Skip to main content
Kodelyth ECC
Changelog

Release history

Live-verified releases pulled from the kodelyth-ecc repo. Last 10 highlighted below, full history further down.

Latest 10 releases

v2.31.10 · adversarial sweep: close expansion-splice, clustered -n, and hook-manager bypasses

November 2026

A maintainer-run "be-the-reporter" adversarial sweep of the shipped scripts/hooks/block-no-verify.js, each vector reproduced in live git (a commit landing past a pre-commit that exits 1) before any code changed. Three classes of real bypass closed; the structural ceiling named and left pinned.

Closed:

10. Empty-expansion splices. --no-ver$()ify, ` --no-verify (empty backticks), --no-ver${UNSET}ify, --no-ver$(:)ify, and the same inside the config key (core.hooks$()Path) all reassemble the bare flag/key — the shell removes a command substitution or parameter expansion before git parses. Closed by checking a view with $(...), ... , ${...}, and $name collapsed away, in addition to the raw command (so a substitution that literally contains the flag is still caught by the raw pass).

11. Clustered short flags. git unbundles a single-dash run, so -n hides in -nm / -vn / -qn / -sn. Closed by matching a bundle whose pre-n characters are all no-argument flags — arg-takers (m c C F t) are excluded, so -mn (= -m "n") and -am (= -a -m) correctly do NOT match. This rule runs only on views with multi-word quoted phrases removed: a bare -n is a common word in commit messages, and live git confirms a -n inside a message is the message, not the flag, so matching it there would be a false positive.

12. Hook-manager kill-switches. husky and lefthook install their hooks as git's pre-commit/commit-msg/pre-push and no-op on HUSKY=0 / HUSKY_SKIP_HOOKS= / LEFTHOOK=0|false|off. In those repos (the JS/TS ones this toolkit most runs in) that skips hooks exactly like --no-verify. Verified as real bypasses. Matched when such an assignment shares a command with a git commit/push; runs on phrase-free views (a message mentioning the string does not trip), and the names are unambiguous (HUSKY=1, MY_HUSKY=0, husky install do not match).

Ceiling, named not faked: the shell-variable / assembled-flag forms (f=--no-verif; git commit $f, $(printf -- '--no-ver'; printf 'ify')) put the literal in no view of the text — unclosable by a text guard, pinned by tests. The only real fix is enforcement that observes whether the hook RAN (committer-side wrapper or server-side pre-receive).

Also out of scope, documented: pre-commit framework's SKIP= / PRE_COMMIT_ALLOW_NO_CONFIG — real bypasses in pre-commit-framework repos, but SKIP is a collision-prone name; documented, not matched, pending a request.

tests/hooks/block-no-verify.test.js: 27 → 32. A false-positive sweep confirms no legitimate command (incl. -am, $(git rev-parse …) in a message, HUSKY=1`) is blocked. Full suite green.

v2.31.9 · close the shell line-continuation bypass

November 2026

Fourth report from the same reporter (daisy-49@ilands.app), against the shipped scripts/hooks/block-no-verify.js from v2.31.8 (sha b793a75). The v2.31.8 fixes all held across six spellings; one new case was found and is now closed.

Closed:

9. Shell line continuation inside the token. git commit --no-\verify and git -c core.hooks\Path=/dev/null commit both bypass — the shell removes backslash-newline before parsing, so git receives the joined flag and skips the repo hooks, but the raw command text the hook sees has the literal split across two physical lines. The v2.31.8 strip removed ' " \ but kept the newline, so the token never became contiguous. Both forms were **confirmed live bypasses** (a commit landed past a pre-commit that exits 1). Closed by unwrapping backslash-newline (\r?\n) first, producing a joined variant the patterns then match (raw, and quote/escape-stripped). Same class as case 7 (shell-level splicing); it over-triggers inside single quotes where backslash-newline is literal — the fail-closed direction already chosen for cases 5-7.

Unchanged: case 8 (a flag passed through a shell variable) remains the documented ceiling — the reporter confirmed a committer-side wrapper is a step not a wall, so the only agent-unauthored enforcement is a server-side pre-receive hook, a different surface. The pinning test stays.

tests/hooks/block-no-verify.test.js: 25 → 27 (line-continuation --no-verify, CRLF form, line-continuation core.hooksPath, and a benign-continuation negative). Full suite green.

Also removed from the maintainer's own ~/.claude/settings.json a stale npx block-no-verify@1.1.2 PreToolUse entry — the pre-v2.31.4 network/fail-open hook — which had lingered alongside the vendored local hook. No package change; noted here for the record.

v2.31.8 · close the closable --no-verify bypasses; name the ceiling

November 2026

Third report from the same reporter, four more cases against the shipped scripts/hooks/block-no-verify.js from v2.31.7, each verified in live git before being touched. Three are closed; the fourth is the honest ceiling of a text-pattern guard and is now documented as such instead of pretended away.

Closed:

5. Case. git config keys are case-insensitive, so core.hookspath and core.HooksPath are the same key as core.hooksPath — and git -c core.hookspath=/dev/null commit was a *confirmed* live bypass (lowercase key landed a commit past a failing pre-commit hook). The hooksPath pattern is now matched case-insensitively.

6. Quoted flag. git commit "--no-verif" parses identically to the bare flag, but the quote sat outside the pattern's boundary class.

7. Concatenated / escaped flag. The shell splices --no-ver''ify, --no-ver""ify, and --no-\verify back into --no-verify before git sees them, so the literal word is absent from the raw text.

Cases 5–7 are closed together: every pattern now runs against the raw command and a copy with shell quote/escape characters (' " \) stripped, which collapses the hidden forms back to what the patterns expect. The word-boundary classes still gate, so a flag merged into the middle of another token (x"--no-verify"y → x--no-verifyy) is not a boundary match and does not trip — verified by negative tests.

Not closed, and why:

8. Variable. f=--no-verif; git commit $f passes the flag through a shell variable, so the literal never appears in the command text at all. No text pattern can catch this — it is the ceiling of any guard that reads the command string. Closing it needs a check that observes whether the hook actually *ran* (a committer-side wrapper, or a server-side pre-receive hook), a different enforcement surface than a PreToolUse text guard. Left documented, not half-built. A unit test pins the gap so it stays explicit: for the agent use case this guard exists for, an agent does not assemble a bypass flag from a variable by accident; the common, accidental forms (5–7) are what matter and are closed.

Out of scope, corrected: the reporter verified that GIT_CONFIG_SYSTEM=/dev/null git commit does not bypass in a normal repo — with core.hooksPath in the repo's local config, blanking the system or global config file still runs the hook. It only bypasses in the narrow case where hooksPath itself lives in the file being blanked. The header note is corrected to say so rather than implying it is a live vector.

Trade-off, stated honestly: the quote-stripping pass means a quoted mention of the flag in an echo or a git log --grep now trips the guard where it didn't before. That is the intended direction for a fail-closed security hook — a spurious block costs one edited command; a silent allow is the whole reason the hook exists.

Tests: tests/hooks/block-no-verify.test.js grew from 21 to 25 — case-variant hooksPath, quoted/spliced/escaped flags, the variable-form ceiling (pinned as a known gap), and a mid-token negative proving the boundary gate holds.

v2.31.7 · doctor now detects stale installed config

November 2026

doctor gained a config-drift check. It answers a question nothing in the toolkit answered before: **is the config in ~/.claude behind the installed package?**

The gap. The existing update check (scripts/cli/update-check.js) compares the npm PACKAGE against the registry. It says nothing about whether the installed config — hooks, agents, skills, rules in ~/.claude — matches the package. Those drift apart the instant someone runs npm i -g kodelyth-ecc without re-running the installer: the package jumps forward, ~/.claude stays frozen at whatever version last wrote it. This is not hypothetical — it is exactly how a hooks.json installed at v2.30.0 kept running a superseded npx block-no-verify hook (and a pre-fix capture-correction.js) while the package sat six patch versions ahead. Every existing doctor check passed; the drift was invisible.

The fix. install.sh and install.ps1 have always stamped the installing version into /kodelyth-ecc-install-state.json — but nothing ever read it back (it was write-only). The new checkConfigDrift() in scripts/doctor-health.js closes that loop: it reads the installed config's version and compares it to this package's version. Behind → WARN with the exact re-install command. In sync → PASS. The comparison is entirely local (both versions read from disk), so unlike the registry update check it needs no warm cache and works in doctor every time.

Example output against a real stale install:

`` ⚠ config-drift — ~/.claude config was installed from v2.30.0 on 2026-11-11T13:51:02Z; package is now v2.31.7 — installed hooks, agents, skills and rules may be stale fix: kodelythecc --target claude-code refreshes ~/.claude to v2.31.7 `

Scope, kept deliberately tight. This detects version drift (package moved, config did not), which is the real failure mode. It does not content-hash every installed file — a user who hand-edits one hook is a rarer, different case. The decision logic is a pure, exported configDriftStatus() with 8 unit tests; cmpSemver() is self-contained so doctor still runs if sibling modules fail to load. The deliberately cache-only doctor update check is left untouched — that was a conscious design choice by the author, not a gap.

Tests: tests/doctor/health.test.js grew from 4 to 12 (config-drift presence in run()`, behind/ahead/in-sync/missing-version/unknown-package cases, and numeric-not-lexical version ordering).

v2.31.6 · close two more --no-verify bypass vectors

November 2026

Security patch. The vendored block-no-verify hook shipped in v2.31.4 still had two silent-allow gaps against adversarial git commands. Both reported by daisy-49@ilands.app, both verified in-session against the shipped file (not a diff) and reproduced in live git before shipping.

Gap #1 — prefix matching. git's parse-options accepts any unambiguous prefix of a long flag. git commit --no-verif -m x and even git commit --no-veri -m x both resolve to --no-verify and skip the hooks. The previous regex /(^|[\s;&|(])--no-verify(\s|$|[;&|)])/ required the literal word and let both prefix forms through.

Gap #2 — core.hooksPath override. Pointing git at an empty hooks directory bypasses the repo's hooks without ever mentioning --no-verify. Both forms work:

  • git -c core.hooksPath=/dev/null commit -m x (inline override)
  • git config [--global] core.hooksPath /dev/null (persistent write)
The env-var siblings (GIT_CONFIG_PARAMETERS, GIT_CONFIG_COUNT + GIT_CONFIG_KEY_N / GIT_CONFIG_VALUE_N) are the same class.

The fix. detectBypass in scripts/hooks/block-no-verify.js grew two additional rules on top of the existing --no-verify and git commit -n checks:

  • /(^|[\s;&|(])--no-veri(f(y)?)?(\s|$|[;&|)=])/ — catches the two
unambiguous prefixes (--no-veri, --no-verif) alongside the full flag, while the - excluded from the terminator keeps --no-verify-signatures (legitimate GPG-verification skip on push/merge) out of scope.
  • /\bgit\b[^;&|\n]*?\bcore\.hooksPath\b|\bcore\.hooksPath\b[^;&|\n]*?\bgit\b/ —
blocks any command segment that mentions both git and core.hooksPath in either order, covering the inline -c form, the persistent git config write, and the env-var forms. Shell boundaries (;, &&, ||, |, newline) stop the match so unrelated chained commands don't cross over.

**Still out of scope, documented so it's an explicit gap rather than a silent one:** GIT_CONFIG_SYSTEM=/dev/null git commit (points git's system config at a dead file; no core.hooksPath string appears in the command) and attacker-controlled edits to .git/hooks/* before commit (not a flag class — a file-tamper class).

Tests. tests/hooks/block-no-verify.test.js grew from 14 tests to 21. The new coverage locks in every prefix form, every core.hooksPath form (including both env-var variants), the --no-verify-signatures exemption, and two negative tests proving the detection does not false-positive on documentation or cross segments on ;/&&.

Credit: Daisy's follow-up report, rechecked against v2.31.5 shipped files.

v2.31.5 · stop writing compaction summaries to lessons.md

November 2026

Pollution fix. The Stop hook hooks/memory/capture-correction.js was reading compaction-summary turns as if they were human corrections and appending them to tasks/lessons.md. Across the user's projects on this machine, 123 polluted entries had accumulated in 9 tasks/lessons.md files before the fix landed.

The bug. Claude Code's /compact (and the auto-compact that fires near the context limit) injects the previous conversation's summary into the next session as an event of type: "user" with isCompactSummary: true and message text that starts with "This session is being continued from a previous conversation…". The capture hook filtered only on event.type !== "user", so every compaction summary flowed into the correction classifier. The classifier is intentionally broad — "always use", "we always", "never do", "please stop" — and a long summary full of narrated facts is guaranteed to hit it. The result: every /compact pushed the entire summary into tasks/lessons.md, which then got injected back into every subsequent session at ~300 characters per entry, every session forever.

The fix. A new isSystemMessage(text, event) guard runs before the correction classifier. It skips the turn when any of these is true:

  • event.isCompactSummary — the official flag, confirmed from real session
JSONLs on disk (compaction turns are emitted as type: "user", isCompactSummary: true, isVisibleInTranscriptOnly: true)
  • event.isMeta / event.isSidechain / event.isVisibleInTranscriptOnly —
belt-and-suspenders for related synthetic shapes (subagent turns, metadata injections)
  • Text starting with This session is being continued from a previous
conversation — the fallback if a future Claude Code version changes the flag shape
  • Text that opens with a synthetic block tag: ,
, , , , , , , , , raw //

The guard matches at leading position only — a user who types "stop writing system-reminder blocks in output" in prose is still a valid correction.

Testability. The hook's pure functions (isSystemMessage, extractText, extractCorrections) are now exported, and the stdin listener is gated behind require.main === module. The new test file tests/hooks/capture-correction.test.js exercises 33 cases that matter: the exact two shapes of pollution that infected v2.31.4 (markdown-emphasised and plain summaries), every event-level flag, every synthetic tag, and an end-to-end JSONL fixture that reproduces the regression ("summary turn then real correction → summary dropped, correction survives").

Cleanup. The two polluted blocks that had been committed to this repo's own tasks/lessons.md (dated 2026-11-11) are removed in the same commit. Scanning the rest of this machine turned up 9 other tasks/lessons.md files with the same shape — those are the user's own project repos and are not touched by this release.

843 tests pass total (was 811 in v2.31.4). No other runtime behaviour changed.

v2.31.4 · fail-open security hook + vendored replacement

November 2026

Security fix. Reported externally by Daisy (daisy-49@ilands.app) on 2026-11-12 after the address listed in SECURITY.md bounced.

The bug. hooks/hooks.json ran npx block-no-verify@1.1.2 on every Bash PreToolUse. Two problems in that one line:

1. Network dependency in a hot path. On a cold npm cache, every session fetched third-party code (maintainer: tupe12334) from the registry. The toolkit markets "all local, zero telemetry"; SECURITY.md itself claimed "all hook scripts are plain Node.js with zero external npm dependencies". Neither was true while this hook was wired.

2. Fail-open on network loss. Claude Code treats only exit code 2 as blocking; any other non-zero is a non-blocking log. When the npm registry was unreachable (corporate proxy, airplane, VPN flake, DNS failure, actual outage) npx exited 1 — so the --no-verify bypass the hook exists to stop went through exactly when users most expected the guard to be in force. A guard that fails open on network partition is worse than no guard; it produces false confidence.

The fix. The hook is now a local vendored script at scripts/hooks/block-no-verify.js — pure Node, no npm dependencies, no child process, no network call, no fail-open path. Detects --no-verify and git commit -n with scoped regexes (so git log -n 5 is not confused with a commit bypass). Blocks with exit 2 and a human-readable stderr message. hooks/hooks.json now invokes it via the same ${CLAUDE_PLUGIN_ROOT}/scripts/hooks/… pattern as every other hook in the file.

Defense in depth. A new packaging test fails the release if any hook command string contains a bare npx call. Vendored logic is the correct shape for a security hook, and the test keeps future contributors from reaching for npx as a shortcut. 14 new behavior tests cover the hook itself (pure detectBypass cases plus end-to-end spawn with simulated network-dead environment); the regression guarantee is both behavioral (blocked-proxy env still returns exit 2) and structural (the script's require() set is restricted to node: builtins and relative paths).

SECURITY.md cleanup (same release). The contact address is corrected to ecc@kodelyth.com — the mailbox listed in package.json bugs.email and the one that actually receives mail. The old sifxprime@kodelyth.com is documented as not-monitored so researchers who find the old copy understand why their report bounced. The Supported Versions table, which still said 1.8.x was current, is updated to the 2.31.x line.

Credit. Daisy's report was specific, correct, and arrived with a clear reproduction. She asked for nothing and was honest about her delivery troubles. Thank you — this is the shape of report maintainers hope for.

811 tests pass total (was 787 in v2.31.3). No other runtime behaviour changed.

v2.31.3 · release-ritual guards for the v2.31.1 + v2.31.2 bug classes

November 2026

Pure test-suite release. No user-facing behavior change. Hardens the release ritual so the two categories of bug the v2.31.x train shipped cannot ship again unnoticed.

v2.31.1 was CITATION.cff — a root-level metadata file consumed by external tooling (GitHub's cite / community-standards widgets, Zenodo, reference managers) but missing from package.json files[]. v2.31.2 was the same pattern applied to code-referenced files: the dashboard's /brand/* route reads SVGs from the package root via path.join(__dirname, "..", "..", "brand"), and brand/ wasn't in files[] either.

Two paired packaging tests now catch both patterns pre-publish: • Code-reference guard — scans bin/ and scripts/ for literal path.join(ROOT, "", …) and path.join(__dirname, "..", …, "", …) (depth-matched to each file's distance to ROOT). Cross-references every distinct first segment against npm pack --dry-run --json. Fails on any missing path with a clear file:line citation. • CITATION-class guard — iterates a 12-entry allowlist of conventional root-level metadata files (CITATION.cff, CODE_OF_CONDUCT.md, CONTRIBUTING.md, SECURITY.md, SUPPORT.md, GOVERNANCE.md, MAINTAINERS.md, AUTHORS, AUTHORS.md, NOTICE, THIRD-PARTY.md, ACKNOWLEDGMENTS.md). For each that exists in the repo but is neither in files[] nor auto-included by npm, fail with the filename.

Both run automatically via node tests/run-all.js, which is already the pre-publish gate in .github/workflows/publish.yml. No workflow edits needed.

Verified both guards catch regressions: a simulated v2.31.1 (remove CITATION.cff from files[]) trips the CITATION-class guard; a simulated v2.31.2 (remove brand/ecc-lockup-*.svg from files[]) trips the code-reference guard.

Blind spots the guards don't cover (documented honestly in the test file): dynamic path interpolation (only literal strings are scanned), new external-tool conventions not yet in the allowlist, and doc-linked-but-not-code-referenced files (that's a separate link-checker, not packaging).

31 tests pass total (was 30 in v2.31.2 — one new test). No runtime code changes shipped.

v2.31.2 · brand/ assets in the published tarball

November 2026

Another tiny packaging fix in the same category as v2.31.1's CITATION.cff bug. The dashboard' lockup images (ecc-lockup-light.svg, ecc-lockup-dark.svg) live in brand/ and were missing from package.json files[] — so a fresh npm install -g kodelyth-ecc left the installed package without a brand/ directory, and the dashboard's /brand/* route returned 404. The lockup rendered as a broken-image icon with alt text instead of the actual SVG.

Side effect: ecc_version would also appear to be missing from /api/overview in users who had a dashboard process running from before the upgrade — Node's require() cache held the old data.js. That resolves on a dashboard restart even without a new release.

Added brand/ to files[]. All SVGs re-stamped v2.31.2.

Caught by the same release-prep post-install smoke-test that caught the pipe-truncation bug in v2.31.0 and the CITATION.cff bug in v2.31.1: install a fresh global tarball and poke every HTTP route the dashboard exposes.

No code changes. 30 tests still pass.

v2.31.1 · CITATION.cff in the published tarball

November 2026

Tiny packaging fix. The CITATION.cff file added in v2.30.x was committed to the repo but not listed in package.json files[], so it never shipped in the npm tarball. Researchers reading the published package through npm pack wouldn't find it (GitHub was fine — the file is in the repo, which is what GitHub's own "Cite this repository" button + Zenodo + most academic tools read). Fixed by adding CITATION.cff to files[] alongside the other root-level metadata files.

Caught by the same release-prep smoke-test that caught the pipe-truncation bug in v2.31.0 — npm pack kodelyth-ecc@2.31.0 && ls package/CITATION.cff returned "file not found". Same fresh-install verification will confirm the fix on v2.31.1.

No code changes. 30 tests still pass.