Hermes Web Tools

Hermes Web Tools (web_search / web_extract)

Overview

How backend selection works for Hermes' two web tools, and how to diagnose a broken web_extract/web_search. All providers now ship as bundled plugins (web-firecrawl, web-tavily, web-exa, web-parallel, web-keenable, web-searxng, web-brave-free, web-ddgs, web-xai). Plugins showing "not enabled" in hermes plugins list is NORMAL — they load on demand; the config selection is what matters.

USER RULE — investigate first, change nothing without explicit command

The operator requires read-only investigation for Hermes config/troubleshooting tasks: diagnose, present fix options with exact commands, and WAIT for him to pick one. Do not run hermes config set, edit config.yaml, or enable/disable anything on your own initiative, even when the fix is obvious. This overrides the general "prefers autonomous execution" tendency.

Backend selection (priority order, per capability)

1. web.search_backend / web.extract_backend (per-capability override)
2. web.backend (shared fallback; nous = managed Tool Gateway)
3. Auto-detect from env vars — ONLY if no selection was ever stored

Strict selection: a stored name is used as-is, no availability probe, no silent swap. An unknown name surfaces an honest error naming the bad selection. Once any selection exists, adding keys to .env does NOT reroute traffic — a stale selection silently blocks otherwise-working credentials.

Keyless free-tier ring: zero-credential installs rotate across Exa, Parallel, Tavily, Firecrawl, Keenable public free tiers with automatic failover. Plus a one-shot keyless rescue for keyed backends on failure (never sticky). Disable via web.keyless_fallback: false / web.keyless_rescue: false.

Valid backend names

Anything else (e.g. trafilatura — a Python extraction library, never a valid Hermes backend name) errors with "no registered web extract provider has that name".

Diagnosis recipe

0. Differential test FIRST — is the tool broken, or just this site? Call the same tool on a known-good URL (https://example.com, https://en.wikipedia.org/wiki/Web_scraping). If it returns full content there, the original failure is site-specific, NOT a tool fault — stop diagnosing the tool/config and investigate the target site instead. This one test prevents pivoting into hours of needless config poking or SPA-bundle reverse-engineering when the only real issue is that the page renders client-side. (Session lesson: an "empty content" web_extract on one URL prompted a full Docusaurus-bundle reverse-engineering detour that a 2-URL differential test would have made unnecessary. The user explicitly redirected: "why is the web tool failing? you should check it out" — i.e. diagnose the failure, don't silently work around it.)
1. Reproduce: call the tool on the failing URL; read the exact error string.
2. grep -n -A5 '^web:' ~/.hermes/config.yaml — check stored selections.
3. grep -oE '^(TAVILY|EXA|FIRECRAWL|PARALLEL|KEENABLE|BRAVE_SEARCH)_API_KEY' ~/.hermes/.env — what credentials exist (names only; values are secrets).
4. Validate the stored name against the table above. Unknown name = root cause.
5. Check hermes plugins list for the matching web-* plugin (informational only).
6. Present fix options; wait for the operator's pick. Fixes need a NEW session (/reset) — tool config is not re-read mid-conversation.

Common pitfalls

• Stale/typoed selection is the #1 cause of "web tool broken" reports. The strict-selection design (commit d7119ea2) means it errors instead of falling back — the error names the bad value directly.
• Search can work while extract is broken: per-capability keys resolve independently (e.g. search_backend: ddgs fine, extract_backend: <invalid> broken).
• hermes config set web.<key> "" clears a selection and re-enables auto-detect.
• Tool/toolset config changes take effect on next session, never mid-session.
• Empty web_extract on a JS-SPA site is a site limitation, not a tool bug. Sites built with Docusaurus / Next.js / Vite / plain SPAs serve an HTML shell (<div id="__docusaurus">) whose article text lives in JS chunks the fetcher can't execute. web_extract (Tavily) legitimately returns empty there. Confirm with the Step-0 differential test; if you still need the content, fetch via curl -sSL in the terminal — a Docusaurus page's real markdown is referenced by the route map inside main.*.js (search for @site/docs/...md keys). Do NOT conclude the tool is broken from a single empty extract.

References

• references/backend-selection-internals.md — code paths in tools/web_tools.py, resolution chain, incident transcript.