/arbiter-compact

When user invokes /arbiter-compact:

PHASE 1: EXTRACTION

Analyze the current conversation and extract all factual knowledge in ARBITER SYNTAX.

Follow .claude/rules/arbiter-syntax.md strictly:

• Use snake_case identifiers only
• One statement per line
• Operators: & (and), | (or), ! (not), -> (implies), <-> (iff)
• NO prose, NO explanations, ONLY logic statements
• Group with # comments for readability

Categories to Extract

Extract facts across these dimensions:

1. Project Setup

# Technologies/frameworks in use
use_<technology>
requires_<dependency>
version_constraint_<tool>

2. Architecture Decisions

# System design choices
api_is_<type>
database_is_<type>
auth_is_<method>
pattern_is_<pattern>

3. Coding Conventions

# Style and patterns
use_type_hints
use_<naming_convention>
require_<practice>

4. Access Control / Business Logic

# Permission rules
<condition> -> <capability>
<role> & <status> -> <permission>

5. API Contracts

# Endpoint behaviors
endpoint_<name>_returns_<format>
endpoint_<name>_requires_<condition>

6. Dependencies / Tooling

# Available tools and configs
has_<tool>
<tool>_<config>_enabled

7. File Locations / Structure

# Where things live (if relevant to future work)
tests_in_tests_dir
config_in_<location>

8. Decisions Made

# Explicit choices during conversation
decided_<decision>
rejected_<alternative>

PHASE 2: VALIDATION & COMPRESSION

1. Write extracted facts to /tmp/arbiter_extracted.txt

2. Run arbiter compression:

   python3 .claude/tools/arbiter.py /tmp/arbiter_extracted.txt > /tmp/arbiter_compressed.txt
   

3. Compare results:

   • Show original fact count
   • Show compressed fact count
   • Show compression ratio
   • Highlight any contradictions detected

PHASE 3: OUTPUT

Present to user:

## Extracted Facts (N statements)
[show /tmp/arbiter_extracted.txt contents]

## Compressed Facts (M statements, X% reduction)
[show /tmp/arbiter_compressed.txt contents]

## Analysis
- Tautologies removed: [count]
- Redundancies removed: [count]
- Contradictions detected: [list if any]

## Next Steps
You can now use this compressed fact set as context seed.
To reset conversation with these facts, manually use:
  /compact focus=<paste compressed facts>

EXAMPLE OUTPUT

Given conversation about setting up a FastAPI project with pytest:

Extracted (8 statements):

use_python
use_fastapi
use_pytest
use_sqlalchemy
database_is_postgres
authenticated -> can_read
admin & authenticated -> can_write
admin -> authenticated

Compressed (7 statements, 12.5% reduction):

use_python
use_fastapi
use_pytest
use_sqlalchemy
database_is_postgres
authenticated -> can_read
admin & authenticated -> can_write

Analysis:

• Redundancies removed: 1
  • admin -> authenticated is implied by admin & authenticated -> can_write

IMPORTANT NOTES

• Be comprehensive: Extract ALL decisions, conventions, constraints discussed
• Be precise: Use exact arbiter syntax (see arbiter-syntax.md)
• Be minimal: Arbiter compression will remove redundancies, but start clean
• Check yourself: Before running arbiter, verify your syntax is valid

WORKFLOWS

Manual Workflow (Current)

Use /arbiter-compact when you want to manually compress conversation context:

1. Claude extracts facts in arbiter syntax
2. Facts are validated and compressed
3. User sees results and can manually use /compact focus=<facts>

Automatic Workflow (Available via Hook)

To enable automatic compression on every context compaction:

1. Enable PreCompact hook in .claude/settings.json:

{
  "hooks": {
    "PreCompact": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "prompt",
            "prompt": "Extract conversation facts in arbiter syntax per .claude/rules/arbiter-syntax.md. Output ONLY logical statements, no prose."
          }
        ]
      }
    ]
  }
}

2. How it works:

   • When Claude Code hits context limit → PreCompact hook fires
   • Claude automatically extracts facts in arbiter syntax
   • Facts are compressed via arbiter.py
   • Compressed facts become the new context seed
   • Conversation continues with ~95% token reduction

3. Benefits:

   • Zero manual intervention
   • Mathematically precise compression
   • No information loss (only redundancy removal)
   • Seamless long conversations