Feature request: Auditor session mode — parallel session that critically analyzes another session's work

Feature Request

An auditor session mode where two Claude Code sessions work in tandem:

• Worker session: executes the task (writing code, running skills, generating content)
• Auditor session: watches the worker's actions in real-time, critically analyzes decisions, and either generates a report or sends live feedback to the worker session

How it would work

1. Start a worker session normally: claude
2. Start an auditor session attached to it: claude --audit <session-id> (or similar)
3. The auditor session receives a live stream of the worker's tool calls, outputs, and decisions
4. The auditor can:

• Passive mode: observe silently and produce a final quality report (mistakes, missed edge cases, better approaches, security issues)
• Active mode: inject messages/warnings into the worker session on the fly ("you're about to overwrite a file that has uncommitted changes", "this approach has an N+1 query problem", "the skill output doesn't match the spec")

Best use case: Skill analysis

When a skill runs a long autonomous pipeline across many phases, there's no built-in quality gate. The skill generates dozens of files and makes hundreds of decisions. An auditor session could:

• Verify generated code actually compiles/renders
• Check that content matches the spec/brief at each phase
• Detect hallucinated data, broken imports, missing files
• Flag when the skill diverges from its own SKILL.md instructions
• Catch repetitive/generic output that needs a design pass
• Produce a structured report at the end: what was good, what needs rework, severity levels

Other use cases

• Code review on the fly: auditor catches bugs, security issues, or style violations as code is written, before commit
• Pair programming: one session writes, the other pressure-tests assumptions
• Compliance/security audit: auditor checks for OWASP issues, secrets in code, unsafe patterns in real-time
• Teaching/learning: auditor explains why the worker made each decision

Why this matters

Right now, quality control is manual — you wait for the session to finish, then review everything yourself. For long-running autonomous skills that generate many files, this doesn't scale. An auditor session would catch issues while they're cheap to fix (mid-generation) instead of after the full pipeline has run.

Suggested UX

# Start a task
claude "build the project"
# Session ID shown: session_abc123

# In another terminal, attach auditor
claude --audit session_abc123

# Or with specific focus
claude --audit session_abc123 --focus "security,code-quality,spec-compliance"

# Auditor output modes
claude --audit session_abc123 --mode passive    # Report at the end
claude --audit session_abc123 --mode active     # Live feedback to worker