Ten Merged PRs in Two Weeks: 12,279 False Errors and a Hindi Full Stop

· 2 min readoriginal postPost views
Open SourcePythonSystemsNLPDebugging

In the first two weeks of November I got ten pull requests merged across five repositories: six in microsoft/TokenEconomizer, one in deepset-ai/haystack, one in pgalko/BambooAI, one in BistaDinesh03/critique, and one in crewAIInc/crewAI. Seven fixed real runtime or parser bugs; three were small cleanups, including a three-line README link fix. Here is each one at its real size.

Most of my time went into microsoft/TokenEconomizer, Microsoft's standard-library Python runtime that parses JSONL session traces to learn repository environment facts, prune unused tools, and block repeated failure loops. Walking through the trace parsers and loop guard alongside recorded JSONL logs turned up six separate bugs.

The biggest one was PR #152 (fixing issue #151). Successful Edit and Write tool payloads return structured file-diff dictionaries rather than stdout. Because response_text fell back to json.dumps, any edit touching a file that contained words like ValueError tripped the loop guard's error regex. Skipping structured diff dicts took false error hits across 37,797 recorded payloads (964 trace logs) from 12,279 down to 0.

Two earlier loop and trace fixes set that up. In PR #110, TurnGuard._track used a raw regex search instead of status_of_tests, so grep -rn pytest src or git commit -m "fix pytest" reset last_verify_failed back to green, while piped python manage.py test runs were missed. In PR #126, Go, Rust, and RSpec test summaries were skipped during ArgRepair, and multi-block tool_result outputs were flattened onto one line so line-anchored failure patterns stopped matching.

Rounding out TokenEconomizer: PR #127 added the absent_tool micro-skill family for missing tools like ruff with no working fallback (and stopped make build from counting as a repair for ruff, verified with 0 false positives across 12 local repos). PR #153 (fixing issue #144) taught the Windows hook launchers to bypass the WindowsApps Python Store stub that triggers the python/pymanager #396 AppXSvc leak. And PR #154 stopped _apply_codex from discarding _toml_insert's boolean and logging Codex capabilities that were never inserted.

Outside TokenEconomizer, Haystack PR #13646 (fixing issue #13645) fixed Indic sentence splitting. NLTK's Punkt tokenizer only treats ., ?, and ! as sentence endings, so Hindi, Marathi, and Bengali text ending with a Purnaviram danda (। or ॥) came back from DocumentSplitter as one giant chunk. Adding DandaPunktLanguageVars fixed Indic splits across both whitespace modes with all 533 preprocessor tests passing.

The other three were quick maintenance fixes: BambooAI PR #64 replaced four deprecated datetime.utcnow() calls with datetime.now(datetime.UTC) for Python 3.14; critique PR #729 (fixing issue #159) added <lastmod> dates to sitemap URLs in the existing SQL query; and crewAI PR #9303 (fixing issue #9243) fixed three broken README FAQ anchor links after a heading rename.

All ten PRs together add about 560 lines. Keep each PR to one bug, read the callers around the function before patching, and replay ugly real logs before trusting unit tests.