Thanks for the clear report. I found the cause, and it was a bug in Objo Studio, not the model, Anthropic, or your credits.
For each message you send, Studio prepares the list of tools the assistant may use for that reply. That preparation was checking the wording of each single message and only including the file-editing tools when the message looked like a change request - in practice, when it contained an English action verb such as add, create, edit, fix, update, or a short approval like "do it". A follow-up such as "continue", "keep going", or "the same for the other windows" matched none of those words, so for that reply the editing tools were silently left out. The assistant wasn't refusing to write - it genuinely had no write tool available in that round, which is exactly what it told you. Your next message happened to include a verb, and the tools returned, which made the problem look intermittent. Prompts written in other languages were affected too, since the wording check only recognised English verbs.
This is fixed for the next release (bug/1666): while Allow AI to edit code is enabled in Settings > AI Assistant, the editing tools are now available for every message in the conversation, however each message is worded and in any language. The only time the assistant cannot edit code is when that permission is turned off.
To answer your question about what to post for debugging: nothing was needed this time, but for future reference the Diagnostics button in the AI panel toolbar exports a report showing exactly which tools were offered for each round - that's how you can verify behaviour like this yourself. Note the export includes your conversation text, so review or redact it before posting anything publicly.