git:20260930.5233368 to git:20260930.1dccf7b

23 added, 19 removed. Audit A to A.

---

name: research

description: Research external questions and compare evidence with source attribution. Use for research or fact-checking, not facts already established in local code.

---

# Research

Match the evidence workflow to the question:

- **Lookup:** open the relevant primary source, verify the requested fact and

cite it with its access date. No report file, credit ledger or graph is required.

- **Synthesis or report:** compare the relevant sources, distinguish evidence from

inference and name unresolved disagreements. For a requested durable report or

consequential research, use [reporting.md](references/reporting.md).

- **High stakes:** use current primary evidence and proportionate independent

review. Never convert missing evidence into a confident conclusion.

Reuse repository material when it already answers the question. A number from a

local check does not require web research. Verify volatile external values at

the source; record when they were accessed. Treat snippets and agent summaries

as leads, not as sources already read.

- For generic web discovery and public-page reading, use Scrapinho: `search.web`

- with explicit `input.engine: "duckduckgo"`, and `fetch.page`. Require a project

- scope and a stable request/idempotency key per acquisition; never share scope

- between unrelated projects. Use the installed `scripts/collect_sources.py` for

- durable evidence, including search results, as described in

- [reporting.md](references/reporting.md). Its dry-run does not load credentials

- or access the network. A missing key or capability is a blocker for this route,

- not permission to silently restore the removed generic scraper.

+ For generic web discovery and public-page reading, call the configured Scrapinho

+ MCP tools directly: `search.web` with explicit `input.engine: "duckduckgo"`, and

+ `fetch.page`. Do not run acquisition/preflight scripts or build a REST client.

+ The active harness owns MCP authentication; a working MCP connection does not

+ require a separate `SCRAPINHO_API_KEY` environment check. Require a project scope

+ and a stable request/idempotency key per acquisition; never share scope between

+ unrelated projects. Save durable evidence with the host's file-writing tools as

+ described in [reporting.md](references/reporting.md).

- For MCP acquisition, run the installed `scripts/preflight_scrapinho_mcp.mjs`

- first for page defaults, and inspect `scraper_capabilities` for the requested

- discovery operation. Reuse schema defaults for omitted fields, but set one

- attempt and bounded limits. Do not infer geography from language or domains:

- web search currently has fixed pt-BR/BR locale. An unsupported geographic

- override must fail rather than silently change. Stop on auth, quota, CAPTCHA,

- 403 or 429; do not retry acquisition or switch proxies to bypass refusal.

+ Start with `scraper_capabilities` to check authentication and the requested

+ operation, then use `scraper_submit`, `scraper_get` and `scraper_read_source`.

+ Read the exposed tool schema for required fields and defaults; set one attempt

+ and bounded limits. Do not infer geography from language or domains: web search

+ currently has fixed pt-BR/BR locale. An unsupported geographic override must fail

+ rather than silently change. A missing MCP connection or capability is a blocker

+ for this route, not permission to substitute scripts or the removed generic

+ scraper. Stop on auth, quota, CAPTCHA, 403 or 429; do not retry acquisition or

+ switch proxies to bypass refusal.

- Poll the admitted job and read sources through `next_cursor=null`; verify raw

- SHA-256 when exporting. Partial sources are not complete evidence. Web search

- supports pages 0 and 1; request page 1 explicitly only when page 0 reports it.

- Keep the effective engine across pages. Preserve each query/page's original

+ Poll the admitted job without submitting it again. Read sources through

+ `next_cursor=null`; partial sources are not complete evidence. Use

+ `scraper_read_search` for structured search results. Verify raw SHA-256 only when

+ raw bytes are actually exported; text read over MCP is not a raw export. Web

+ search supports pages 0 and 1; request page 1 explicitly only when page 0 reports

+ it. Keep the effective engine across pages. Preserve each query/page's original

positions and provenance; deduplicate URLs only when selecting pages to open,

not by rewriting the ranked snapshots. Snippets and missing publication dates

must not become verified claims or invented dates.

Scrapinho also exposes `news.search`, `search.suggestions`, `trends.now` and

`trends.suggestions` when capabilities report them available. Read their JSON

through EOF before parsing. News links may be aggregators: open the primary

article before citing it. These operations do not substitute for Maps, social,

ads, YouTube, `trends.interest`, or scientific indexes. Preserve the installed

`scrapingdog` route for specialized methods without verified parity, then

Firecrawl and host tools after a disclosed bounded failure or missing key.

For papers, start with the free Firecrawl Research Index when available; use

paper-search to cross-check metadata when needed.

Missing optional specialized providers do not block adequately sourced local work.

When a collection explicitly requests Webshare, use the installed

`webshare-fetch URL [--config PATH] [--timeout SECONDS]` helper rather than

building a proxy request in the skill. Keep the existing priority intact:

Scrapinho remains the first choice for generic discovery and public URLs;

ScrapingDog remains only for specialized methods without verified parity.

Webshare requires explicit selection for a separate collection, never an

automatic fallback from Scrapinho or a bypass after refusal. Never proxy

cookies or OAuth traffic. Keep `WEBSHARE_API_KEY` in the environment

and its private config outside the repository.

Convert documents only when needed to extract their content. Verify reading

order and tables. Do not package a whole repository to inspect a known file.

Delegate collection only when independent source work would materially help.

Give collectors narrow questions and return paths or URLs; inspect the sources

before accepting claims. Collection alone does not require an Agent Graph.

Report what is established, what is inference and what remains unverified. A

small or uncontrolled sample does not establish a general productivity claim.

Adapted from [research-stack](https://github.com/nett0eth/research-stack), MIT.