sync(v2.12): model/provider management and streamed usage parity #134

New issue

Open

opened 2026-09-09 14:58:55 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-09 14:58:55 +00:00

Owner

Upstream

Agent Zero v2.12: https://github.com/agent0ai/agent-zero/releases/tag/v2.12

Missing parity

Adopt the v2.12 provider/model-management fixes as one coherent contract.

Required work

Prevent chat-only parameters from leaking into embedding model configuration.
Preserve Codex OAuth prompt-cache affinity using stable session/thread IDs and bounded default cache keys.
Keep OAuth model discovery tied to actual account availability; do not replace empty catalogs with generic registry models.
Report provider usage for streamed Chat Completions and retry once without stream_options when the provider rejects it.
Filter provider encrypted-reasoning markers from readable output.
Preserve Codex Chat bridge usage and terminal outcomes with correct token-count translation.
Validate model-parameter JSON before save and return preset/slot editor context with errors.
Avoid repeated preset reads during embedding comparisons/save paths.
Discard stale model-search results when the draft/query changes.
Use the reviewed v2.12 tested Codex client version (0.153.3 upstream) as the default discovery identity unless this fork intentionally pins a newer reviewed version.

Acceptance

Embedding calls receive only embedding-compatible params.
Streamed usage accounting is present when supported and gracefully retries providers that reject stream_options.
Empty OAuth account catalogs stay empty rather than being silently filled from generic registries.
Invalid parameter JSON cannot be saved and identifies the affected preset/slot.
Search result races cannot overwrite newer draft/query state.
Provider/model regression tests cover Codex OAuth, cache affinity, usage translation, encrypted reasoning filtering, embedding comparison performance, and stale-search rejection.

## Upstream Agent Zero v2.12: https://github.com/agent0ai/agent-zero/releases/tag/v2.12 ## Missing parity Adopt the v2.12 provider/model-management fixes as one coherent contract. ## Required work - Prevent chat-only parameters from leaking into embedding model configuration. - Preserve Codex OAuth prompt-cache affinity using stable session/thread IDs and bounded default cache keys. - Keep OAuth model discovery tied to actual account availability; do not replace empty catalogs with generic registry models. - Report provider usage for streamed Chat Completions and retry once without `stream_options` when the provider rejects it. - Filter provider encrypted-reasoning markers from readable output. - Preserve Codex Chat bridge usage and terminal outcomes with correct token-count translation. - Validate model-parameter JSON before save and return preset/slot editor context with errors. - Avoid repeated preset reads during embedding comparisons/save paths. - Discard stale model-search results when the draft/query changes. - Use the reviewed v2.12 tested Codex client version (0.153.3 upstream) as the default discovery identity unless this fork intentionally pins a newer reviewed version. ## Acceptance - Embedding calls receive only embedding-compatible params. - Streamed usage accounting is present when supported and gracefully retries providers that reject `stream_options`. - Empty OAuth account catalogs stay empty rather than being silently filled from generic registries. - Invalid parameter JSON cannot be saved and identifies the affected preset/slot. - Search result races cannot overwrite newer draft/query state. - Provider/model regression tests cover Codex OAuth, cache affinity, usage translation, encrypted reasoning filtering, embedding comparison performance, and stale-search rejection.