[LukasParke]

Why

Every weekly port run since 10/12 has been killed by the 2h converter alarm (perl alarm, exit 142 "Alarm clock") partway through the doom-loop (78c562e/75271c3) or async-tools (e8d7d6d) delta. Because the port step failed, the commit/PR steps were skipped and all work was lost. On 10/05 the opposite happened: opencode ended on Error: Tool execution aborted but exited 0, so a half-finished tree was opened as eval-failed, as if the eval had judged it.

What

scripts/upstream (same change in python-agent and go-agent):

• Captures opencode's exit status instead of letting set -e exit on it.
• 142 = timeout; any other non-zero status is passed through; 4 = opencode exited 0 but its last output line is Error: ... (abort / terminal tool error). Mid-run tool errors the agent recovers from, like a failed edit it retries, are still treated as normal.

.github/workflows/upstreamer-port.yaml:

• The port step records rc and doesn't fail, so the partial tree is still committed and pushed to upstreamer/sync, and the PR is opened or updated.
• The outcome now has three values, each with its own title and label:
  • passed
  • eval-failed: the converter finished but state.yaml didn't advance
  • port-incomplete: the converter was cut off, so the eval result means nothing
• An incomplete run reverts state.yaml so it can never advance the pointer.
• create-pull-request only adds labels, so a new step removes stale eval-failed / port-incomplete labels from a reused PR.
• A final step fails the job when the converter didn't finish, so the run still shows red.
• New timeout_minutes dispatch input (default 120, max 300). The job timeout-minutes goes from 150 to 330 so the converter's alarm always fires before the job is killed. If the job limit hit first, nothing would be saved.
• PR body now links the run and states the outcome and exit code.
• PORTING.md documents port-incomplete.

Testing

• Ran the new script tail with a fake opencode through four cases: clean exit (with a recovered mid-run Error:) gives 0, Tool execution aborted as the last line gives 4, exit 7 gives 7, and a hang under UPSTREAMER_TIMEOUT_SECONDS=2 gives 142.
• actionlint (with shellcheck) is clean. The script's shellcheck shows only one warning, at line 140, which this PR doesn't touch.
• The repo's CI checks pass locally.

Not in scope / follow-ups

• A later run still starts from main and re-ports the whole range. Partial work is now preserved for a human to look at, but it isn't resumed from. The stepwise ref dispatches are the way through large ranges.
• ${{ inputs.ref }} is still interpolated straight into a shell script (pre-existing). The new inputs go through env:.

🤖 Generated with Claude Code