analyze

Fix Claude -- Analyze

Analyze an existing CLAUDE.md file against the 7 structural bottlenecks discovered in the Claude Code source code leak. Produce a gap report explaining what each issue is, why it matters, and whether the current directives mitigate it.

Workflow

Step 1: Resolve and Read Target

TARGET="CLAUDE.md"
if [ -L "$TARGET" ]; then
  TARGET=$(readlink -f "$TARGET")
  echo "Resolved symlink to: $TARGET"
fi

Read the full target file. If it doesn't exist, tell the user to run fixclaude:init first and stop.

Step 2: Load the Findings

Read the source leak findings reference:

${CLAUDE_PLUGIN_ROOT}/skills/analyze/references/source-leak-findings.md

This contains all 7 findings with their source locations, problems, check criteria, and overrides.

Step 3: Analyze Each Finding

For each of the 7 findings, evaluate the existing CLAUDE.md using the "What to check" criteria from the reference. Rate each finding:

• MITIGATED -- The CLAUDE.md contains a directive that fully addresses this issue
• PARTIAL -- The CLAUDE.md touches on this but is missing key aspects
• UNADDRESSED -- The CLAUDE.md does not address this issue at all

Step 4: Produce the Report

Present the analysis inline (never hide behind a file path). Use this format:

# CLAUDE.md Gap Analysis

Analyzed: <file path>
Lines: <line count>
Date: <today>

## Summary

<X>/7 findings mitigated | <Y>/7 partial | <Z>/7 unaddressed

## Findings

### 1. Employee-Only Verification Gate
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** Claude Code's success metric for file writes is: did bytes hit
disk? Anthropic's internal verification loop (compile check, linter, test suite)
is gated behind `process.env.USER_TYPE === 'ant'`. Their own docs cite a
29-30% false-claims rate.

**Source:** `services/tools/toolExecution.ts`

**Your coverage:** <Explain what the CLAUDE.md does or doesn't do. Quote
relevant lines if present. Be specific about what's missing.>

**Recommended override:** <If not fully mitigated, provide the specific
directive text they should add.>

---

### 2. Context Death Spiral
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** Auto-compaction fires at ~167K tokens. Keeps 5 files (5K tokens
each), compresses everything else into a 50K-token summary. Every file read,
reasoning chain, and intermediate decision is destroyed. Dead code in the
project accelerates this.

**Source:** `services/compact/autoCompact.ts`

**Your coverage:** <...>

**Recommended override:** <...>

---

### 3. The Brevity Mandate
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** System prompt contains: "Try the simplest approach first,"
"Don't refactor code beyond what was asked," "Three similar lines of code is
better than a premature abstraction." These fight user intent when structural
fixes are needed.

**Source:** `constants/prompts.ts`

**Your coverage:** <...>

**Recommended override:** <...>

---

### 4. The Agent Swarm Nobody Told You About
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** Sub-agents run in isolated AsyncLocalStorage with own memory,
compaction cycle, and token budget. No MAX_WORKERS limit. One agent = ~167K
tokens. Five parallel agents = 835K. Users default to single-agent sequential
processing.

**Source:** `utils/agentContext.ts`

**Your coverage:** <...>

**Recommended override:** <...>

---

### 5. The 2,000-Line Blind Spot
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** File reads hard-capped at 2,000 lines / 25,000 tokens.
Everything past that is silently truncated. The agent hallucinates the rest.

**Source:** `tools/FileReadTool/limits.ts`

**Your coverage:** <...>

**Recommended override:** <...>

---

### 6. Tool Result Blindness
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** Tool results exceeding 50,000 characters are persisted to disk
and replaced with a 2,000-byte preview. The agent works from the preview and
doesn't know results were truncated.

**Source:** `utils/toolResultStorage.ts`

**Your coverage:** <...>

**Recommended override:** <...>

---

### 7. grep Is Not an AST
**Status: <MITIGATED|PARTIAL|UNADDRESSED>**

**The issue:** GrepTool is raw text matching. It can't distinguish function
calls from comments or differentiate identically named imports from different
modules. Renames that rely on grep miss dynamic imports, re-exports, and
string references.

**Source:** GrepTool implementation

**Your coverage:** <...>

**Recommended override:** <...>

Step 5: Recommend Next Steps

Based on the gap count:

• 0 unaddressed: "Your CLAUDE.md covers all known issues. Run fixclaude:update if you want to strengthen any partial mitigations."
• 1-3 unaddressed: "Run fixclaude:update to add the missing directives."
• 4+ unaddressed: "Run fixclaude:update for a comprehensive upgrade, or fixclaude:init to start fresh with the full template (backs up your existing file first)."

Important Rules

• Always show the full report inline. Never write it to a file and point to it.
• Quote specific lines from the user's CLAUDE.md when rating coverage.
• Be honest about partial coverage -- don't inflate ratings.
• The analysis must be useful to someone who has never heard of the source leak. Each finding section should be self-contained and educational.