Agent WAL (Write-Ahead Log)

Write important state to disk before responding. Prevents the #1 agent failure mode: losing corrections and context during compaction.

Core Rule

Write before you respond. If something is worth remembering, WAL it first.

When to WAL

Trigger	Action Type	Example
User corrects you	INLINECODE0	"No, use Podman not Docker"
You make a key decision

decision | "Using CogVideoX-2B for text-to-video" | | Important analysis/conclusion | analysis | "WAL/VFM patterns should be core infra not skills" | | State change | state_change | "GPU server SSH key auth configured" | | User says "remember this" | correction | Whatever they said |

Commands

All commands via scripts/wal.py (relative to this skill directory):

CODEBLOCK0

Integration Points

On Session Start

1. 1. Run replay to get unapplied entries
2. Read the summary into your context
3. Mark entries as applied after incorporating them

On User Correction

1. 1. Run append with action_type correction BEFORE responding
2. Then respond with the corrected behavior

On Pre-Compaction Flush

1. 1. Run flush-buffer to persist any buffered entries
2. Then write to daily memory files as usual

During Conversation

For less critical items, use buffer-add to batch writes. Buffer is flushed to WAL on flush-buffer (called during pre-compaction) or manually.

Storage

WAL files: ~/clawd/memory/wal/<agent_id>.wal.jsonl
Buffer files: INLINECODE13

Entries are append-only JSONL. Each entry:
CODEBLOCK1