feat(models): cross-assistant effort selector (low/medium/high/xhigh/max) for all providers

[umi-appcoder]

Split out from #336 (item D3). Verdict: confirmed missing — and the current knob is the wrong axis.

Ask: effort sliders mapped to real-world options.

No UI control exists (no effort anywhere in charts/workspace/web/src). The only knob is values.yaml:286-290 — claude.maxThinkingTokens -> MAX_THINKING_TOKENS, a raw integer token count, with no labels at all.

The token-count axis is deprecated. The real control on current models is output_config.effort, with exactly five levels — low / medium / high / xhigh / max — defaulting to high. xhigh sits between high and max, is the best setting for coding and agentic work, and is Claude Code's own default. The fixed thinking-token-budget concept it replaces is deprecated on 4.6-era models and removed outright on Opus 4.8 / 4.7 / Fable 5 / Sonnet 5, where budget_tokens returns a 400.

So the request maps cleanly: a 5-stop selector labelled low/medium/high/xhigh/max, mapping 1:1 onto the real levels, defaulting to high (or xhigh for coding-heavy defaults) — replacing, not supplementing, the token-count knob.

(Note: MAX_THINKING_TOKENS is a Claude Code env var, not a raw API parameter, so whether Claude Code translates or ignores it on current models is worth a quick check before ripping it out — but effort is unambiguously the correct axis going forward.)

Related: the model-list/picker work in #361 touches the same server/SPA config surface and should probably share the picker UI.

---

Update 2026-07-21 (research pass, details in this comment):

• Scope broadened: the selector must work across ALL assistants (claude, codex, ante, opencode-openrouter, opencode-deepseek, librefang, kc-harness), not just Claude — per @imran31415's comment.
• Correction: claude.maxThinkingTokens / MAX_THINKING_TOKENS does not actually exist in this repo (values.yaml:286-290 is just the claude.apiKey block). This is a greenfield feature, not a replacement.
• Design direction: clone the Hypervisor: Make it so the user can change the provider model from a UI switcher in the Hypervisor chat window. Should work for new or switching during session #308/feat(models): enumerated Claude model list + picker; add Fable 5 (not a drop-in entry) #361 model-picker machinery — canonical 5-level axis (low/medium/high/xhigh/max, default high) → server-side per-assistant effort registry + resolve_effort() → per-thread API param + SPA selector → per-adapter translation to each CLI's native knob (CLAUDE_CODE_EFFORT_LEVEL, codex model_reasoning_effort, opencode reasoningEffort, Ollama think), clamping where an assistant's native level set is narrower and hiding the control where none exists.

[imran31415]

This needs to be supporte for all providers not just claude and needs to be scoped to ensure it will work accross them cleanly.

[umi-appcoder]

Did a research pass on this (cross-provider scoping per the comment above). Three findings change the shape of the work:

1. Correction to the issue body: the knob we planned to replace doesn't exist

claude.maxThinkingTokens / MAX_THINKING_TOKENS appears nowhere in the repo — values.yaml:286-290 is just the # Claude Code configuration header and the claude.apiKey block. git log -S finds no trace on any branch either. So there is nothing to rip out; #362 is a greenfield feature, not a replacement. (Moot anyway: MAX_THINKING_TOKENS is a deprecated Claude Code env var — ignored on 4.6+ models, where fixed thinking budgets were removed in favor of output_config.effort.)

2. Cross-provider effort survey — every major harness has a knob, but the level sets and config surfaces differ

Assistant	Native knob	Levels	How to set per-launch
claude (Claude Code)	effort level	low / medium / high / xhigh (+ /effort slider, auto)	CLAUDE_CODE_EFFORT_LEVEL env var, --effort flag, or effortLevel in settings.json (env wins)
codex (Codex CLI)	model_reasoning_effort	minimal / low / medium / high / xhigh	codex -c model_reasoning_effort="high" or ~/.codex/config.toml
opencode-openrouter	reasoningEffort model option	minimal → xhigh (OpenRouter unified reasoning.effort)	opencode.json → provider.openrouter.models.<m>.options.reasoningEffort (config file, no CLI flag)
opencode-deepseek	same option, but DeepSeek support is partial	deepseek-reasoner historically had no CoT-length control; reasoning_effort is only now landing in DeepSeek's API	config file, if/when the model honors it
ante	inherits OpenRouter reasoning.effort if ante exposes it	—	open question — needs a check of ante's config surface; not documented
librefang	none documented	—	hide the control
kc-harness (Ollama)	think parameter	true/false/low/medium/high (no xhigh/max)	harness.py → Ollama API think field

Two caveats worth encoding:

• OpenRouter + Anthropic models: reasoning.effort is reported to be a silent no-op for Claude 4.6+/Fable over OpenRouter (the model runs its adaptive default regardless). So for ante/opencode-openrouter the knob is only real for OpenAI-family models.
• Level sets don't line up. Codex/OpenRouter have minimal but no max; Ollama tops out at high; Claude Code tops out at xhigh in settings while the raw Anthropic API goes to max.

3. Proposed design: clone the model-picker machinery (#308/#361), don't invent a new axis

The repo already solves this exact shape problem for models: helm default (assistant.* → KC_*_MODEL env in deployment.yaml:103-139) → server per-assistant registry (_CLAUDE_MODELS / _MODEL_LIST_ENV, available_models() / resolve_model() at server.py:1198-1270) → surfaced via /api/hypervisor/config → per-thread choice in the SPA (selectedModel in web/src/store/hypervisor.ts, PUT /api/hypervisor/threads/{id}/model) → translated to each CLI's flags in assistant_command() (server.py:1312-1377), the hypervisor adapters (hypervisor_session.py:433-735), and the orchestrator ladder (mcp_agent_orchestrator.py:287-344).

Effort should be the twin of that, with a per-assistant translation table absorbing the differences:

1. UI axis: the 5 canonical levels low / medium / high / xhigh / max, default high — labels stay stable across assistants.
2. Server registry: per-assistant supported effort levels + resolve_effort() mirroring resolve_model(). Each adapter maps the canonical level to its native flag, clamping where the native set is narrower (Ollama: xhigh/max → high; Codex/OpenRouter: max → xhigh; Claude Code: max → xhigh unless the CLI accepts it). Assistants with no knob (librefang, ante-until-verified) report an empty list and the SPA hides the selector — same pattern as assistants with no model list.
3. Per-thread API param (effort on POST /api/hypervisor/threads + PUT …/threads/{id}/effort), read fresh each turn like model is, so mid-thread switching works.
4. Helm value only as default/curation: assistant.<name>.effort → KC_*_EFFORT env, matching the KC_*_MODEL pattern. No raw token counts anywhere.
5. Delivery mechanism per assistant: env var for claude (CLAUDE_CODE_EFFORT_LEVEL — cleanest, no argv change), -c model_reasoning_effort= for codex, config-file write for opencode (needs care — it's a shared file), think param for kc-harness.

Since #361's model picker touches the identical server/SPA surface, the effort selector should land after (or with) it and reuse the same dropdown component in the Hypervisor tab.

Open items before implementation:

• Verify whether ante exposes OpenRouter's reasoning.effort (config or env)
• Verify current opencode supports a per-invocation override (flag/env) vs. config-file-only
• Decide clamp-vs-hide for levels an assistant doesn't support (recommend clamp + show effective level in the UI)