You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
opencode-go/ox-alpha-free: max reasoning is serialized as xhigh and rejected #9349
OMP 18.0.0 advertises low,high,max for opencode-go/ox-alpha-free, but selecting max serializes the request as reasoning_effort: "xhigh". The OpenCode Go gateway rejects xhigh for this model.
Environment
OMP: 18.0.0
Platform: Windows 11 x64
Provider/model: opencode-go/ox-alpha-free
API selected by OMP: openai-completions
Endpoint shape: /zen/go/v1/chat/completions
Reproduction
omp models opencode-go/ox-alpha-free
OMP reports:
thinking: low,high,max
Then:
omp -p --no-tools --no-session `--model opencode-go/ox-alpha-free `--thinking max `--max-time 2m`"Reply with exactly OK."
After retries, the provider returns:
[1210] This model always engages in thinking and cannot be disabled; please use low, high, or max
omp -p --no-tools --no-session `--model opencode-go/ox-alpha-free `--thinking high `"Reply with exactly OK."
This also succeeds through OpenCode directly and requests the provider's real max tier:
opencode run --pure `--model opencode-go/ox-alpha-free `--variant max `"Reply with exactly OK."
Both controls returned OK.
Expected behavior
Selecting max for a model whose catalog entry advertises max should send reasoning_effort: "max", not xhigh. Alternatively, the model entry should supply an explicit route-specific effort map.
Workaround
Use --thinking high in OMP. Direct OpenCode --variant max reaches actual max successfully.
This appears to be the same catalog/wire-mapping class previously fixed for other OpenCode Go models in #2864, but affects the current Ox Alpha entry and the newly distinct max tier.
The bundled opencode-go/ox-alpha-free catalog entry advertises thinking.efforts = [minimal, low, medium, high, xhigh] with no effort map. Resolving --thinking max clamps to the top supported tier (xhigh) and serializes it verbatim:
The OpenCode Go gateway rejects xhigh for this model ([1210] ... please use low, high, or max), so the top reasoning tier is unreachable through OMP even though OpenCode's own --variant max reaches it. --thinking high works because high is a valid wire value on both sides.
The gateway's actual reasoning vocabulary for this model is the wire-exact low/high/max ladder (thinking is mandatory) — the same ladder the OpenCode Go gateway already uses for GLM-5.3 and Kimi K3. The catalog entry's minimal..xhigh vocabulary is wrong for this route. Fixing the effort ladder for OpenCode gateway ox-alpha now.
Fix up in #9352. opencode-go/ox-alpha-free now uses the gateway's wire-exact low/high/max ladder (mandatory thinking), so --thinking max sends reasoning_effort: "max" and reaches the real top tier; minimal/medium/xhigh selections clamp down to accepted values instead of 400'ing. Scoped to the OpenCode gateways only — Kilo/NanoGPT/Venice/OpenRouter ox-alpha entries are untouched.
Activity
roboomp commentedon Aug 22, 2026
Looking into this, will report back with a repro.
roboomp commentedon Aug 22, 2026
Reproduced against the current tree.
The bundled
opencode-go/ox-alpha-freecatalog entry advertisesthinking.efforts = [minimal, low, medium, high, xhigh]with no effort map. Resolving--thinking maxclamps to the top supported tier (xhigh) and serializes it verbatim:The OpenCode Go gateway rejects
xhighfor this model ([1210] ... please use low, high, or max), so the top reasoning tier is unreachable through OMP even though OpenCode's own--variant maxreaches it.--thinking highworks becausehighis a valid wire value on both sides.The gateway's actual reasoning vocabulary for this model is the wire-exact
low/high/maxladder (thinking is mandatory) — the same ladder the OpenCode Go gateway already uses for GLM-5.3 and Kimi K3. The catalog entry'sminimal..xhighvocabulary is wrong for this route. Fixing the effort ladder for OpenCode gatewayox-alphanow.roboomp commentedon Aug 22, 2026
Fix up in #9352.
opencode-go/ox-alpha-freenow uses the gateway's wire-exactlow/high/maxladder (mandatory thinking), so--thinking maxsendsreasoning_effort: "max"and reaches the real top tier;minimal/medium/xhighselections clamp down to accepted values instead of 400'ing. Scoped to the OpenCode gateways only — Kilo/NanoGPT/Venice/OpenRouter ox-alpha entries are untouched.fix(catalog): map opencode-go ox-alpha to low/high/max effort ladder
fix(catalog): map opencode-go ox-alpha to low/high/max effort ladder
fix(catalog): normalized aliased OpenCode Zen ox-alpha effort ladders