Repository navigation
Conversation
Recheck dead-owner turns against native terminal status, keep delayed end handlers within their own session epoch, and only settle reuse evidence actually queued. Continue reading the completion handshake after stdout capture reaches its limit.
Keep verified report payloads in the injection ledger after a local queue failure. Retry using the original check evidence and leave unverified injections out of delivery.
Forward the supplied record after native terminal confirmation and handle a missing status reader. Give a dead owner a bounded grace for terminal metadata and omit any stale worktree. Isolated model declarations default to an empty environment.
Keep the cross-shell calibration regression already landed upstream in EvoMap#55.
Require an explicit regular ZCode CLI bundle for background completion. Install through CLI plugin directories, and exercise the package without a Desktop installation.
Copying the index advances its mtime, which can make Git trust a cached stat entry for a same-size edit. Use the source index timestamp rounded down before adding untracked entries. A deterministic regression checks both diff parts and the untouched source index. (cherry picked from commit bbd7c68)
Preserve the pending native-terminal deadline after its owner exits, pass the private configuration path to unbound environment callbacks, reject escaping scratch-file paths, and tolerate a vanished dead-owner state lock. Run the fixture Proxy independently of synchronous host work, preserve its request getter through the DSH wrapper, and track fixture-owned MiniMax workers for scoped teardown. Add regressions for the runtime contracts and native ZCode database discovery. Validation on this exact tree: full Linux tests on Node 22.23.4 and 26.13.0 each passed 908 with one Windows-only skip. Windows test:minimax passed 528 with ten platform skips and test:zcode passed 14 on both Node versions. Packaging all ten hosts passed on both platforms. Assertions, suite workloads, and existing deadlines remain unchanged.
A throwing closeInputOn callback escaped the stdout EventEmitter, which could terminate the host before model completion cleaned its private files. Keep the original error, stop the owned process tree, and normally finish when the child closes. Preserve the same error through the existing timeout and process-error paths, and do not start a Windows handshake process when tree termination is unavailable. Add regressions for child/grandchild termination, model error identity and private-file cleanup, and the missing Windows termination backend. The two exception tests failed uncaught before the fix. All five model-executor tests pass on Windows Node 22 and 26; each full Windows gate passes 531 common/MiniMax tests (10 platform skips), 14 ZCode tests, and packaging of all ten hosts. Existing assertions and deadlines are unchanged.
An older turn waiting for its host record could consume a later turn's untagged check notes and edit baselines. Store evidence by the plugin epoch and turn counter, snapshot only the opened generation, and leave failed-reader claims recoverable with their original hashes. Keep ambiguous legacy files in their own namespace and bound abandoned evidence claims with the existing TTL; report-queue claims are unchanged. Carry prompt-opened and tool-recorded identities into monitor tasks instead of adopting the state current after asynchronous work. Pin that identity through polling and atomic turn close, separate generation lock keys, and prevent a delayed Git HEAD read from updating a replacement epoch. Regressions exercise real prompt/tool/turn flows, missing native turn IDs, same-ID epoch replacement, delayed worker startup, reader rejection and claim recovery, legacy namespace isolation, and evidence-only pruning. Joint driver/app/store/monitor/ZCode tests pass 140/140 on Windows Node 22 and 26. The complete Windows gates pass 542 common/MiniMax tests (10 platform skips), 14 ZCode tests and packaging of all ten hosts on both versions.
|
Sentinel 已审核此变更，未发现已核验问题，但发现需要人工确认的风险。 另有 11 个需要人工确认的风险；黄色 advisory 不阻塞合并，建议开发 agent review advisory 评论并逐条确认。 审核覆盖提示： 开发 agent 查询详情： 调用 MCP 工具 由 Sentinel 审核 · 目标提交 95c5b21352fa |
|
Sentinel 保留了 11 个尚未确认的审核候选。这些黄色条目不阻塞合并；红色 P0/P1 与其他独立硬门禁不受影响。 单一猜测、缺少变更行证据或纯 checklist 的候选会统一降为 P3。
需要完整证据时，可用行末 由 Sentinel 审核 · 目标提交 95c5b21352fa |
imxyanua
marked this pull request as ready for review
November 9, 2026 06:44
imxyanua
marked this pull request as draft
November 9, 2026 06:53
Bugbot is paused — on-demand spend limit reachedBugbot uses usage-based billing for this team and has hit its on-demand spend limit. A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue. |
|
Sentinel 已审核此变更，未发现已核验问题，但发现需要人工确认的风险。 另有 15 个需要人工确认的风险；黄色 advisory 不阻塞合并，建议开发 agent review advisory 评论并逐条确认。 审核覆盖提示： 开发 agent 查询详情： 调用 MCP 工具 由 Sentinel 审核 · 目标提交 07b44c5381d0 |
|
Sentinel 保留了 15 个尚未确认的审核候选。这些黄色条目不阻塞合并；红色 P0/P1 与其他独立硬门禁不受影响。 单一猜测、缺少变更行证据或纯 checklist 的候选会统一降为 P3。
需要完整证据时，可用行末 由 Sentinel 审核 · 目标提交 07b44c5381d0 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Draft; depends on #54. Please review the Step-only changes here. The branch currently includes #54; sync with main after that dependency merges.
Add a native resident extension for Step Code CLI 0.1.7 on Windows:
The adapter reuses the plugin channel, verification, reports, session ranges, private drafts and explicit publication commands. It pairs native calls/results, latches cancellation across model steps, and keeps Recipe context out of the saved transcript and subsequent prompts.
Private completion pins provider/model/thinking level/account, uses no tools or prior session transcript, rejects changed routes and late answers, and never persists credentials. Local package install/remove preserves plugin data. No new dependency or Desktop adapter is included.
Evidence
After: All four cases suppress reports/candidates; the final Step suite passes 37/37 on Windows Node 22.23.4 and 26.13.0, without skipped tests.
95c5b21352fae35f474c1f093f06d7fb047520d4: Linux Node 22 and 26 each ran 963 tests, 961 pass, 2 existing skips, then packed all 11 hosts. Windows/release jobs correctly skipped for this PR event.Remaining acceptance gates: paid-provider and real Proxy/Hub acceptance; native transport/OAuth-refresh cleanup. Separate foreground
run_command/bashchecks are supported; inlinethen_runand background execution exclude the whole turn. Native double Ctrl-C can exit Step without a terminal event and leave its tool process alive: the adapter suppresses publication evidence but does not kill Step-owned processes. These limits keep this PR in draft.Merge Danger
Door: two-way
Remove the extension to stop new work; its profile data remains. Do not merge before #54 and the remaining acceptance gates are reviewed.
Blast Radius: host
New Step host and packaging/CI coverage. Shared production code only re-exports the existing headless guard and host-command catalog from the plugin channel. PR CI runs the existing Linux matrix; Windows host suites remain in the existing tag gate and were run locally above.