feat(ui): run the Lite preset's medium and complex tiers at their documented efforts

[tin-berri]

TLDR

Problem this solves:

• Lite runs Muse Spark and Kimi K3 at provider defaults
• Kimi K3's own default is max, which it never got
• No bundled preset ever set a per-tier effort

How it solves it:

• Medium runs Muse Spark 1.2 at xhigh
• Complex runs Kimi K3 at max
• Both efforts are what those models document

User Flow

Before: an admin picking the Lite template gets two mid-tier models on whatever thinking their provider defaults to

1. They open https://litellm-domain/ui/?page=llm-playground, go to Add Model, then the Auto Router tab
2. They pick the Lite template
3. Medium fills in with muse-spark-1.2 and Complex with kimi-k3
4. Every Reasoning effort dropdown reads Default
5. They try to set Complex to max by hand and the dropdown does not offer it

After: both tiers arrive on the effort their model documents

1. They open the same page and pick the Lite template
2. The Reasoning effort dropdown beside muse-spark-1.2 reads xhigh, and the one beside kimi-k3 reads max
3. They save the router
4. A complex prompt reaches Kimi K3 with reasoning_effort: max even though the caller sent none
5. A medium one reaches Muse Spark 1.2 with reasoning_effort: xhigh

Relevant issues

• Stacked on feat(model_prices): let a map entry declare its exact reasoning_effort levels #38481, which is what lets the map declare max for kimi-k3 at all
• Sibling change for the Anthropic Family preset is feat(ui): run the Anthropic Family preset's reasoning tier on Opus 5 at high thinking #38490 and is independent of both

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5

Screenshots / Proof of Fix

Shared setup. A local proxy holding the four model groups this preset names, each pointed at a local stub upstream that echoes the body it received. The auto-router is built by reading autorouter_presets.json straight off the branch and handing its complexity_router_config to the proxy verbatim, so what is under test is the shipped preset data rather than a hand-copied version of it.

LITELLM_LOCAL_MODEL_COST_MAP=True litellm --config rig/preset_config.yaml --port 4481

No request below sends reasoning_effort. Every value the stub reports came from the preset. Tiers are reached by prompt complexity, with LITELLM ESCALATE bumping the scored tier one step up.

Substitution declared: no provider on this account sells Kimi K3, Muse Spark or Claude Opus 5, so the upstream is a local stub. The stub only ever receives what the gateway decided to send, which is the whole question here.

Before (49655da, the parent commit)

1. Trivial prompt routes to fireworks_ai/deepseek-v4-flash, the stub receives no effort
2. Escalated trivial prompt routes to meta/muse-spark-1.2, the stub receives no effort
3. Hard prompt routes to fireworks_ai/kimi-k3, the stub receives no effort
4. Escalated hard prompt routes to anthropic/claude-opus-5, the stub receives no effort

After (98a7b4a)

1. The same four prompts, same commands:

SIMPLE      routed=fireworks_ai/deepseek-v4-flash  stub received: nothing (provider default)
MEDIUM      routed=meta/muse-spark-1.2             stub received: {'reasoning_effort': 'xhigh'}
COMPLEX     routed=fireworks_ai/kimi-k3            stub received: {'reasoning_effort': 'max'}
REASONING   routed=anthropic/claude-opus-5         stub received: nothing (provider default)

2. Medium and Complex carry the efforts the preset declares, on requests that asked for none
3. Simple and Reasoning send nothing, so the effort is scoped to its tier rather than applied router-wide

Type

🆕 New Feature

Caveats (if any)

Medium

• Kimi K3 at max needs feat(model_prices): let a map entry declare its exact reasoning_effort levels #38481 to merge first
  • Without it the map does not declare max for kimi-k3, so the dropdown will not offer it
  • A stored value still renders and stays clearable, so the preset works, but an admin cannot re-pick max after clearing it

Low

• Muse Spark 1.2 needed no map change
  • It already carries supports_xhigh_reasoning_effort: true, confirmed on the live rig

Final Attestation

• The tests check the right things, including the edge cases

---

Note

Low Risk
Bundled preset JSON and dashboard tests only; behavior change is default routing params for admins who choose the Lite template, with no auth or core gateway logic edits in this diff.

Overview
The Lite auto-router preset now declares per-tier reasoning_effort for its medium and complex models instead of leaving them on provider defaults.

autorouter_presets.json adds tier_model_configs for MEDIUM (muse-spark-1.2 → xhigh) and COMPLEX (kimi-k3 → max), and the preset description is updated to match. SIMPLE and REASONING tiers are unchanged.

Tests lock in those tier_model_configs values and assert buildPresetPrefill maps them into tier_model_params so the Add Model / Auto Router UI shows the correct effort dropdowns and saved routers forward the params on routed requests.

^{Reviewed by Cursor Bugbot for commit 542649b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[tin-berri]

@greptileai can you reply here with a score and why that score?

[tin-berri]

bugbot run

[greptile-apps]

Score: 4/5

The feature is well-scoped and convincingly covered. The explicit supported_reasoning_efforts capability correctly takes precedence over per-level flags, preserves canonical ordering, supports prefixed/bare model fallback, and survives model-info hydration. The preset changes keep reasoning efforts tier-specific, and the tests cover resolver behavior, mixed-group intersection, preset prefill, and the final create payload.

I’m holding back 5/5 because rig/stub.py is not portable repository test infrastructure: it imports FastAPI outside proxy/ and writes to a hard-coded absolute path. It should be removed or made repository-safe before merge. I found no functional issue in the reasoning-effort or preset implementation itself.

[greptile-apps]

Greptile Summary

The Lite auto-router preset now assigns documented reasoning efforts to its medium and complex tiers.

• Configures Muse Spark 1.2 with xhigh effort and Kimi K3 with max.
• Adds tests for the preset data and its conversion into UI prefill parameters.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
ui/litellm-dashboard/src/autorouter_presets.json	Updates the Lite preset description and adds per-tier reasoning-effort configuration for its medium and complex models.
ui/litellm-dashboard/src/lib/autorouter_presets.test.ts	Verifies the Lite preset’s effort declarations and their conversion into tier model parameters.

_{Reviews (5): Last reviewed commit: "feat(ui): run the Lite preset's medium a..." | Re-trigger Greptile}

[greptile-apps]

Out-of-boundary FastAPI handler

This file constructs custom FastAPI request handlers outside litellm/proxy/, bypassing the repository's established dependency and handler boundary.

Rule Used: What: Do not allow fastapi imports on files outsid... (source)

[greptile-apps]

Machine-specific output path

Both handlers write to an author-specific directory that the repository never creates, so requests in a normal checkout raise FileNotFoundError instead of returning the expected stub response.

[greptile-apps]

Narrative test commentary

These tests add extensive narrative comments and docstrings beyond the repository's allowed comment categories, duplicating implementation history that can drift from the self-describing assertions.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[tin-berri]

@greptileai rig/stub.py is gone; it was a stale parent commit, now rebased. This PR is lite-only. Score?

[tin-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 98a7b4a. Configure here.}

[tin-berri]

Pushed: trimmed overlong lines and condensed the test prose per the repo comment rule. All within 120 now.

[tin-berri]

bugbot run

[tin-berri]

@greptileai line lengths and test comments addressed. Score?

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 7399548. Configure here.}

[tin-berri]

bugbot run

[tin-berri]

@greptileai can you reply here with a score and why that score?

[greptile-apps]

Confidence score: 4/5.

No blocking issues found. The change is appropriately scoped to the Lite preset, sets xhigh and max only on the MEDIUM and COMPLEX tiers, keeps SIMPLE/REASONING unchanged, and the description matches the shipped behavior. The added tests verify both the bundled tier_model_configs and the resulting tier_model_params, plus guard against tier/model mismatches across all presets. The supplied stub-upstream evidence also demonstrates that the values reach provider requests when the caller omits reasoning_effort.

I’m not giving 5/5 because the max option depends on stacked PR #38481, and the end-to-end forwarding evidence is provided as manual validation rather than an automated regression test.

[tin-berri]

bugbot run

[tin-berri]

@greptileai can you reply here with a score and why that score?

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit a4fef59. Configure here.}

[greptile-apps]

Confidence score: 4/5

The change is appropriately scoped and correctly uses the existing tier_model_configs → tier_model_params → serialized router-config flow. The Lite preset sets muse-spark-1.2 to xhigh only for MEDIUM and kimi-k3 to max only for COMPLEX, while leaving SIMPLE and REASONING unchanged. The tests cover the preset values, tier/model consistency, and UI prefill behavior, including model-name resolution.

I’m not giving 5/5 yet because the max capability depends on #38481 and the current e2e check is still pending; those should be confirmed before merge.

[greptile-apps]

Confidence score: 4/5.

The change is focused and correctly uses the existing tier_model_configs → tier_model_params path. The Lite preset now declares xhigh for Muse Spark 1.2 and max for Kimi K3, while leaving SIMPLE and REASONING unchanged. The added tests verify both the shipped preset values and the prefill transformation, and the provided stub-upstream evidence confirms the values reach routed requests when callers omit reasoning_effort.

I’m not giving 5/5 because the max dropdown availability depends on stacked PR #38481, and the tests added here stop at preset/prefill transformation rather than directly exercising the final request serialization for both tiers.

[tin-berri]

@greptileai can you reply here with score & why

[tin-berri]

bugbot run

[greptile-apps]

Confidence score: 5/5

I found no blocking issues. The change is narrowly scoped to the Lite preset and uses the existing tier_model_configs → tier_model_params hydration path, so xhigh and max are applied only to the MEDIUM and COMPLEX tiers rather than globally. The tests cover both the bundled declarations and the UI prefill transformation, while the existing invariant checks ensure each configured model belongs to its tier. SIMPLE and REASONING remain unchanged. The dependency on #38481 for exposing max in the capability map is an integration prerequisite, not a defect in this PR.

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 542649b. Configure here.}

[tin-berri]

@greptileai rereview