live validation of the imported implementation and test assertions (assertion helper extraction and lint cleanup preserve the original test coverage): shared-runtime OpenSandbox fixtures passed against two tenant configurations; real OpenCode 1.14.24 default-provider configuration readback passed; the real qwen3.8-max endpoint control passed with two distinct judge sessions and cleanup confirmed.
• Native OpenAI model invocation remains unverified: direct api.openai.com connectivity timed out from the test sandbox. No native-model success or service-egress isolation is claimed.

The writing-evals guide includes reproduction commands for the shared environment, real provider calls and configuration-only readback.

Follow-up: OpenCode judge connection ordering

Fixed the review regression where a legacy judge.model member could reuse the preceding independent OpenCode judge's endpoint. Every OpenCode Agent judge now reapplies its resolved provider connection before invocation; independent-engine install/preflight and shared runtime/fresh-session behavior are retained.

• Regular race tests cover native/custom providers, explicit/default runner endpoints, legacy-alone, independent → legacy, and legacy → independent → legacy, including custom key binding and preservation of MCP/unrelated providers. The affected order cases fail before the fix and pass after it.
• A gated local CLI E2E uses OpenCode 1.14.24 and two loopback HTTP recorders. Before: independent → legacy routes B → B. After: B → A; reverse order remains A → B. It uses fixture keys and fixed HTTP 400 responses, so this proves HTTP routing, not successful model grading.
• Local make fmt verify test build e2e test-action, including race tests, passed. E2E-tag lint passed with zero issues. All standard CI checks passed for 83583ba after rebase onto main with fix(ci): bound GoReleaser build concurrency on self-hosted runners #323 (runs linked above); live-service/model-gated skips are not counted as service validation.

Reproduce the local HTTP test with SKILL_UP_OPENCODE_ORDER_E2E=1 go test -count=1 -tags e2e -run '^TestOpenCodeJudgeConnectionOrder$' -v ./e2e and OpenCode installed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T07:44:55.544167Z	3796e6e	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[JHWang-1997]

Reproduced an order-dependent P2 regression in shared-runtime judging: a legacy OpenCode judge can retain the preceding independent judge’s endpoint instead of its resolved runner connection. Verified actual HTTP request routing with OpenCode 1.14.24; the same reproduction routes correctly on base 87f86b5. Details and reproduction are in the inline comment.

[JHWang-1997]

LGTM