fix(webfetch): keep body read inside the timeout scope to stop "context canceled"

[ysyneu]

Problem

webfetch against large/slow pages (reported on Aliyun help docs, e.g. https://help.aliyun.com/zh/openapi/user-guide/openapi-mcp-server-guide) fails with:

remote webfetch operation failed: failed to read response: context canceled

The user set timeout: 60, yet the error is context canceled, not context deadline exceeded — so it was never a real timeout. It's a premature context cancel.

Root cause

fetchURL created a per-request timeout context and defer cancel()'d it, then returned the *http.Response:

func (e *Environment) fetchURL(...) (*http.Response, error) {
    httpCtx, cancel := context.WithTimeout(ctx, timeout)
    defer cancel()                 // fires when fetchURL RETURNS
    ...
    return fetchClient.Do(req)     // cancel() runs here, killing httpCtx
}

WebFetch then called readResponseBody(resp) after fetchURL returned. net/http ties the lifetime of resp.Body to the request context, so the io.ReadAll aborts with context canceled for any response still streaming when Do() returned. Large pages trip this every time; small pages got lucky only because their body was already buffered when Do() returned.

Safari's own localWebFetch (the support-agent path) does it correctly — Do() and ReadAll share a single defer cancel() scope. The runner had split them across the cancel boundary.

Fix

Merge the request and the body read into one timeout-scoped function (fetchBody), so the cancel cannot fire before the read. This makes the bug class structurally impossible. processor.Process still runs on the outer ctx (body is already in memory, must not be bound to the per-fetch timeout).

Test

TestFetchBody_ReadsSlowStreamedBody streams a body slowly (so it cannot be pre-buffered) and asserts the full body is read. Verified it reproduces the exact failure against the old split structure (failed to read response: context canceled) and passes against the fix.

🤖 Generated with Claude Code