[FEATURE REQUEST] Context tracking & decision registry to prevent approved-feature regression in long sessions

Problem

During a multi-day production development session for a healthcare platform, Claude Code repeatedly lost track of user-approved decisions and regressed working features when making fixes. This resulted in:

• 10+ unnecessary APK rebuilds to fix what was a single-line bug (wrong JSON key name)
• ~6 hours of wasted developer time going back and forth
• Features that were explicitly approved being removed when Claude rewrote files from scratch instead of making incremental edits
• Significant user frustration and loss of trust in the tool

Root Cause

Claude Code does not maintain a persistent decision registry across long conversations. When context grows large:

1. Earlier approvals get "forgotten" or deprioritized
2. Claude defaults to rewriting entire files instead of surgical edits
3. Bug fixes inadvertently remove approved features
4. The user has to re-explain and re-approve the same decisions multiple times

Real Example (anonymized)

1. User approved a screen (ScreenA v1.2.0) with Feature X (multi-item focus with timed rotation)
2. A crash bug appeared (caused by heavy SDKs, not by ScreenA code)
3. Instead of just removing the problematic SDKs, Claude rewrote the entire ScreenA from scratch — losing Feature X and 5 other approved features
4. User pointed out all features were missing
5. Claude rewrote it again — still wrong
6. Finally had to git-restore the original commit verbatim
7. 5 versions wasted on unnecessary rewrites

Proposed Solution

1. Decision Registry

Claude should maintain a machine-readable file tracking:

• What the user approved (with commit hash reference)
• What needs to be changed (specific issues only)
• What must never be modified without explicit re-approval

2. Incremental-Only Edit Mode for Approved Code

When a file has been marked as approved:

• Default to surgical edits (Edit tool), never full rewrites (Write tool)
• Show a diff preview before modifying approved files
• Warn: "This file was approved at commit X. Making targeted fix to line Y only."

3. Regression Detection

Before committing, Claude should:

• Compare new version against approved version
• Flag any removed functionality (deleted functions, removed imports, reduced line count)
• Ask: "I removed [Feature X] which was previously approved. Should I keep it?"

4. Session Continuity Metadata

Store approved decisions in a format that survives context compaction:

• .claude/decisions.json or similar
• Read automatically at session start
• Never overridden by conversation context

Impact

• Cost: Each unnecessary rebuild = ~5 min compute + 10 min developer time
• Trust: Users lose confidence when approved work disappears
• Scale: Problem gets worse as projects grow larger
• Revenue risk: Enterprise customers will not tolerate this level of rework

Environment

• Claude Code (Opus, 1M context)
• Multi-day session with context compaction
• ~30 commits across 3 repos
• Production system