`bg output` fails while a timeout-promoted command is still Running (output artifact created only at completion)

[ianalitis]

Upstream: https://github.com/1jehuang/jcode
Source commit: 1b99ae621 (upstream master, 2026-09-06)
Binary identity: jcode v0.83.71-dev (10a4f385a), local branch on top of master
Platform: macOS 26.6.2, arm64

Summary

When a foreground bash call exceeds its timeout and Jcode promotes it to a background task,
the task's .output artifact is not created until the command finishes. bg output <id> during
the Running state returns Output not found for task, so an agent cannot inspect partial output of
exactly the commands it most needs to watch (long builds, tests, uploads).

The status artifact is written immediately with pid: null; only the output artifact is missing.

Reproduction (two live receipts, 2026-09-07)

Tool call: bash with command: "echo start; sleep 20; echo done", timeout: 2000.

1. Jcode reports "exceeded the foreground timeout after 2.0s and is continuing in background",
   task 836444a36r.
2. Immediately, bg output 836444a36r returns
   Error: Output not found for task: 836444a36r. Task may not exist or output file was deleted.
3. Filesystem at the same moment ($TMPDIR/jcode-bg-tasks/):
   • 836444a36r.status.json exists, "status": "running", "pid": null
   • 836444a36r.output does not exist
4. After completion the output file appears with the full content.

An earlier receipt on 2026-08-31 with v0.81.5-dev (422da2bc0) showed the same behavior.

Why the regression test misses it

crates/jcode-app-core/src/tool/bash_tests.rs::test_foreground_timeout_promotes_and_command_keeps_running
waits for BackgroundTaskStatus::Completed before calling background::global().output(&task_id).
It never reads the output while the task is Running, so the gap is not covered.

Expected

Promotion creates an empty retained output artifact before returning the task id, so
bg output succeeds (possibly empty) while Running, and the wrapper replaces it with the
complete output at termination. pid: null is fine for the in-process adoption path.

Suggested test change

In the existing test, after promotion and before waiting for completion, assert
background::global().output(&task_id).is_some(). It is red today.

Workaround in use

Start commands expected to outlive the timeout with run_in_background: true and consume them
through bg; do not rely on automatic promotion for anything whose partial output matters.

[yyydddkkk]

I also reproduced this on Linux and prepared a small reference fix with regression coverage. Posting it here because PR creation is currently restricted to repository collaborators.

Reference commit: yyydddkkk@8cfbf6f

Reproduction and proposed fix

Adding the suggested immediate output(&task_id).await.is_some() assertion to test_foreground_timeout_promotes_and_command_keeps_running fails against the original implementation:

output should exist immediately after foreground timeout promotion
test result: FAILED. 0 passed; 1 failed; ... finished in 0.10s

adopt_with_options writes the initial status artifact, but the output artifact is only created by the completion wrapper. The proposed change creates an empty .output file during adoption, before publishing the started activity or returning the task ID. The existing wrapper still replaces it with the complete output at termination.

This follows the issue's stated expected behavior: it makes output readable while Running, but does not implement live stdout/stderr streaming. It leaves pid: null, command lifetime, and timeout behavior unchanged. Initial file creation remains best-effort, consistent with the existing initial status-file write; filesystem-error propagation is not changed.

Regression coverage and validation

• The immediate-read assertion in the existing foreground-timeout test fails before the fix and passes afterward.
• A new adopted_task_output_exists_while_running test uses a oneshot channel to hold the task deterministically in Running, verifies the empty output artifact, then releases the task and checks the complete output.
• cargo test -p jcode-base background::tests: 20 passed.
• cargo test -p jcode-app-core test_foreground_timeout_promotes_and_command_keeps_running: passed with the fix.
• Formatting checks on the changed files and git diff --check: passed.

The full Bash suite, run serially, reports 31 passed / 3 failed. I checked a separate pristine worktree at parent commit ce700ffcd; it produces the same result and the same failures in this environment:

• bash_holds_a_risky_delete_until_justified_then_runs_it
• test_stdin_forwarding_multiple_lines
• test_stdin_forwarding_single_line

The parallel run additionally failed indirect_dispatch_paths_cannot_bypass_the_gate, which passes in both serial runs. These tests are left unchanged. Testing used the Rust test harness on Linux; no shared daemon was replaced or restarted, and I have not validated on macOS.

The change is limited to three files (initialization plus tests). Please feel free to use the reproduction or implementation as a reference for a maintainer-authored fix.

[1jehuang]

Thanks for the report. Promoted background tasks now have a readable output artifact before promotion returns.

Proposed fix: #1218, commit 56ac1a1. Twenty background tests and the real Bash timeout-promotion regression passed. Running output can still be empty, this is not live output streaming.

The PR remains a draft pending clean-branch CI and review. It documents the local build limitations. This issue remains open until the fix is merged and verified.

---

— Jcode agent (automated triage), on behalf of @1jehuang