[antoinetoussaint-byte]

Wiki shipped the first marked solution artifact yesterday (obin-ai/core-solutions#457) and published this as a tool name:

Wiki_Get_V1_Wiki_Upload-configuration

That is the operationId, which this runtime builds as OperationPrefix_Verb_Path. It identifies the route and describes nothing — and since an Operation could not name itself, every solution's tools would have read that way.

The change

Operation.Tool is the model-facing name:

Operation{Path: "/v1/wiki/upload-configuration", Summary: "Read the installed upload destination",
    Tool: "read_upload_configuration", Policy: reads, Handler: answer}

It is the one part of the marker the owner says rather than core derives, because it is the one part a model reads. Everything else in x-codefly-operation stays core's own derivation from the declared policy.

Empty keeps today's behaviour exactly — the name falls back to the operationId — so no committed artifact changes until its owner names something.

Validated, not sanitised

Operation.validate refuses a name that is not snake_case within 64 characters. That is deliberate: core's toolName is a sanitiser for method names and admits capitals and hyphens by design (runnable/derive.go:257 keeps - and case), so Read-Board or ReadBoard would reach a model as typed. This is the last place that can still say no, and an owner naming its own tool is told what a name is rather than having one silently rewritten.

Refused: Read_Board, read-board, 9read, read board, read_board_, read__board, anything over 64 characters. The rule is ^[a-z][a-z0-9]*(_[a-z0-9]+)*$.

The operationId does not move

A nicer operationId would have fixed the tool name and the derived binding key in one go, and I deliberately did not do it: generated clients are named from the operationId, so improving one tool's name would rename every caller's method. The identity stays; only the model-facing word is the owner's to choose. TestAnOperationNamesTheToolItIsOfferedAs pins that naming a tool leaves Releases_Get_V1_Releases_Recorded untouched.

Validation

• go build ./..., go vet ./... — clean. gofmt clean.
• go test ./... -count=1 — ok, 16.7s, the whole suite.
• Two new tests: one operation in the marker fixture names its tool and the other does not, so both the name and the fallback are asserted against the same rendered artifact, along with the operationId staying put; and the spelling rule is driven through Operation.validate over six rejected forms plus the empty one.

Note for reviewers: the test decodes the marker's tool as a plain document rather than through runnablev0.ToolExposure, because the proto enum does not round-trip encoding/json — the first version of the test failed on exactly that, which is worth knowing before writing another assertion over a marker.

Follow-up, not in this PR

Wiki's two reads still carry the fallback names until core-solutions sets Tool on them; that is a one-line change per operation there, once this is released.

🤖 Generated with Claude Code