Qualify one third-party native Responses path for the 0.1.7 release matrix #220

Activity

1. NOirBRight

   added this to the 0.1.7 — Trustworthy (stability chain) milestone on Jul 25, 2026
2. NOirBRight

   added

   enhancementNew feature or requestNew feature or request

   ready-for-humanRequires human implementationRequires human implementation

   wayfinder:taskWayfinder implementation taskWayfinder implementation task

   tier:heavyOrchestrator provider tier: heavyOrchestrator provider tier: heavy

   P1Priority 1: current iteration goalPriority 1: current iteration goal

   on Jul 25, 2026
3. NOirBRight

   added a parent issue on Jul 25, 2026

   • Wayfinder: CodexHub 0.1.6–0.3.0 Gateway and release roadmap #147
4. NOirBRight

   marked this as blocking on Jul 25, 2026

   • Replace the 0.1.7 Volc Chat qualification leg with native Responses #226
5. NOirBRight commented on Jul 25, 2026

   

   NOirBRight

   on Jul 25, 2026

   OwnerAuthor

   More actions

   Bounded provider-boundary reconcile (2026-07-25):

   • Volc Ark officially documents a native Responses endpoint at
     POST https://ark.cn-beijing.volces.com/api/v3/responses.
   • The official Function Calling example uses native Responses objects,
     function_call / function_call_output, call_id, and
     previous_response_id, so Volc Ark is a valid first qualification target:
     https://www.volcengine.com/docs/82379/1958524?lang=zh
   • The official model/API index also lists Responses API as a distinct supported
     API family:
     https://www.volcengine.com/docs/82379/66619f8df281250274ef4f88?lang=zh
   • No authoritative evidence was found that the currently configured
     /api/coding/v3 glm-5.2 alias accepts native /responses. Do not infer
     support from the existence of Ark Responses or silently substitute Chat.
   • Preserved run local-r32-volc-responses-8ec9d27c-20260723 contains only a
     pending manual-evidence template and no sanitized result artifacts; it proves
     no acceptance criterion.

   Next human qualification is therefore bounded:

   1. Try the exact available Volc Ark Provider/Endpoint/model credential against
      the canonical gateway-mediated /responses path.
   2. Capture sanitized text, SSE, function-call/result, cancellation, and
      terminal-error evidence with route/model identity and zero fallback.
   3. If the available credential is Coding-Plan-only or the exact model rejects
      native Responses, record that exact boundary and qualify Ollama Native
      Responses instead.

   Issue remains ready-for-human; no acceptance box is complete yet.
6. NOirBRight commented on Jul 25, 2026

   

   NOirBRight

   on Jul 25, 2026

   OwnerAuthor

   More actions

   Ollama fallback boundary is now pinned from primary documentation:

   • Native OpenAI-compatible endpoint:
     POST http://localhost:11434/v1/responses
   • Minimum documented version: Ollama 0.13.3
   • Documented Responses features include SSE streaming and function tools.
   • The implementation is explicitly non-stateful: do not rely on
     previous_response_id or conversation; the Gateway qualification must send
     the required portable input/tool history itself.
   • Official reference:
     https://docs.ollama.com/api/openai-compatibility

   This makes Ollama a concrete fallback Model Access Path if the available Volc
   credential/model cannot prove native Responses. It does not itself satisfy
   #220: the selected exact Ollama version, model, Gateway endpoint binding, text,
   SSE, tool/result, cancellation, terminal-error, route identity, and zero
   fallback still require sanitized runtime evidence.
7. NOirBRight commented on Jul 25, 2026

   

   NOirBRight

   on Jul 25, 2026

   OwnerAuthor

   More actions

   2026-07-25 Ollama Cloud Native Responses qualification update

   This probe used Ollama Cloud directly. It did not use the local Ollama
   daemon or 127.0.0.1:11434.

   Candidate Gateway tree:
   717c27254f6597373ce354c8288346da2c796069.

   Model Access Path:

   • Provider: ollama-cloud
   • Cloud base URL: https://ollama.com/v1
   • Upstream endpoint: https://ollama.com/v1/responses
   • Canonical model: ollama-cloud/glm-5.2
   • Gateway endpoint: /v1/providers/ollama-cloud/responses
   • Protocol: Native Responses Transparent Passthrough

   Sanitized observations:

   • Authenticated Cloud /v1/models returned glm-5.2.
   • Direct Cloud non-streaming /v1/responses completed with the expected
     sentinel and truthful usage.
   • Gateway-mediated non-streaming Responses completed with the expected
     sentinel. Gateway telemetry recorded provider_hint=ollama-cloud,
     upstream_format=responses, and wire_format_adapter=transparent.
   • Gateway-mediated SSE returned 95 Responses events, contained the expected
     streamed sentinel and response.completed, and contained no Chat [DONE]
     sentinel. Retained response-body SHA-256:
     ddb0c77bb748a44277032a172d70c4c0ef9b48671bdb7c42f04eab871eabeb14.
   • A standard Responses function_call / function_call_output round trip
     completed: get_probe_value({"key":"cloud_probe"}) produced the expected
     final sentinel.
   • A provider-side invalid max_output_tokens=-1 request returned HTTP 400 on
     the same canonical route with zero retry and zero fallback.
   • Mid-event client cancellation after 5,083 streamed bytes produced one
     downstream_stream_closed event with ConnectionResetError,
     close_phase=during_event, and the same transparent Responses route.

   Qualification remains incomplete, rather than being represented as pass:

   1. The cancellation request still records request_complete=200 instead of
      the parent Handle downstream client disconnects consistently for all SSE writes #18 contract's 499 disposition.
   2. The upstream HTTP 400 path records request_error but no final
      request_complete.

   Those are request-scoped downstream ledger/accounting gaps owned by #223, not
   Ollama Cloud protocol or endpoint failures. Re-run only the bounded
   cancellation/error assertions after #223; do not repeat provider discovery,
   text, SSE, or tool qualification.
8. NOirBRight commented on Jul 26, 2026

   

   NOirBRight

   on Jul 26, 2026

   OwnerAuthor

   More actions

   2026-07-27 post-#223 bounded ledger qualification — pass

   Reused the already retained Ollama Cloud Native Responses evidence from the 2026-07-25 qualification comment for Provider/model/route discovery, non-streaming text, native SSE, and the function_call / function_call_output round trip. Only the two previously blocked request-ledger assertions were rerun on current dev candidate 1758b830130ac3bf4a0d55f08b69fce342f4a969.

   Exact Model Access Path:

   • Provider: ollama-cloud
   • Upstream: https://ollama.com/v1/responses
   • Gateway route: /v1/providers/ollama-cloud/responses
   • Model: ollama-cloud/glm-5.2
   • Wire contract: native Responses, wire_format_adapter=transparent

   Bounded results:

   • Provider-side invalid max_output_tokens=-1: HTTP 400, exactly one request_complete(status=400).
   • Mid-event downstream RST after streaming began: one downstream_stream_closed(status=499, close_phase=during_event) and exactly one request_complete(status=499).
   • Aggregate: two requests, exactly two terminal request_complete records, zero retry events, zero fallback events.
   • Sanitization audit: API key absent; isolated absolute path absent; no raw provider payload retained in this Issue.
   • Sanitized event-log SHA-256: be56094af5d3d09bd0cb959e992f4348eefaf785f94356096ac4fce769d83668 (6 bounded records).
   • Gateway shut down cleanly with exit code 0.

   This closes the only two gaps recorded in the prior qualification. Together with the retained text/SSE/tool evidence, ollama-cloud/glm-5.2 is the selected third-party Native Responses path for the 0.1.7 matrix. No local Ollama daemon or localhost:11434 was used.
9. NOirBRight commented on Jul 26, 2026

   

   NOirBRight

   on Jul 26, 2026

   OwnerAuthor

   More actions

   Qualification complete; the exact Ollama Cloud Native Responses path and bounded post-#223 ledger evidence are recorded above.
10. NOirBRight

    closed this as completedon Jul 26, 2026