Introspection

Activity logs and session analysis for the rhi ecosystem.

Log Structure

• log/ — weekly snapshots, named by end date (e.g. 2026-02-25.md). Read the most recent first when evaluating direction or focus.
• log/daily/ — daily session summaries, YYYY-MM-DD.md, one day across all projects.
• log/synthesis-*.md — cross-cutting pattern analysis over a date range.

Check these before asking "what should we work on?" or "what were we focused on?"

Updating Daily Logs

Run via a subagent (Sonnet, general-purpose). The main session cannot execute these commands directly.

1. Backup sessions:

   rsync -a --update ~/.claude/projects/ /mnt/ssd/ai/claude-sessions/projects/ && rsync -a --update ~/.claude/history.jsonl /mnt/ssd/ai/claude-sessions/history.jsonl && rsync -a --update ~/.claude/usage-data/ /mnt/ssd/ai/claude-sessions/usage-data/

   --update skips destination files newer than source (safe for incremental runs). usage-data/ holds pre-computed /insights facets — incremental, worth preserving.

   Also archive /insights reports into the append-only history dir (run every backup pass, right after the rsync above):

   mkdir -p /mnt/ssd/ai/claude-sessions/insights-history/ && cp -n ~/.claude/usage-data/report-*.html /mnt/ssd/ai/claude-sessions/insights-history/

   /insights reports are otherwise ephemeral — the source dir only retains the latest timestamped file. cp -n (no-clobber) into the sibling insights-history/ dir builds an append-only archive that the --update mirror of usage-data/ cannot clobber. The report-*.html glob intentionally excludes the report.html alias (no dash), so only unique timestamped reports are archived.

2. Find missing days: list docs/automated-introspection/log/daily/ and diff against today.

3. Spawn haiku agents in parallel, one per missing day. Never invoke normalize directly for this — go through the private-project exclusion wrapper instead, so excluded projects' sessions are never read into the generating agent's context in the first place (not filtered/redacted after generation):

   CLAUDE_SESSIONS_DIR=/mnt/ssd/ai/claude-sessions/projects tooling/normalize-excluding-private.sh sessions messages --all-projects --role user --since YYYY-MM-DD --until YYYY-MM-DD+1 --limit 0 --show-usage

   Each writes to docs/automated-introspection/log/daily/YYYY-MM-DD.md. Quiet days: note as such. Include ## Token Usage with per-session output tokens and cache hit ratios.

   Private-project exclusion. tooling/normalize-excluding-private.sh wraps the normalize binary: it reads directory basenames from .git/info/private-names (machine-local, gitignored-by-design — see that file for the current list), resolves each to its mangled ~/.claude/projects/-style dir name, and builds a scratch dir of symlinks to every other project before calling normalize with CLAUDE_SESSIONS_DIR pointed at that scratch dir. --all-projects then only ever sees the non-excluded dirs — normalize has no idea the private ones exist, so the agent generating the log never sees their content either. This is the default; only bypass it (call normalize directly) for known one-off ecosystem-only investigations where you've separately confirmed no private project is in scope.

   Effect on daily logs: a day where a private project is the only activity gets no mention of it at all — same as if nothing happened that day (no "1 private session" placeholder, no count that includes it). A day with both ecosystem and private-project activity still gets a log, but every count/list in it is built only from the non-excluded sessions — the private activity is fully absent, not summarized down.

   Before writing the draft to disk, pipe it through the hash guard: echo "$draft" | tooling/check-private-name-hashes.js - (or tooling/check-private-name-hashes.js - fed the draft on stdin). Exit 0 means clean; exit 1 prints the offending token(s) — redact and recheck before writing the file. This catches protected names pre-write, ahead of (and independent of) the commit-time hook check in .githooks/pre-commit. It stays in place as a second layer even with the exclusion wrapper — the wrapper stops private sessions from being read, the hash guard catches private names if they get typed/pasted into a draft some other way.

   If synthesis insights feel thin: re-run agents on existing logs with --show-usage output, instructing them to flag token outliers (debugging churn, cold-start cache inefficiency, architectural output spikes). Then re-run opus synthesis.

4. Add new days to sidebar in docs/.vitepress/config.ts under Daily Logs.

5. If a week or more of new days: spawn an opus agent to read all daily logs and write/update docs/automated-introspection/log/synthesis-<start>-<end>.md. Tell it CLAUDE.md conventions may have evolved over the period. Same as step 3: go through tooling/normalize-excluding-private.sh for any session re-querying, and pipe the draft through tooling/check-private-name-hashes.js - before writing to disk.

Scope note: this exclusion procedure applies going forward only. Daily logs already published before it was adopted are not being retroactively regenerated under it.

6. Commit and push.

Session Data

Claude Code deletes session .jsonl files based on cleanupPeriodDays in ~/.claude/settings.json (default 30). Currently 999999 to prevent deletion. Cannot use 0 — bug #23710 silently disables transcript persistence.

Backup location: /mnt/ssd/ai/claude-sessions/.

Before any session analysis (run via a subagent):

1. Re-backup:

   rsync -a --update ~/.claude/projects/ /mnt/ssd/ai/claude-sessions/projects/ && rsync -a --update ~/.claude/history.jsonl /mnt/ssd/ai/claude-sessions/history.jsonl && rsync -a --update ~/.claude/usage-data/ /mnt/ssd/ai/claude-sessions/usage-data/
   mkdir -p /mnt/ssd/ai/claude-sessions/insights-history/ && cp -n ~/.claude/usage-data/report-*.html /mnt/ssd/ai/claude-sessions/insights-history/

2. Run with CLAUDE_SESSIONS_DIR=/mnt/ssd/ai/claude-sessions/projects prefix, not from ~/.claude/.

Session analysis (same private-project exclusion as daily logs — see above):

CLAUDE_SESSIONS_DIR=/mnt/ssd/ai/claude-sessions/projects tooling/normalize-excluding-private.sh sessions stats --all-projects --limit 0 --group-by project,day --since YYYY-MM-DD --until YYYY-MM-DD --compact

Structural exploration:

~/git/rhizone/normalize/target/debug/normalize view <file>     # structural outline with line numbers
~/git/rhizone/normalize/target/debug/normalize view <dir>      # directory structure