| name | codenames-feedback |
| description | Pull and summarize user feedback (๐/๐ on clues) for the Hebrew Codenames co-pilot from the private HF dataset shmulc/codenames-feedback. Use when asked to "check feedback", "any new feedback", "grab feedback", review clue thumbs-up/down, or see what users complained about. |
Codenames feedback
Live ๐/๐ feedback on clues is written by the deployed Space
(shmulc/hebrew-codenames-copilot) and mirrored to the private HF dataset
shmulc/codenames-feedback (file data/feedback.jsonl). The truth is on HF โ
the local feedback/feedback.jsonl is only a seed/mirror, so always query HF.
Run
python .agents/skills/codenames-feedback/scripts/pull_feedback.py
Requires an HF token with read access to the dataset (the stored ~/.cache/huggingface/token
is used automatically; no HF_TOKEN env needed). Uses force_download so you always get the
latest, not a cached snapshot.
What it reports
- Verdict counts (๐/๐), time span, and how many rows are new vs the local mirror.
- Every written comment (these are the highest-signal rows โ real complaints).
- Self-test rows (comment contains
ignore/test) are flagged so you can discount them.
Each row's fields
ts, verdict, clue, count, intended, mode, risk, comment, board {words, roles}, revealed, focus, why.
The full board + focus + revealed cards are stored, so any row is reproducible โ feed it back
through the engine (see the codenames-qa skill's bench_feedback.py) to see what the current
engine would do on that exact board.
Turning feedback into fixes
Downvotes cluster into: illegal clues (shared root โ now caught), over-reach (a stretched
extra word), and taste (e.g. a more general clue preferred). See the project memory
(spymaster-scoring-design, codenames-feedback-loop) for the roadmap these drove.
