Log Rotation

log-rotation

Add a richer memory summary for a rotated Lazar stream archive.

The kernel auto-rotates stream.jsonl when it exceeds LAZAR_LOG_MAX_BYTES (default 10MB). That is the safety floor. The kernel owns logs/: sandboxed tools and hooks may read logs, but must not write, compress, prune, or repair them.

The kernel writes a minimal summary into memory/log-summaries/<ts>.md whenever it rotates. This skill is the polish layer on top: read one archive with bounded commands and append useful human navigation details to memory.

When to use

• After kernel auto-rotation, if you want richer summaries than the minimal kernel header.
• When a memory/log-summaries/<ts>.md file exists but needs top prompts, top commands, or a small file-touch index.
• The user explicitly asks "summarize the last session" or similar.

Do NOT run this on every prompt. Once per archive is plenty.

Hard rules

• logs/ is trusted kernel-owned state. Do not mv, gzip, rm, truncate, or write anything under $LAZAR_HOME/logs from a sandboxed skill or hook.
• Rotation itself is kernel-owned and non-destructive at the moment it happens: the live log moves to .bak, and the next append creates a fresh stream.jsonl.
• Full-fidelity archive retention is kernel policy, controlled by LAZAR_LOG_ARCHIVE_KEEP. Summaries in memory/log-summaries/ are the durable navigation layer.
• Write enriched summaries only under $LAZAR_MEMORY, usually memory/log-summaries/ or memory/distilled/.
• Always bound archive reads. Never cat a whole archive.

Recipes

Check current stream size

LOG=$LAZAR_HOME/logs/stream.jsonl
[ -f "$LOG" ] && wc -c "$LOG"

Pick the newest rotated archive

ARCHIVE=$(ls -t $LAZAR_HOME/logs/stream.jsonl.*.bak 2>/dev/null | head -n 1)
[ -n "$ARCHIVE" ] || { echo "no raw archive found"; exit 0; }

If an archive has already been pruned by the kernel, use the matching memory/log-summaries/<ts>.md summary as the durable record.

Enrich the summary for one archive

ARCHIVE=$LAZAR_HOME/logs/stream.jsonl.<TS>.bak
TS=$(basename "$ARCHIVE" .bak | sed 's/^stream\.jsonl\.//')
SUM=$LAZAR_HOME/memory/log-summaries/${TS}.md
mkdir -p "$LAZAR_HOME/memory/log-summaries"

{
  echo ""
  echo "## User prompts (last 30)"
  echo ""
  jq -r 'select(.kind == "user") | .content' "$ARCHIVE" 2>/dev/null \
    | tail -n 30 \
    | sed 's/^/- /' \
    || true

  echo ""
  echo "## Tool commands (top 20 by frequency)"
  echo ""
  jq -r 'select(.kind == "tool_result") | .command' "$ARCHIVE" 2>/dev/null \
    | awk '{print $1}' \
    | sort | uniq -c | sort -rn \
    | head -n 20 \
    | sed 's/^/    /' \
    || true
} >> "$SUM"

echo "summary: $SUM"

Pairing with load-context and archive-search

The three tiers of memory:

• L1: $LAZAR_HOME/logs/stream.jsonl (current, raw) — bounded tail/rg via _meta/load-context.
• L2: $LAZAR_HOME/memory/log-summaries/<ts>.md — minimal entries written by the kernel, optionally enriched by this skill.
• L3: $LAZAR_HOME/logs/stream.jsonl.*.bak (recent full-fidelity archives) — searched via _meta/archive-search only when L1+L2 don't have what's needed. Kernel retention may prune old L3 archives.

Never cat an archive. Always use rg, jq, head, or tail with explicit bounds.

Principle

The kernel records and rotates. Skills interpret and summarize. Keep it that way: no log mutation from sandboxed skills.