feat: Ollama Cloud model picker dropdown + minimax-m3 backend validation#166
Open
ankitvgupta wants to merge 4 commits into
Open
feat: Ollama Cloud model picker dropdown + minimax-m3 backend validation#166ankitvgupta wants to merge 4 commits into
ankitvgupta wants to merge 4 commits into
Conversation
Validate the new minimax-m3:cloud model end-to-end through both agent harnesses and characterize its quality/speed on Ollama Cloud: - test-minimax-claude-sdk.mjs Claude Agent SDK + Ollama routing smoke (tool call) - test-minimax-opencode.mjs OpenCode + Ollama routing smoke (tool call) - bench-minimax.mjs single-shot quality + TTFT/throughput, both transports - agentic-minimax.mjs multi-step tool-chaining + cross-model speed comparison Each mirrors the exact routing the providers build (env remap for the Claude SDK path; ollama-cloud provider registration for OpenCode). Not wired into CI — they require OLLAMA_API_KEY and hit real Ollama Cloud. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Switching open-source models required hand-typing model ids in two places. Replace both free-text inputs — Extensions → Ollama Cloud "Default Model", and the per-feature override in General → feature routing — with a curated dropdown of common Ollama Cloud models, defaulting to Kimi K2.6. A "Custom…" option reveals a text field for any other model id, so new models (e.g. minimax-m3) stay reachable without maintaining an allowlist. When a saved value isn't in the curated list, the field opens in custom mode showing what's stored, so existing hand-typed configs don't break. - COMMON_OLLAMA_MODELS curated list in shared/types.ts (verified-live :cloud ids) - reusable OllamaModelSelect component (dropdown + custom escape hatch) - wired into ExtensionsTab (default model) and SettingsPanel (per-feature) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
Greptile SummaryThis PR adds an
Confidence Score: 5/5Safe to merge — the UI changes are well-scoped and all three previously flagged issues have been addressed. The OllamaModelSelect component correctly handles all three initial states (curated value, unknown non-empty value, empty string), the useEffect properly resets custom mode when the parent supplies a curated value, and the three previously flagged issues are all resolved. The only finding is a port leak in dev-only benchmark scripts on error paths, which has no impact on production code or the test suite. The agentic benchmark and opencode smoke-test scripts have cleanup-on-error gaps, but these are developer-run scripts with no CI or production impact.
|
| Filename | Overview |
|---|---|
| src/shared/types.ts | Adds COMMON_OLLAMA_MODELS readonly array with 7 curated :cloud model ids; well-typed and clearly documented as a convenience list, not an allowlist. |
| src/renderer/components/OllamaModelSelect.tsx | New reusable dropdown with Custom… escape hatch; correctly handles pre-existing hand-typed values on mount and uses a useEffect to sync back to dropdown when value is externally reset to a curated model. |
| src/renderer/components/ExtensionsTab.tsx | Replaces the free-text Default Model input with OllamaModelSelect; onChange signature correctly adapted from event-based to value-based. |
| src/renderer/components/SettingsPanel.tsx | Replaces per-feature Ollama model text inputs with OllamaModelSelect; value fallback to DEFAULT_OLLAMA_MODEL preserved via ?? operator on the call site. |
| scripts/lib/load-env.mjs | Shared minimal .env loader extracted from the four script files; handles double-quoted values and skips keys already in process.env. |
| scripts/agentic-minimax.mjs | Multi-step agentic harness test; search_emails now uses proper term-based filtering. runOpencode starts httpServer before createOpencodeServer, so a failure in server setup leaves the HTTP bridge port open. |
| scripts/bench-minimax.mjs | Quality + TTFT/throughput benchmark across Anthropic and OpenAI-compat transports; median aggregation and per-trial error capture are correct. |
| scripts/test-minimax-claude-sdk.mjs | Claude Agent SDK smoke test for minimax-m3; correctly mirrors ClaudeAgentProvider's childEnv construction and MCP tool setup. |
| scripts/test-minimax-opencode.mjs | OpenCode smoke test for minimax-m3; cleanup (server.close, httpServer.close) only happens on the happy path — an unexpected throw before those lines would leave both servers open. |
Flowchart
%%{init: {'theme': 'neutral'}}%%
flowchart TD
A[OllamaModelSelect mounts] --> B{value in COMMON_OLLAMA_MODELS?}
B -- Yes --> C[custom = false, inCustomMode = false, Show dropdown curated model]
B -- No, value length > 0 --> D[custom = true, inCustomMode = true, Show dropdown CUSTOM + text input]
B -- No, value empty --> E[custom = false, inCustomMode = false, browser renders first option visually]
C --> F{User changes dropdown}
F -- Picks curated model --> G[onChange called, custom = false]
F -- Picks Custom --> H[custom = true, inCustomMode = true, Text input appears]
H --> I{User types in text input}
I -- Types known id --> J[onChange called, useEffect: isKnown=true, custom=false, Dropdown mode restored]
I -- Types novel id --> K[onChange called, custom stays true, inCustomMode stays true]
D --> F
E --> F
Prompt To Fix All With AI
Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.
---
### Issue 1 of 1
scripts/agentic-minimax.mjs:396-418
**`httpServer` not closed on `createOpencodeServer` failure**
`httpServer` is started and listening (line 397) before `createOpencodeServer` is awaited (line 400). If `createOpencodeServer` throws — e.g. a port conflict or missing binary on first install — the HTTP bridge server is never closed. Because `runOne` catches the error and continues to the next scenario, multiple leaked ports can accumulate across a full multi-model run. The same pattern exists in `test-minimax-opencode.mjs` before its cleanup block.
Wrapping the post-listen body in a `try/finally` that calls `httpServer.close()` would prevent the leak regardless of the failure point.
Reviews (3): Last reviewed commit: "address Greptile review feedback (review..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Owner
Author
✅ Pre-PR verification — PASS
Agentic verification — summaryAgentic verification — verify-diff
SummaryThe primary user-visible change is the OllamaModelSelect component that replaces a plain text field with a curated model dropdown when 'Ollama Cloud' is selected as a provider. Verified end-to-end: switching Email Analysis to Ollama Cloud shows the 7-model dropdown (Kimi K2.6 selected by default); choosing 'Custom…' correctly reveals a text input pre-populated with the current model id; switching back to another provider is unaffected. The Extensions tab renders correctly with built-in and toggleable extensions. No JS errors observed. Agentic verification — literal trace[2026-06-01T22:32:04.134Z] text: No errors. All checks passed. The primary regression checks are complete:
{"verdict":"pass","summary":"The primary user-visible change is the OllamaModelSelect component that replaces a plain text field with a curated model dropdown when 'Ollama Cloud' is selected as a provider. Verified end-to-end: switching Email Analysis to Ollama Cloud shows the 7-model dropdown (Kimi K2.6 selected by default); choosing 'Custom…' correctly reveals a text input pre-populated with the current model id; switching back to another provider is unaffected. The Extensions tab renders correctly with built-in and toggleable extensions. No JS errors observed.","anomalies":[],"actions_taken":10} |
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- OllamaModelSelect: clear custom mode when value is externally reset to a curated model id, via an effect on [value]. Previously the sticky custom flag could leave the select showing "Custom…" with a curated id in the text box. Verified live: typing/selecting a known id now collapses to the dropdown. - agentic-minimax search_emails: make the fixture genuinely query-driven (match query terms against subject/sender) instead of a hardcoded "acme" branch that ignored the query. - extract the duplicated loadEnv helper into scripts/lib/load-env.mjs and import it from all four minimax scripts. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two related pieces of work that came out of evaluating the new minimax-m3 model on Ollama Cloud:
Changes
Feature (UX):
shared/types.ts—COMMON_OLLAMA_MODELScurated list (7 verified-live:cloudids; Kimi first/default)OllamaModelSelect.tsx— reusable dropdown + custom-id escape hatchExtensionsTab.tsx— Ollama Cloud "Default Model" → dropdownSettingsPanel.tsx— per-feature Ollama override → dropdownScripts (not in CI; require
OLLAMA_API_KEY):test-minimax-claude-sdk.mjs/test-minimax-opencode.mjs— routing smoke tests (tool-calling) for each harnessbench-minimax.mjs— single-shot quality + TTFT/throughput across both transportsagentic-minimax.mjs— multi-step tool-chaining + cross-model speed comparisonWhy
The curated list is a convenience, not an allowlist — new models ship on Ollama Cloud constantly, so the "Custom…" option keeps any id reachable. When a saved value isn't in the list (e.g. a previously hand-typed model), the field opens in custom mode showing what's stored, so existing configs don't silently break.
Screenshots
UI states (captured live via CDP in demo mode; local copies in
.context/):Default Modelshowing "MiniMax M3 — newest, agentic", default is Kimi(Will attach inline — these were verified in the running app.)
Test plan
npm run typecheck✅ ·eslint✅ ·npm run test:unit→ 1425 passed ✅Pre-PR verdict: PASS
fulldab2f1e