Claude finished @jonathangreen's task in 6m 16s —— View job

Code Review

• Read the changed files (summary.py, tests, identifier.py, build files)
• Analyze the new scoring heuristic for correctness
• Check test coverage and edge cases
• Post review findings

Summary

Clean, well-executed dependency removal. The consensus-vocabulary rewrite is mathematically sound: the coverage score is provably bounded to [0, 1] for any summary (a summary's content words are always a subset of the corpus's, and out-of-corpus words contribute 0 via Counter's __missing__), the zero-denominator case falls back to a neutral score, and the length term is correctly bounded to a ≤5% nudge that only flips genuine near-ties. I traced the four ranking tests and the test_edition.py integration test by hand and they all hold under the new logic; the old API (noun_phrases, blobs, _nltk_installed, parser=) has no remaining references, and the textblob/nltk/joblib/regex/tqdm removal is threaded cleanly through pyproject.toml, uv.lock, the Dockerfile, and tox.ini. The expanded test suite is a real improvement — it also fixes the pre-existing missing assertion in test_non_english_is_penalized. Just one minor tokenization refinement below.

Details

Nit: src/palace/manager/util/summary.py:107-111

Because _word_re ([a-z][a-z']*) keeps a trailing possessive apostrophe, a possessive and its base form tokenize to different content words — "Harry's" → harry's, "Harry" → harry. Character-name possessives are common in book blurbs and are exactly the recurring entities the consensus signal is meant to reward, so this splits their document-frequency (harry df=2 → harry/harry's df=1 each), weakening the very weighting the rewrite is built on. A small normalization that still preserves contractions:

for word in SummaryEvaluator._word_re.findall(text.lower()):
    word = word.removesuffix("'s").rstrip("'")  # "harry's" -> "harry", "cats'" -> "cats"; "don't" unchanged
    if len(word) > 2 and word not in ENGLISH_STOPWORDS:
        ...

Greptile Summary

This PR removes the textblob and nltk dependencies (along with their transitive joblib, regex, and tqdm packages) by replacing the single point of use — SummaryEvaluator — with a pure-Python heuristic that scores candidate descriptions by how well they cover the work's consensus vocabulary (content words weighted by document-frequency²).

• Core algorithm change (summary.py): noun-phrase extraction via TextBlob is replaced with a bag-of-words coverage score computed over ENGLISH_STOPWORDS-filtered tokens; NLTK sentence tokenization is replaced with a regex heuristic; a bounded length nudge (≤5%) is added as an explicit tiebreaker.
• Infrastructure cleanup (Dockerfile.baseimage, tox.ini, uv.lock): corpus-download steps are removed and the five removed packages fall out of the lock file, shrinking image build time and eliminating the NLTK path-traversal CVE from the dependency surface.
• Test suite (test_summary_evaluator.py): 7 new behavioural tests added, TextBlob mock tests removed, and a previously silent missing assertion in test_non_english_is_penalized is now actually checked.

Confidence Score: 5/5

Safe to merge; the change only affects which description is displayed when a work has multiple candidates, and the new algorithm is well-tested with no correctness concerns.

The algorithmic replacement is logically sound: coverage is always in [0, 1], the df² weighting naturally elevates recurring words, the length nudge is provably bounded, and the zero-content-word edge case is handled explicitly. The test suite exercises all meaningful paths including the previously untested non-English penalty assertion. Dependency removal is clean across pyproject.toml, uv.lock, Dockerfile, and tox.ini with no residual references.

No files require special attention.

Important Files Changed

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["add(summary)"] --> B["_content_words(summary)\n(regex tokenize → lowercase → filter stopwords + len≤2)"]
    B --> C["word_sets[summary] = words"]
    B --> D["content_words[word] += 1\n(document frequency counter)"]
    E["ready()"] --> F["total_word_weight = Σ count²\n(sum of df-squared weights)"]
    G["score(summary)"] --> H{total_word_weight > 0?}
    H -- Yes --> I["covered = Σ count[w]²\nfor w in summary words"]
    I --> J["coverage_score = covered / total_word_weight"]
    H -- No --> K["score = 1.0 (neutral)"]
    J --> L["Apply sentence-length penalty\n÷ off_from_optimal^1.5"]
    K --> L
    L --> M["Apply bad-phrase penalties\n× 0.5^bad_count"]
    M --> N["Apply language penalty\n× 0.5^(difference-1)"]
    N --> O["Apply length nudge\n× (1 - nudge × saturation)"]
    O --> P["Return final score"]

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["add(summary)"] --> B["_content_words(summary)\n(regex tokenize → lowercase → filter stopwords + len≤2)"]
    B --> C["word_sets[summary] = words"]
    B --> D["content_words[word] += 1\n(document frequency counter)"]
    E["ready()"] --> F["total_word_weight = Σ count²\n(sum of df-squared weights)"]
    G["score(summary)"] --> H{total_word_weight > 0?}
    H -- Yes --> I["covered = Σ count[w]²\nfor w in summary words"]
    I --> J["coverage_score = covered / total_word_weight"]
    H -- No --> K["score = 1.0 (neutral)"]
    J --> L["Apply sentence-length penalty\n÷ off_from_optimal^1.5"]
    K --> L
    L --> M["Apply bad-phrase penalties\n× 0.5^bad_count"]
    M --> N["Apply language penalty\n× 0.5^(difference-1)"]
    N --> O["Apply length nudge\n× (1 - nudge × saturation)"]
    O --> P["Return final score"]

_{Reviews (4): Last reviewed commit: "Make generic-copy test literally cover n..." | Re-trigger Greptile}

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.50%. Comparing base (87d1bfa) to head (8332c98).
⚠️ Report is 8 commits behind head on main.

@@            Coverage Diff             @@
##             main    #3571      +/-   ##
==========================================
+ Coverage   93.47%   93.50%   +0.03%     
==========================================
  Files         512      512              
  Lines       46619    46640      +21     
  Branches     6353     6358       +5     
==========================================
+ Hits        43575    43612      +37     
+ Misses       1968     1959       -9     
+ Partials     1076     1069       -7     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.