feat: Ollama Cloud model-picker dropdown in settings (+ backend validation scripts)

[ankitvgupta]

Summary

Switching the open-source (Ollama Cloud) model used to require hand-typing model ids into free-text boxes in Settings. This PR replaces those inputs with a curated dropdown of common Ollama Cloud models, defaulting to GLM 5.2 (the current DEFAULT_OLLAMA_MODEL), with a "Custom…" escape hatch for any other id. The same picker is used in both places that take an Ollama model.

It also adds the backend validation/benchmark scripts that were used to evaluate candidate models (GLM, MiniMax, Kimi) end-to-end through both agent harnesses.

Changes

Feature (UX):

• shared/types.ts — COMMON_OLLAMA_MODELS curated list (verified-live :cloud ids), kept in sync with DEFAULT_OLLAMA_MODEL. GLM 5.2 is the default-first entry; MiniMax M3 / M2.7, Kimi K2.6 / K2-Thinking, Qwen3 Coder, DeepSeek are also selectable.
• OllamaModelSelect.tsx — reusable dropdown + custom-id escape hatch. Falls back to custom mode when the stored value isn't curated (so pre-existing hand-typed configs don't break), and drops back to the dropdown when the value is reset to a curated id.
• ExtensionsTab.tsx — Ollama Cloud "Default Model" → dropdown.
• SettingsPanel.tsx — per-feature Ollama model override → dropdown.

Scripts (local only; require OLLAMA_API_KEY, not in CI):

• scripts/test-minimax-claude-sdk.mjs / test-minimax-opencode.mjs — routing smoke tests (tool-calling) for the Claude Agent SDK + OpenCode harnesses over Ollama Cloud.
• scripts/bench-minimax.mjs — single-shot quality + TTFT/throughput across both transports.
• scripts/agentic-minimax.mjs — multi-step tool-chaining + cross-model speed comparison.
• scripts/lib/load-env.mjs — shared .env loader for the above.

Why a curated list + custom hatch

The list is a convenience, not an allowlist — new models ship on Ollama Cloud constantly, so "Custom…" keeps any id reachable. The default-first entry tracks DEFAULT_OLLAMA_MODEL so the active default always renders as a real option rather than dropping into custom mode.

Screenshots

Verified live via CDP in demo mode (local copies in .context/):

• ollama-dropdown-glm-default.png — Default Model = "GLM 5.2 — z.ai (default)", MiniMax still in the list
• ollama-dropdown-custom.png — "Custom…" reveals a free-text field
• ollama-perfeature-dropdown.png — General → feature routing renders the same dropdown

Test plan

• npm run typecheck ✅ · eslint ✅ · prettier --check ✅ · npm run test:unit → 1425 passed ✅
• npm run pre-pr (full): PASS — evals 80/80, agentic-verify clean, real-gmail 5/5
• Live-verified both dropdown locations: default = GLM 5.2, 7 options + Custom, escape hatch reveals/persists a text id, and typing/selecting a known id collapses back to the dropdown
• Candidate models validated E2E through both harnesses (tool-calling + streaming + completion) via the included scripts

Notes

• The Security Audit CI check is red on a pre-existing, repo-wide issue (npm audit --omit=dev --audit-level=high flags transitive prod deps hono + protobufjs; main's own latest CI fails the same check). This PR changes no dependencies — out of scope here; best fixed on main.

---

Pre-PR verdict: PASS

• mode: full
• sha: 7660486
• generated: 2026-06-16T23:10:47.968Z

Phase	Status	Duration
eval:analyzer	✅ exit 0	14.5s
eval:features	✅ exit 0	64.1s
agentic-verify	✅ exit 0	133.0s
real-gmail:cached	✅ exit 0	8.5s

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.

[greptile-apps]

Greptile Summary

This PR replaces free-text Ollama model id inputs in both Settings locations with a curated dropdown (OllamaModelSelect) backed by a new COMMON_OLLAMA_MODELS list, while preserving a "Custom…" escape hatch for unlisted ids. It also adds four validation/benchmark scripts for minimax-m3 and extracts the shared loadEnv helper into scripts/lib/load-env.mjs.

• UI: OllamaModelSelect handles the known/custom-mode toggle cleanly, including a useEffect that drops custom mode when an external settings reload resets the value to a curated model id.
• Scripts: All four scripts now import loadEnv from the shared library instead of duplicating the logic; the search_emails fixture in agentic-minimax.mjs was updated to filter by query terms rather than always returning all entries.
• Types: COMMON_OLLAMA_MODELS comment is internally consistent with DEFAULT_OLLAMA_MODEL = "glm-5.2:cloud"; the only stale reference is the JSDoc inside OllamaModelSelect.tsx which still says "defaulting to Kimi".

Confidence Score: 5/5

Safe to merge — the UI changes are additive and both callsites supply non-empty defaults, so the new dropdown has no empty-value edge cases in production. The scripts are not in CI and require a key to run.

The component logic is correct, the shared loadEnv extraction is straightforward, and both callsites guard against empty values. The only finding is a stale JSDoc comment ("defaulting to Kimi") that contradicts the actual first entry in COMMON_OLLAMA_MODELS (GLM 5.2), which has no runtime impact.

No files require special attention. The stale comment is in OllamaModelSelect.tsx but does not affect behavior.

Important Files Changed

Filename	Overview
src/shared/types.ts	Adds COMMON_OLLAMA_MODELS curated list (7 models); DEFAULT_OLLAMA_MODEL stays glm-5.2:cloud and is correctly kept as the first entry. Types and doc comment are consistent.
src/renderer/components/OllamaModelSelect.tsx	New dropdown + custom-mode escape-hatch component. Logic is sound; useEffect correctly resets custom state when value externally becomes a known model. JSDoc "defaulting to Kimi" is stale — actual first/default is GLM 5.2.
src/renderer/components/ExtensionsTab.tsx	Swaps bare text input for OllamaModelSelect in the Ollama Cloud Default Model field; initializes with DEFAULT_OLLAMA_MODEL so value is never empty.
src/renderer/components/SettingsPanel.tsx	Replaces bare text input with OllamaModelSelect for per-feature Ollama model override; falls back to DEFAULT_OLLAMA_MODEL when no saved value, so value is always non-empty.
scripts/lib/load-env.mjs	New shared .env loader extracted from the four benchmark scripts. Parses KEY=VALUE lines, strips outer double-quotes, respects existing env vars. Single-quoted values are not stripped but this is an intentional scope limitation.
scripts/agentic-minimax.mjs	Multi-step agentic test across both harnesses and 3 models. Fixture search_emails now correctly filters by query terms. Not part of CI; requires OLLAMA_API_KEY.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User opens Settings / Extensions] --> B{Saved value\nin COMMON_OLLAMA_MODELS?}
    B -- Yes --> C[Select shows curated option]
    B -- No + non-empty --> D[inCustomMode = true\nSelect shows Custom…\nText input revealed]
    C --> E{User picks Custom…}
    E -- Yes --> D
    D --> F{External settings reload\nresets value to known id?}
    F -- Yes --> G[useEffect fires\ncustom = false\nDropdown resumes]
    F -- No --> D
    C --> H{User picks different\ncurated model}
    H --> I[onChange fires with new id\nParent state updated]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[User opens Settings / Extensions] --> B{Saved value\nin COMMON_OLLAMA_MODELS?}
    B -- Yes --> C[Select shows curated option]
    B -- No + non-empty --> D[inCustomMode = true\nSelect shows Custom…\nText input revealed]
    C --> E{User picks Custom…}
    E -- Yes --> D
    D --> F{External settings reload\nresets value to known id?}
    F -- Yes --> G[useEffect fires\ncustom = false\nDropdown resumes]
    F -- No --> D
    C --> H{User picks different\ncurated model}
    H --> I[onChange fires with new id\nParent state updated]

Loading

_{Reviews (4): Last reviewed commit: "feat: default Ollama dropdown to GLM 5.2..." | Re-trigger Greptile}

[ankitvgupta]

✅ Pre-PR verification — PASS

• mode: full
• sha: 7660486
• generated: 2026-06-16T23:10:47Z

Phase	Status	Duration
eval:analyzer	✅ exit 0 (80/80)	14.5s
eval:features	✅ exit 0	64.1s
agentic-verify	✅ exit 0	133.0s
real-gmail:cached	✅ exit 0	8.5s

Agentic verification — summary

The primary user-visible change is the OllamaModelSelect component that replaces the plain text field with a curated model dropdown when "Ollama Cloud" is selected as a provider. Verified end-to-end against the post-merge state:

• The dropdown defaults to "GLM 5.2 — z.ai (default)" (glm-5.2:cloud, the current DEFAULT_OLLAMA_MODEL) — captured DOM: currentValue: "glm-5.2:cloud", selected: true.
• All 7 curated models are present, including minimax-m3:cloud and minimax-m2.7:cloud.
• Selecting "Custom…" reveals a text input pre-populated with the current model id; selecting a known model again hides the text input.
• Extensions tab renders correctly. No JS errors observed.

ℹ️ Correction: an earlier revision of this comment (commit dab2f1e, pre-merge) reported "Kimi K2.6 selected by default". That reflected the default before merging main — #175 made GLM 5.2 the default DEFAULT_OLLAMA_MODEL. This verification reflects the current 7660486 state (GLM 5.2 default). The earlier comment couldn't be auto-updated because an invalid GH_TOKEN shadowed the gh keyring auth during that pre-pr run.

[ankitvgupta]

Update: merged main, default Ollama model → GLM 5.2

• Merged origin/main (brings in feat: make GLM 5.2 the default Ollama Cloud model #175 GLM-5.2 default, fix: migrate persisted Ollama default kimi-k2.6 → glm-5.2 #177 persisted-config migration, fix(evals): reminder-service fixture should expect needs_reply=true #174 eval-fixture fix).
• Updated the new model-picker dropdown to match: GLM 5.2 is the default-first option; glm-5.2:cloud added (was absent, which would've forced the default into "Custom…" mode); (default) label moved off Kimi; superseded glm-4.6 dropped. MiniMax M3 / M2.7 remain selectable. Verified live via CDP.
• Re-ran full pre-pr: PASS — evals now 80/80 (the reminder-service fixture passes after fix(evals): reminder-service fixture should expect needs_reply=true #174), agentic-verify clean, real-gmail 5/5.

CI note — Security Audit (pre-existing, not from this PR)

The Security Audit job is red on npm audit --omit=dev --audit-level=high for two transitive production deps — hono and protobufjs — whose advisories were published recently. This PR changes no dependencies (no package.json / package-lock.json diff), and main's own latest CI is failing the same check. This is a repo-wide fix best done on main (a focused npm audit fix) rather than bundled into this UI change. Flagging so it isn't mistaken for a regression here.