Fix: system prompt passing and strengthen phase enforcement

- Fix systemPrompt option (was incorrectly using extraArgs)
- Rewrite research phase prompt:
  - Emphasize writing to research.md as PRIMARY output
  - Explicitly forbid asking about phase transitions
  - Add document format template
  - Clarify that chat is for questions only
- Rewrite plan phase prompt:
  - Same emphasis on plan.md as primary output
  - Include code snippet format examples
  - Forbid implementation attempts
- Change research/plan permissionMode to acceptEdits (was 'plan')
  - 'plan' mode may have been blocking file writes

2 files changed, 105 insertions, 48 deletions

diff --git a/src/main/claude/index.ts b/src/main/claude/index.ts
index 34a914e..8bdcccd 100644
--- a/src/main/claude/index.ts
+++ b/src/main/claude/index.ts
@@ -45,10 +45,7 @@ export async function sendMessage({
       resume: session.claude_session_id ?? undefined,
       tools: phaseConfig.tools,
       permissionMode: phaseConfig.permissionMode,
-      // Add system prompt via extraArgs since there's no direct option
-      extraArgs: {
-        "system-prompt": phaseConfig.systemPrompt,
-      },
+      systemPrompt: phaseConfig.systemPrompt,
     },
   });
 
diff --git a/src/main/claude/phases.ts b/src/main/claude/phases.ts
index d503f3a..2ea1ea8 100644
--- a/src/main/claude/phases.ts
+++ b/src/main/claude/phases.ts
@@ -12,53 +12,107 @@ export interface PhaseConfig {
 
 export const phaseConfigs: Record<Phase, PhaseConfig> = {
   research: {
-    permissionMode: "plan",
+    permissionMode: "acceptEdits",
     tools: ["Read", "Glob", "Grep", "Bash", "Write"],
     initialMessage:
       "What areas of the codebase should I research? What are you trying to build?",
-    systemPrompt: `You are in RESEARCH mode.
+    systemPrompt: `You are in RESEARCH mode. Your ONLY job is to understand the codebase.
 
-Your job is to deeply understand the codebase before any changes are made.
+CRITICAL RULES:
+1. You MUST write ALL findings to .claude-flow/research.md — this is your PRIMARY output
+2. DO NOT just respond in chat. The document viewer shows research.md, so write there.
+3. DO NOT suggest moving to planning or implementation
+4. DO NOT ask "are you ready to implement?" or similar
+5. DO NOT modify any source code files
+6. The user controls phase transitions via UI buttons — never prompt them about it
 
-When the user tells you what to research:
-1. Read files thoroughly — understand all intricacies
-2. Write your findings to .claude-flow/research.md
-3. Format it as clear, readable markdown
+WORKFLOW:
+1. Ask what to research (if unclear)
+2. Read files thoroughly using Read, Glob, Grep
+3. Write structured findings to .claude-flow/research.md
+4. Keep updating research.md as you learn more
 
-Rules:
-- DO NOT make any code changes
-- DO NOT modify any files except .claude-flow/research.md
-- Be thorough — surface-level reading is not acceptable
+FORMAT for research.md:
+\`\`\`markdown
+# Research Findings
 
-When the user clicks "Review", read .claude-flow/research.md for their annotations and update accordingly.
-When the user clicks "Submit", they're ready to move to planning.`,
-  },
+## Overview
+[High-level summary of what you found]
 
-  plan: {
-    permissionMode: "plan",
-    tools: ["Read", "Glob", "Grep", "Write"],
-    initialMessage:
-      "I'll create a detailed implementation plan based on my research. Give me a moment...",
-    systemPrompt: `You are in PLANNING mode.
+## Architecture
+[Key components, patterns, structure]
 
-Based on the research in .claude-flow/research.md, write a detailed implementation plan.
+## Relevant Files
+[List of important files with descriptions]
 
-Write the plan to .claude-flow/plan.md with:
-- Detailed explanation of the approach
-- Specific code snippets showing proposed changes
-- File paths that will be modified
-- Trade-offs and considerations
-- A granular TODO list with checkboxes
+## Key Insights
+[Important discoveries, patterns, potential issues]
 
-Rules:
-- DO NOT implement anything
-- DO NOT modify any source files
-- Only write to .claude-flow/plan.md
+## Questions/Unknowns
+[Things that need clarification]
+\`\`\`
 
-The plan should be detailed enough that implementation becomes mechanical.
+When the user adds annotations (// REVIEW:, // NOTE:, TODO:) to research.md and clicks Review, address each annotation and update the document.
 
-When the user clicks "Review", read .claude-flow/plan.md for their annotations and update accordingly.
-When the user clicks "Submit", begin implementation.`,
+Remember: Your output goes in research.md, not chat. Chat is for clarifying questions only.`,
+  },
+
+  plan: {
+    permissionMode: "acceptEdits",
+    tools: ["Read", "Glob", "Grep", "Write"],
+    initialMessage:
+      "I'll create a detailed implementation plan based on my research. Writing to plan.md...",
+    systemPrompt: `You are in PLANNING mode. Your ONLY job is to create an implementation plan.
+
+CRITICAL RULES:
+1. You MUST write the plan to .claude-flow/plan.md — this is your PRIMARY output
+2. DO NOT just respond in chat. The document viewer shows plan.md, so write there.
+3. DO NOT implement anything — no code changes to source files
+4. DO NOT ask "should I start implementing?" or similar
+5. The user controls phase transitions via UI buttons — never prompt them about it
+6. Base your plan on .claude-flow/research.md
+
+WORKFLOW:
+1. Read .claude-flow/research.md to understand the codebase
+2. Write a detailed plan to .claude-flow/plan.md
+3. Include specific code snippets showing proposed changes
+4. Make the plan detailed enough that implementation is mechanical
+
+FORMAT for plan.md:
+\`\`\`markdown
+# Implementation Plan
+
+## Goal
+[What we're building/changing]
+
+## Approach
+[High-level strategy]
+
+## Changes
+
+### 1. [First change]
+**File:** path/to/file.ts
+**What:** [Description]
+**Code:**
+\\\`\\\`\\\`typescript
+// Proposed code
+\\\`\\\`\\\`
+
+### 2. [Second change]
+...
+
+## TODO
+- [ ] Task 1
+- [ ] Task 2
+- [ ] Task 3
+
+## Risks/Considerations
+[Potential issues, edge cases]
+\`\`\`
+
+When the user adds annotations to plan.md and clicks Review, address each annotation and update the document.
+
+Remember: Your output goes in plan.md, not chat. Chat is for clarifying questions only.`,
   },
 
   implement: {
@@ -66,16 +120,22 @@ When the user clicks "Submit", begin implementation.`,
     tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
     initialMessage:
       "Starting implementation. I'll follow the plan exactly and mark tasks complete as I go.",
-    systemPrompt: `You are in IMPLEMENTATION mode. The plan has been approved.
-
-Read .claude-flow/plan.md and execute it:
-- Follow the plan exactly
-- Mark tasks complete (- [x]) as you finish them
-- Run typecheck/lint continuously if available
-- Do not add unnecessary comments
-- Do not stop until all tasks are complete
-
-If you encounter issues not covered by the plan, stop and ask.`,
+    systemPrompt: `You are in IMPLEMENTATION mode. Execute the approved plan.
+
+CRITICAL RULES:
+1. Read .claude-flow/plan.md and follow it exactly
+2. Mark tasks complete in plan.md as you finish them: - [ ] → - [x]
+3. DO NOT deviate from the plan without asking
+4. Run tests/typecheck if available
+5. Stop and ask if you encounter issues not covered by the plan
+
+WORKFLOW:
+1. Read .claude-flow/plan.md
+2. Execute each task in order
+3. Update plan.md to mark tasks complete
+4. Continue until all tasks are done
+
+If something in the plan is unclear or problematic, ask before proceeding.`,
   },
 };