[TheGreatAxios]

Summary

• Custom OpenAI-compatible providers explicitly select enabled reasoning levels (none/low/medium/high/xhigh/max) and one enabled default; empty ladders and disabled defaults cannot save.
• Persisted ladders/defaults govern startup, resume, model switching, local/profile validation, worker role clamping, displayed effort, and cycling. Legacy and protocol-specific providers keep their existing behavior.
• Toggle focus and hidden-field isolation remain stable. Reasoning and sampling options compose, with current settings overriding stale discovery catalogs so disabled efforts and cleared sampling values cannot leak onto requests.

Verification

• Full bun run check with all four PRs composed passes: 8,314 tests, zero failures.
• Failing-then-passing regressions cover custom defaults, empty selections, local/profile pins, role cascading, cycling, toggle focus, and stale settings/catalog precedence.
• Combined submit → settings reload → source assembly → adapter request verifies context/maxTokens, enabled default, temperature zero, and omission of cleared top-p.
• Independent review findings are resolved; no live provider API smoke runs.

Fixes CL-9833

[linear-code]

CL-9833

[TheGreatAxios]

Review · Comment

Custom OpenAI-compatible providers persist an operator effort ladder and default; Shift+Tab, the prompt label, and the openai-compatible fall-through use that set instead of the family table.

Findings

• src/config/index.ts:1150 — loadConfig still calls validateEffort(model, local.reasoningEffort, isCodex) with no providerEfforts. A custom ladder that includes none (or any level the family table rejects for that model id) cannot round-trip through .corbits/settings.json: restart throws Invalid reasoningEffort in local settings. The same miss is at src/subagent/agent-fleet.ts:1039, so spawn_agent rejects an inherited custom-only pin against the family table. Pass entry.reasoningEfforts into both calls. Verified: validateEffort("fp-custom-model", "none") is {ok:false} while the same call with ["none","low","high"] is {ok:true}.

• src/config/inference-sources.ts:191 and src/provider/reasoning-effort.ts:244 — the prompt label resolves defaultReasoningEffort via resolveSessionEffort(..., providerDefault), but the custom source only computes customEffort when configured !== undefined, and cycleReasoningEffort never forwards providerDefault. A fresh session shows the operator default and sends nothing; the first Shift+Tab starts at supported[0], not the picked default. Hydrate session reasoningEffort from resolveSessionEffort at startup (including providerDefault), and pass providerDefault through cycleReasoningEffort.

• src/tui/runner/wiring.ts:459 — Shift+Tab and the model label read settings.providers[name].reasoningEfforts with no custom-path guard. The second commit closed that leak on the wire for first-class builders; the TUI still honors a hand-edited ladder on Go/Claude/Codex and can cycle onto a level the source will clamp away. Gate the TUI the same way as buildInferenceSourceForRef (only the openai-compatible fall-through).

[TheGreatAxios]

All five initial reasoning findings are resolved: defaults serialize without session pins; the form represents the explicit six-level set and rejects an empty ladder; config/profile and role resolution honor enabled levels; cycling advances from the effective default; toggles keep cursor focus. Further test-first regressions establish current-settings precedence over stale catalogs and complete reasoning/sampling serialization without resurrecting disabled/cleared values. Twelve overlap conflicts are composed without dropping either feature or its tests. Original changed files and follow-ups are reviewed. Full bun run check with all four PRs composed passes: 8,314 tests, zero failures. No remaining actionable findings or live API smoke claims.