Add deep-research skill#127
Conversation
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
| For each sub-question: | ||
|
|
||
| 1. Run 2-3 search query variations. Use parallel tool calls when the host supports them. | ||
| 2. Fetch the top 3-5 relevant unique URLs. Prefer primary sources, official docs, filings, company pages, reputable reporting, and recent material. |
There was a problem hiding this comment.
Quick depth contradicts plan
Medium Severity
quick allows 20 steps for a single narrow question, but Phase 1 still requires 3–7 sub-questions and Phase 2 still mandates 2–3 searches plus 3–5 fetches per sub-question. Following all sections exceeds the budget or forces incomplete research.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 981008c. Configure here.
| await opened.browser.close().catch(() => {}); | ||
| } | ||
|
|
||
| console.log(JSON.stringify({ ok: true, mode: opened.mode, output: outputPath }, null, 2)); |
There was a problem hiding this comment.
PDF success after load failure
Medium Severity
The renderer catches every page.setContent error and still runs page.pdf, then prints ok: true. Real load failures (not just late timeouts) can produce blank or truncated PDFs while the skill treats PDF generation as successful.
Reviewed by Cursor Bugbot for commit fb7387b. Configure here.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
There are 3 total unresolved issues (including 2 from previous reviews).
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 814f748. Configure here.
| const browser = await chromium.launch({ headless: true }); | ||
| return { browser, mode: 'local-playwright' }; | ||
| } | ||
| } |
There was a problem hiding this comment.
Browserbase sessions not released
Medium Severity
Remote PDF mode creates a Browserbase session via the SDK but never retains its id or calls REQUEST_RELEASE. Cleanup only invokes browser.close() on the CDP connection, which in-repo Browserbase guidance says can leave cloud sessions running and billable, including when connectOverCDP fails after sessions.create succeeds.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 814f748. Configure here.
1 participant
Summary
Changes
Test plan
Notes
Full Browserbase-backed test report artifacts were generated under .deep-research/full-reports-2026-06-04/ and intentionally left untracked.
Note
Low Risk
Adds documentation, marketplace metadata, and an optional Node renderer; no changes to existing runtime skills or auth paths beyond requiring BROWSERBASE_API_KEY for remote PDF sessions.
Overview
Adds a deep-research skill that turns agents into a Browserbase-backed planner–researcher–synthesizer: plan sub-questions, search/fetch via
bb/browse cloud/ APIs, fall back to remote browser when fetch is dynamic or blocked, maintain a citation ledger, and write cited markdown (plus optional PDF).skills/deep-research/SKILL.mddocuments depth budgets, untrusted-web discipline, dynamic-fetch → browser rules, prospect report shape, and a completion checklist.scripts/render-report.mjs(Playwright + Browserbase CDP by default,--localfor smoke tests) converts final markdown to A4 PDF with escaped HTML and safe link handling;package.json/ lockfile pin@browserbasehq/sdkand Playwright.README and
.claude-plugin/marketplace.jsonregister the skill and add a usage example for cited competitive research.Reviewed by Cursor Bugbot for commit 7895021. Bugbot is set up for automated code reviews on this repo. Configure here.