Comparing changes

Port upstream #27 (drop ctx.abort, deliver as steer) and #21 (assert the turn is never aborted; correct the bash.ts comments). #27 keeps #21 abort fix and additionally switches delivery from followUp to steer: a follow-up is drained only once a turn produces NO tool calls, so steering otherwise waited for the entire tool-chaining run to settle. #21 tests asserted followUp delivery and were retargeted to steer.

Port upstream #26. attachAction fell through to "finished" even when the job was still running (wait:false, or a job outliving its awaited donePromise), telling both the agent and the user that a live job had ended.

Port upstream #15. The timeout timer returned early when the session was non-interactive, so the bash timeout never fired under -p / RPC.

Port upstream #28. Monitor-follow cleanup and log-path resolution failures were swallowed rather than surfaced.

@ESSO0428

…am emit Port upstream #19. Preferred over the equivalent ad-hoc fix carried in the @ESSO0428 fork (1ea0e6d), which is not brought over.

Port upstream #20. v2.0.0 kills jobs on session_shutdown, but spawn.ts unrefs the children, so a crashed or SIGKILLed pi left them running forever. Wires process exit plus SIGINT/SIGTERM/SIGHUP.

The notice ends with the absolute install directory, so anchoring on a trailing "pi-patty-bg-tasks" could never match. This suite was already red at v2.0.0, before any of the ported fixes.

v2.0.0 was developed against pi 0.83 and pinned peers at >=0.79.0. Compiles and tests clean against 1.1.0; drop the old-version floor.

Renamed from pi-patty-bg-tasks so the fork never shadows the upstream npm package, and republished from Apothic-AI/bizzy. Upstream is untouched and nothing is sent back. - package.json: name, version 2.0.0 -> 2.1.0, repository URL, and a fork note in the description. `author` stays "Patrick Rho" — MIT requires the original copyright and he wrote this code. - LICENSE: keeps (c) patty.io, adds (c) Apothic for the fork changes. - READMEs (en/ko/zh): title, npm badge, install and clone commands, and a fork banner; author sections credit both Patty (original) and Apothic (fork maintainer). English README gains a "Fork deltas" section. - src/index.ts + src/commands.ts: header and the /bg-version fallback name. - commands.test.ts now matches an npm name@semver shape instead of hardcoding the package name, so a future rename cannot break it. - docs/superpowers/* deliberately left alone: dated design and plan records that name the package as it was, with the original authors paths.

spawnWithFileOutput calls proc.unref() so a background job never keeps pi alive. That is correct for production, but it means that once a spawn has returned nothing references the pending child, so a test awaiting that childs exit sees Node resolve the loop first and is cancelled with "Promise resolution is still pending but the event loop has already resolved". Because the unaffected tests still exit 0, this hid itself as a silently shrinking suite. Affected suites, measured before and after: spawn 0 pass / 6 cancelled -> 6 pass / 0 cancelled bash-results 1 pass / 5 cancelled -> 6 pass / 0 cancelled monitor 4 pass / 2 cancelled -> 6 pass / 0 cancelled Whole suite: 150 pass / 1 fail / 12 cancelled -> 171 pass / 0 fail / 0 cancelled. Holding the loop once per file rather than wrapping each await means a test added later cannot silently regress back into cancellation. Production code is untouched; removing proc.unref() would make pi hang until every background job finished. The test script also gains --test-timeout=30000. Previously a spawn whose exit never resolved surfaced as a fast cancellation; with the loop held it would hang forever, so the timeout turns that into a visible failure. Notably this now actually executes the non-interactive bash-timeout case from #15 and the timeout-kill path, neither of which had ever run.

The Fork deltas section described the spawn/bash-results/monitor cancellations as a known gap left open. That was true when written and is now false: the previous commit fixed them. Rewrite the bullet to record the fix, including that the non-interactive timeout case had never actually executed, and to state the new 171/0/0 result.

Three defects found by audit, each verified with a probe before and after. 1. Job logs were world-readable. LOG_DIR was a hardcoded /tmp/pi-bg, and mkdirSync gave 0755 with openSync giving 0664, so every background job log was readable by any local user. Logs carry whatever a command printed, including anything echoed out of its own environment, and a predictable path in a shared directory is also pre-creatable by another user, whose ownership we would then write into. Logs now live under $XDG_STATE_HOME/pi-bg (absolute check, else ~/.local/state/pi-bg) at 0700/0600, matching where pi-bg-bash keeps them. 2. A vanished log directory wedged the process permanently. A module-level logDirCreated flag meant that once anything removed the directory (tmp cleaner, reboot, an agent running rm -rf) every LATER spawn in that process threw ENOENT for the rest of its life, because the flag said the directory existed so it was never recreated. openLogFd now creates it on ENOENT and retries, paying one mkdir only on the failure path. Proven: a second spawn after removing the dir went ENOENT before, works now. 3. agent_bg children inherited this extension. The worker was spawned as a plain `pi -p`, so it loaded us too: it could nest further agent_bg calls, auto-background its own commands into a turn that had already ended (losing their completion notice), and install process signal handlers on a process it does not own. pi cannot exclude a single extension from a child (--no-extensions also strips the user own skills and MCP), so the child is marked via env and the extension stands down inside it. This is the concern behind upstream #7, which could not be ported because its file no longer exists upstream. Also: the model-facing tool descriptions hardcoded /tmp/pi-bg, so they now derive from LOG_DIR rather than going stale silently. 7 regression tests added. Each was verified to FAIL against the unfixed code. The first attempt at the /tmp guard asserted against os.tmpdir() and passed against the broken value, because this box sets TMPDIR=/tmp/user/1000; it now compares the literal /tmp and pins the resolved root. Suite: 178 pass / 0 fail / 0 cancelled. Version 2.2.0 (log location is user-visible).

Navigation Menu

Repository navigation

Comparing changes

Open a pull request

Uh oh!

Commits on Oct 8, 2026

This comparison is taking too long to generate.

Uh oh!