Check

Context Check

Diagnose what is consuming this session's context, then act on it.

Step 1: Measure

python3 "${CLAUDE_PLUGIN_ROOT}/tools/session_cost.py" --session-id "${CLAUDE_SESSION_ID}" --cwd "$(pwd)"

If the session id is unavailable, find the newest transcript for this directory:

BASE="${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
DIR="$BASE/projects/$(pwd | sed 's/[\/.]/-/g')"
LATEST=$(ls -t "$DIR"/*.jsonl 2>/dev/null | head -1)
if [ -n "$LATEST" ]; then
  python3 "${CLAUDE_PLUGIN_ROOT}/tools/session_cost.py" --transcript "$LATEST"
else
  echo "No transcript yet for this directory."
fi

Step 2: Scan the preamble

BASE="${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
python3 -c "
import sys; sys.path.insert(0, '${CLAUDE_PLUGIN_ROOT}/tools')
import scanner, os
base = '$BASE'
for r in scanner.scan_claude_md(os.getcwd()):
    print('%-8s %6d  %s' % (r['verdict'], r['est_tokens'], r['path']))
u = scanner.scan_unused(base + '/skills', base + '/projects')
print('unused skills: %d (~%d tokens)' % (len(u['unused_skills']), u['est_recoverable']))
"

CLAUDE.md sits at position zero and is re-read every turn, so its cost is size x turn count. Anything flagged is the first thing to trim.

Step 3: Choose the cheapest action that fits

The governing relation is cost ~= context size x remaining turns. Content entering context early is multiplied by every turn after it.

• Next task is independent of this one -> /clear. Drops context to zero. Compaction would preserve a summary that is not needed. This is the highest-leverage habit.
• Same task, history is long -> /compact <instructions>. Always instructed; never leave it to auto-compact.
• About to explore broadly -> dispatch a subagent. Reads land in the subagent's context and die there; only the conclusion returns. This flattens the growth curve instead of trimming after the fact.

Step 4: If compacting, say what to keep

Preserve what cannot be re-derived:

• decisions and their rationale
• constraints found empirically ("fzf needs --filter; fails without a TTY")
• current task state: what is done, what remains
• corrections, so a wrong path is not retried

Drop aggressively, starting with file contents. Re-reading files costs about 0.04% of billed tokens, so discarding bodies and re-reading on demand is very nearly free. Keep the paths, discard the text. Also drop tool output already acted upon, and abandoned approaches beyond a one-line note of what failed.

This is deliberately the opposite of what auto-compact tends to preserve.