diff --git a/.gemini/skills/regression-finder/SKILL.md b/.gemini/skills/regression-finder/SKILL.md
new file mode 100644
index 0000000000..0f9e1b8946
--- /dev/null
+++ b/.gemini/skills/regression-finder/SKILL.md
@@ -0,0 +1,55 @@
+---
+name: regression-finder
+description: Identify the root cause of a regression by scanning recent PRs and commits or using automated bisecting. Use when a user reports a bug that "used to work before", "broke", "stopped working", or asks "which PR caused this regression".
+---
+
+# Regression Finder
+
+## Overview
+
+This skill helps you find the exact change (PR or commit) that introduced a bug. It uses a combination of metadata analysis (searching PR titles and file changes) and behavioral analysis (automated `git bisect`).
+
+## Workflow
+
+### 1. Reproduction Assessment
+First, determine if the bug is reliably reproducible.
+- Ask the user for a reproduction command or script.
+- Attempt to write a minimal unit test or shell script.
+- **CRITICAL**: If you cannot create a **reliable** test (e.g., due to complexity, external deps, or flakiness), **DO NOT** create an unreliable one. Skip `git bisect` and proceed to Step 2 to rely solely on metadata and code analysis.
+
+### 2. Metadata Culprit Search (PR Focus)
+Scan recent history to find the PR that likely caused the issue.
+- **Prioritize finding the Pull Request (PR)**. PRs provide context (why a change was made) that commits lack.
+- **Targeted File Search**: If you have high confidence in which files are involved (e.g., a UI bug in `packages/cli`), search for PRs touching those files first:
+    - Use `git log -n 30 --pretty=format:"%h %s" -- <relevant_files>` to see the most recent commits to those files.
+    - Look for commit messages that mention PR numbers (e.g., "Merge pull request #123" or "feat: ... (#456)").
+- **Broad Search**: If the location is unknown, use `gh pr list --limit 30 --state merged` to get a general list of recent PRs and filter by title/description.
+- Always try to link a suspicious commit back to its originating PR for full context.
+
+### 3. Candidate Selection & Verification
+Your goal is to identify the **PR** that caused the regression.
+
+**Scenario A: You HAVE a reliable test**
+- **Verify**: Checkout the merge commit of a candidate PR and run the test.
+- **Bisect**: If candidates are unclear, use `git bisect` with the test script (see `scripts/bisect_run.sh`).
+
+**Scenario B: You DO NOT have a reliable test**
+- **Manual Analysis**: Read the code diffs of potential candidates. Look for logic changes that match the bug description.
+- **Diff Check**: `git show <commit_hash>` or view the PR diff.
+- **Selection**:
+    - If one PR is a **Strong Match** (obvious logic error matching the bug), select it as the result.
+    - If ambiguous, select the **Top 3 Candidates** based on file relevance and recency.
+- **Constraint**: Do **NOT** run `git bisect` without a reliable test.
+
+### 4. Reporting
+- Present your findings:
+  - **Single Strong Candidate**: If identified.
+  - **Top 3 Candidates**: If the exact cause is uncertain.
+- **Do NOT** automatically attempt to fix, revert, or run further verification unless explicitly asked. The user will decide the next step (e.g., revert locally, investigate further).
+
+## Tips for Efficiency
+- **Limit File Scope**: When scanning metadata, always provide file paths to `git log` to ignore unrelated changes.
+- **Sanity Check**: Always verify the "Good" and "Bad" commits manually before starting a long `bisect run`.
+
+## Common Regression Patterns
+See [references/patterns.md](references/regression-patterns.md) for a guide on interpreting "breaking" changes like state synchronization issues or dependency mismatches.
diff --git a/.gemini/skills/regression-finder/references/regression-patterns.md b/.gemini/skills/regression-finder/references/regression-patterns.md
new file mode 100644
index 0000000000..1c740986bc
--- /dev/null
+++ b/.gemini/skills/regression-finder/references/regression-patterns.md
@@ -0,0 +1,41 @@
+# Common Regression Patterns
+
+When analyzing a "bad" commit, look for these common patterns that often cause
+regressions in this codebase.
+
+## 1. State Synchronization Issues
+
+**Symptoms**: UI doesn't update, stale data, "disappearing" history.
+**Pattern**: State is derived from multiple sources that are updated at
+different times, or state is updated via side-effects (e.g., `useEffect` or
+callbacks) instead of being purely derived. **Fix**: Use `useMemo` to derive
+state or ensure atomic updates.
+
+## 2. Missing Hook Dependencies
+
+**Symptoms**: Logic runs with old state, variables seem stuck. **Pattern**: A
+`useCallback`, `useMemo`, or `useEffect` has an incomplete dependency array.
+**Check**: Look for ESLint suppression comments like
+`// eslint-disable-next-line react-hooks/exhaustive-deps`.
+
+## 3. Bypassed Logic
+
+**Symptoms**: Feature works sometimes but fails in specific paths (e.g., after
+cancellation). **Pattern**: A refactor introduced a new submission or update
+path that bypasses existing validation or cleanup logic. **Check**: Search for
+direct calls to `onSubmit` or `setState` that should have gone through a wrapper
+function (like `handleSubmit`).
+
+## 4. Environment/Platform Specifics
+
+**Symptoms**: Works on MacOS but fails on Windows or Linux. **Pattern**: Usage
+of path delimiters (`/` vs
+``), terminal escape sequences, or platform-specific CLI flags. **Fix**: Use `node:path`
+and verify terminal capability detection.
+
+## 5. Mock Mismatch in Tests
+
+**Symptoms**: Tests pass but application fails (or vice versa). **Pattern**: A
+mock in a unit test was not updated to reflect a change in the real
+implementation's interface or behavioral expectations. **Check**: Verify that
+mock return values and implementations match the current code.
diff --git a/.gemini/skills/regression-finder/scripts/bisect_run.sh b/.gemini/skills/regression-finder/scripts/bisect_run.sh
new file mode 100755
index 0000000000..e055f6b0d7
--- /dev/null
+++ b/.gemini/skills/regression-finder/scripts/bisect_run.sh
@@ -0,0 +1,26 @@
+#!/bin/bash
+
+# bisect_run.sh
+# Automated test script for git bisect run.
+# Expects the test command as the first argument.
+
+TEST_COMMAND=$1
+
+if [ -z "$TEST_COMMAND" ]; then
+  echo "Error: No test command provided."
+  exit 125 # Skip this commit
+fi
+
+echo "Running test: $TEST_COMMAND"
+
+# Execute the test command
+eval "$TEST_COMMAND"
+EXIT_CODE=$?
+
+if [ $EXIT_CODE -eq 0 ]; then
+  echo ">>> COMMIT IS GOOD"
+  exit 0
+else
+  echo ">>> COMMIT IS BAD (Exit Code: $EXIT_CODE)"
+  exit 1
+fi
diff --git a/packages/core/src/config/config.ts b/packages/core/src/config/config.ts
index 6dfc62f322..84cb96a78f 100644
--- a/packages/core/src/config/config.ts
+++ b/packages/core/src/config/config.ts
@@ -56,6 +56,7 @@ import {
   DEFAULT_GEMINI_MODEL,
   DEFAULT_GEMINI_MODEL_AUTO,
   isAutoModel,
+  isGemini2Model,
   isPreviewModel,
   PREVIEW_GEMINI_FLASH_MODEL,
   PREVIEW_GEMINI_MODEL,
@@ -809,9 +810,9 @@ export class Config {
       params.truncateToolOutputThreshold ??
       DEFAULT_TRUNCATE_TOOL_OUTPUT_THRESHOLD;
     // // TODO(joshualitt): Re-evaluate the todo tool for 3 family.
-    this.useWriteTodos = isPreviewModel(this.model)
-      ? false
-      : (params.useWriteTodos ?? true);
+    this.useWriteTodos = isGemini2Model(this.model)
+      ? (params.useWriteTodos ?? true)
+      : false;
     this.enableHooksUI = params.enableHooksUI ?? true;
     this.enableHooks = params.enableHooks ?? true;
     this.disabledHooks = params.disabledHooks ?? [];
diff --git a/regression-finder.skill b/regression-finder.skill
new file mode 100644
index 0000000000..5ac78500db
Binary files /dev/null and b/regression-finder.skill differ