[man4ish]

Why

A static and live audit of all 1,716 HTTP tools that TES registers here found that most fail inside this executor before the upstream API sees a valid request:

• 1,004 tools declare inputs only in inputs_schema, with no inputs list. Their {placeholders} were never substituted, so the URL went out literally, e.g. /vep/human/region/%7Bregion%7D/%7Ballele%7D. Their required fields were never validated either.
• 24 tools put their POST body under http.body. Only http.body_map was read, so they POSTed an empty body.
• 40 tools get a 3xx from upstream. httpx doesn't follow redirects by default, so every such run failed.

What

1. _input_fields() falls back to inputs_schema.properties for any name the inputs list doesn't declare. Where both exist, the inputs entry wins. Schema secret: true is honoured for log masking.

2. http.body is accepted as an alias for http.body_map. Nested JSON values, such as GraphQL variables, keep their structure instead of being sent as a Python repr.

3. Header values are templated, so API keys can go in a header rather than the query string.

4. Redirects:

   • Followed only when they stay on the same host or are an http→https upgrade.
   • Never downgraded to http; at most 5 hops.
   • When an allowed redirect changes host, credential headers (Authorization, Cookie, *api-key*, *token*, …) and secret-valued query params are stripped first.
   • Refused redirects fail the run with a message that names hosts only.

5. Response-kind guard on every response, not only redirects. Many upstreams return 200 with a landing page, login wall or error page, and the run used to finish COMPLETED with {"text": "<html>…"}. Now:

   • JSON tool (Accept mentions json, or a response_map that reads anything other than the {"text": …} wrapper): the body must parse as JSON. The body is parsed rather than trusting the content type, so mislabelled JSON still passes.
   • Text tool (response_map reads only text, e.g. KEGG, efetch, Cactus): the body must not be HTML. Without this exception, 16 working text tools would have broken.
   • No expectation: unchanged.

   From the live probe: this fails 19 redirect-to-HTML runs, plus 11 tools that previously "succeeded" with HTML or XML (DisGeNET, UniChem, DepMap, Orphanet, …).

Tests

428 passed (26 new: schema fallback, body alias and nesting, header templating and masking, 8 redirect-policy cases, 9 response-kind cases). The ruff error count is unchanged from main (25, all pre-existing).

Merge order

Merge and deploy this before the TES PR (OmniBioAI/omnibioai-tes, branch fix/http-tool-definitions). TES's new CI gate assumes this executor behaviour.

🤖 Generated with Claude Code