Fix review: H-L11-01, H-L11-02 (branch A) and H-L11-03 (branch B)

From mnema-talk v3.3.13, rendered on this site.

Reviewer: an independent, read-only reviewer. I wrote neither the findings nor the fixes. I changed nothing in the repository and pushed nothing. Each branch was checked out in its own temporary worktree, and both worktrees were removed afterwards. The probes ran in throwaway repositories and clones.

  • A: msch128/fix-h-l11-01-02, commits 9574f981b, 7f8d5234f, 329e85077 on e68fe3271. Finding H-L11-01 (S4) and H-L11-02 (N).
  • B: msch128/fix-h-l11-03, commit 72f28943b on e68fe3271. Finding H-L11-03 (N).

Verdicts

FixVerdict
A (security floor carried forward; misplaced marker refused)MERGE AFTER FIXES: A-1 and A-2 (two one-line changes)
B (strict reader for signed JSON)MERGE

No BLOCKER on either branch. The core of A is right. The floors match the expected values for every 3.0.x release. The floor never falls back and never goes down through the pipeline. A misplaced marker can no longer be dropped silently, because the release decision fails loudly. The two SHOULD-FIX items are both in the new pull-request gate (--check-message / --check-commits). In one direction that gate disagrees with the release decision, and in another it fails open. Neither problem weakens the release-time check.


Branch A

What I ran

  1. Real history, every 3.0.x tag plus the next release. I ran sh and dash with scripts/security-floor.sh v<X> <X> in a full clone with tags (not shallow, 124 tags, all strict vX.Y.Z):

    rev3.0.03.0.13.0.23.0.33.0.43.0.53.0.63.0.73.0.83.0.9HEAD as 3.0.10origin/main as 3.0.10
    floor(empty)(empty)(empty)(empty)3.0.43.0.43.0.63.0.63.0.63.0.93.0.93.0.9

    Every run exited 0 with empty stderr, and sh (bash on macOS) and dash gave the same output. An empty line before 3.0.4 is expected. release.yml:546-548 writes floor= to the step output and the summary says "never raised (web/desktop-compat.json applies)". The Sign the bundle step passes -security-floor "". internal/webbundle/cli.go:87,141-146 treats an empty flag as "no raise" and otherwise takes max(flag, compat), so it never lowers the floor. That is the same fallback to the compat floor (0.7.23) as before. For older tags, v1.1.0, v2.0.0 and v2.9.15 print an empty line. v1.2.0 now exits 1 ("c50b6886 has a 'Security-Floor: raise' line that is not a git trailer"). That is the finding itself. It only matters if someone rebuilds 1.2.0 from its tag (NOTE A-9).

  2. Squash reality. In 93502a10e, cd402e0bf and e2c3afa89 the body ends with Security-Floor: raise as its own last paragraph, and %(trailers:key=Security-Floor) returns it for all three. The script reads them. I also ran --check-message over the bodies of the last 200 merged pull requests (gh pr list --state merged). It refused exactly one: #448, the 1.2.0 PR. Its body is the squash message of c50b6886e, so in this repository the merge body really is the PR description. There were no false positives in 200 real bodies. --check-commits over the whole history (v0.1.0..HEAD, 625 commits) flags only c50b6886e. The release-please shape is a fix commit, then a release commit on top, with the tag on the release commit. I reproduced it in a throwaway repo (probe 26): v0.2.0 gives 0.2.0, v0.2.1 gives 0.2.0, and the untagged HEAD gives 0.2.0. git tag --contains <sha> --merged <rev> --sort=v:refname picks the lowest version tag that contains the commit, which is the release that shipped it (3.0.9 for 93502a10e).

  3. Can the floor go down or be wrong? (throwaway repos, dash)

    • Numeric comparison: raises at v3.0.9 and v3.0.10 give 3.0.10 for 3.0.11 (sort -t. -kN,Nn, and v:refname for the tag order). This is correct.
    • Extra tags on the 3.0.9 commit (v3.0.10-rc.1, desktop-v9.9.9, v0.9) are ignored by the strict semver grep. The result is still 3.0.10.
    • When the tag v3.0.10 is deleted, the raise moves up to the version being released (3.0.11), never down.
    • Two raises in one release give one floor, the release's own version.
    • A raise reachable only through a merge commit (--no-ff side branch) counts at the tag and in every later release (probe 25).
    • A shallow clone is refused (test test_a_shallow_clone_is_refused). The release checkout (release.yml:496-501) uses fetch-depth: 0 with ref: <tag>, and actions/checkout then fetches all branches and tags.
    • Ways down, both needing a maintainer's tag write: a mistaken lower vX.Y.Z tag on a newer commit that contains the raise lowers the floor to that tag (probe 24: v3.0.2 on HEAD gives 3.0.2) (NOTE A-5). A clone without tags has a gap too (NOTE A-4).
  4. Misplaced-marker check (probes 1-20, m1-m8). These must pass, and do: the trailer as the last paragraph; prose quoting inline (Use the `Security-Floor: raise` trailer.); CRLF bodies (commit and stdin); the trailer followed by Co-authored-by: and Signed-off-by: in the same block, and Signed-off-by before it; a git revert message of a raising commit (git does not copy the body); Raise, Security-Floor:raise, Security-Floor : raise; a # heading line; a --- rule in a commit body before the trailer paragraph. These are refused, as they should be: a marker followed by BREAKING CHANGE: (LF and CRLF); a marker followed by a prose paragraph; a marker in the same paragraph as a prose line; an indented marker in the last paragraph (git reads it as a continuation line, not a trailer). These are refused too, and that is acceptable because it fails loudly and never silently: a marker on its own line inside a fenced or indented code block, with or without a real trailer at the end (NOTE A-6), and an HTML comment holding a marker line in a PR body. The PR-gate mismatches are SHOULD-FIX A-1, and A-2 covers fail-open input. Silent false negatives that remain are typos in the value or the key (NOTE A-3).

  5. The very next release after this merge. I simulated it with git commit-tree in a throwaway clone. If the merge commit carries the trailer, tag v3.0.10 gives 3.0.10 and an untagged commit on top gives 3.0.10. If it does not, the tag gives 3.0.9 and the commit on top gives 3.0.9. Both cases are at least 3.0.9. The c50b6886e marker in the old history does not fail it, because only prev..rev is checked.

  6. Portability. All runs and the whole test_security_floor.py (21 tests) pass with dash as sh (PATH shim). I read the awk at scripts/security-floor.sh:74 against mawk rules: /^@@sha /, $2, tolower, and gsub(/^[ \t]+|[ \t]+$/, "", v) are POSIX awk that mawk supports (I verified with BWK awk). ${sha%"${sha#????????}"}, grep -Eic, tr -d '\r', sort -t. -k1,1n and head -n 1 are all POSIX. set -e stops the release path on a failing check_commits call (it is a simple command), as the tests show.

  7. Nightly trust job (sim-sweep.yml:83-102). Actions are pinned to the same full SHAs as elsewhere (checkout v7.0.1, setup-go v7.0.0), with persist-credentials: false. Permissions are contents: read at the top level, there are no secrets and no untrusted input. The trigger is the existing schedule plus workflow_dispatch, with timeout-minutes: 30 and go test -timeout 25m. Runtime: 2,000 seeds of shard 1/4 took 17 s here (load about 12). So a 50,000-seed shard is about 7 minutes locally, which leaves room under 25 minutes even on a slower runner. TestTrustChainSequences honours WEBBUNDLE_TRUST_SWEEP and WEBBUNDLE_TRUST_SHARD (trustchain_test.go:325-340) and fails on a malformed value.

  8. SECURITY.md. The diff is one hunk, and one sentence is replaced (SECURITY.md:1077-1082), with nothing else changed. The new sentence is true. The signed 1.2.0 manifest declares 0.7.23 (phase-2 measurement), and doc/desktop-trusted-client.md:29,359-363 defines the per-web_client_api high-water mark, which refuses any bundle older than one already verified. That rule was documented on 2026-10-17 (4c060836c, #164), before 1.2.0 (2026-10-26). The matching corrections in doc/api-stability.md:172 and doc/upgrade.md:486,853-863 are consistent with it.

Suites: make test-scripts passes, including test_security_floor.py with 21 tests OK and all other script suites OK. go test -p 2 ./internal/webbundle/ ./internal/repo/ passes. go vet is clean. actionlint on ci.yml, release.yml and sim-sweep.yml is clean. make docs-check passes. node scripts/check-english-only.mjs passes. make complexity-check passes. No allowlist was touched. Commit types are ci: (no feat, no !) and there is no attribution.

Revert checks:

  • With the script reverted to e68fe3271, 13 of 21 tests fail.
  • With only the release-path check_commits call removed, 5 fail (all the misplaced-marker tests, including the RealHistory one).
  • With the marks > trailers comparison disabled, 7 fail.
  • The carry-forward is pinned by test_a_raise_is_carried_into_every_later_release, test_a_later_raise_wins... and RealHistory.test_releases_from_3_0_4_on, which all fail against the old script.

Findings

A-1 SHOULD-FIX: the PR gate and the release read the trailer block differently when the text contains a --- line. scripts/security-floor.sh:55 uses git interpret-trailers --parse, which stops at a --- line (the patch divider). The release reads %(trailers) from the commit (lines 43 and 73), and that ignores the divider. I reproduced both directions:

  • False negative: the PR body Text\n\nSecurity-Floor: raise\n\n---\n\nNotes after a rule.\n passes --check-message (rc 0). The same text as the commit message makes the release decision exit 1. So the gate lets through exactly the failure it exists to stop, and that release gets no signed bundle.
  • False positive: the PR body Text\r\n\r\n---\r\n\r\nMore\r\n\r\nSecurity-Floor: raise\r\n is refused (rc 1), but as a commit it raises correctly (probe 16 gives 0.2.0). That blocks a valid security-fix PR.
  • --- lines are common here: 39 of the last 100 merged PR bodies have one.

Fix: git interpret-trailers --parse --no-divider. With that flag, both inputs above give the release's answer (I checked both). Also add both inputs to test_check_message_mode.

A-2 SHOULD-FIX: --check-commits passes when the range cannot be read. scripts/security-floor.sh:41 uses for sha in $(git log --format=%H "$1"), so a failing git log is ignored. --check-commits 0123…4567..HEAD (an unknown base) and --check-commits nonsense both print git's fatal error and exit 0. In ci.yml:199 both SHAs are present today (fetch-depth: 0; head.sha is a parent of the merge ref, and base.sha is on the fetched base branch), so I know of no reachable case. Still, a gate in a protected area should fail closed. Fix: shas=$(git log --format=%H "$1") before the loop. Under set -e the assignment exits on failure.

A-3 NOTE: typos in the marker are still dropped silently. Security-Floor: raise. (a trailer whose value is "raise.") and Security-floors: raise both give keep with no error and no warning, in both the release and the PR gate. That is the same "meant but dropped" class as H-L11-01. test_other_value_keeps deliberately keeps Security-Floor: lower silent. My suggestion: refuse any line matching ^\s*security-floor\s*: that is not a parsed trailer with the value raise, since there is no other legitimate value.

A-4 NOTE: a checkout without tags fails closed only by accident. In a repo without tags, prev is empty, so the check runs over the whole history and every raise maps to the version being released. In a throwaway repo (probe 27: git clone --no-tags) the script exited 0 and printed the release's own version, a silent over-raise. In this repository it fails, but only because c50b6886e is in the full range. The release checkout fetches tags, so this is not reachable today. Suggestion: refuse when git describe finds no v* tag below <rev> while git tag --list 'v*' is not empty, or refuse when no tag exists at all.

A-5 NOTE: "the pipeline has no way to lower the floor" (doc/desktop-trusted-client.md, new text) is true for the pipeline's own decisions. A maintainer who pushes a wrong lower vX.Y.Z tag on a newer commit lowers it (probe 24). That needs write access to tags, so it is a documentation nuance and not a defect.

A-6 NOTE: a marker line quoted in a code block fails the release. A commit that documents the marker on its own line (fenced or indented block) cannot be told apart from a misplaced marker, and that release gets no bundle. This fails loudly, and the PR gate catches it first when the merge body is the PR description. Acceptable, and worth one sentence in AGENTS.md item 11 ("quote it inline, never on its own line").

A-7 NOTE: the PR gate only sees the description at the last push. pull_request runs on the default types, so editing the description to add the trailer at merge time does not re-run the check. Also, the repository's squash default is COMMIT_MESSAGES (gh api), so the ci.yml:191-194 comment "(the squash message)" holds only because maintainers merge with the description as the body, which is current practice (#448 and c50b6886e). The release-time check is the backstop in both cases. A comment correction is enough.

A-8 NOTE: shellcheck. shellcheck -s sh scripts/security-floor.sh reports SC1087 (error severity, a false positive for $marker[[:space:]]) at lines 31-32. It is not a CI gate here, but other scripts in scripts/ are kept clean with directives. Writing ${marker} silences it.

A-9 NOTE: other effects.

  • RealHistory and DocumentedRaises in test_security_floor.py need the full history with tags, so make test-scripts fails in a shallow or tagless local clone. The failure message says so, and CI uses fetch-depth: 0.
  • Rebuilding v1.2.0 from its tag now fails (see run 1).
  • doc/upgrade.md:865 (unchanged) still says "(the floor is a high-water mark)" in the 1.2.0 rollback note. It now reads better as "(the high-water mark)".

Branch B

What I ran

  1. Published manifests. I downloaded the manifest.json of every release that has one (86 releases, v0.7.18 to v3.0.9, including v1.0.0, v2.0.0, v3.0.0 and v3.0.9), each with gh release download <tag> -p manifest.json into its own new empty directory. I ran strictDocument(manifest, strictManifest) plus the unchanged typed decode on all of them in a temporary test file (removed afterwards). All 86 are accepted. I also downloaded the bundle zips of v0.7.18, v1.0.0, v2.0.0, v3.0.0 and v3.0.9 and ran the full webbundle verify -zip … -pub web/desktop-signing.pub, built once from branch B and once from the base. v3.0.0 and v3.0.9 are ok on both builds. v0.7.18, v1.0.0 and v2.0.0 fail identically on both builds with invalid or duplicate path ".gitkeep". That is the existing dotfile rule in m.check, which runs after the strict reader, so the strict reader accepted those manifests too. This is unrelated to this fix.

  2. What the tools write, round-tripped through the new reader:

    • webbundle build (with and without -security-floor, a non-ASCII file name, a short and a 40-character commit) followed by webbundle verify gives ok.
    • trust keygen gives a root key. trust sign round-trips plain, with -prev + -revoke-bundle-key + -max-version 1=… -max-version 2=… + -serial + -valid-for, and with -prev -renew. trust verify gives ok for all of them, and the output has [] and {} and never null.
    • The writer still refuses -max-version 01=….
    • The committed desktop-trust/ statements at f034695fc (serial 3) and bfbd8306e (serial 2) are ok on both builds. The first one at 414fa1275 (serial 1, an older format) is refused by both builds. Only the message differs: it is now missing field "expires_at".
  3. My own inputs to strictDocument (a temporary test, removed). All 30 verdicts are as expected:

    • Duplicate keys after unescaping are refused: "serial" with "serial" at the top level, "1" with "1" in max_version, and "size" with "size" in a file record. A duplicate inside an unknown nested object is refused too. An escaped key spelled alone is accepted.
    • Number ranges: u64 max is accepted and u64 max+1 refused. The schema at 2^32 is refused. -0, 5.0, 5e0, 05 and "5" are refused.
    • Whitespace around numbers and commas, \/, \", \\ and an escaped surrogate pair are accepted. A BOM, a second document, a lone low half, a high half followed by a non-low escape, a high half at the end of a string and invalid UTF-8 are refused. "\\ud800" (an escaped backslash) is accepted.
    • max_version: null and lists that are null, missing or contain null are refused, while max_version: {} is accepted. I checked this against both clients: the Linux client's parseStatement (linux/src/core/trust.ts, which refuses null) and the Rust host's unique_map visitor (experimental/app/src/web_json.rs, which has only visit_map). Both agree.
  4. Cap keys and the deliberate stricter rule. When reading, 1, 01, +1, 0001 and 4294967295 are accepted. ++1, +, +0, 4294967296, 1_0, 1, -1 and 0x1 are refused, which matches Rust u32::parse and the Linux regex ^\+?[0-9]+$. When writing, only canonical keys are accepted. {"1","+1"} is refused when reading with "names web client API 1 twice". Both clients accept that, but they would keep different values, so refusing it is the safe direction. It can only make the owner's check go red, never green. It is documented in doc/desktop-trusted-client.md:97-106, and the CLI never writes it.

  5. Times. 1970-01-01T00:00:00Z is accepted. 1969-12-31T23:59:59Z and 0001-01-01T00:00:00Z are refused. 2106-02-07T06:28:16Z and 9999-12-31T23:59:59Z are accepted, as before.

Suites:

  • go test -p 2 ./internal/webbundle/ ./internal/repo/ ./cmd/webbundle/ passes.
  • With MNEMA_DESKTOP_VECTORS set to a local mnema-talk-desktop checkout at d5cde742, all of these pass: TestDesktopBundleVectorsAgreeWithGo (18 versions, 20 paths, 42 bundles), TestDesktopTrustVectorsAgreeWithGo (23 times, 28 statements), TestDesktopJSONVectorsAgreeWithGo (64 bundles, 38 statements) and TestVerifyTrustReadsJSONLikeTheDesktopClients.
  • go vet and gofmt are clean. make complexity-check, make docs-check and the English-only check pass.
  • File sizes: strictjson.go 250, trust.go 331, desktop_vectors_test.go 249 and strictjson_test.go 125 lines, all within limits. No allowlist was touched.

Revert check: with verify.go and trust.go reverted to e68fe3271 (strictjson.go left in place), 8 tests fail: TestCapKeysAreReadLikeTheClientsAndWrittenCanonically, TestDesktopBundle/Trust/JSONVectorsAgreeWithGo, TestEverythingSignTrustWritesVerifies, TestTimesBefore1970AreRefused, TestVerifyReadsManifestJSONLikeTheDesktopClients and TestVerifyTrustReadsJSONLikeTheDesktopClients.

Behaviour, API, cost:

  • internal/webbundle is imported only by cmd/webbundle, so there is no server hot path, and the second parse runs only in the CLI.
  • The parse still happens after the signature check, both in Verify (verify.go:29) and in VerifyTrust (trust.go:295).
  • Validate() keeps its signature and the writer's rule. validate(reading) is unexported.
  • SignTrust now writes [] and {} for nil lists and maps. Before, a library caller got null, which the clients refuse. The CLI always set them, so no published file changes.
  • The commit type is fix: (no feat, no !) and there is no attribution.

Findings

B-1 NOTE: the client vectors still only run with an environment variable. TestDesktop*VectorsAgreeWithGo skip in CI without MNEMA_DESKTOP_VECTORS. CI coverage of the agreement is strictjson_test.go plus the nine cases in TestVerifyTrustReadsJSONLikeTheDesktopClients, which do run. The finding's suggested direction (commit a copy of test-vectors/) is not done. That is fine for this fix, and a candidate follow-up.

B-2 NOTE (outside this fix): published bundles 0.7.18 to 2.0.0 fail webbundle verify today because of .gitkeep, before and after this branch. That is the existing dotfile rule, and it is unrelated to the strict reader.

Not checked

  • I did not run the release workflow itself or a real GitHub squash merge. Whether GitHub appends Co-authored-by: as a separate paragraph after a hand-written body that ends in the trailer is untested (it would make the release refuse loudly, never silently). This repository has a single author.
  • I did not run the Rust host. Its behaviour was read from source (web_trust/statement.rs, web_json.rs) and taken from the vectors. The Linux client was read, not run.
  • I did not test ubuntu-latest's exact git version. The format atoms used (%(trailers:key=…,valueonly,separator=…), interpret-trailers --parse) have been in git for years, and I tested with git 2.56.
  • I did not run the nightly sweep at full size (200,000 seeds). I only extrapolated it from one 2,000-seed run.

Round 2: branch A at 817330e4e (delta on 329e85077)

Verdict for A: MERGE AFTER FIXES, one small item (R2-1). Branch B is unchanged and stays MERGE.

R2-1 is a false positive of the pull-request gate. It refuses a valid PR and never lets a bad one through. Apart from it, the release decision and the pull-request check agreed on every input I tried.

Round 1 items against the commit

ItemStatusEvidence
A-1 --- divider mismatchfixedraises() uses git interpret-trailers --parse --no-divider after tr -d '\r' (security-floor.sh:32-34). Rule before the trailer: PR check passes, release raises. Trailer then --- then notes: both refuse. --- directly above the trailer in one paragraph: both refuse.
A-2 unreadable range fails openfixedshas=$(git log …) || { …; return 1; } (line 53). An unknown SHA range exits 1 and nonsense exits 1, both printing "cannot read the commits". New test test_check_commits_refuses_an_unreadable_range.
A-3 near misses dropped silentlyfixedAny line starting with Security-Floor (any case, after blanks) must be exactly Security-Floor: raise and read as a trailer. These are refused by both PR check and release: raise., lowercase key, a tab or blank before the colon, a folded value, a leading NBSP.
A-4 checkout without tagsunchanged (NOTE)A tagless repo still prints the release's own version with rc 0. In this repo the check still fails closed through c50b6886e and now also 4c060836c. The release checkout fetches tags.
A-5 "no way to lower"unchanged (NOTE)Wording kept. A maintainer's wrong lower tag still lowers the floor.
A-6 code-block quotes fail the releaseunchanged (NOTE), now widerSee R2-2.
A-7 PR check timing / squash defaultunchanged (NOTE)ci.yml is not touched by this commit.
A-8 shellcheckfixedshellcheck -s sh scripts/security-floor.sh exit 0.
A-9 upgrade.md wordingfixed"(the high-water mark)". AGENTS.md item 11 and doc/desktop-trusted-client.md describe the new rule correctly.

check_commits now also compares the count from %(trailers:key=Security-Floor) with raises() (lines 59-63). I found no message where the two counts differ, CRLF included.

Agreement probes

For each case I fed the PR check the body alone, as CI does (printf '%s\n' "$PR_BODY" | sh … --check-message). The release decision ran on a commit built the way GitHub stores a squash commit: title, blank line, body. I used git commit-tree in throwaway repos and ran everything with dash. 25 of 26 cases agree:

  • Merge-train bodies: an override block (BEGIN_COMMIT_OVERRIDE … END_COMMIT_OVERRIDE) with the trailer as its own last paragraph, in LF and in CRLF, passes and raises. The trailer directly under END_COMMIT_OVERRIDE in the same paragraph is refused by both.
  • Other bodies that pass and raise in both:
    • CRLF line endings, with or without Co-authored-by.
    • A Signed-off-by line, a prose line and the trailer in one paragraph (git's 25% rule applies to both readers the same way).
    • Trailing spaces, two trailers, a # heading line above the trailer, an HTML comment above it, Release-As followed by the trailer, and trailing blank lines.
  • Bodies both refuse:
    • A BREAKING CHANGE: paragraph after the trailer.
    • Prose above the trailer in the same paragraph.
    • A wrapped prose line that starts with the key.
  • CR-only line endings: both pass and neither raises. Removing CR joins the lines, so the marker disappears without an error. GitHub never stores CR-only messages, so this is a NOTE only.

R2-1 SHOULD-FIX: the PR check refuses a valid body that is a single paragraph. A PR description consisting only of the trailer block (Security-Floor: raise, or Release-As: 3.0.10 followed by Security-Floor: raise) is refused by --check-message, because interpret-trailers reads the only paragraph as the subject. The squash commit has the PR title above it, so the release reads the trailer and raises the floor. This is the only disagreement I found. It blocks a correct security-fix PR, and it never lets a bad one through. It slipped past the agreement test because test_release_and_pull_request_checks_agree (test_security_floor.py:157-160) feeds --check-message the full message including fix: x, while CI feeds it only the body.

Fix: in ci.yml, pass the title through env and check printf '%s\n\n%s\n' "$PR_TITLE" "$PR_BODY", or prepend a subject line inside --check-message. Then make the test feed the body alone, as CI does, and add the body-only case.

R2-2 NOTE: the stricter line rule refuses prose and code-block quotes. A line that starts with Security-Floor, including a wrapped, indented list item, now fails the commit. Run over the whole history, --check-commits flags c50b6886e (expected) and also 4c060836c (v0.7.25). Its body has the wrapped line Security-Floor: raise trailer signs its bundle with security_floor set.

  • That commit is long released, so the check never reads it again.
  • The same wording in a future PR fails loudly, and the PR check catches it before main whenever the description is the merge body.
  • This trade-off is acceptable: it never fails silently. AGENTS.md item 11 could add "never start a line with Security-Floor except the trailer".

Run (exit codes read directly)

  • Real history with sh and dash: v3.0.3 to v3.0.9 give (empty), 3.0.4, 3.0.4, 3.0.6, 3.0.6, 3.0.6, 3.0.9. HEAD as 3.0.10 gives 3.0.9. v1.2.0 exits 1 (expected). --check-commits e68fe3271..HEAD (the branch's own commits) exits 0.
  • make test-scripts exit 0. All 11 suites are OK, including test_security_floor.py with 22 tests. Run again with dash as sh: 22 tests OK.
  • shellcheck -s sh scripts/security-floor.sh exit 0. actionlint on ci.yml, release.yml and sim-sweep.yml exit 0.
  • Revert check: with the script at 329e85077, the delta's tests show 10 failures. Nine are agreement subtests (rule before, rule after, rule between paragraphs, CRLF, full stop, capitalised with a bang, key case and a blank before the colon, another value, no blank after the colon). The tenth is test_check_commits_refuses_an_unreadable_range.
  • The worktree was removed afterwards. Nothing was changed in the repository or pushed.

Round 3: branch A at 40d4dc0c3 (delta on 817330e4e)

Verdict for A: MERGE. R2-1 is fixed, and I found no new defect. Branch B stays MERGE.

R2-1

--check-message <title> now checks the title, a blank line, then the description from stdin, using the same misplaced() the release uses (security-floor.sh:73-79). ci.yml:199-204 passes PR_TITLE and PR_BODY only through env, and the run block quotes them as "$PR_TITLE" and "$PR_BODY". The step has no ${{ }} inside run:.

The agreement test now feeds the description alone, with the title as the argument, the way CI calls it.

Revert check: with the script at 817330e4e, 5 tests fail: test_check_message_mode, plus four agreement subtests (description is only the trailer; Release-As and the trailer; only the trailer in CRLF; trailer as the title).

My 26 bodies again, in the CI shape

The PR check got the body on stdin with the title as the argument. The release decision ran on a commit built the way GitHub stores a squash: the title with (#900), a blank line, then the body. Everything ran under dash. All 26 agree. The former disagreement, a body made only of the trailer, now passes and raises in both.

Five more cases also agree:

  • The marker as the title, with a plain body: both refuse.
  • The marker as the title, with the trailer in the body: both refuse.
  • A prose title that starts with the key ("Security-Floor handling: …"): both refuse, the same accepted trade-off as R2-2.
  • A body of Release-As then the trailer, in CRLF: both pass and raise.
  • A title holding quotes, backticks and $(…): both pass and raise.

Injection

I ran the step's run block under bash, with sh pointing to dash, using these hostile values:

  • A title holding "…", backticks, $(touch <canary>), ', and ; touch <canary>.
  • A body holding backticks, $(…), "'\, ${PATH}, and newlines, ending in the trailer.

The step exited 0 and the canary file was not created.

Three more title cases:

TitleResult
Contains newlines plus the markerrefused (rc 1). GitHub titles are one line anyway.
--check-commitstreated as a title, rc 0.
-n, with an empty descriptionrc 0.

It runs on fork pull requests because it needs no secrets.

Run (exit codes read directly)

  • make test-scripts exited 0, with 10 suites OK, including test_security_floor.py (22 tests). The same tests also pass with dash as sh (exit 0).
  • shellcheck -s sh scripts/security-floor.sh exited 0.
  • actionlint on ci.yml, release.yml and sim-sweep.yml exited 0.
  • Floors on the real history are unchanged: 3.0.4, 3.0.4, 3.0.6, 3.0.6, 3.0.6, 3.0.9 for v3.0.4 to v3.0.9, and 3.0.9 for the next release.
  • --check-commits e68fe3271..HEAD (the branch's own commits) exited 0, and the commit messages carry no attribution.

The temporary worktree was removed. Nothing was changed in the repository or pushed.

Notes that still stand from earlier rounds, none blocking:

  • A-4: a checkout without tags.
  • A-5: a maintainer's wrong lower tag can still lower the floor.
  • A-7: the PR check only reruns on a push, not on a description edit.
  • R2-2: a line in prose that starts with the key is refused.
  • CR-only line endings drop the marker silently.

Something wrong on this page? Tell us.