Deep research
A research run that plans a question into parts, searches your files and the web across several rounds, and comes back with a cited report and a ledger that separates what your files back from what only the web says.
/deepresearch is for questions that need more than one search pass. Perch plans the question into sub-questions, searches your indexed files and the public web, reads what it finds across several rounds, and writes a report with its sources and a claim ledger.
/research and /deep-research are the same command.
/deepresearch what changed in EU AI Act enforcement this year
/deepresearch --rounds 5 does creatine improve strength in trained adults
/deepresearch --web only use the public web for this merger timelineReach for it when you want a researched answer with sources you can check. Do not reach for it for a one-line lookup, which web search handles in a single turn, or for a finish-line task, which belongs on a goal.
Flags
| Flag | What it does |
|---|---|
--rounds N | Set the round budget, 1 to 6. The default is 3. -r and --rounds=N work too. |
--web | Skip your workspace files and research the public web only. --web-only is the same. |
--resume | Continue a run parked by a hard interrupt. --continue is the same. |
Bare /deepresearch with no question prints help and spends nothing. If a run is waiting to be resumed, it says so.
How a run goes
A run takes a few minutes, and it is deliberately broken into phases so you can watch it rather than stare at a spinner.
- Plan. The question becomes a set of sub-questions, each routed to the kind of source that suits it: a general web search for most, a specialized index where one fits better.
- Rounds. Each round searches, opens the pages worth reading, and indexes what it found so later rounds can search inside it. What an earlier round turned up shapes what the next one asks.
- Synthesize. The findings become a report with numbered sources.
- Sweep. The temporary index built for this run is deleted. Nothing from it stays in your workspace.
A footer keeps the cost visible while it runs: searches made, pages read, and model calls spent.
The claim ledger
This is the part worth understanding, because it is the reason the command exists. Every claim in the report lands in one of three tiers:
| Tier | What it means |
|---|---|
| Backed by your files | A document in your workspace corroborates it, and the ledger names the file. |
| Web only | A web source says it, and the ledger points at the numbered citation. Nothing in your own material confirms it. |
| Unsupported | Neither one backs it up. |
Unsupported claims are not hidden or quietly dropped. The chat reply highlights the backed and the unsupported ones, and the full record lists all three tiers itemized.
With --web there is no first tier at all, because the workspace pass never runs. The same is true if nothing in your folder is indexed yet, and the run says so rather than implying your files agreed.
The report file
On the desktop app and the CLI, Perch also writes a markdown file under perch-research/ in your working folder, holding the full report, the plan, every source, and the complete ledger. That file is the durable record of the run; the chat reply is the summary.
Resuming
A hard interrupt parks the run so you can pick it up with /deepresearch --resume. You have about 45 minutes, after which the park expires and you start again from the question. A graceful cancel writes a partial report instead and leaves nothing to resume.
Where it works
Deep research runs in Perch Terminal, which is the desktop app and the CLI. Web chat does not have slash commands.
Background tasks
Long commands and delegated agents run in the background while you keep the chat, and Perch can pick the work back up on its own when one of them finishes.
Artifacts
Create documents, interactive pages, code, and tables in chat, then open, revise, download, and share them from the Artifacts library.

