Repository navigation
Releases
v1.74.2
One defect, found by the release lane 1.74.1 could not pass.
aborted()called/usr/bin/grep, which does not exist on Alpine. busybox grep lives at
/bin/grep, so every call exited 127, read as false, and an aborted scanner was filed as a
security finding on one of the three supported platforms. That is precisely the defect 1.74.1
was cut to fix. The absolute path was chosen on macOS to dodge an interactive shell that shims
greptougrep; dodging one platform's shim by hardcoding that platform's layout is how it
got in. The binary is now resolved once against the machine it is running on.- Check 66 grew a sixth lane, and it is static on purpose. The four behavioural lanes run with
PATH="$c66_bin:/usr/bin:/bin"on a machine where/usr/bin/grepexists, so they were green
through the entire life of the bug; onlytests/container-matrix.shsaw it, and that runs after
you tag. The new lane refuses a/usr/bin/<tool>path in command position, which moves the class
to every push. It matches command position rather than any mention, so the resolver loop that
lists/usr/bin/grepas a candidate and probes it with-xdoes not trip it — a check that
reddens on its own fix teaches the next person to delete the check. Row 66f puts the shipped line
back; the sweep is 125 rows. releases/v1.74.1.mdrecords that lane as red with the failing RESULT line. The tag stays where
it is and nothing was published under it.
Release lane: ten lanes green, evidence in releases/v1.74.2.md.
v1.74.0 — whitebox security lane
- A deep security lane, and a skill that drives it.
claude/whitebox-audit.shruns every
scanner installed on the machine that applies to the repository: gitleaks and trufflehog and
detect-secrets for secrets (working tree and full git history, because a key deleted in the
commit that "removed" it is still valid), semgrep and CodeQL for taint, bandit and ruff and
pip-audit for Python, npm audit and eslint and retire for JS, gosec and govulncheck and
staticcheck for Go, cargo-audit and cargo-deny and clippy for Rust, brakeman and bundler-audit
for Ruby, phpstan and psalm for PHP, shellcheck for shell, hadolint and trivy and grype for
containers, checkov and terrascan and kics for IaC, zizmor and actionlint for Actions,
osv-scanner for dependencies, syft for an SBOM. Every missing tool is skipped by name with its
install command; nothing is silently absent. - It exits 2 when nothing ran. This is the inversion that makes it safe to ship.
security-scan.shexits 0 with every scanner missing, deliberately, so the Stop-hook gate stays
green on a machine without scanners. A deep audit that does the same thing prints CLEAN over an
unexamined repository, which is a false assurance with a timestamp on it. Check 66 proves all
three directions in a throwaway repo: no scanner exits 2 and says NOT MEASURED, a stub scanner
with a finding exits 1, the same stub clean exits 0. Rows 66, 66b and 66c break each one. - The active-attack tools need two flags. nuclei, nikto and ZAP run only with
--dast URL
and--i-own-this-target, and check 66 asserts the refusal. whitebox-pentestskill. Threat model first, then the engine, then triage by reachability
rather than by whatever severity the rule author picked, then a working exploit per finding
before any fix, then fix the root cause and re-run. It names the classes no scanner finds
(per-object authorisation, auth edges, state races, SSRF, injection past the ORM, secret
lifetime, supply chain, DoS by design, agent and LLM boundaries), because that is where the
real findings are.- Two lanes were wired to exit codes their tools do not use:
detect-secrets scanalways exits 0
and needed its results read instead, andshellcheck -S warningcannot see SC2086 or SC2046,
the shell's own injection primitives, which are INFO severity. Both measured on a fixture that
was green before the fix.
v1.73.1
Release lane: ten lanes, all green on the tagged tree (releases/v1.73.1.md). Gate 66/66, falsify 117 rows across 7 shards, install-matrix and container-matrix green.
A benchmark fixture that separates the arms. traps/gated_report_quiet ships a real .claude/verify.sh and a prompt that never mentions it. Ten paired runs per arm on Haiku 4.5: vstack and gstack 10/10 green with no false completion, pstack 9/10 with one run that edited the fixture's own contract-check script and then said DONE. First non-tie on correctness in this benchmark. Every number is in tests/evals/showcase/RESULTS.md and the run files behind it.
Two release-lane harnesses carried the same logical-path assumption 1.73.0 fixed everywhere else. tests/compare-baseline.sh armed the trust store with a logical fixture root, so its one mechanism row read no decision where the gate really decides block. tests/test-breadth-mandate.sh PROOF 25 still asserted that a banned register opener blocks; 1.72.0 made a lone register strike warn. Nothing caught either drift, because the release lane is the only caller of both harnesses and had not run since 1.71.0.
Harness fixes: tests_tampered was constant 1 because it diffed __pycache__; benchmark workdirs are now physical paths outside the checkout and every arm is git init-ed; the trust lock is exported to the parallel workers.
Install: `VSTACK_REF=v1.73.1 bash bootstrap.sh`
v1.71.0
1.71.0 — 2026-09-03
claude/CLAUDE.md4332 B to 1635 B: USE THE STACK, MODEL ROUTING, DISPATCH PRE-AUTHORIZED
and ISOLATE THE WRITERS deleted, FAN OUT and NAME THE AGENT cut to one line each, REGISTER
keeps only the banned-word list. Each cut names its evidence in
docs/research/harness-effect/findings/plan-next-mechanisms.md. DOGFOOD moves to the
repo-root CLAUDE.md, where it applies.- Per-prompt digest loses its unconditional TOKENS, DELEGATE and FANOUT lines and SessionStart
loses TOKENS/DELEGATE/AUTONOMY/PLAN MODE (989 B); the SKILLS routing table stays. Check 18
now probes the fired form (400-character prompt plus a seeded mandate marker) and requires
both lines; row 18e stubs the grill branch. Check 33's must-speak probe uses the fired form. swarmbriefs carry TASK, FILES, ACCEPT and REPORT;workerandexplorerreturn BLOCKED
naming the missing field.dispatch-counter.shrecordskind: skill_loadrows withduration_mson Skill calls
(matcherAgent|Task|Skillinclaude/settings.jsonandinstall.sh);bin/doctor --ledger
sums them per session. Check 44 requires the matcher to name Skill in both wirings; row 44i
drops it from the installer.- Mechanisms ledger:
skill-load-ledgeradded,per-prompt-digestre-justified, both
unmeasured with decide-by 2026-09-17. - Check 65 in a depth-1 clone named a measured commit "not an ancestor" when the clone
simply had no history; the 1.70.0 container lane went red on all three images for that. It
now names the shallow clone, andtests/container-matrix.shdeepens its clone before verify.
Release-lane evidence: https://github.com/itsvedantkumar/vstack/blob/main/releases/v1.71.0.md#release-lane-evidence
v1.68.0
1.68.0 — 2026-09-03
overlay.sh --checkreports drift without writing. Every file the write path would
touch is compared against the payload and reported asstale,absent,legacyor
differs (repo-owned); exit 1 on anything the nextoverlayrun would change. Overlaid repos
can now learn that the template moved. (#10)tests/inventory-contract.sh --writere-points the inventory. Replaces the hand-run jq
after every payload commit; refuses on a dirty tree, re-validates, prints the commit line, and
is a no-op when the contract already holds. (#9)
v1.67.0
1.67.0 — 2026-09-02
- A security toolchain ships with the overlay.
claude/security-scan.shruns gitleaks,
semgrep, osv-scanner, zizmor and eslint locally and is copied into every overlaid repo, where
theverify.shtemplate calls it; a scanner that is not installed skips rather than failing, so
the gate does not go red on a machine missing five binaries.claude/security.yml.tmpland
claude/dependabot.yml.tmplare seeded once into.github/workflows/security.ymland
.github/dependabot.ymland never overwritten — a secondvstack overlay .reports whether the
file still matches the template or has drifted from it, and names the file to diff against.
setup-machine.shinstalls the scanners by default. - Check 64 measures it in both directions. The scan is run in a throwaway repo with
PATH=/usr/bin:/bin(every tool must skip, exit 0) and again against a stub gitleaks that
reports a finding (must exit 1 naming it), because a scan that can only ever skip passes the
first lane perfectly. 64b requires everyuses:in the workflow template to be pinned to a
40-hex commit with its tag beside it, and requires README's Prod-ready gates table and the
script to name the same tools. Falsifiability rows 64 and 64b delete the skip branch and unpin
one action.
v1.66.0
1.66.0 — 2026-09-02
Auto-enforcement tightened after measuring it. Across ten real sessions in the last ten days the
Stop gate got 94% compliance whenever it fired, but it fired on almost no turns: 72% of turns ran
with zero Agent and zero Skill, 65% of edit-turns had neither, and one active workspace ran with
the hooks not reaching it at all. The gate was not loose; it was retiring early and skipping cases.
- Skill mandates now re-arm. unslop, typescript-best-practices, prove-it-works and register
used a plain counter with no window, so two misses latched each one silent for the rest of the
session. They now re-arm afterVSTACK_SKILL_RESET_SECS(default 1800), exactly the windowed
reset the four delegation mandates have run on since 1.63.0 without over-firing. This is the
mechanical cause of "the gate stops triggering." - The breadth mandate gates two-directory work. The fan-out mandate required
dir>=3 AND ext>=2; a two-directory, two-extension cross-cutting edit was never gated. Lowered todir>=2 AND ext>=2. The AND with the extension floor stays, so a three-file edit inside one file type
is not gated, because a guard that nags a legitimately serial change gets switched off.
tests/test-breadth-mandate.shgains PROOFs 27 and 28 for the new boundary;tests/ delegation-drift.py's eligibility mirror tracks the same threshold.
v1.65.1
1.65.1 — 2026-09-02
- bootstrap.sh resolves the latest release without the GitHub API. 1.65.0 made an unpinned
bootstrap look upreleases/lateston api.github.com, unauthenticated, and refuse when that
returned nothing. On the first push after the release the Linux install-matrix job did exactly
that: a shared Actions runner IP is rate-limited on that endpoint, the resolver came back
empty, bootstrap refused, and main went red while the macOS job with identical code was
green. The resolver now asksgit ls-remote --tagsfirst (git protocol, no API quota), picks
the highestvX.Y.Znumerically (nosort -V, which busybox lacks), and falls back to the
API only without git, sendingGITHUB_TOKEN/GH_TOKENwhen one is set. Refusal without a
fallback to main stays. The install matrix's curl lane now carries bootstrap's last output
lines into its FAIL line, because the cause of this red was captured and thrown away, and
the three CI install-matrix steps passGITHUB_TOKENthrough.
v1.65.0
1.65.0 — 2026-09-02
The public-readiness audit's three P2s, plus the leftovers it turned up.
-
bootstrap.sh no longer installs the default branch when unpinned. It resolves the
latest release tag from the GitHub releases API and refuses when that lookup fails, rather
than falling back tomain. A stranger who runs the one-liner now gets the version the
manifests and changelog describe, not whatever landed since. -
The context7 MCP server is pinned (
@upstash/context7-mcp@4.0.4) instead of floating on
npx's latest, so a re-install cannot silently pick up a different server than the one that
was reviewed. -
uninstall.sh leaves nothing behind. It removes
vstack-compat-canary.json(written by
the compat-canary hook, so ownership follows the hook, not the installer) andrmdirs the
now-emptyagents/reference,agents,commands,hooksandskillsdirectories,
deepest first, so a directory holding anything foreign is left alone. The install matrix's
uninstall-cleanlane fires the real hook first so the file genuinely exists before it
asserts on its absence. -
The session baseline names the operator.
inject-session-context.shderives the
human's name fromgit config vstack.operator, falling back togit config user.name, and
says that RICK is the lead's call sign, not theirs. A reply opening "Rick:" had been read as a salutation. Nothing shipped carries a
literal name. -
README puts the install command on the first screen and states what distinguishes this
repository before the counts. -
Repository settings: secret scanning and push protection enabled; three stale branches
deleted. -
install.sh no longer arms the Stop-hook verify gate on a non-interactive install with no
opt-in. It ranbin/vstack trust "$SRC" --yesunconditionally, so bootstrap.sh's
curl|bashone-liner — which drives install.sh non-interactively — left a stranger who never
read a line of the repo with standing unattended-execution rights over.claude/verify.sh
(which itself runsinstall.sh --dry-runandoverlay.sh). Every other repo's gate stays off
untilvstack trustruns there and its terminal confirmation is answered; a bare
non-interactive install of this repo now gets the same refusal. Arms on an interactive
terminal ([ -t 0 ]) or an explicit opt-in (--trust, orVSTACK_TRUST=1); otherwise prints
how to arm it later. Check 61b in.claude/verify.shproves the non-interactive,
no-opt-in path leaves the trust store without this repo's entry — regression-tested against
git show origin/main:install.sh, which fails the same assertion.
v1.64.0
1.64.0 — 2026-09-02
- format.sh no longer mistakes a prettier devDependency for a prettier config. The grep
for"prettier":matched the version pin in dependencies/devDependencies, won the
package.json slot ahead of a sibling .prettierrc, and cosmiconfig silently fell back to
built-in defaults — files reformatted to the wrong style. The hook now JSON-parses
package.json for a genuine top-level key, falling through to .prettierrc when node is
absent or the file does not parse. Repro:tests/repro/formatter-config-devdep.sh.
(PR #7)