feat(agent): add plan mode, plan resumption, and context overflow fixes

## Plan Mode & Planning Workflow
- Add PlanMode state (Standard/Planning) with Shift+Tab toggle
- Add plan tools: plan_create, plan_next, plan_update, plan_list
- Add plan_menu.rs for interactive post-plan action selection
  - Execute with auto-accept changes
  - Execute with review for each change
  - Modify plan with feedback
- Add plan resumption on startup - shows incomplete plans and suggests "continue"
- Add /plans command to list incomplete plans with progress info
- Update prompts with plan_execution_protocol explaining resumable plans

## Context Overflow Fix ("Input Too Long" Error)
- Fixed critical bug: was compacting empty conversation_history instead of raw_chat_history
- Add estimate_raw_history_tokens() to estimate tokens from rig::Message directly
- Truncate raw_chat_history directly when context too large (10→6→4 messages progressively)
- Add pre-request check to detect large context before hitting API limits
- Add find_most_recent_plan_file() for reliable plan path detection

## Auto-Accept Mode Improvements
- Show full code diff in auto-accept mode (same as confirmation prompts)
- Use render_diff() for existing files, render_new_file() for new files
- Apply to both WriteFileTool and WriteFilesTool

## Rate Limiting Mitigation
- Reduce extended thinking budget from 16000 to 8000 tokens
- Helps avoid API rate limits during heavy tool-use sessions

## New Modules
- src/agent/tools/plan.rs - Plan creation and execution tools
- src/agent/tools/truncation.rs - Output truncation utilities
- src/agent/tools/diagnostics.rs - Diagnostic tools
- src/agent/ui/plan_menu.rs - Plan action menu UI

## Vendor Changes
- Move rig-bedrock from patches/ to vendor/ directory
- Clean up obsolete patch files

## Prompts Updates
- Add is_plan_continuation_query() to detect "continue" requests
- Update plan_execution_protocol with resumption instructions
- Explain that plan_next automatically finds where you left off

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: 2 people

.gitignore

@@ -16,6 +16,7 @@ node_modules/
 
 # MSVC Windows builds of rustc generate these, which store debugging information
 *.pdb
+docs/**
 
 # Generated by cargo mutants
 # Contains mutation testing data

Cargo.lock

@@ -73,7 +73,7 @@ futures-util = "0.3"
 
 # Agent dependencies (using Rig - LLM application framework)
 rig-core = { version = "0.27", features = ["derive"] }
-rig-bedrock = "0.3"  # AWS Bedrock provider for Rig
+rig-bedrock = { path = "vendor/rig-bedrock" }  # Vendored with fix for extended thinking + tool calls
 
 # Diff rendering for file confirmation UI
 similar = "2.6"
@@ -104,10 +104,4 @@ path = "examples/check_vulnerabilities.rs"
 
 [[example]]
 name = "security_analysis"
-path = "examples/security_analysis.rs"
-
-# Patch rig-bedrock to fix extended thinking support with tool use
-# The upstream version drops Reasoning blocks when a tool call is present.
-# See: patches/rig-bedrock/src/types/assistant_content.rs for the fix.
-[patch.crates-io]
-rig-bedrock = { path = "patches/rig-bedrock" }
+path = "examples/security_analysis.rs"