[TheGreatAxios]

Summary

• Custom OpenAI-compatible providers configure context window, max output tokens, temperature, and top p; blank fields keep existing defaults and temperature zero remains valid.
• Saved settings feed request defaults and refresh live context-window/compaction limits, including model switching, discovery, edits, removal, and duplicate-model precedence.
• Token counts require positive integers; sampling ranges and mutually exclusive temperature/top-p are validated before connection or persistence. Preset and OAuth paths ignore hidden custom fields.

Verification

• bun run check passes against current main: 8,254 tests, zero failures.
• Failing-then-passing regressions cover invalid settings, persistence, hidden-field isolation, live context limits, worker/candidate isolation, and model switching during delayed discovery.
• Independent review findings are resolved; no live provider API inference runs.

Fixes CL-9834

[linear-code]

CL-9834

[TheGreatAxios]

Review · Comment

Custom OpenAI-compatible providers persist optional context window, max output tokens, temperature, and top-p; the form rejects setting both sampling knobs.

Findings

• src/config/index.ts:280 — buildOpenAISource writes both temperature and top_p into providerOptions when both fields are set. Mutual exclusion exists only in src/tui/provider/submit.ts:84. Hand-edited settings.json (or a catalog row) with both knobs reaches the outbound request. Verified: buildInferenceSourceForRef with temperature: 0.5 and topP: 0.8 yields { temperature: 0.5, top_p: 0.8 }. Refuse the pair in buildOpenAISource / the settings schema, not only at submit.

• src/tui/provider/form.ts:69 — stepReady has no lower bound for maxTokens or contextWindow. "0" and "-1" are ready, submit persists them, and fields.maxTokens ?? SOURCE_MAX_TOKENS keeps 0 (it is not undefined). A zero/negative output cap starves generation; a zero/negative context window poisons compaction and the status-bar meter. Reject non-positive integers at stepReady and at submit, matching the 0–2 / 0–1 checks already used for temperature / top-p.

[TheGreatAxios]

Review findings resolved at 9f16963: live context limits refresh without a restart, hidden custom numeric fields cannot affect preset/OAuth submissions, and invalid/noninteger token counts are rejected by shared ArkType validation. A further integration regression proves delayed discovery cannot undo a model switch. Test-first evidence: original regressions fail before fixes, focused checks pass afterward, and full bun run check against current main passes with 8,254 tests and zero failures. All original PR files and follow-up changes are reviewed; no remaining actionable findings. No live API smoke is claimed.