[devin-ai-integration]

Summary

• Fix OR-543: reuse one native batch SSH shell per target within each Windows orx process for staging, launch, status and log commands; Unix ControlMaster behavior is unchanged.
• Serialize binary-safe framed commands/uploads, invalidate dead or reconfigured sessions, drain stderr continuously, and bound configuration probes with cancellation-safe Windows process-tree cleanup using existing Job Object helpers.
• Retry startup timeouts with 30/60/120-second backoff before command submission, never replay a submitted command. Add ORX_SSH_POLL_INTERVAL_SECS and completion-triggered final log draining.

Test plan

• Reproduced on unchanged main: manual native SSH succeeded, then public orx exp run … --backend ssh --host or543-rate --no-container failed under the genuine Ubuntu server's 3-new-connections/minute, burst-3 kernel limiter.
• Final head d64bc705039cfaacdb9c5f53fbc5bd1b328cb838: independent scope Yes, Greptile 5/5 with no findings, all eight CI checks green, including Windows and Linux x86_64/aarch64.
• Native Windows formatting/clippy, focused framing/lifecycle/configuration-timeout/cancellation/log-drain tests, existing Codex process-tree tests, real-SSH reuse/concurrency test, UI i18n/style/typecheck.
• Fresh final-head native rebuild, genuine subscription-authenticated Codex/Claude live preflights, restarted isolated backend with inherited credentials, and fresh dashboard Codex reply.
• Same limiter re-armed: manual SSH and public launch exited 0; detached supervision completed run 8060faac-c8d0-4411-80fe-6c4568a77d07 in 46 seconds. CLI/dashboard preserved final output without a trailing newline. Authenticated SSH count increased by exactly three (manual, launcher, supervisor), with zero drops and no per-poll churn.
• No stranded SSH/supervisor after completion or backend/listeners after Ctrl+C. The interrupted backend shell returned code 1; this is not a claimed normal backend exit 0.
• Fresh origin/main fetched and merge-tree check clean; PR remains open for human review, assigned to Daniel based on related SSH sharing/recovery work.

The full local Rust suite has four failures also reproduced on unchanged main (Git identity/remote fixtures and a Git Bash Python shim assertion). Changed-area tests and final-head CI pass.

Bug reproduction on main

Windows Server 2022 / native OpenSSH 9.5p2 / orx 0.2.19 against genuine Ubuntu SSH with kernel rate limiting. Main failed before detached supervision.

Post-fix functionality

Same controlled Windows environment and limiter, final PR head. Both recordings are animated WebP conversions made with ffmpeg.

Caveats and remaining limits

• Background skill-import screening logged Codex (gpt-6.1-sol) request failed: batch file arguments are invalid. The real dashboard chat and SSH flow passed; whether this adjacent failure predates the PR was not investigated.
• Exact Windows 11/university infrastructure, ProxyJump and live timeout/backoff were not exercised. Framing, cancellation and configuration edge cases have focused test coverage, not recorded end-to-end coverage.
• Separate orx processes still create separate transports; concurrent launch bursts can exceed server limits.

Automation improvements

• Add Windows blueprint knowledge for the optional genuine Ubuntu/Docker SSH limiter fixture, re-arm/counter checks without extra SSH traffic, and bounded native ssh -n -T/console paste. This would reduce fixture setup friction; the suggestion could not be saved because this session lacks ManageRepoBlueprints permission.

Link to Devin session: https://app.devin.ai/sessions/314bfb3947db4ee88510e83379b34811
Open in Devin Desktop: https://app.devin.ai/desktop/session/314bfb3947db4ee88510e83379b34811?variant=devin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Confidence Score: 5/5

[High impact] The latest changes appear safe to merge; no blocking issue remains.

Summary

This PR reuses Windows SSH connections for experiment launch and supervision. The latest change makes configuration-probe cleanup safe when a request is cancelled.

• Windows orx commands share one SSH shell per target.
• Timed-out SSH connections retry before the remote command starts.
• SSH supervisors can change how often status and logs are checked.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Start SSH suspended] --> B{Assign kill-on-close job}
  B -->|Failure| C[Return error and drop child]
  B -->|Success| D[Resume SSH]
  D --> E[Await configuration output]
  E --> F[Success, error, timeout, or cancellation]
  F --> G[Drop job and stop remaining descendants]

Loading

_{Reviews (6) · Last reviewed commit: "Make SSH probe process-tree cleanup canc..." · Reviewed by Greptile}

[devin-ai-integration]

@greptileai Please review the latest head f55340f. Follow-up commits address the four previous findings and add native Windows framing/draining, diagnostics, effective-config, dead-child and final-log-wakeup regressions.

[devin-ai-integration]

@greptileai Please review latest head b60aac5. Only change since the previous review request: the opt-in real-SSH test explicitly disables ControlMaster on every platform instead of being excluded on Unix. This resolves Unix dead-code clippy and tests the Windows transport; production code is unchanged.

[devin-ai-integration]

@greptileai Final test-only follow-up is a65133c: keep the integration test and reconnect call on all platforms, but assert a different TCP connection only on Windows. This preserves Unix ControlMaster behavior without ineffective duplicate options. Native Windows clippy and the real-SSH test pass. Please review this exact latest head.

[devin-ai-integration]

@greptileai Please review final head cbbf45c. The remaining configuration-probe deadline finding is fixed with a native-SSH regression; Windows also terminates hung Match exec descendants before dropping SSH to release inherited pipes. All five Windows session unit tests, the opt-in real-SSH transport test, formatting and clippy pass.

[devin-ai-integration]

@greptileai Please review latest head d64bc70. Follow-up makes configuration-probe cleanup cancellation-safe: reuse the existing Windows kill-on-close Job Object helpers, start SSH suspended, enroll before resume, retain the job through output/deadline. Strict enrollment failure stops the probe. Native-SSH regressions cover both timer expiry and cancellation after Match exec starts; all six session tests, existing Codex wrapper teardown tests and Windows clippy pass.