Model-ID version gate excludes dotless IDs, so gpt-6-astra is treated as older than gpt-5.5 #48247

Description

packages/opencode/src/provider/transform.ts:1331 reads a model's version with a regex that requires a dotted minor:

const [, gptMajorVersion, gptMinorVersion] = input.model.api.id.match(/gpt-(\d+)\.(\d+)/) ?? []
const isGpt55OrNewer = Number(gptMajorVersion) > 5 || (Number(gptMajorVersion) === 5 && Number(gptMinorVersion) >= 5)

gpt-6-astra has no dotted minor, so the match fails and Number(undefined) > 5 evaluates NaN > 5. A GPT-6 model reports isGpt55OrNewer === false. gpt-5-pro fails the same regex.

The only consumer of that flag is the Azure early-return at :1333-1338, and the comment above it says "Any gpt version above 5.4 in combination with azure does not support reasoningEffort". So on Azure + useCompletionUrls, GPT-6 currently gets reasoningEffort: "medium" set precisely where the code intends to skip it.

Separately at :1340, "gpt-6-astra".includes("gpt-5") is false, so that block is skipped and textVerbosity: "low" never applies.

gpt-6-astra    regexMatch=false  isGpt55OrNewer=false  entersBlock1340=false
gpt-5.6-luna   regexMatch=true   isGpt55OrNewer=true   entersBlock1340=true
gpt-5-pro      regexMatch=false  isGpt55OrNewer=false  entersBlock1340=true

Suggested fix

Correcting the regex alone is safe and fixes the Azure path. Version-gating :1340 naively is not — the block contains four interlocking substring tests, and today gpt-6-pro / gpt-6-chat / gpt-6-codex are excluded only as a side effect of the outer includes("gpt-5") being false. Generalize just the outer test and they all fall in while includes("gpt-5-pro") and includes("gpt-5-chat") stay false:

• gpt-6-pro → reasoningEffort: "medium", but OPENAI_GPT5_PRO_EFFORTS = ["high"] at :577 — an out-of-range effort rather than a lost optimization.
• gpt-6-chat → reasoning params on a model the file deliberately treats as non-reasoning (gpt5ChatReasoningEfforts returns [] at :619).

All four tests need to move together, or none.

Note also that fixing the regex does not by itself restore textVerbosity for Astra: :1359 gates on includes("gpt-5."), and "gpt-6-astra" does not contain that either. And includes("gpt-5.") currently means "has an explicit minor" — generalizing it so a minor-less GPT-6 id qualifies is a deliberate behavior change worth arguing separately.

Rather than adding a second parser, this file already has anchored helpers at :594-609 used by variants():

const GPT5_FAMILY_RE = /(?:^|\/)gpt-5(?:[.-]|$)/
const GPT5_VERSION_RE = /(?:^|\/)gpt-5[.-](\d+)(?:[.-]|$)/
const GPT5_PRO_RE = /(?:^|\/)gpt-5[.-]?pro(?:[.-]|$)/

options() is the odd one out, hand-rolling substring tests and an unanchored inline regex. Generalizing those helpers to be family-agnostic and reusing them in options() seems like the right shape. The (?:^|\/) anchors matter — a bare /gpt-(\d+)/ reads gpt-50 as major 50.

Any test matrix should include gpt-6-pro, gpt-6-chat, gpt-6-codex, and gpt-50, since those are exactly the ids a partial fix breaks.

Plugins

@cortexkit/[email protected], @cortexkit/[email protected], [email protected], @cortexkit/[email protected], @cortexkit/[email protected], [email protected], @fro.bot/[email protected]

OpenCode version

1.18.29, confirmed present at v1.18.30

Steps to reproduce

1. Pick a model ID with no dotted minor — gpt-6-astra or gpt-5-pro.
2. Evaluate the conditions at transform.ts:1331 and :1340 against it.
3. isGpt55OrNewer is false for GPT-6, so the Azure early-return at :1333 sets reasoningEffort: "medium" where the comment says it should not.
4. The :1340 block is skipped, so textVerbosity: "low" never applies.
5. Repeat with gpt-5.6-luna — both pass.

Screenshot and/or share link

n/a

Operating System

macOS 15.7.9 (arm64)

Terminal

Ghostty