release: v6.8.4

[code-crusher]

Release v6.8.4

Fixed

• Dynamic model catalog no longer prunes the user's current selection. fetchDynamicModels reconciles the registry against each backend response, but previously pruned any managed model not in the latest fetch — including the model the user had just selected. A transient backend gap (or a model temporarily missing from one /v1/models response) would delete it from AXON_MODELS, and the next loadSettings() would silently fall back to DEFAULT_MODEL_ID. The reconciliation now skips the currently selected model so a user's choice survives across /new, /resume, and mid-session usage refreshes.

Changed

• TIP text under the spinner now renders in an info-purple color (#C792EA dark / #7C4DFF light) so it's clearly distinct from the thinking spinner. Added a new info theme token.
• README screenshot now hosted on Cloudinary instead of a local assets/ file.

Chore

• Version bumped to 6.8.4.

Checklist

• npm run typecheck and npm run build pass
• Version 6.8.4 not already on npm
• Merge, then tag v6.8.4 on main to trigger publish

[matterai-app]

Summary By MatterAI

🔄 What Changed

• Bumped package version to v6.8.4 in package.json.
• Implemented loadSettingsModel() in src/api/models.ts to synchronously read the user's selected model from disk (config.json) without triggering full settings-load side effects.
• Updated model fetching reconciliation logic to ensure the user's current model selection is never pruned during dynamic model updates.
• Added an info color token to branding and theme definitions (src/branding.ts, src/ui/theme.tsx) and integrated it into the Spinner component for tips.

🔍 Impact of the Change

• Prevents user-selected AI models from being inadvertently wiped or swapped out during transient backend catalog refreshes.
• Enhances TUI visual feedback by introducing a distinct information color token for loading tips.

📁 Total Files Changed

Click to Expand

File	ChangeLog
Version Bump package.json	Updated package version to 6.8.4.
Model Persistence src/api/models.ts	Added loadSettingsModel() to protect user model selection from pruning.
Theme Branding src/branding.ts	Added info color branding token.
UI Spinner src/ui/components/Spinner.tsx	Applied theme info color to loading tips.
Theme Definitions src/ui/theme.tsx	Defined info color values for dark and light themes.

🧪 Test Added/Recommended

Recommended

• Add unit tests for loadSettingsModel() covering missing config files, invalid JSON, and valid model selection reads.
• Add integration tests verifying that dynamic model reconciliation respects the user's current model selection.

🔒 Security Vulnerabilities

• None detected.

[matterai-app]

🧪 PR Review is completed: Release bump plus a fix that preserves the user's selected model during dynamic-catalog pruning, and a new info theme token for spinner tips. The pruning guard only reads config.json, so selections made via settings.json or MATTERAI_MODEL are still vulnerable to being pruned.

Skipped files

• CHANGELOG.md: Skipped file pattern
• README.md: Skipped file pattern
• assets/orbcode-screenshot.webp: File hunk diff too large

⬇️ Low Priority Suggestions (1)

src/api/models.ts (1 suggestion)

Location: src/api/models.ts (Lines 12-21)

🟡 Logic Error

Issue: loadSettingsModel() only reads config.json, but per loadSettings() in src/config/settings.ts, the effective model selection can also come from settings.json (user or project scope — model is in SETTINGS_KEYS) or the MATTERAI_MODEL env var, which takes precedence over all files. If a user's selected model is set via either of those paths and the backend catalog no longer lists it, the pruning loop at line 476 will still delete it from AXON_MODELS. The next loadSettings() then fails isValidAxonModel() and silently resets the user to DEFAULT_MODEL_ID — exactly the behavior this PR intends to prevent.

Fix: Check MATTERAI_MODEL first (it has highest precedence), then walk the same file precedence as loadSettings(): config.json, user settings.json, project .orbcode/settings.json, returning the first model found. Wrap each read in its own try/catch so a malformed config.json doesn't skip the settings.json candidates.

Impact: Guarantees the user's actual effective model selection is never pruned, regardless of where it was configured.

-  function loadSettingsModel(): string {
-    try {
-      const dir = process.env.MATTERAI_CONFIG_DIR || path.join(os.homedir(), ".orbcode")
-      const raw = fs.readFileSync(path.join(dir, "config.json"), "utf8")
-      const parsed = JSON.parse(raw)
-      return typeof parsed.model === "string" ? parsed.model : ""
-    } catch {
-      return ""
-    }
-  }
+  function loadSettingsModel(): string {
+    if (process.env.MATTERAI_MODEL) return process.env.MATTERAI_MODEL;
+    const dir = process.env.MATTERAI_CONFIG_DIR || path.join(os.homedir(), ".orbcode");
+    const candidates = [
+      path.join(dir, "config.json"),
+      path.join(dir, "settings.json"),
+      path.join(process.cwd(), ".orbcode", "settings.json"),
+    ];
+    for (const file of candidates) {
+      try {
+        const parsed = JSON.parse(fs.readFileSync(file, "utf8"));
+        if (typeof parsed.model === "string") return parsed.model;
+      } catch {
+        // missing or malformed file — try the next candidate
+      }
+    }
+    return "";
+  }