grainger-fetch-route

name	grainger-fetch-route
description	Decide how to reach grainger.com and prove the route works before spending requests on it. Use at the start of any Grainger task, and whenever a fetch returns something that looks like an empty or missing page.
allowed-tools	Read, Bash(python3 *), Bash(ls *), Bash(stat *)

Choosing a route into Grainger

The fact that governs everything here: Grainger refuses requests with HTTP 200. Three distinct refusals, none with an error status, one of which is a convincing "not found" page. A client that trusts the status code reports an empty result set and sounds sure of itself.

Order of preference

1. scripts/grainger_fetch.py from this machine — typeahead, category search, facets, result SKUs, branch stock. Cheap, fast, no browser.
2. claude-in-chrome — price, spec table, live availability, branch names. Everything the script tier cannot reach. Costs a real browser session.

There is no third route. Verified 2026-08-17: anonymous curl, geo-egress from Israel, geo-egress from the US Windscribe exit, and gateway Playwright all return EDGE_BLOCK. A US IP alone is not enough. Do not spend calls rediscovering this — see reference/access-routes.md.

Always probe first

python3 scripts/grainger_fetch.py probe

Read three fields off the result before doing anything else:

• session.signed_in — false means the cookie jar has no gcom.access_token
• session.token_expires_in_minutes — negative means re-export before continuing
• status — OK, or one of the three refusals

When the probe says no

status	Meaning	Do
NO_SESSION	No exported cookies found	Ask the user to export grainger.com cookies from their signed-in Chrome to ~/.local/share/claude-plugins/grainger/cookies.json. Do not attempt to sign in yourself
EDGE_BLOCK	Session missing/stale, or UA mismatch	Re-export. Check --ua matches the exporting Chrome's version
BOT_WALL	DataDome flagged the IP	Stop. Tell the user, and say that retrying extends the block. Switch to the browser tier or wait
HTTP_ERROR	A genuine HTTP error	Report it as itself — distinct from a block
NOT_JSON	200, parseable as neither JSON nor a known block page	The endpoint shape may have changed. Capture the body and look before concluding anything

What never happens here

• Captchas are not solved or bypassed. A DATADOME_403 ends the script tier for that request. Move to the browser or report the limit.
• No retry loops. Two attempts, then change route or ask.
• No parallel fan-out. This is a real account. Serial requests, with pauses.
• Cookie values are never printed and never written into the repo.

Reporting

Say which tier answered. A price from the script tier does not exist; if the browser tier did not read it this session, report it as not retrieved rather than carrying one over from earlier or inferring it from a list page.