Good rebase — this feature has been wanted for a while and the implementation is clean. A few observations:

What works well

• The 70% word-overlap dedup threshold is reasonable for catching near-duplicate drawer summaries
• 2000-char cap is a sensible default for a "catch me up" context injection
• Dispatching to L2 for the underlying retrieval is correct (fast, no embedding cost)
• CLI integration follows the existing pattern exactly

One correctness concern — dedup math

The overlap check:

if len(words & prev_words) > 0.7 * max(len(words), len(prev_words), 1):

This is asymmetric in a way that can produce false positives. A short line (3 words) that shares 3 words with a long line (20 words) will pass because 3 > 0.7 * 3, even though the long line has lots of unique content. Consider Jaccard similarity instead:

intersection = len(words & prev_words)
union = len(words | prev_words)
if union > 0 and intersection / union > 0.6:  # Jaccard threshold

Or keep the current approach but use min() instead of max():

if len(words & prev_words) > 0.7 * min(len(words), len(prev_words), 1):

Minor: parsing L2 output

brief() parses the string output of l2.retrieve() to get individual lines — this creates a text→parse→text pipeline that's fragile if L2's output format changes. If l2.retrieve() can return raw drawer data (or if there's a retrieve_raw() method), using that would be more robust. Not a blocker, just worth considering for the long term.

Tests

test_brief checks the output contains "Brief" and "topics" — that's the minimum needed to pass. A test for the dedup behavior (e.g., two near-identical drawers → only one in output) would give more confidence. Also worth a test for the empty-palace case since brief() has a special code path for raw.startswith("No ").

568 tests passing is solid. LGTM with the minor dedup note — would be comfortable approving after the math is verified.

All feedback addressed — dedup math confirmed correct (max() is intentional and right), test gaps filled with two new tests, L2 string parsing noted for follow-up PR. 570 passing.