Repository navigation
fix(ai): treat OpenAI gateway aliases as reasoning models #1474
Description
HappyVertical's Bifrost gateway now exposes stable OpenAI tier aliases (openai/luna, openai/terra, openai/sol, openai/astra) that resolve to the current GPT-6 model (happyvertical/iac#2435), so apps never name a version. Anytown will switch to them.
usesCompletionTokenLimit in packages/ai/src/shared/providers/openai.ts recognises reasoning models only by name (gpt-5*, gpt-6*, o1/o3/o4). An alias's last segment (terra) matches none of these, so the request goes out in the legacy shape: max_tokens plus temperature, which GPT-6 reasoning models reject. top_p is also always forwarded, and reasoning models reject it too.
Change
- For an
openai/-prefixed id, treat the model as a reasoning model unless its last segment is a known legacy family (gpt-3*,gpt-4*,chatgpt-*, andft:fine-tunes of those). Aliases and future OpenAI families then work without an SDK change. Bare ids and other vendors (anthropic/…,ollama/…) keep today's name-based behaviour. - Omit
top_pfor reasoning-shaped requests, in both chat and streaming.
Acceptance
openai/terra(and the other tiers) getmax_completion_tokens, with notemperatureortop_p.openai/gpt-4.1-mini,openai/gpt-4o-mini,anthropic/claude-sonnet-4-6and baremistral:latestare unchanged.- Unit tests cover both directions. Package typecheck, lint and tests pass.
Activity
willgriffin commentedon Oct 31, 2026
Claimed by Claude Code · session
anytown-terra-6-1-47a0e2· branchclaude/ai-openai-reasoning-aliases· 2026-10-30Status: ready for review in #1475 (validation green, review-cycle clean after one accepted fix). Next: owner merge; the release workflow publishes the patch, then Anytown bumps
@happyvertical/aifrom 0.108.0.willgriffin commentedon Oct 31, 2026
Review budget ledger (maintained by
hv-review-dispatch; do not edit).{ "dispatches": [ { "counted": false, "dispatch_id": "bdedf1b6-2675-4ad5-8d4d-988e62d6735c", "disposition": "unavailable", "escalation_reason": null, "finished_at": "2026-10-31T05:36:39Z", "head": "789a203bbf17deb5dbc394c22c8f2eb2f66ccfd3", "past_cap": null, "phase": "final", "reconciliation": null, "review_base": "64a1b9c04bd4b07ee66657028bb9a689fe955c97", "review_cycle": "6539608522", "review_round": 1, "target": "codex-sol", "token_cap": 30000000, "total_tokens": null, "user_approval": null }, { "counted": true, "dispatch_id": "42f0c4c8-7407-4f21-b8de-4fa0d85c9ab9", "disposition": "findings", "escalation_reason": null, "finished_at": "2026-10-31T05:36:56Z", "head": "789a203bbf17deb5dbc394c22c8f2eb2f66ccfd3", "past_cap": null, "phase": "final", "reconciliation": null, "review_base": "64a1b9c04bd4b07ee66657028bb9a689fe955c97", "review_cycle": "6539608522", "review_round": 1, "target": "claude-sonnet", "token_cap": 30000000, "total_tokens": 12748, "user_approval": null }, { "counted": true, "dispatch_id": "4f0964a2-cb8c-4f96-bba9-6aec2c65a537", "disposition": "clean", "escalation_reason": null, "finished_at": "2026-10-31T05:37:55Z", "head": "2aa37e6ce8fcebe92693660c5a630a39e4050436", "past_cap": null, "phase": "recall", "reconciliation": null, "review_base": "64a1b9c04bd4b07ee66657028bb9a689fe955c97", "review_cycle": "6539608522", "review_round": 2, "target": "claude-sonnet", "token_cap": 30000000, "total_tokens": 15575, "user_approval": null } ], "final_pass_cap": 6, "issue": 1474, "repository": "happyvertical/sdk", "schema": "https://happyvertical.com/schemas/review-budget/v1" }