Use Fetch API for retrieval-only templates

[shrey150]

Summary

• use Browserbase Fetch API for five retrieval-only workflows in both TypeScript and Python
• remove browser sessions, Stagehand, and outer-agent/model dependencies where no page interaction is required
• keep each shipped template focused on the product workflow; correctness checks remain out-of-band
• update setup, environment examples, dependencies, and the root template index

Scope

Template	Why Fetch is the right primitive
website-link-tester	Fetch the homepage, parse links locally, and check destinations without one browser session per URL
business-lookup	Query the official SF Open Data JSON endpoint and select the exact DBA match deterministically
company-value-prop-generator	Use Fetch API's structured JSON response for landing-page extraction
council-events	Fetch the year-filtered, server-rendered Legistar calendar as schema-constrained JSON
sec-filing-research	Resolve and retrieve filings from the SEC's official JSON datasets

This intentionally does not rewrite interactive or rendered-content workflows. cerebras-docs-checker still demonstrates rendered accessibility content, company-address-finder is a Search API candidate for a separate change, and interactive templates remain browser/Stagehand based.

The link tester includes bounded retries for transport errors and 5xx responses after a real Fetch request transiently returned 503 during verification.

Before/after semantic comparison

The exact pre-Fetch parent commit and this branch were both rerun with live credentials. All ten pre-Fetch language entrypoints completed successfully. The new implementations preserve the workflow outputs with these deliberate differences:

Template	Fresh comparison
website-link-tester	Still discovers and health-checks links; removes the probabilistic content-relevance judgment and reports deterministic HTTP metadata instead
business-lookup	Returns the same exact official DBA record and normalized business fields, now directly validated against the source schema
company-value-prop-generator	Returns the same value-proposition and short personalized-opener content
council-events	Both old and new TypeScript/Python runs returned all 91 current-year events; Fetch uses ?Mode=2026 to request the full-year HTML
sec-filing-research	Returns the same company/CIK and five-recent-filing metadata contract from the more direct official SEC datasets

E2E Test Matrix

All results below were produced from the latest branch with a real BROWSERBASE_API_KEY; no historical runs were reused.

Command / flow	Observed output	Confidence / sufficiency
TypeScript website-link-tester: MAX_LINKS=5 pnpm start	60 unique links discovered; 5/5 returned HTTP 200; 0 failures	Proves homepage parsing, bounded concurrent Fetch calls, and report generation on a public target
Python website-link-tester: MAX_LINKS=5 uv run python main.py	60 unique links discovered; 5/5 returned HTTP 200; 0 failures	Proves Python parity for the same public target and limit
TypeScript business-lookup: pnpm start	Exact Jalebi Street record with ownership, account, location, address, and classification fields	Proves direct schema validation and deterministic exact-match mapping
Python business-lookup: uv run python main.py	Same exact official record and normalized fields as TypeScript	Proves Python SDK and Pydantic parity
TypeScript company-value-prop-generator: pnpm start	Schema-valid Browserbase value proposition and Your... opener	Proves structured Fetch extraction; wording remains model-dependent
Python company-value-prop-generator: uv run python main.py	Schema-valid Browserbase value proposition and Your... opener	Proves Python structured-extraction parity; wording remains model-dependent
TypeScript council-events: pnpm start	91 events for 2026 from Calendar.aspx?Mode=2026	Proves the full-year filter and restores the fresh pre-Fetch event count
Python council-events: uv run python main.py	91 events for 2026 from Calendar.aspx?Mode=2026	Proves Python full-year parity and restores the fresh pre-Fetch event count
TypeScript sec-filing-research: pnpm start	Apple resolved to CIK 0000320193; 5 official recent filings returned	Proves company resolution and submissions retrieval against live SEC data
Python sec-filing-research: uv run python main.py	Apple resolved to CIK 0000320193; 5 official recent filings returned	Proves Python SDK parity against the same live SEC datasets

Static verification

• pnpm check
• TypeScript tsc --noEmit for the changed TypeScript entrypoints
• Python compileall for the changed Python entrypoints
• git diff --check

Policy audit

• no interactive Stagehand primitive was swapped for a page method; these workflows no longer require a browser at all
• no provider model name or provider API key is required
• no verification, judging, or eval harness is embedded in shipped template code
• no custom Browserbase API URL is required
• TypeScript and Python use their first-class Browserbase SDK Fetch surfaces

---

Note

Low Risk
Changes are limited to example templates and documentation, with simpler deterministic code paths and no auth or production service logic.

Overview
This PR replaces browser automation and outer-agent stacks with Browserbase Fetch API across five paired TypeScript/Python templates: business-lookup, company-value-prop-generator, council-events, sec-filing-research, and website-link-tester.

Setup and docs now only require BROWSERBASE_API_KEY (plus optional env like BUSINESS_NAME, TARGET_URL, SEARCH_QUERY). Stagehand, Vercel AI SDK / Deep Agents, and AI_GATEWAY_API_KEY are removed where they were only used for HTTP retrieval. Dependencies shift to @browserbasehq/sdk / browserbase, with cheerio or BeautifulSoup for HTML link parsing in the link tester.

Behavior changes by template: SF business lookup fetches the Open Data JSON endpoint and picks an exact DBA match in code; SEC research resolves CIK via company_tickers.json and reads submissions/CIK*.json instead of EDGAR UI navigation; council events and value-prop flows use schema-constrained Fetch JSON instead of act/extract; the link tester reports HTTP status, content type, and titles (with retries on 5xx) and drops AI “content matches link text” checks. The root README table descriptions are updated to match.

Interactive Stagehand-based templates are unchanged by design.

^{Reviewed by Cursor Bugbot for commit 34194a2. Bugbot is set up for automated code reviews on this repo. Configure here.}

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	npm/​cheerio@​1.2.0

View full report

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 27a1001. Configure here.}