Bug report

Bug description:

A two-pass review of Lib/idlelib turned up 69 findings — 37 FIX (a wrong result, silent data loss, or an unreachable feature) and 32 CONSIDER. Every one carries a file:line, a concrete failure scenario, and — where idlelib already handles the same situation correctly somewhere else — the guarded twin that shows it is a defect by the project's own standard and suggests the fix.

I'm filing them under one umbrella so they can be picked off individually rather than flooding the tracker.

Please don't start work on a row until a maintainer has cleared it. These are a reviewer's findings, not triaged bugs: some will be intentional behaviour, some not worth the churn, and some wrong. Every row starts at Under analysis, which means not yet cleared — do not work on it. A maintainer moving a row off Under analysis is what opens it up; a PR against an uncleared row may be wasted effort, and several of these touch code where the current behaviour is load-bearing.

To take a cleared one: open a normal issue or PR for it and drop a comment here with the link — I'll put it in the Status column. If any row is a duplicate, a non-bug, or intentional, say so and I'll annotate it.

Found with code-review-toolkit. Reports were drafted with AI assistance (Claude Code) and then re-verified by hand — see Disclosure at the end.

Scope and verification

• Everything below is against main at 993a0c65a7b (2026-07-27). The review itself was carried out on the 3.14 branch at 6080c866096; every finding was then re-checked against main by comparing the enclosing function (46 byte-identical, 13 changed and read individually, 11 module-level checked by hand), and every line number here is main's. Line numbers drift; the cited function and mechanism will not.
• 18 findings were reproduced against a running interpreter (real Tk where the widget matters); the rest were confirmed by reading the code and its tests. Each gist entry says which. The reproductions were re-run on a release build of main at the commit above.
• python -m test test_idle passes on this commit (305 run, 83 skipped headless). Every finding below is one the suite cannot see — and four of them (26, 71, 72, 73) are about why.
• Prior art was checked for all 73 findings against the whole tracker: 1626 topic-IDLE issues, 281 unlabelled issues mentioning idlelib, 573 PRs and 86 idle.rst items (2503 unique). Three were already reported and have been removed from this report — they are listed at the bottom so nobody re-derives them. Where an open issue covers the area without covering the defect, the table names it.

Three systemic roots

These are worth reading first: each explains several of the individual rows below, and fixing the root is a smaller change than fixing its instances one at a time.

• 31 — Five independent hand-written recognizers for Python syntax with no shared source of truth
  idlelib answers “is this Python syntax?” in five hand-written places with no shared definition, and new syntax reaches some and not others. colorizer is consistently the one that gets updated — four of the rows below are drift between it and a sibling.
• 32 — The -n execution backend is missing five cross-cutting protections the RPC backend has
  use_subprocess selects between two interchangeable execution backends, and every cross-cutting protection lives inside the RPC one rather than in a shared driver. SystemExit is the sharpest cell and the two are opposites: the RPC backend contains it and returns to the prompt, while -n asks “Do you want to exit altogether?” with default="yes" and raises out of mainloop(), so unsaved editor buffers get no save prompt. Ctrl-C, Restart and the recursion headroom are missing from -n too.
• 34 — WidgetRedirector swallows every TclError, so text.index() can never raise on any IDLE widget
  except TclError: return "" in WidgetRedirector.dispatch intercepts every subcommand of every percolated Text — which is every EditorWindow, PyShell and OutputWindow. On the widgets idlelib actually uses, text.index() therefore never raises and never returns None, callers cannot tell “no selection” from “the call failed”, and three incompatible guard idioms have grown in one file.

Start here

Lowest effort to confirm, highest cost if left alone — all silent:

• 35 — Format Paragraph silently deletes characters from every non-comment line of a selection (format.py:59)
• 36 — A literal % in any setting strips every open window's keybindings and leaves them off (config.py:48)
• 27 — Copy-with-prompts drops the last line whenever the selection ends at column 10, 20, 100... (pyshell.py:1034)
• 60 — Saving a file whose coding cookie cannot encode its text writes a BOM and keeps the cookie (iomenu.py:324-326)
• 51 — restore_file_breaks appends without clearing, so Save As carries the old file's breakpoints over (pyshell.py:268)
• 59 — _study1 has no f-string state, so a PEP 701 multi-line f-string is read as three statements (pyparse.py:245)
• 43 — Three keysyms shipped in every keyset cannot be spelled by the key-configuration dialog (config_key.py:15)
• 34 — WidgetRedirector swallows every TclError, so text.index() can never raise on any IDLE widget (redirector.py:116)

Findings

The # column is a row number in this issue, not an identifier. Please cite this issue plus the row (“#‹this issue› row 35”) rather than the bare number on its own — a bare number does not autolink and does not lead anyone back here.

Status: Under analysis = awaiting a maintainer's call, not cleared for work · #N = filed, in progress · #N FIXED = filed and fixed · won't fix / intentional = closed by a maintainer.

Prior art names an existing issue where one is relevant: from = this is residue of that merged change; see = that issue covers the area but not this defect, so a comment there may be better than a new issue.

Shell, the -n backend, and breakpoints (15)

pyshell.py, run.py, rpc.py — full reports

Configuration and the Settings dialog (9)

config.py, configdialog.py, config_key.py — full reports

Editor: formatting, undo, search/replace, save and run (14)

format.py, undo.py, replace.py, searchengine.py, iomenu.py, runscript.py, zoomheight.py — full reports

The five hand-written syntax recognizers (7)

pyparse.py, hyperparser.py, codecontext.py, colorizer.py — full reports

Completions and calltips (5)

autocomplete*.py, calltip.py — full reports

Infrastructure: redirector, file list, debugger, trees (9)

redirector.py, filelist.py, debugger*.py, tree.py, stackviewer.py — full reports

Documentation and the generated in-app help (6)

Doc/library/idle.rst, help.py, help.html, config-main.def — full reports

Test-suite integrity (4)

idle_test/ — full reports

Excluded — please don't pursue these

Found by the review and then dropped — three because they are already reported, one because main already fixed it. Listed only so the same ground isn't covered twice:

Disclosure & caveats

• The reports were drafted with AI assistance (Claude Code); each gist carries an explicit disclaimer. Before publishing, every cited line was re-read in the tree at the stated commit, and every "Verified" transcript was produced by running the code.
• That re-verification changed three findings, which is the reason it is worth mentioning: 59 was upgraded from an argument to a reproduction by a differential its author had not run; a fourth claim (about how many test files return 0 on failure) was wrong by an order of magnitude and is corrected in 73; and two line numbers were off.
• 33 is deliberate in part. The commit that introduced the -n exclusion says so in its own message. What is reported is only the part that commit did not consider: under -n the name is not absent, it is IDLE's own entry point.
• Severities are the reviewer's, not a maintainer's. CONSIDER items in particular are judgement calls and several may be intentional; say so and I'll annotate them.

Each finding is also classified against a catalogue of recurring Python bug shapes; the shape name in each gist entry (zip-truncates-on-length-mismatch, per-line-property-derived-from-the-aggregate, …) is what let the review hunt siblings of a confirmed defect rather than stop at the first instance. Several rows here were found that way.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux