Investigate and respond to GitHub issue about Excel add-in context handling
Session from pi-for-excel-sessions on Hugging Face by thomasmustier · MIT · condensed by Coders Talk
Click a segment to jump to that moment
13msession
2interventions
GOAL
Review a GitHub issue reporting a discrepancy in context-window behavior between OfficeAI and Pi for Excel, determine which parts were user error versus real bugs, and respond with a comment plus a follow-up issue.
- +0Prompt · +0Review this GitHub issue (#558) about context handling differences between OfficeAI and Pi for Excel — seems like RTFM but check what I'm missing
- +0mAgent did · +0mTraced through the codebase (context-management-policy.md, compaction.md, selection.ts, auto-compaction.ts, read-range.ts, action-queue.ts) and pi-coding-agent docs to understand context window, selection auto-context, compaction triggers, and overflow handling
- +3mAgent did · +3mSummarized findings: ~70% of the issue's questions are already answered in the docs (context scope, files workspace, serialization), but identified a genuine gap around overflow handling for small-context models
- +4mPrompt · +4mOk fine can you draft the comment please
- +4mIntervention · +4mAsked for the draft without bolding or em-dashesWhy: I wanted a plain-text comment I could actually post on GitHub without markdown flourishes.
- Break · 30m away, not counted
- +5mIntervention · +5mRewrote the draft themselves to be much more concise, pointing to docs directly and cutting the fluff, then had the agent continue from that tighter versionWhy: It's open source with docs — I wanted the reporter to actually read them, so I tightened it myself rather than let the agent pad it out.
- +12mPrompt · +12mI posted. let's open the other issue and we can comment the link to the issue
- +12mAgent did · +12mChecked labels, drafted and opened a new issue (#566) for the small-context-window overflow gap, relabeled #558 from bug to question, and posted a comment on #558 linking to #566
- +13mOutcome · +13mOpened issue #566 tracking the overflow-handling gap, relabeled #558 to 'question', and posted a linking comment on #558
VERDICT
Next time I'd ask the agent up front to separate 'what's already documented' from 'what's an actual gap' so I get that split without a second pass. The correction that mattered was rewriting the comment draft myself — the agent's version was accurate but too padded with bold/em-dashes and explanation; once I tightened it and pointed straight at the docs, it was actually postable.