fix(tools): align serper, firecrawl, langsmith and qdrant declared shapes with the APIs

[waleedlatif1]

Declared tool outputs that do not match what the provider actually returns. Every shape change below is backed by a primary doc, cited inline. Each fix has a test that was verified to fail against the old code first.

1. Serper places mapped fields /places does not return

places.toResult emitted link, snippet and reviews. A /places item has none of those keys, so all three resolved to undefined on every result, while the fields Serper does return (ratingCount, category, phoneNumber, website, latitude, longitude, cid) were dropped. ratingCount was a declared output that no code path could ever populate — the review count arrives under that name, not as reviews.

Documented /places item (serper.dev, via the published response examples):

{ "position": 1, "title": "Whole Foods Market", "address": "2001 Market St, San Francisco, CA 94114",
  "latitude": 37.7687616, "longitude": -122.42701559999999, "rating": 4.2, "ratingCount": 1500,
  "category": "Grocery store", "phoneNumber": "(415) 626-1430", "website": "https://…", "cid": "6353588238324409422" }

link and snippet stay declared but are now optional — they are real on every other vertical, so they are documented as absent on places rather than removed. reviews was never a declared output (only a TS field commented "not an advertised output"), so dropping it removes no advertised surface.

Also: news items always carry source (declared on the unified result as "Source name (news/videos/shopping)") and the mapper dropped it. Now mapped.

Docs: https://serper.dev (places + news response examples). Cross-checked against the News type in the community client at https://github.com/tkdkid1000/serper/blob/main/docs/types/News.html.

2. Serper /search panels were undeclared

peopleAlsoAsk and relatedSearches are top-level keys on the /search payload and were never surfaced. Both are now declared and populated, and stay absent (not empty arrays) on verticals that do not return them.

Docs: https://serper.dev (search response example showing both blocks verbatim).

knowledgeGraph is not included — see "Unverifiable" below.

3. Firecrawl search declared the wrong container

SearchResponse.output.data was Array<…>. POST /v2/search answers with a source-keyed envelope: data.web, data.news, data.images, present only for the requested sources. News items carry snippet (not description); image items put the containing page in url and the image in imageUrl. Declared outputs now model all three item shapes, and an absent body resolves to {} rather than undefined.

Docs: https://docs.firecrawl.dev/api-reference/v2-openapi.json (paths./search.post.responses.200). Cross-checked against SearchData / SearchResultWeb / SearchResultNews / SearchResultImages in the official JS SDK: https://github.com/firecrawl/firecrawl/blob/main/apps/js-sdk/firecrawl/src/v2/types.ts.

Note the drift's origin: the repo's apps/api/openapi.json and v1-openapi.json both still describe v1, where data genuinely was an array. The tool calls /v2.

4. Firecrawl map — timeout param renamed, and links are objects

request-transport.ts:191 reads params.timeout as the outbound fetch deadline for every tool. firecrawl_map declared a param of that name intended for Firecrawl's own map deadline, so setting it also armed the local abort at the same instant — the client abort wins the race and the caller gets an opaque AbortError instead of Firecrawl's response. Renamed to mapTimeout and mapped to the body's timeout.

No subBlock id changed. The block's timeout subBlock is conditioned on ['scrape', 'search', 'parse'] and the map branch of tools.config.params never forwarded it, so no saved workflow state referenced this param — the rename is confined to the tool's direct-call surface. Nothing needed adding to tools.config.params.

firecrawl/parse.ts's timeout is left alone as instructed, and is unchanged in this PR.

Separately, /v2/map returns links as SearchResultWeb[] — objects with a required url plus optional title/description — while the tool declared items: { type: 'string' }. Corrected.

Docs: https://docs.firecrawl.dev/api-reference/v2-openapi.json (components.schemas.MapResponse, and paths./map.post.requestBody for timeout: "Timeout in milliseconds. There is no timeout by default."). Cross-checked against MapData.links: SearchResultWeb[] in the official JS SDK.

5. LangSmith feedback score coercion only ran in the block

blocks/blocks/langsmith.ts did the Number(value) + NaN check inside tools.config.params, so only the block path got it — LLM tool calls and direct tool calls posted the raw value. Moved into create_feedback.ts's body, so every caller is covered; the block now passes the value straight through. Behavior is preserved exactly, including the Invalid score: "…" is not a number message and treating ''/null/undefined as omitted.

6. Qdrant block declared outputs no tool emits

QdrantBlock.outputs declared matches and upsertedCount. All three Qdrant tools return exactly { status, data } — no code path has ever produced either key, so every reference resolved to undefined. Both removed; data's description now says what it holds per operation.

This one needed no API doc — it is provable from the repo. For completeness the point shapes were checked against Qdrant's OpenAPI (Record, ScoredPoint, UpdateResult at https://github.com/qdrant/qdrant/blob/master/docs/redoc/master/openapi.json) and every declared per-point field (version, score, shard_key, order_value, operation_id) is real — no phantom fields found in the qdrant tool layer.

Fields removed, and the evidence

Field	Where	Evidence it is never returned
reviews	serper places mapper + SearchResult	/places items report the review count as ratingCount. reviews appears in no serper.dev example. Was not a declared output.
matches	QdrantBlock.outputs	No Qdrant transformResponse writes it; all three return { status, data }.
upsertedCount	QdrantBlock.outputs	Same. Qdrant's UpdateResult has only operation_id and status.

link and snippet on places results are narrowed to optional, not removed — they remain populated on every other vertical.

Unverifiable

• Serper knowledgeGraph. SearchResponse has carried a dead knowledgeGraph? TS field and a KNOWLEDGE_GRAPH_OUTPUT_PROPERTIES constant, neither declared as an output nor populated. I could not retrieve a primary serper.dev example of the object, so I did not add it. To finish it I need a captured /search payload for an entity query (e.g. q: "OpenAI") or a live Serper key. Left exactly as-is.
• Same for answerBox and topStories: dead TS-only fields, no advertised output, untouched.

Tests

Every assertion was run against the unmodified source first and confirmed red (13 failures), then green after the fix, then the source changes were stashed and the same 13 confirmed red again.

• tools/serper/search.test.ts — places payload copied from serper.dev; asserts no link/snippet/reviews; news source; the two /search panels present on search and absent on news.
• tools/firecrawl/search.test.ts (new) — source-keyed envelope, news snippet not description, image url vs imageUrl, empty-envelope fallback, declared-output shape, mapTimeout rename and body mapping, map link objects.
• tools/langsmith/langsmith.test.ts — string coercion, 0 preserved, blank omitted, non-numeric rejected, all at the tool layer.
• tools/qdrant/search_vector.test.ts — block declares only data and status; no tool emits matches/upsertedCount.

Gates

tool-metadata:generate, generate-docs.ts (generated diff committed), lint, check:audits (39/39, includes docs:check), check-block-registry.ts all pass. Type-check is clean for these files; the two @aws-sdk/client-lambda errors are pre-existing and unrelated.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
docs	Ready	Preview	Aug 29, 2026 4:45am

[greptile-apps]

Greptile Summary

This PR aligns several integration declarations and transformations with their provider response and request contracts.

• Models Firecrawl search results as source-keyed data, exposes source selection, corrects map links, and separates the provider map deadline from the transport timeout.
• Corrects Serper vertical mappings and exposes search panels.
• Centralizes LangSmith feedback-score coercion and removes unsupported Qdrant block outputs.
• Regenerates tool metadata and integration documentation and adds focused regression tests.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
apps/sim/tools/firecrawl/map.ts	Renames the provider map deadline, normalizes map links into declared objects, and preserves valid edge-case timeout values.
apps/sim/tools/firecrawl/search.ts	Exposes source selection and transforms Firecrawl v2 search responses into the source-keyed output envelope.
apps/sim/tools/firecrawl/types.ts	Defines source-specific Firecrawl result shapes, nullable scraped fields, metadata, and object-based map links.
apps/sim/tools/serper/search.ts	Corrects place and news mappings and conditionally surfaces the web-search question and related-query panels.
apps/sim/tools/serper/types.ts	Aligns unified Serper result and panel declarations with the mapped provider payloads.
apps/sim/tools/langsmith/create_feedback.ts	Applies finite numeric score coercion consistently at the tool request boundary.
apps/sim/blocks/blocks/qdrant.ts	Removes block outputs that no Qdrant tool emits and documents the actual operation-specific payload.
apps/sim/tools/firecrawl/search.test.ts	Covers Firecrawl v2 envelopes, source forwarding, map timeout behavior, link normalization, and declared output details.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Inputs[Block or direct tool inputs] --> Normalize[Tool request normalization]
  Normalize --> Providers[Firecrawl / Serper / LangSmith / Qdrant]
  Providers --> Transform[Provider-specific response transformation]
  Transform --> Shapes[Declared tool output shapes]
  Shapes --> Metadata[Generated metadata and integration docs]

Loading

_{Reviews (2): Last reviewed commit: "fix(tools): reject non-numeric langsmith..." | Re-trigger Greptile}

[cubic-dev-ai]

7 issues found across 18 files

Confidence score: 3/5

• apps/docs/content/docs/integrations/firecrawl.mdx documents Search requests for news and images, but the tool schema and Firecrawl block do not expose sources, so those documented requests cannot work; expose sources before relying on the documentation.
• apps/sim/tools/langsmith/create_feedback.ts can serialize infinite scores as null and converts whitespace-only scores to 0, producing incorrect LangSmith feedback; reject non-finite values and treat trimmed-empty input as undefined.
• apps/sim/tools/firecrawl/types.ts does not fully match Firecrawl’s nullable or missing response fields, so valid search results can fail validation or misrepresent optional content; make metadata and nullable scraped fields optional in the advertised schema.
• apps/sim/tools/firecrawl/map.ts drops an explicit mapTimeout of 0, and apps/sim/blocks/blocks/qdrant.ts leaves the registered skill describing an upsert count that the output no longer exposes; preserve zero-valued timeout configuration and align the skill with the returned data.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="apps/sim/tools/firecrawl/map.ts">

<violation number="1" location="apps/sim/tools/firecrawl/map.ts:93">
P2: When `mapTimeout` is explicitly `0`, this guard omits the Firecrawl body timeout and silently applies the provider default. Preserve zero while still omitting empty values, for example by checking `params.mapTimeout || params.mapTimeout === 0`. 

(Based on your team's feedback about preserving explicit zero numeric parameters.) .</violation>
</file>

<file name="apps/sim/blocks/blocks/qdrant.ts">

<violation number="1" location="apps/sim/blocks/blocks/qdrant.ts:263">
P2: After this change, the upsert output exposes only `data` containing Qdrant’s operation result, but the registered `upsert-points` skill still tells agents to report an upserted count. Update that skill to describe and report the operation ID and status instead.</violation>
</file>

<file name="apps/docs/content/docs/integrations/firecrawl.mdx">

<violation number="1" location="apps/docs/content/docs/integrations/firecrawl.mdx:166">
P2: When users use the documented Search action, `news` and `images` cannot be requested because neither the tool schema nor the Firecrawl block exposes `sources`. Expose `sources` in the tool and block before documenting these verticals, or remove the unreachable result arrays here.</violation>
</file>

<file name="apps/sim/tools/langsmith/create_feedback.ts">

<violation number="1" location="apps/sim/tools/langsmith/create_feedback.ts:16">
P2: When the optional score contains only whitespace, `Number(value)` converts it to `0`, so the tool records a score instead of omitting the blank input. Treat trimmed-empty strings as undefined before coercion.</violation>

<violation number="2" location="apps/sim/tools/langsmith/create_feedback.ts:18">
P2: When a caller supplies `Infinity` or `-Infinity` as the score, `Number.isNaN(parsed)` allows it through and JSON serialization sends `null` to LangSmith. Reject non-finite parsed scores with `Number.isFinite`.</violation>
</file>

<file name="apps/sim/tools/firecrawl/types.ts">

<violation number="1" location="apps/sim/tools/firecrawl/types.ts:214">
P2: When Firecrawl returns a search item without scraped metadata, the output declaration still requires `metadata`. Mark metadata optional in both web and news item definitions, because `transformResponse` forwards items unchanged.</violation>

<violation number="2" location="apps/sim/tools/firecrawl/types.ts:458">
P2: When Firecrawl returns `null` for optional scraped content, the new TypeScript shape allows it but the advertised output schema says only `string`. Mark `markdown`, `html`, `rawHtml`, `screenshot`, and metadata `error` nullable, or normalize nulls before returning.</violation>
</file>

<sub>Tip: instead of fixing issues one by one fix them all with cubic

Re-trigger cubic</sub>

[waleedlatif1]

@cubic-dev-ai review

[cubic-dev-ai]

@cubic-dev-ai review

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

4 issues found across 18 files

Confidence score: 3/5

• apps/docs/content/docs/integrations/firecrawl.mdx: mapTimeout is documented but cannot be configured or forwarded by the map operation, so user timeout settings have no effect; expose and pass the setting through the integration block.
• apps/sim/tools/firecrawl/types.ts: Search web and news results can contain audio and video, but the declared interfaces and output properties omit them, potentially dropping fields for consumers; add nullable fields to the shared interface and both output property sets.
• apps/sim/tools/firecrawl/types.ts: Crawled pages may return a null screenshot while the response interfaces reject null, creating a type mismatch for valid crawl responses; use string | null throughout the shared screenshot shape.
• apps/sim/tools/firecrawl/types.ts: metadata.url contains the final redirected URL but is absent from SEARCH_METADATA_OUTPUT_PROPERTIES, so generated metadata consumers cannot access it; add the url property.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="apps/sim/tools/firecrawl/types.ts">

<violation number="1" location="apps/sim/tools/firecrawl/types.ts:95">
P2: The crawl output now allows null screenshots, but the corresponding response interfaces reject null. Update the crawled-page screenshot types to `string | null` wherever this shared shape is used.</violation>

<violation number="2" location="apps/sim/tools/firecrawl/types.ts:240">
P2: Search scraping can return `audio` and `video`, but the new web and news result shapes do not declare them. Add nullable `audio` and `video` fields to the shared interface and both output property sets.</violation>

<violation number="3" location="apps/sim/tools/firecrawl/types.ts:482">
P2: Search results preserve `metadata.url`, but the declared output omits it, hiding the final redirected URL from generated tool metadata and consumers. Add `url` to `SEARCH_METADATA_OUTPUT_PROPERTIES`.</violation>
</file>

<file name="apps/docs/content/docs/integrations/firecrawl.mdx">

<violation number="1" location="apps/docs/content/docs/integrations/firecrawl.mdx:314">
P2: When users configure the Firecrawl integration block, `mapTimeout` has no input path and is never forwarded by the block's `map` operation, so this documented setting cannot affect the request. Expose `mapTimeout` through the block mapping or remove it from this integration documentation.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[waleedlatif1]

@cubic-dev-ai review

[cubic-dev-ai]

@cubic-dev-ai review

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

2 issues found across 18 files

Confidence score: 4/5

• In apps/sim/tools/langsmith/create_feedback.ts, malformed boolean input such as false can be coerced into score 0, polluting evaluation data; restrict coercion to number and string values before calling Number.
• In apps/sim/tools/firecrawl/types.ts, categorized web or news results cannot expose the provider’s category field through generated metadata and exported interfaces; add an optional category property to the relevant declarations.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="apps/sim/tools/langsmith/create_feedback.ts">

<violation number="1" location="apps/sim/tools/langsmith/create_feedback.ts:18">
P2: Malformed direct or LLM input such as `false` is silently recorded as score `0` instead of rejected, producing incorrect evaluation data. Restrict coercion to `number` and `string` values before calling `Number`.</violation>
</file>

<file name="apps/sim/tools/firecrawl/types.ts">

<violation number="1" location="apps/sim/tools/firecrawl/types.ts:163">
P3: When Firecrawl returns a categorized web or news result, these declarations omit `category`, so generated tool metadata and exported result interfaces cannot describe that provider field. Add optional `category` to both result property maps and both TypeScript result interfaces.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[waleedlatif1]

@cubic-dev-ai review

[cubic-dev-ai]

@cubic-dev-ai review

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

No issues found across 18 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Re-trigger cubic}

[waleedlatif1]

@greptile review

[waleedlatif1]

Closing for now — not because of a defect. This batch grew to 17 PRs across ~700 changed call sites, and we would rather revisit it as smaller, independently testable pieces than merge this much at once.

Nothing here is lost: the branch fix/declared-output-shapes is preserved and this PR can be reopened. Review state, the reasoning on every thread, and the red-first verification all stay attached.