Skip to content
Pi · gpt-5.5infraCLI & tooling

Share pi coding-agent sessions to Hugging Face dataset

Session from gradio-pi-sessions on Hugging Face by abidlabs · MIT · condensed by Coders Talk

14msession
1interventions
GOAL

Upload the developer's local pi coding-agent session logs to a Hugging Face dataset repo, matching the format of an existing reference dataset (synthtraces).

  1. Prompt · +0
    Share my sessions to HF using the extension I downloaded for it
  2. Agent did · +0m
    Located the hf-sessions extension data (.pi/hf-sessions), reviewed approved/denied session manifests, and found one approved session ready to share
  3. Agent did · +2m
    Verified HF login, created dataset repo abidlabs/gradio-pi-sessions, uploaded the approved session and reported success with a commit link
  4. Fail · +5m
    Dataset viewer failed with a CastError because uploaded files (manifest.json, reviews, sessions) had mismatched schemas
  5. Agent did · +5m
    Normalized upload into a single data/sessions.jsonl with consistent columns and proper dataset card YAML metadata; re-uploaded and removed old mixed-schema files
  6. Intervention · +13m
    Still didn't look right; developer pointed to a reference dataset (julien-c/synthtraces) and showed a screenshot of the actual viewer output they wanted
    Why: It still doesn't look like what I was expecting, so I showed the agent the exact dataset I wanted mine to look like.
  7. Agent did · +13m
    Downloaded and inspected synthtraces' raw file layout (sessions/<repo>/<timestamp>.jsonl of raw pi events), then restructured the dataset to match: removed normalized table, uploaded raw trace file under sessions/gradio/, updated dataset card
  8. Outcome · +14m
    Dataset abidlabs/gradio-pi-sessions restructured to match synthtraces format, raw session trace uploaded under sessions/gradio/, latest commit 5fd7f5b3
VERDICT

Next time I'd start by pointing the agent at the exact dataset format I want (like synthtraces) instead of letting it guess a schema first. Showing it the reference dataset's file layout and a screenshot of the viewer I expected was what actually fixed things — everything before that was trial and error with mismatched schemas.

I tried this · 0

Discussion · 0

How Builds work