docs / history / FIXHANG5_REDOS_RESULTS.md
docs / history / FIXHANG5_REDOS_RESULTS.md
Branch spike/ssh2-transport @ 43d0934 + UNCOMMITTED FIX-HANG-5 / 5b / 5c in
src/tools/fileSystemTools.ts only. Live bundle verified before first test:
dist/index.js + dist/index.mjs mtime 2026-08-30T08:04:3Z; worker source literal contains
require("worker_threads").parentPort (1× each), pre-fix anchored L123 literal count 0,
unanchored prefix form at 5 sites; the single self.onmessage occurrence in each bundle is an
inert comment line. Host runtime: Node v24.15.0.
((a+){3}){4}x vs 15k-a payload) — PASS (definitive)grep_files(pattern ((a+){3}){4}x, path _redos_test/, include t1b_worker_redos.txt,
max_file_size 60000). Returned fast; filesScanned: 1; matches 0; skipped_files reason:
regex evaluation terminated in isolated worker after 2000ms — likely ReDoS-prone pattern.
main.log (C:\Users\root.MPITS\AppData\Roaming\LM Studio\logs\main.log):
"exceeded 2000ms budget" can only be emitted by the watchdog firing on a LIVE, spinning worker
→ all three fixes proven live end-to-end for the first time: 5b routing (worker engaged),
5c boot + real regex work (no self is not defined in new window), and the load-bearing claim
that worker.terminate() preempts an unpreemptible .test(). Host never froze.
_redos_test/t2_classifier.txt (lines: ab / aab / xab / (ab) / c) — PASS| Pattern | Gate verdict (source-traced) | Path | Live result | Verdict |
|---|---|---|---|---|
(a+)b | false (single split, no explosion) | INLINE | matches lines 1–4 exactly as predicted (ab,aab,xab,(ab)); c correctly not matched | PASS |
(ab){2} | false (fixed {n}, unique partition; null-safe advance) | INLINE | clean zero matches — NO TypeError (latent null![0] crash class confirmed dead) | PASS |
ab\1a+(\d+) vs _redos_test/t2c_backrefs.txt) — PASSGate: sawBackreference=true → WORKER path (deterministic source trace; live call showed no
anomaly). Result: 0 matches, fast return. Live ground-truth probe in the host runtime
(Node v24, flags i, non-global): pattern matches NONE of the three payload lines — including
line "x 42" (earlier expectation that it would match was WRONG; modern V8 backreference
semantics: group-1 capture undefined at \1 position → no match). Zero is therefore the correct
engine result. SCOPE NOTE (added with D2 resolution): the literal-demotion class behind FINDING-1/D2 is narrower than first implied here — only patterns with a word char directly abutting an unescaped * were affected; (ab){25} and ([a-z]*){3}x never demoted. The ~2.9k-q line (ReDoS amplifier) caused no hang — containment holds even if routing were ever bypassed on this shape.
(a*){50}-class — offline probe PASS (see FINDING 1)Verbatim shipped worker source run in a real worker_threads Worker (same Node runtime),
T2 payload lines, pattern /(a*){50}/i: returned [0,1,2,3,4] in 17 ms — all five non-empty
lines match via zero-length matches; first-match-per-line semantics intact.
Setup: unique token t3-end-marker in _redos_test/t3_fra_scratch.txt (collision-checked
against pre-existing fixtures before firing). Call: directory _redos_test/, ext txt,
replacement T3-END-MODIFIED, dry_run=false, confirm=true, backup=true.
PASS legs: replacement applied exactly once in the one intended file; all other lines byte-identical; the two >100 KB payloads were correctly size-gated and skipped; no deadline/abort anomaly — F1/F2/F3 phases completed cleanly ("Changes applied successfully").
FAIL leg: no .bak exists anywhere after a successful FRA write. Root cause (source, FRA
implementation): the backup is created pre-write for rollback only and then deleted on success:
This contradicts: (1) FRA's own schema description ("Create .bak backup before modification"),
(2) TOOLS_REFERENCE.md L112 (find_replace_all advertised with ".bak backups"), and
(3) every sibling file tool in the same module (replace_text_in_file, insert_at_line,
append_file, delete_lines_in_file persist .bak + announce via createBackupAnnouncement for
restore_from_bak). No existing test pins either behavior (tests only mock unlink).
Git history unavailable inside the sandbox; user can check intent with:
git log -S "fs.unlink(backupPath)" -- src/tools/fileSystemTools.ts.
Scratch artifacts cleaned after verification. Pre-existing _redos_test/fra_scratch.txt untouched.
Branch spike/ssh2-transport @ 43d0934 + UNCOMMITTED FIX-HANG-5 / 5b / 5c in
src/tools/fileSystemTools.ts only. Live bundle verified before first test:
dist/index.js + dist/index.mjs mtime 2026-08-30T08:04:3Z; worker source literal contains
require("worker_threads").parentPort (1× each), pre-fix anchored L123 literal count 0,
unanchored prefix form at 5 sites; the single self.onmessage occurrence in each bundle is an
inert comment line. Host runtime: Node v24.15.0.
((a+){3}){4}x vs 15k-a payload) — PASS (definitive)grep_files(pattern ((a+){3}){4}x, path _redos_test/, include t1b_worker_redos.txt,
max_file_size 60000). Returned fast; filesScanned: 1; matches 0; skipped_files reason:
regex evaluation terminated in isolated worker after 2000ms — likely ReDoS-prone pattern.
main.log (C:\Users\root.MPITS\AppData\Roaming\LM Studio\logs\main.log):
"exceeded 2000ms budget" can only be emitted by the watchdog firing on a LIVE, spinning worker
→ all three fixes proven live end-to-end for the first time: 5b routing (worker engaged),
5c boot + real regex work (no self is not defined in new window), and the load-bearing claim
that worker.terminate() preempts an unpreemptible .test(). Host never froze.
_redos_test/t2_classifier.txt (lines: ab / aab / xab / (ab) / c) — PASS| Pattern | Gate verdict (source-traced) | Path | Live result | Verdict |
|---|---|---|---|---|
(a+)b | false (single split, no explosion) | INLINE | matches lines 1–4 exactly as predicted (ab,aab,xab,(ab)); c correctly not matched | PASS |
(ab){2} | false (fixed {n}, unique partition; null-safe advance) | INLINE | clean zero matches — NO TypeError (latent null![0] crash class confirmed dead) | PASS |
ab\1a+(\d+) vs _redos_test/t2c_backrefs.txt) — PASSGate: sawBackreference=true → WORKER path (deterministic source trace; live call showed no
anomaly). Result: 0 matches, fast return. Live ground-truth probe in the host runtime
(Node v24, flags i, non-global): pattern matches NONE of the three payload lines — including
line "x 42" (earlier expectation that it would match was WRONG; modern V8 backreference
semantics: group-1 capture undefined at \1 position → no match). Zero is therefore the correct
engine result. SCOPE NOTE (added with D2 resolution): the literal-demotion class behind FINDING-1/D2 is narrower than first implied here — only patterns with a word char directly abutting an unescaped * were affected; (ab){25} and ([a-z]*){3}x never demoted. The ~2.9k-q line (ReDoS amplifier) caused no hang — containment holds even if routing were ever bypassed on this shape.
(a*){50}-class — offline probe PASS (see FINDING 1)Verbatim shipped worker source run in a real worker_threads Worker (same Node runtime),
T2 payload lines, pattern /(a*){50}/i: returned [0,1,2,3,4] in 17 ms — all five non-empty
lines match via zero-length matches; first-match-per-line semantics intact.
Setup: unique token t3-end-marker in _redos_test/t3_fra_scratch.txt (collision-checked
against pre-existing fixtures before firing). Call: directory _redos_test/, ext txt,
replacement T3-END-MODIFIED, dry_run=false, confirm=true, backup=true.
PASS legs: replacement applied exactly once in the one intended file; all other lines byte-identical; the two >100 KB payloads were correctly size-gated and skipped; no deadline/abort anomaly — F1/F2/F3 phases completed cleanly ("Changes applied successfully").
FAIL leg: no .bak exists anywhere after a successful FRA write. Root cause (source, FRA
implementation): the backup is created pre-write for rollback only and then deleted on success:
This contradicts: (1) FRA's own schema description ("Create .bak backup before modification"),
(2) TOOLS_REFERENCE.md L112 (find_replace_all advertised with ".bak backups"), and
(3) every sibling file tool in the same module (replace_text_in_file, insert_at_line,
append_file, delete_lines_in_file persist .bak + announce via createBackupAnnouncement for
restore_from_bak). No existing test pins either behavior (tests only mock unlink).
Git history unavailable inside the sandbox; user can check intent with:
git log -S "fs.unlink(backupPath)" -- src/tools/fileSystemTools.ts.
Scratch artifacts cleaned after verification. Pre-existing _redos_test/fra_scratch.txt untouched.
.bak policy — (A) persist backup after success to match docs/siblings
(clutter managed by cleanup_backups; restores documented workflow), or (B) keep rollback-only
semantics and correct the schema description + TOOLS_REFERENCE.md wording. Tests can pin either.isSafeRegex gate in security.ts demotes (a*){50}-class patterns to
LITERAL mode BEFORE FIX-HANG-5 triage, so such searches silently return literal results (matches lost).
Worker mechanism proven correct for this class if ever routed. RESOLVED by user decision D2 this session (~10:3x): isSafeRegex LOOSENED in src/security.ts —
codesig clause alternative [\w][*&] → [\w]& (drop '' from word-char adjacency; the redundant \* escape
in the following class also dropped, semantically identical). Surviving rejection paths: ::, →,
word-char+'&', and [*&]-whitespace-word. VERIFIED against a verbatim disk extraction of the NEW function:
all pinned REV-24/safe/UNSAFE cases hold; D2 targets ((a*){50}, foo*bar) now SAFE; codesig positives
(List* ptr, std::vector<int>*, ptr→val*) still rejected. One matrix mismatch was a wrong guard
expectation, not a regression: bare-& patterns with with no unescaped , + or ? (const int& x) were SAFE pre AND
post — unchanged. Regression tests pinned in tests/security.test.ts (D2 blocks). SCOPE PRECISION: the
over-rejection class is NARROWER than first documented here — only word char DIRECTLY ABUTTING an unescaped ''
((a){50}, foobar); ([a-z]){3}x and (ab){25} were never demoted. USER-SIDE VERIFICATION CONFIRMED ALL GREEN ~10:41 30.08 (jest incl. new D2 blocks + tsc). Remaining: tsup rebuild + full restart (if not yet done) → live re-fire '(a*){50}' (expect patternMode regex, count=5, fast worker completion), then D3 commit ruling for all uncommitted work.src/tools/fileSystemTools.ts modified vs 43d0934).[2026-08-30 09:55:57.121] [error] ... stderr: [grep_files] FIX-HANG-5: worker error (self is not defined) — treating as skipped ← pre-rebuild bundle, known failure mode
[2026-08-30 10:09:37.145] [error] ... stderr: [grep_files] FIX-HANG-5: worker exceeded 2000ms budget — terminated (possible ReDoS) ← live T1b, fresh bundle
if (backup) { backupPath = fullPath + '.bak'; await fs.copyFile(fullPath, backupPath); }
try { await atomicWriteFile(fullPath, newContent); } catch (err) { /* restore from .bak */ throw ...; }
if (backupPath) { try { await fs.unlink(backupPath); } catch {} } // ← deleted after success
.bak policy — (A) persist backup after success to match docs/siblings
(clutter managed by cleanup_backups; restores documented workflow), or (B) keep rollback-only
semantics and correct the schema description + TOOLS_REFERENCE.md wording. Tests can pin either.isSafeRegex gate in security.ts demotes (a*){50}-class patterns to
LITERAL mode BEFORE FIX-HANG-5 triage, so such searches silently return literal results (matches lost).
Worker mechanism proven correct for this class if ever routed. RESOLVED by user decision D2 this session (~10:3x): isSafeRegex LOOSENED in src/security.ts —
codesig clause alternative [\w][*&] → [\w]& (drop '' from word-char adjacency; the redundant \* escape
in the following class also dropped, semantically identical). Surviving rejection paths: ::, →,
word-char+'&', and [*&]-whitespace-word. VERIFIED against a verbatim disk extraction of the NEW function:
all pinned REV-24/safe/UNSAFE cases hold; D2 targets ((a*){50}, foo*bar) now SAFE; codesig positives
(List* ptr, std::vector<int>*, ptr→val*) still rejected. One matrix mismatch was a wrong guard
expectation, not a regression: bare-& patterns with with no unescaped , + or ? (const int& x) were SAFE pre AND
post — unchanged. Regression tests pinned in tests/security.test.ts (D2 blocks). SCOPE PRECISION: the
over-rejection class is NARROWER than first documented here — only word char DIRECTLY ABUTTING an unescaped ''
((a){50}, foobar); ([a-z]){3}x and (ab){25} were never demoted. USER-SIDE VERIFICATION CONFIRMED ALL GREEN ~10:41 30.08 (jest incl. new D2 blocks + tsc). Remaining: tsup rebuild + full restart (if not yet done) → live re-fire '(a*){50}' (expect patternMode regex, count=5, fast worker completion), then D3 commit ruling for all uncommitted work.src/tools/fileSystemTools.ts modified vs 43d0934).[2026-08-30 09:55:57.121] [error] ... stderr: [grep_files] FIX-HANG-5: worker error (self is not defined) — treating as skipped ← pre-rebuild bundle, known failure mode
[2026-08-30 10:09:37.145] [error] ... stderr: [grep_files] FIX-HANG-5: worker exceeded 2000ms budget — terminated (possible ReDoS) ← live T1b, fresh bundle
if (backup) { backupPath = fullPath + '.bak'; await fs.copyFile(fullPath, backupPath); }
try { await atomicWriteFile(fullPath, newContent); } catch (err) { /* restore from .bak */ throw ...; }
if (backupPath) { try { await fs.unlink(backupPath); } catch {} } // ← deleted after success