[RowanAldean]

Adds Parallel as an explicit provider for the existing web_search tool, so users can research public CVEs, vendor advisories, and documentation with a Parallel API key.

export STRIX_WEB_SEARCH_PROVIDER="parallel"
export PARALLEL_API_KEY="your-parallel-api-key"

The adapter uses the Parallel Search API (POST /v1/search, fast mode) and formats up to five source URLs with excerpts into the existing {success, query, provider, content} response. Excerpts are shortened to fit a 12,000-character output budget plus a truncation notice; source titles and URLs are always preserved, including when those headers alone exceed the budget. It includes configuration persistence, missing-key guidance, error handling, documentation, and mocked tests; no additional dependencies.

Existing auto precedence is unchanged, and Parallel requires explicit selection. Page extraction remains Exa-only.

Related to #1179. This is the native-provider option discussed there; it complements @AyushDubey23's optional MCP documentation in #1185.

Validation

• Search/configuration tests: 96 passed, including seven additional truncation regression cases.
• Full suite before the truncation follow-up: 2,223 passed, 3 expected failures, 1 unrelated failure (see below). The follow-up was validated with the focused search/configuration suite.
• Formatting, lint on changed files, mypy, and Bandit pass. Pyright passes for the search tool and settings; repository-wide Pyright diagnostics are unchanged from main.
• make check-all stops at a pre-existing RUF100 in tests/test_tool_call_ids_providers.py:109, also reproduced from unmodified main.
• The unrelated tests/test_pricing.py::test_resolves_common_bare_model_names fails for the grok-4.5 provider mapping; reproduced on clean main at 007ed1a with the same locked dependencies.
• HTTP behavior is covered by mocked tests; no live API smoke test was run.

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Adds opt-in Parallel web search provider.

The PR appears safe to merge, though long Parallel results should keep source links intact.

Findings

1.  Source links can be cut ▶

Fix with agent prompt

### Issue 1
strix/tools/web_search/tool.py:180-181
When five Parallel results together exceed 12,000 characters, this slice can cut through a source URL or excerpt. The rest is discarded, so the agent may get an unusable link with no way to recover the missing text. Keep each result’s title and URL intact when shortening excerpts.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds Parallel as an explicitly selected provider for the existing web_search tool without changing auto provider choice. Parallel returns source URLs and excerpts, while page extraction remains Exa-only.

• Adds Parallel API settings, setup guidance, and saved-config support.
• Sends searches to Parallel and formats up to five results with a 12,000-character cap.
• Adds mocked coverage for provider selection, formatting, limits, and errors.

_{Reviews (1) · Last reviewed commit: "feat(web_search): add opt-in Parallel pr..."}

[mintlify]

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit e366036.

Approve preview