workflow: unwrap MCP text-block results + discover replicate tools by suffix#448
workflow: unwrap MCP text-block results + discover replicate tools by suffix#448FMXExpress wants to merge 1 commit into
Conversation
… suffix
Two field bugs against the live Replicate MCP:
1. Replicate returns a tool's payload as a JSON STRING inside a text content
block, so the bridge's ResultJSON is the wrapper { content:[{text:"…"}] } --
id/status/output are NOT at the top level. The replicate node's poll then
failed with `{{self.id}}: selector "id" did not resolve`, and the model
search "returned nothing". Add UnwrapResult(RawJSON, Text): when the
flattened Text is itself a JSON object, THAT is the payload -- use it. Apply
it to each node result (downstream selectors + the poll's self) and to every
poll result (status/output). New test drives the real text-block shape end
to end (create->poll->finished URL).
2. The /v1/replicate/{search,model} endpoints hardcoded `replicate__search` /
`replicate__get_models`, which only resolve if the MCP server is named
exactly "replicate". Discover the tool from the registry by suffix
(`*__search`, `*__get_models`), preferring a server whose name contains
"replicate", so it works whatever the operator named it. On no match, return
a clear error (surfaced in the UI) instead of silent-empty; endpoints also
return the UNWRAPPED payload so the picker parses models/schema.
Verified: text-block unit test green; search with no MCP returns a clear
"no *__search tool registered" diagnostic; binary + gateway/skills tests build.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGQKz579j1ZnDRwmr6h1V6
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9fb2d07315
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| begin | ||
| Q := Trim(ARequest.Params.Values['q']); | ||
| if Q = '' then begin WriteJSON(AResp, 400, '{"error":"missing q"}'); Exit; end; | ||
| Tool := FindMCPToolBySuffix('search'); |
There was a problem hiding this comment.
When the Replicate MCP server is named anything other than exactly replicate, this endpoint can now successfully discover and return models via e.g. custom__search, but the workflow node the UI creates still runs as the synthetic tool:"replicate"; PlanNode then hardcodes execution to replicate__create_predictions and polling to replicate__get_predictions. That means the newly-supported renamed-server path lets users build/select a model, but the workflow fails at runtime because those hardcoded tools are not registered under that namespace.
Useful? React with 👍 / 👎.
2 participants
Two field bugs against the live Replicate MCP (both fixed here, one branch off main)
1.
{{self.id}}: selector "id" did not resolve+ "search returned nothing"Root cause: the Replicate MCP returns a tool's payload as a JSON string inside a text content block, so the bridge's
ResultJSONis the wrapper{content:[{text:"{…prediction…}"}]}—id/status/outputare not at the top level. My REST probe showed top-level fields, but the MCP wraps them, so the replicate node's poll ({{self.id}}) and the model-search parser both missed.Fix:
UnwrapResult(RawJSON, Text)— when the flattenedTextis itself a JSON object, that is the payload, so select against it. Applied to:{{nodes.ID.sel}}+ the poll'sself),status/outputselectors),/v1/replicate/*endpoint responses.New unit test drives the real text-block shape end-to-end (create → poll
processing→succeeded→ finished URL) — this is what the earlier top-level-only stub was hiding.2. Model search silently returned nothing
The
/v1/replicate/search+/modelendpoints hardcodedreplicate__search/replicate__get_models, which only resolve if the MCP server is named exactlyreplicate. Now they discover the tool from the registry by suffix (*__search,*__get_models), preferring a server whose name containsreplicate— so it works whatever you named the server. On no match they return a clear error (now surfaced in the search box) instead of an empty list, and they return the unwrapped payload so the picker parses the models/schema.Validation
make test-workflow); gateway + skills tests build; binary links.no MCP "*__search" tool registered — is the Replicate MCP connected & authorized?(clear, not silent-empty).Live search/run still needs your OAuth-connected Replicate MCP (the
r8_REST key can't drive the hosted MCP headlessly), but this makes the whole path shape- and name-robust. After merging, yourgenerate → upscalereplicatenodes should poll to completion and chain the URL.🤖 Generated with Claude Code
https://claude.ai/code/session_01LGQKz579j1ZnDRwmr6h1V6
Generated by Claude Code