Summary
The task-file shows a coarsened summary of acceptance criteria (e.g. three bullets for an Imbue-study task), while the authoritative AC list lives in the proposal file (e.g. AC1–AC6). When the coder's AC self-check enumerates the task-file bullets rather than the proposal ACs, the reviewer flags the gap as blocking — exactly what happened in gh-ocannl-270 round 2. This is a recurring failure mode worth a structural fix.
Data Points
- "AC self-check walks the proposal's ACs, not the task-file bullets. The task-file shows a coarsened summary; the authoritative ACs live in the proposal file. When the proposal expands the task-file's N bullets into M > N ACs, the self-check must enumerate all M with separate evidence lines. Skipping the proposal-vs-task delta is a recurring failure mode and the reviewer treats it as blocking." (from gh-ocannl-270 / coder, 2026-05-05)
Raw Excerpts
Original feedback
- AC self-check walks the proposal's ACs, not the task-file bullets. The task-file shows a coarsened summary; the authoritative ACs live in the proposal file. When the proposal expands the task-file's N bullets into M > N ACs, the self-check must enumerate all M with separate evidence lines. Skipping the proposal-vs-task delta is a recurring failure mode and the reviewer treats it as blocking.
— gh-ocannl-270--workflow-feedback-coder.md
- "Issue is updated" means post a comment, not "the linked file mentions the issue." AC6-style criteria of the form "the GH issue is updated to link to X" are satisfied only by an actual GH-side update (comment or body edit) — a one-way cite from the docs to the issue does not close the loop.
— gh-ocannl-270--workflow-feedback-coder.md
(Note: separately captured in docs/swe-textbook.md Entry A; cited here as same precipitating round.)
Suggested Action
Add a coder-skill instruction (pair-coder-work.md or worker-conventions.md) explicitly directing the AC self-check to read the proposal file and enumerate every AC there, with a one-line guidance that the task-file bullets are non-authoritative. Could also add a tiny mechanical lint that, in the coder's .peer-sync/ self-check artefact, the count of AC<N> evidence lines is >= the count of ^- AC or ^### AC lines in the linked proposal.
Related: #338 (traceability artifacts belong in proposal, not task file) — same root cause, sibling consequence.
Filed by ludics-feedback-digest
Summary
The task-file shows a coarsened summary of acceptance criteria (e.g. three bullets for an Imbue-study task), while the authoritative AC list lives in the proposal file (e.g. AC1–AC6). When the coder's AC self-check enumerates the task-file bullets rather than the proposal ACs, the reviewer flags the gap as blocking — exactly what happened in gh-ocannl-270 round 2. This is a recurring failure mode worth a structural fix.
Data Points
Raw Excerpts
Original feedback
Suggested Action
Add a coder-skill instruction (
pair-coder-work.mdorworker-conventions.md) explicitly directing the AC self-check to read the proposal file and enumerate every AC there, with a one-line guidance that the task-file bullets are non-authoritative. Could also add a tiny mechanical lint that, in the coder's.peer-sync/self-check artefact, the count ofAC<N>evidence lines is>=the count of^- ACor^### AClines in the linked proposal.Related: #338 (traceability artifacts belong in proposal, not task file) — same root cause, sibling consequence.
Filed by ludics-feedback-digest