zarvis: default model (gpt-5-mini) is very inefficient on simple coding tasks

[edwin-zvs]

Summary

On its default model (openai:gpt-5-mini), a zarvis session given a straightforward "edit this single HTML file" coding task was very inefficient: 178 transcript events, with repeated find_files loops, vs ~31 events for the equivalent claude task and ~5 for codex. It still completed correctly after approvals, but the cost/latency was disproportionate.

Ask

Consider a stronger default model for zarvis coding tasks (or document guidance / AGENTD_ZARVIS_MODEL recommendations), and/or investigate the find_files looping behavior.

Found during an isolated dogfood run building a browser game.

[edwin-zvs]

Resolved by #333 (merged): the zarvis default model now bumps to claude-opus-4-8 when ANTHROPIC_API_KEY is set and gpt-5 (non-mini) when only OPENAI_API_KEY is set. Explicit --model / AGENTD_ZARVIS_MODEL still take precedence.

Note: this addresses the "stronger default model" ask. The separate find_files looping observation isn't tracked by this change — please reopen or file a focused issue if that behavior still needs investigation.