Eslopes
14 de julio de 2026, 02:13
Ollama Cloud calls work — the provider pointed at the wrong host (api.ollama.com); the service lives at ollama.com (/api native, /v1 OpenAI-compatible). The default endpoint is corrected and every outbound path (requests, connection test, model listing) heals endpoints saved with the old host, so existing configs work without re-selecting the provider. Verified live: HTTP 200 with Bearer auth against https://ollama.com/v1/chat/completions.
Local Ollama native responses parse — the orchestrator only understood the OpenAI reply shape (choices[0].message.content); it now also parses Ollama-native (message.content), and picks the right wire format from the endpoint suffix (/api → native, /v1 → OpenAI-compatible).
Agent requests carry the real prompt again — the mesh rewrite dropped the composed system prompt, skills, and per-request context on the floor (parameters received but discarded), running every request on a canned one-liner. The full spec-025 composition (system prompt → skills → history → prompt+context) is restored for the dev agent, the editor assistant, and history compaction.
Agentic AI log is real and live — the orchestrator now streams every step through a log callback into the AI activity pane: specialist routing, resolved URL and wire format, HTTP status with duration and payload size, and token usage (both OpenAI and Ollama accounting). The connection log always records a compact request/response trace (full bodies with the verbose setting), so a "successful" call can no longer be invisible — or faked.
Gemini Codex Agent
Local Ollama native responses parse — the orchestrator only understood the OpenAI reply shape (choices[0].message.content); it now also parses Ollama-native (message.content), and picks the right wire format from the endpoint suffix (/api → native, /v1 → OpenAI-compatible).
Agent requests carry the real prompt again — the mesh rewrite dropped the composed system prompt, skills, and per-request context on the floor (parameters received but discarded), running every request on a canned one-liner. The full spec-025 composition (system prompt → skills → history → prompt+context) is restored for the dev agent, the editor assistant, and history compaction.
Agentic AI log is real and live — the orchestrator now streams every step through a log callback into the AI activity pane: specialist routing, resolved URL and wire format, HTTP status with duration and payload size, and token usage (both OpenAI and Ollama accounting). The connection log always records a compact request/response trace (full bodies with the verbose setting), so a "successful" call can no longer be invisible — or faked.
Gemini Codex Agent