Files
gstack/docs/state-root.md
T
669e59d65b v1.91.206.0 feat(pregate): user-owned policy file and local lanes (port of #3286) (#3398)
* feat(pregate): user-owned policy file and local lanes

A repo that keeps .gstack/ out of git can hold its pregate policy at
<GSTACK_HOME>/projects/<slug>/pregate.json. gstack-pregate run reads it only
under --policy-from base when the base carries no .gstack/pregate.json, so a
reviewed base copy always wins and a PR still cannot reach the policy. A
missing file falls back; any other read error, and a non-null sealed or
price_file (read only from the git copies), refuse with PREGATE_CONFIG_INVALID
and a fix that names the file.

A new `local` key lists <workflow.yml>/<job> lanes a local run satisfies: a
platform lane named there runs here (tier unit) instead of becoming a
requires-remote obligation. Lane ids share one LANE_ID check with selection.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(pregate): the user-owned policy file and local lanes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): policy fixes name the user-owned file when it is the active policy

With no .gstack/pregate.json on the base, an edit to that path is never read,
so the no-lane, runner, stall_timeout_s, remote_gate and strays fixes now name
ctx.configFile: the user-owned path when it is the active policy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(pregate): the remote_gate, strays and stall fixes name the active policy file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): pre-landing review fixes for the user-owned policy

Ignore a user policy path that is relative or inside the checkout, fingerprint
the active user file into check inputs, name the user file in its JSON errors
and the refusal lead, add --explain lines for an absent user file and an
unknown local id, and document the state-root location and slug lookup.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): containment edges and explain wording for the user policy

A state root under an in-repo directory named ..something passed the
inside-the-repo check, and realpath kept the caller's casing on macOS.
The unknown-local explain line now says workflow lane, and the state-root
chain in explain and docs names every source.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): a base copy git cannot read, or a dangling user symlink, refuses

Any git show failure read as no base policy, which now hands the run to
the unreviewed user file; it is absent only when cat-file says so, and a
spawn error (stdout cut at the buffer limit) counts as a failure. A user
pregate.json that is a dangling symlink refuses instead of reading as
absent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): decide a missing base policy from the tree, and catch dangling parent links

cat-file -e checks the blob, so a missing or corrupt object still read as
no base copy; ls-tree reads the tree entry, and any failure refuses. The
dangling-symlink check now covers every path component, not just the file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(result-codes): POLICY_SOURCE_UNAVAILABLE names an unreadable committed pin file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(pregate): the active policy may be the user-owned file; new refusals

Guards, remote_gate, stall timeouts and lane fixes name the user-owned file
when it is the active policy; the trust boundary says run alone reads it;
the dangling-symlink and unreadable-base-copy refusals are documented, and
state-root lists the bucket's pregate.json.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): a checkout's .env cannot pick the user policy; verify re-checks it

Bun loads .env from the working directory, so a branch could set
GSTACK_PROJECT_SLUG or a state root and choose which pregate.json governs
it. The shebang now passes --no-env-file, scriptInvocation carries that
flag over when it launches the bin as bun <script>, and a bare bun launch
beside any dotenv file refuses (POLICY_SOURCE_UNAVAILABLE).

A recorded pass also outlived its policy: pregate.json now records the
user file's digest (or, under the defaults, the path that was absent), and
verify and the ship receipt treat an edited, deleted or newly present user
file as PREGATE_STALE, reading both paths from the record.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): keep the POLICY_SOURCE_UNAVAILABLE summary's pin-files wording

The added '(or .gstack/pregate.json)' broke describeRefusal's rewrite to
'the reviewed pin files cannot be read', which pregate-wtree-refusal pins.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(pregate): pick the host-foreign lane in the user-policy local-lane test

On a Windows host the windows lane is native, so it never became
requires-remote and the test failed on the Windows shards.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): Rust crate lanes and nested JS package runners

lanes reported PREGATE_NO_LANE for every touched .rs file and ran a nested
pnpm+vitest package's tests as `bun test` from the repo root.

- Rust: a .rs file of a crate (nearest Cargo.toml with [package]) is a test
  under <crate>/tests/ or when it holds a #[cfg(test)]; every .rs file of a
  crate and its Cargo.toml select the crate's tests and those of every crate
  depending on it (lib/pregate/rust.ts). Selected tests run per crate as
  `cargo test -p <name>` from the workspace root, judged on the summed
  `test result:` lines (doc-tests included); zero tests stays ZERO-RUN and a
  missing cargo is incomplete, like a missing Dart SDK.
- JS/TS: a test owned by a nested package.json that declares vitest or jest
  runs from that package with package-relative paths, through the package
  manager its packageManager field or nearest lockfile names (pnpm exec,
  yarn, bun x, else npx). Root detection and a declared runner are unchanged.
- init notes that Rust tests run per crate; docs/pregate.md updated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(pregate): parse Cargo manifests as TOML; follow workspace membership

The line scan of Cargo.toml missed dependents (a false green) and picked the
wrong workspace:
- an inherited renamed dependency (`base = { workspace = true }` with
  `[workspace.dependencies] base = { package = "alpha", path = ... }`);
- dotted keys (`alpha.path = "../a"`), `[dependencies.<name>]` tables and
  target-specific tables, matched by key name only;
- a crate the workspace excludes ran `cargo test -p` from the workspace root.

Manifests are now parsed with Bun.TOML. A dependency resolves by path
(relative to the manifest declaring it, or to the workspace root when
inherited), else by package name. A crate runs from the workspace it is a
member of under Cargo's rules (package.workspace, the nearest [workspace],
members globs, exclude, literal members, path dependencies of members),
else from its own directory. Nothing runs cargo.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(pregate): the Rust fixture's cargo shim wins PATH on Windows and beside a real cargo

windows-free-shard (5) failed 'a touched src file runs its crate...' and
'zero tests is ZERO-RUN...' deterministically: the shim PATH was joined with
':' so Git Bash never saw the shim dir, and the runner's real cargo ran
against the fixture (exit 1, no test result lines). On POSIX hosts where
Bun and cargo share a directory (/usr/local/bin on Capy's image), the lane
puts Bun's directory first and the real cargo wins the same way.

Join with path.delimiter, add cargo.cmd for Bun.which on Windows, and on
POSIX put the shim first again in the lane's own bash through BASH_ENV.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): quote the crate name in cargoRunner

The crate name is read from the repo's Cargo.toml and was interpolated into
'cargo test -p <name>', which the lane runs through bash. A name like
'evil;touch pwned' ran the second command. Quote it like every file and cwd
the lane already passes; a new end-to-end case checks the name reaches cargo
as one argument and runs nothing.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(pregate): the nested-package pnpm shim wins PATH beside a real pnpm

'a touched file in a nested pnpm+vitest package runs vitest there' ran the
host's real pnpm wherever Bun and pnpm share a directory (/usr/local/bin on
Capy's image): the lane puts Bun's directory first, so pnpm installed vitest
from the network, dirtied the fixture tree and the check came back
incomplete. Put the shim first again in the lane's bash through BASH_ENV,
as the Rust fixture does.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(changelog): pregate Rust crate lanes and nested package runners (#3344)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): a crate name Cargo would reject never reaches the lane's shell

Port review (outside voice): quoting alone was not enough, because the lane
expands {files} after the runner string is built, so a crate named
'evil{files}' spliced the quoted file list into the quoted name and a file
name like tests/$(cmd).rs then ran. A name outside Cargo's grammar
(letters, digits, _ and -) now gets no runner: its files are unclaimed and
the lane is incomplete, as Cargo itself would refuse the manifest. A valid
name, Unicode included, still runs quoted.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): match Cargo members globs per path component, without a backtracking regex

Port review (outside voice): members globs went through ship-policy's
matchesAny, which escapes [...] classes (crates/[ab] matched nothing, so the
crate was no member, lost its inherited [workspace.dependencies] path and its
dependents' tests were never selected: a false green) and backtracks
exponentially on a hostile glob such as '*a*a...*ab' (over 5 s for one crate,
before any runner deadline). A small dynamic-programming matcher handles *,
?, [..] with negation and ranges, and ** across components, in
O(pattern x path).

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): a plain #[test] in src is a Rust test; bound the cfg attribute scan

Port review (outside voice): only #[cfg(test)] modules and tests/ files
counted, so a dependent crate whose tests are plain #[test] or #[tokio::test]
functions in src never ran when its dependency changed (a false green). The
cfg scan also went quadratic on a run of unclosed '#[cfg(' (about 3 s on
120 KB). Test attributes now count, and the cfg predicate and its strings are
length-bounded so the scan stays linear.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(pregate): Rust test attributes, workspace membership and rejected crate names; changelog for the port review fixes

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): match members globs by code point in one pass, and admit every crate past a size cap

Second port review round (outside voice): the matcher counted UTF-16 units, so
'crates/?' missed a crate directory named with a supplementary character and
dropped its inherited dependencies (a false green); and each unmatched '['
rescanned the rest of the glob (5 s for one crate with a million of them).
Tokenize by code point with a precomputed next-']' table, and past 4096
characters of members globs admit every crate under the workspace: a wrong
member fails loudly in cargo, a dropped one is silent. The new supervised
probe kills a super-linear matcher at 30 s instead of hanging the file.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): scan cfg attributes once without truncating them; an unclosable one counts as a test

Second port review round (outside voice): the length bound from the first
round dropped a valid cfg(any(test, feature = "<257 chars>")), so its crate's
tests were no longer selected (a false green). Read each attribute once to
its closing ']' and resume after it, which keeps the scan linear without a
bound; an attribute the scan cannot close (an unterminated string, no ']')
counts as a test, which selects more, never fewer.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(changelog): second port review round (code-point member globs, unbounded linear cfg scan)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): past the members size cap, a crate's run root is undecided and its tests unclaimed

Third port review round (outside voice): admitting every crate past the cap
made a non-member run as 'cargo test -p <name>' from the workspace root,
where a same-named member answered and its passing tests stood in for the
orphan's failing one (a false green). Keep the conservative membership for
dependency edges, but leave such a crate's workspace undecided (null):
planRuns leaves its files unclaimed and the lane incomplete.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(pregate): the Rust end-to-end runs put the cargo shim under Windows' own Path key

windows-free-shard (5) still failed both end-to-end cases at fe99b200 with
the same symptom (the runner's real cargo, exit 1, no test result lines).
runBin merges process.env, whose Windows key is 'Path', with the shim's
'PATH', so the child got two path variables and read the original. Write the
shim under the key process.env actually uses, as test/helpers/unreachable-gbrain.ts
and sync-gbrain-readiness-fixture.ts do.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(pregate): run the cargo stand-in cases on POSIX only

windows-free-shard (5) still failed both end-to-end cargo-lane cases after the
delimiter, cargo.cmd and Path-key repairs: the lane's Git Bash exits 1 with an
empty log, and that log is not visible from CI. Skip the two cases that run the
bash stand-in on win32 (test.skipIf, as other POSIX-shim pregate fixtures are
handled) and drop the unproven cargo.cmd and Path-key changes. Selection,
planning and the hostile-name case still run on Windows.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(pregate): TODO(#3380) on the two win32-skipped cargo-lane cases

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): refuse a user policy whose link chain passes through the checkout

userConfigRejection checked the lexical path and the final realpath only. A
state root linked into the checkout, or a link hop inside it, let a branch pick
which existing out-of-repo policy governs it. Every hop of the resolution chain
is now checked. Found by the security review of #3286.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): verify goes stale when a link on the way to the absent user policy dangles

A fresh run refuses a user policy path with a dangling parent link, but verify
looked only at the leaf, so the parent's ENOENT read as continued absence and
an old defaults pass stayed publishable. The loader and verify now share one
absence check. Found by the security review of #3286.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(pregate): CHANGELOG entry for #3286 and the link-chain and dangling-parent rules

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* v1.91.206.0 stamp

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

---------

Co-authored-by: arlophoenix <arlo.phoenix@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vilano93 <244097767+vilano93@users.noreply.github.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
2026-10-25 10:50:18 +00:00

10 KiB

Where gstack keeps its state

gstack keeps everything it remembers — config, analytics, sessions, projects, learnings, review logs, the egress ledger, the trust-policy store, hook error logs — under one directory, the state root. By default that is ~/.gstack. Set GSTACK_HOME to put it somewhere else.

One rule picks the root, and every gstack script, hook and skill uses it. The rule lives in two owner files that a parity test keeps identical: bin/gstack-state-root.sh (bash) and lib/state-root.ts (TypeScript).

How the root is chosen

The first variable in this list with a non-empty value wins. Empty values count as unset.

Order Source When to use it
1 GSTACK_STATE_ROOT gstack-paths' own output. It is also honored as input, so re-reading it in the same environment returns the same root. Do not set it by hand.
2 GSTACK_HOME The one variable to set when you want your state somewhere other than ~/.gstack.
3 GSTACK_STATE_DIR Legacy alias, honored for compatibility.
4 CLAUDE_PLUGIN_DATA Only when CLAUDE_PLUGIN_ROOT contains gstack (any case), so another plugin's data directory never captures gstack state.
5 $HOME/.gstack The default. On Windows shells, USERPROFILE stands in when HOME is unset.
6 .gstack Last resort when there is no home directory at all (some containers).

See which root is in use

~/.claude/skills/gstack/bin/gstack-paths --explain

Real output with GSTACK_HOME set while ~/.gstack still holds older state:

state root: /home/you/work-state (selected by GSTACK_HOME)
chain (first non-empty wins):
  GSTACK_STATE_ROOT    unset
  GSTACK_HOME          /home/you/work-state  selected
  GSTACK_STATE_DIR     unset
  CLAUDE_PLUGIN_DATA   unset
  default              /home/you/.gstack  ignored
default root /home/you/.gstack also holds gstack state: yes
merged privacy keys (most restrictive value across roots wins):
  telemetry:        off (from /home/you/.gstack/config.yaml)
  memorable_recall: not set (default applies)
  codex_reviews:    not set (default applies)
  update_check:     not set (default applies)
docs: https://github.com/garrytan/gstack/blob/main/docs/state-root.md

--explain only reads; it writes nothing and exits 0. gstack-config list also prints a one-line note when GSTACK_STATE_ROOT, GSTACK_HOME and GSTACK_STATE_DIR name different directories.

Move your state

gstack never moves state for you. To relocate it:

cp -a ~/.gstack /new/place/gstack-state
export GSTACK_HOME=/new/place/gstack-state   # add this to your shell profile
~/.claude/skills/gstack/bin/gstack-paths --explain   # confirm: selected by GSTACK_HOME

Keep or delete the old ~/.gstack once you have checked the new root works. Until you delete it, its privacy settings still count (next section).

Privacy settings never get looser when the root moves

Four config keys are opt-outs: telemetry (off < anonymous < community), memorable_recall (off < on), codex_reviews (disabled < enabled) and update_check (false < true). gstack reads each of them from the resolved root and from ~/.gstack, and uses the most restrictive value. So turning telemetry off once in ~/.gstack keeps it off after you set GSTACK_HOME. Every other key, including proactive and founder_resources, reads the resolved root only.

When gstack-config set writes a merged key but the other root still holds a more restrictive value, it tells you which root overrides it and prints the command that changes that root, for example:

GSTACK_STATE_ROOT='/home/you/.gstack' ~/.claude/skills/gstack/bin/gstack-config set codex_reviews enabled

gstack-config list shows the winning root for merged keys, for example telemetry: off (set, /home/you/.gstack).

Trust-policy deny tiers merge the same way: a deny or read-only entry in ~/.gstack/gbrain-repo-policy.json still applies when your resolved root is elsewhere. Nothing is ever written to the other root.

Project buckets

Per-project state (checkpoints, learnings, timeline, question preferences) lives in projects/<slug>/. bin/gstack-slug picks the slug from the git origin remote:

  • A 2-segment remote (github.com/owner/repo) files under owner-repo.
  • A remote with 3 or more path segments (GitLab nested groups such as gitlab.com/customer-a/product/repo, Azure DevOps) files under product-repo-<16 hex>. The hex is a digest of the canonical remote, so customer-a/product/repo and customer-b/product/repo get separate buckets, and the ssh, https and ssh-with-port spellings of one repository share one.
  • Without a remote, the project root's directory name is used.

A bucket can also hold a user-owned pregate.json policy, for a repo that keeps .gstack/ out of git (see pregate v1).

The first slug gstack resolves in a directory sticks (it is cached under slug-cache/), so a project keeps its bucket when it adds a remote later.

Upgrading from a version that used one bucket for nested groups. Earlier versions filed every <group>/product/repo under product-repo. That bucket stays where it is; gstack neither moves nor deletes it. The first time gstack resolves the new slug it prints one notice naming both buckets, and /context-restore reports the old bucket while it still holds checkpoints. To bring earlier files across, run this in the project:

~/.claude/skills/gstack/bin/gstack-slug --adopt-legacy

It lists every file in the old bucket. Checkpoints are marked as matching this project, another project (never copied), or unknown (saved before checkpoints recorded their project). Files with no project identity (question preferences, learnings, timeline) may belong to any project that shared the bucket, so they are listed for you to decide. On a terminal it asks per file; otherwise copy chosen files with gstack-slug --adopt-legacy --copy <path> [<path>...]. Nothing is copied without your confirmation, and nothing in the new bucket is overwritten.

Pinning a slug. Set GSTACK_PROJECT_SLUG=<name> to choose the bucket yourself, for example in the project's .claude/settings.local.json ("env": {"GSTACK_PROJECT_SLUG": "<name>"}). Use it when an ssh host alias (git@work-gitlab:group/sub/repo) gives a different host than the https remote, or for a vendored sub-repo that should be its own project. The pin is applied per run and never cached.

Ephemeral machines

A Capy cloud machine, a CI runner or a throwaway VM loses its disk when the run ends, and with it every learning, decision, review-log row and timing record gstack wrote under the default root. gstack does not guess this from the path (a tmpfs and a home directory look alike); the host says so with one environment marker:

GSTACK_EPHEMERAL=1

What the marker changes:

  • gstack-doctor and gstack-doctor --check report the state root row as durable=no (the dashboard warns, the check still passes: an ephemeral root is a fact, not a broken install).
  • Unattended sessions print learnings: skipped (state root is ephemeral; set GSTACK_STATE_ROOT) instead of writing silently to a disk that will vanish.

A durable root is opt-in: the owner sets GSTACK_STATE_ROOT (or GSTACK_HOME) to a directory that outlives the machine. The installer never picks one for you. On Capy the drive is governed by Capy's own drive policy, which excludes memory and verification evidence, so bin/gstack-capy-install only prints the option and never writes there.

A child shell's export cannot configure the parent's later commands (lib/state-root.ts reads each process's own environment), so the installer ends with one machine-readable line the parent supplies to every gstack command it launches:

GSTACK_LAUNCH: GSTACK_SESSION_KIND=unattended GSTACK_STATE_ROOT=<path> GSTACK_EPHEMERAL=1

Prefix each command with those assignments (or export them once in the shell that runs them). See docs/capy-install.md.

Uninstall

gstack-uninstall deletes state only at ~/.gstack.

  • It first refuses (exit 2) when ~/.gstack resolves — following symlinks — to /, your home directory or one of its parents, the gstack checkout, the current git repository, or a parent of either. The message names the reason and a fix: line.
  • When the resolved root is somewhere else (you set GSTACK_HOME, for example), uninstall leaves it in place and prints left in place: <path> (selected by <VAR>=<value>) plus the exact rm -rf -- '<path>' command to run after you have checked it.
  • --keep-state never deletes any state. --force only skips the prompt.

If gstack-paths is missing

Skill blocks resolve the root with

GSTACK_STATE_ROOT=$(~/.claude/skills/gstack/bin/gstack-paths --get GSTACK_STATE_ROOT); : "${GSTACK_STATE_ROOT:?gstack-paths failed; reinstall with ./setup or /gstack-upgrade}"

(No eval: worktree-isolated Claude Code sessions refuse it.)

The second half stops the block instead of writing under / if the resolver is broken. If you see gstack-paths failed or cannot resolve the gstack state root, reinstall: run ./setup in the gstack checkout, or /gstack-upgrade.

Claude Code plugin installs

gstack does not ship an official Claude Code plugin distribution today (no .claude-plugin manifest has ever been on main; some forks publish one), and telemetry does not report plugin-mode installs. When gstack does run as a plugin, CLAUDE_PLUGIN_ROOT names gstack and state lives in CLAUDE_PLUGIN_DATA, while a terminal run outside the plugin session uses ~/.gstack. Those are two different roots; run gstack-paths --explain in each environment to see which one is active. Cross-environment support (a pointer from ~/.gstack to the plugin root) is deferred until a plugin distribution exists.

For contributors

  • Bash: executables source bin/gstack-state-root.sh beside them and call gstack_state_root_select (sets $_gstack_sr_root), or eval gstack-paths with the guard above. Hooks must source the twin; never spawn gstack-paths from a hook.
  • TypeScript: resolveStateRoot() and readConfigKey() from lib/state-root.ts.
  • test/state-root-ratchet.test.ts fails on a hand-rolled chain, an executable ~/.gstack in a template bash block, an unguarded gstack-paths eval or export GSTACK_STATE_ROOT in prose. Its allowlist (test/state-root-ratchet.allowlist.json) needs a reason per entry.