Conversation log

Channel messages are written to disk by the harness, not by the model. The model cannot save memory at session-end (it is no longer running) or after /compact (the content is already gone), so two hooks capture the raw conversation deterministically before either event destroys it.

What it guarantees

Every inbound/outbound channel message lands in an append-only raw layer (memory/conversations/YYYY-MM-DD.jsonl) before the session is garbage-collected, and everything since the last compact marker is captured into the daily log before /compact compresses the context away. Raw data never dies; summaries can be rebuilt from it later.

How it works

• SessionEnd → session-end-flush.py runs extract-conversations.py --today, which scans the transcript JSONL for inbound <channel> messages and outbound __reply tool calls and appends them as structured JSONL ({ts, dir, who, text, message_id}), deduplicated on message id / text.
• PreCompact → pre-compact-flush.py finds the last isCompactSummary entry in the transcript and appends every user/assistant message after it to memory/daily/YYYY-MM-DD.md as a verbatim "Compact Extraction" section. No LLM summarization — just cheap, deterministic capture.

Implement in your project

Reference implementations:

• ../../scripts/hooks/session-end-flush.py
• ../../scripts/hooks/pre-compact-flush.py
• ../../scripts/extract-conversations.py

Wire both events (see ../../.claude/settings.json):

"SessionEnd": [
  { "matcher": "", "hooks": [
    { "type": "command", "command": "EVI_PROJECT_DIR=$CLAUDE_PROJECT_DIR python3 $CLAUDE_PROJECT_DIR/scripts/hooks/session-end-flush.py", "timeout": 90 }
  ]}
],
"PreCompact": [
  { "matcher": "", "hooks": [
    { "type": "command", "command": "EVI_PROJECT_DIR=$CLAUDE_PROJECT_DIR CLAUDE_PROJECT_DIR=$CLAUDE_PROJECT_DIR python3 $CLAUDE_PROJECT_DIR/scripts/hooks/pre-compact-flush.py" }
  ]}
]

extract-conversations.py is configured via env (EVI_CHANNEL_SOURCE, EVI_REPLY_TOOL_SUFFIX, EVI_AGENT_NAME, EVI_OWNER_NAME, EVI_TZ, EVI_TRANSCRIPT_DIRS) — see ../../.env.example.

Adaptation points

• Different channel: set EVI_CHANNEL_SOURCE to your channel keyword and EVI_REPLY_TOOL_SUFFIX to your reply tool's name suffix; the extractor's regex adapts from those env values.
• Different transcript location: point EVI_TRANSCRIPT_DIRS at your harness's session log directory (colon-separated). Default scans all ~/.claude/projects/ subdirs.
• No PreCompact event in your harness: keep SessionEnd alone — you lose mid-session compaction capture but still get end-of-session flush.

Unconfigured = silent no-op. Both hooks return immediately when there is no .env in the project dir, so a fresh clone never scans a transcript or writes to memory/. The behavior switches on once you create your .env.

Verify

1. python3 -m py_compile scripts/hooks/session-end-flush.py scripts/hooks/pre-compact-flush.py scripts/extract-conversations.py exits 0.
2. echo '{}' | python3 scripts/hooks/session-end-flush.py and the same for pre-compact-flush exit 0 (no transcript → graceful skip).
3. End a real session → a file appears/grows in memory/conversations/.

Rollback

Remove the SessionEnd and PreCompact entries from .claude/settings.json.