[sc-source]

Problem and impact

Add authenticated GET /api/v1/ai-vision/skill/spans to recall paginated candidate Skill spans for AI Vision #92. Recall includes native agentic.tool.type=skill, native agentic.skill.name, and argument payloads mentioning SKILL.md; the Python semantic layer rejudges candidates before counting them. Existing agent/tenant filters apply.

The review fix adds native-name recall (otherwise AICoder-style records without a path are missed) and orders timestamp ties by trace ID and span ID for stable pagination.

Validation

Current head: c49b2437d0f401f0ee60b73429bfa307dbecc600. Local real OceanBase-compatible SQL integration used an isolated cloned traces table and exercised the production handler: native-name-only recall plus 501 tied-timestamp candidates across three pages without duplicate/missing identities. The aivision Go module tests and Linux arm64 binary build passed. The same candidate binary served the real Pi/Baidu telemetry acceptance for AI Vision SigNoz#92. The three-Skill round reconciled 66 original spans in 12 traces through both Native/dashboard and SpaceKey APIs; candidate recall returned the exact 12 candidate identities, and the Python layer counted 9 calls (three per Skill). Provider/SSE usage and all original span IDs, parents and precise nanosecond timestamps matched. The candidate count is deliberately distinct from the final call count.

Hosted query-regression passed gofmt and go test -p 2 -count=1 ./pkg/modules/aivision. Hosted CI did not configure a real OceanBase DSN; real SQL evidence is the separate local run. The integration test skips unless SIGNOZ_TEST_OCEANBASE_DSN is supplied and cleans up its uniquely named test table.

Screenshots and interaction

No direct UI change. Source-span navigation and the Skill overview were checked through the compatible frontend in AI Vision SigNoz#92; screenshots accompany its documentation PR.

Compatibility and unverified scope

Deploy this query support together with AI Vision SigNoz#92. Candidate results are intentionally broad and are not final Skill invocation counts. Other architectures, a published image and production deployment were not verified. The stable ordering prevents ambiguous tie ordering; no production data-loss incident is claimed.

[sc-source]

@webup

[webup]

Reviewed with Codex assistance at this commit. Native-name recall and stable pagination match AI Vision SigNoz#92; hosted query-regression passed, and the separate local SQL/integration evidence covers the actual compatible stack. Hosted CI did not use a real OceanBase DSN.