research: nothing says when to start writing the findings file, so an interrupted agent leaves nothing on disk

### Skill `research` ### Harness and version Claude Code. The version at the time was not recorded. The skill was in our repo from 2026-08-05 with the same text as upstream (checked against our commit from 2026-08-18). Upstream's `research/SKILL.md` has changed only by the em-dash removal on 2026-08-19, so the text those runs used is the text that is here now. ### Model and effort Not recorded in our notes. The research agents were dispatched as background sub-agents from a Claude Code session. ### What happened Over 2026-08-17/18 we dispatched four background research agents for one piece of work. The same failure happened three times: the agent was interrupted (user stop, API disconnect, a fetch that hung) while everything it had found was still only in its context, and nothing was on disk. - One read for about 30 minutes and used about 113k tokens. It was stopped with no file written. - One lost its connection while holding a formula and measured results it had already worked out. Nothing was saved. - One lost its connection with all of one angle's findings in context. Nothing was saved. Step 2 of the skill says "Write the findings to a single Markdown file", and step 3 says where to save it. Neither says when to start writing, so an agent can reasonably collect everything and write at the end, and an interruption before that point loses all of it. After we added "create the file within the first few tool calls, with a heading per section, and update it after each source" to our dispatch briefs, the next run was killed twice and still kept its first two sections intact. ### What you expected An interrupted research agent should leave its partial findings on disk. The line in step 2 that says what the file is for could also say when it exists. Something like: > Create the file in your first tool calls, with a heading per section, and update it after each source you finish. Anything held only in your context is lost if you are interrupted. ### Caveats - The dispatch briefs for those runs were hand-written by us, so I cannot show that the skill's own agent spawn would behave the same way. The skill text says nothing either way, which is the gap. - The original transcripts have expired. What is above comes from notes we wrote at the time, not from re-reading the sessions. - This is separate from #530 (a research agent re-spawning agents), which is about nesting, not about losing work on interruption. (Drafted with Claude Code from our notes.)