You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Keep upstream-chosen drift issue text away from the intake triage model #16635
Once #16620 merges, the triage sweep sends a trusted upstream-drift issue (Upstream drift: title, filed by github-actions[bot] 41898282, edited by no one else) to intake. Intake's triage model then reads the issue's whole body. That body's - new unit <path> rows carry file names the upstream repository chooses (scripts/upstream-drift-issues.sh L159-162), so text from an outside party reaches a model run that holds the lane API key and the job's issues: write token. The claude-security-review lane rated this IMPORTANT on #16620 (thread #16620 (comment)).
These limits still apply:
intake can apply only its allowed labels;
apply-verdict reads only the five verdict lines;
admission, the issue lane and pr-verify read the drift issue only through the fixed driftSpecText template, never through the body.
The reviewer found no path to a leaked token, a compromised account or a charge. The worst case is a wrong triage verdict or a misleading comment on the drift issue.
Fix
Keep the upstream-chosen rows away from the intake model for a trusted drift issue. Pick whichever needs fewer moving parts:
Model-free verdict: the sweep or intake gives a trusted drift issue a fixed verdict (fixed Size and Uncertainty values), and the model never runs on it. That needs a verdict source apply-verdict accepts.
Template body: the ci-workflows intake-triage reusable takes an optional body input. This repository passes driftSpecText for a trusted drift issue, the way write-issue already does for the issue lane. This needs a ci-workflows change.
Filer: scripts/upstream-drift-issues.sh writes no upstream-chosen path text into the body. The re-audit agent already re-derives the changes from scripts/check-upstream-drift.sh.
Done when
A trusted drift issue reaches admission without any upstream-chosen text reaching a model that holds a token, shown by a test.
The triage-sweep header comment and docs/contracts/upstream-drift/README.md describe the route as built.
Size: Small
Complexity: Medium
Uncertainty: Knowable
Flags: none
Blocked by: none
I read this as a hardening task. Once #16620 merges, a trusted Upstream drift: issue reaches the intake triage model with its whole body. That body includes upstream-chosen - new unit <path> rows. The model runs with the lane API key and the job's issues: write token. The done-when items are exact. A test must show that no upstream-chosen text reaches a token-holding model. The sweep header comment and docs/contracts/upstream-drift/README.md must describe the route as built.
You list three routes and say to pick the one with the fewest moving parts. That chooses among implementation means, so I treated it as the issue lane's call rather than an owner question. My recommendation is route 3, the filer change in scripts/upstream-drift-issues.sh. It touches only this repository and needs no ci-workflows change or new verdict source.
I found no open issue that duplicates this and no merged pull request that already fixes it. Related: #14215 and #16620, as you noted.
The work-resolve-issue lane stopped on this issue (closed-by-open-pr): an open PR the lanes App or a listed human opened already closes the issue, a race with the earlier-work check, not a finding about the issue; while it stays open the next run stops before any model turn.
It keeps agent-ready, and work-unblock starts it again (retry 1 of 2).
In the run's words:
Open PR #16640 already closes this issue, so no second PR opened from work-resolve-issue/issue-16635-43951792506-1.
Partial work is on the branch work-resolve-issue/issue-16635-43951792506-1. The next run continues from it while it is not behind the default branch, and otherwise starts fresh.
This records a re-fire of an earlier stop, not a new one. The start gate stopped run 43958598829, the work-resolve-issue run that agent-ready started, before any model turn, because the GITHUB_TOKEN budget was under its start-gate share (limits.api github-token start-gate); that run ended at 2026-11-16T21:20:06Z. work-unblock has now removed and added agent-ready back to start a fresh run. A budget stop is the lanes' load, not a fault in this issue: it never counts toward limits.limit-stops-per-issue or the kill-switch wait.
Activity
github-actions commentedon Nov 16, 2026
Size: Small
Complexity: Medium
Uncertainty: Knowable
Flags: none
Blocked by: none
I read this as a hardening task. Once #16620 merges, a trusted
Upstream drift:issue reaches the intake triage model with its whole body. That body includes upstream-chosen- new unit <path>rows. The model runs with the lane API key and the job'sissues: writetoken. The done-when items are exact. A test must show that no upstream-chosen text reaches a token-holding model. The sweep header comment anddocs/contracts/upstream-drift/README.mdmust describe the route as built.You list three routes and say to pick the one with the fewest moving parts. That chooses among implementation means, so I treated it as the issue lane's call rather than an owner question. My recommendation is route 3, the filer change in
scripts/upstream-drift-issues.sh. It touches only this repository and needs no ci-workflows change or new verdict source.I found no open issue that duplicates this and no merged pull request that already fixes it. Related: #14215 and #16620, as you noted.
Automated triage. A maintainer reviews it.
melodic-automation-lanes commentedon Nov 16, 2026
{"v":1,"actor_id":337595936,"title_sha256":"77d230e169f4445862e216e8956f264aa123a5de5585c3a95ece9f53618a5a73","body_sha256":"0ff06507820854e0aa6fc6d5fca91c99cd3f0cbc130f81e8aeff0d1a18b56503"}
github-actions commentedon Nov 16, 2026
melodic-automation-lanes commentedon Nov 16, 2026
The work-resolve-issue lane stopped on this issue (
closed-by-open-pr): an open PR the lanes App or a listed human opened already closes the issue, a race with the earlier-work check, not a finding about the issue; while it stays open the next run stops before any model turn.It keeps
agent-ready, and work-unblock starts it again (retry 1 of 2).In the run's words:
Partial work is on the branch
work-resolve-issue/issue-16635-43951792506-1. The next run continues from it while it is not behind the default branch, and otherwise starts fresh.Run: https://github.com/melodic-software/claude-code-plugins/actions/runs/43951792506
melodic-automation-lanes commentedon Nov 16, 2026
This records a re-fire of an earlier stop, not a new one. The start gate stopped run 43958598829, the work-resolve-issue run that
agent-readystarted, before any model turn, because the GITHUB_TOKEN budget was under its start-gate share (limits.api github-token start-gate); that run ended at 2026-11-16T21:20:06Z. work-unblock has now removed and addedagent-readyback to start a fresh run. A budget stop is the lanes' load, not a fault in this issue: it never counts toward limits.limit-stops-per-issue or thekill-switchwait.github-actions commentedon Nov 16, 2026