The snapshot run for 0.7.0 failed on this one test and published nothing.
`scanPrivateKeys` classifies a file it cannot open as not-a-key, and the test
made the file unopenable with `chmod 0o000` — which stops nobody with uid 0,
and Gitea's act_runner is a container running as root. So the file was read,
recognised as a private key not named id_*, and reported as skipped.
A dangling symlink instead: ENOENT is not a permission anyone can override,
and it is a realistic ~/.ssh inhabitant. Verified by running the gate in a
node:22 container as root, where the whole workspace is now green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two minors over 0.5.0, matching what PLAN.md proposed: 0.6.0 for the
behaviour changes through Phase 4 (recipient verification, 0600 plaintext,
passphrase never handled, overwrite confirmations) and 0.7.0 for the vault
layout finally honouring keysDir/tmpDir everywhere — which is a migration
for anyone on custom names, documented in the README.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Phase 9 of packages/keyman/docs/PLAN.md; closes AUDIT §5.2, §5.3, §5.4, §5.5,
and the §4 one-liners no phase had claimed (§4.1–§4.4, §3.7).
The README is the only document that ships (package.json files: dist,
README.md, LICENSE), and it described four of nine menu entries, invented key
rotation, told the user to run ssh-keygen by hand, asked them to write a
.gitignore keyman now writes, and mentioned none of the command line. It is
rewritten against the code: every operation, the rotate/retire sequence, the
id_ prefix and what happens to keys without it, installation with the scope
mapping (never a bare --registry, which would send 55 transitive dependencies
to a registry that has never heard of them), the configuration semantics
including which relative path resolves against what, and the Phase 5 migration
for a split vault.
The CLI section is helpText() verbatim, with tests/readme.test.ts asserting the
two are identical and that every menu label appears — so a flag or an operation
added later fails the gate instead of shipping undocumented. That is the part
that keeps this from drifting again.
Also: index.ts loses the bin's shebang (it is only ever imported), exports the
config types so a consumer can name what loadConfig returns, and re-exports the
update module wholesale rather than half of it by name — verified by importing
the built dist/index.js and reading its keys. The narrow surface is now a
comment stating the rule rather than an accident.
AUDIT.md marks all 30 findings closed except the second half of §1.8, keeping
each finding's text as the record with what closed it quoted underneath, the
way DOCS-AUDIT.md does. PLAN.md gains a status section naming the three
deviations. Root CLAUDE.md records the keyman architecture as it now is,
including the deliberate `resolution` divergence from nopy.
DOCS-AUDIT.md §2.10, §6.4 and the §7 keyman-config entry are amended in the
working tree but left unstaged, since that file carries unrelated WIP.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Phase 10 of docs/PLAN.md; closes AUDIT §3.6, the README's oldest lie
("Support for key rotation", with no occurrence of "rotat" in src/).
Rotation only ever adds. `rotateKey` generates a replacement under the next
name in the series — prod → prod-2 → prod-3 — and encrypts it *alongside*
the key it replaces, so both are in the vault at once. `retireKey` is a
separate operation, and the only one in keyman that destroys an encrypted
key. The gap between the two is where the new public key gets deployed and
tested: a rotation that replaces the key in one step locks you out of the
host you were rotating for, because the replacement is not on it yet and
the only copy of the one that is has gone.
The name has to change — the vault layout derives the directory from it, so
a replacement also called `prod` *is* the `prod` entry. `nextRotationName`
skips any version already taken in the vault, in tmp or in .ssh, so it
never asks ssh-keygen to overwrite a private key in use. Retirement warns
when nothing in the vault supersedes the key and then makes the user type
its name, since that deletion is unrecoverable.
Three things extracted rather than copied: `listVaultKeys` (vault.ts) now
backs decrypt, rotate and retire; `createKeyPair` and `promptKeyOptions`
(generate.ts) are shared with rotation, which also carries the old key's
comment over as the default. Verified against the real binaries that a
hyphen-suffixed name survives ssh-keygen and age, that the vault entry
round-trips byte-identically, and that ssh-keygen writes the replacement
0600 without help.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four things that each made keyman quietly less useful than it looked.
**Clipboard.** `pbcopy` was spawned unconditionally, with a comment admitting
it. Copy is now a list of commands per platform — pbcopy, clip, and wl-copy /
xclip / xsel tried in order on everything else, because there is no single
answer under Linux and trying them beats detecting the session type. Only an
absent tool advances to the next candidate: one that ran and refused has an
opinion. And if nothing is installed the key is printed, since "give me this
public key" is answerable without a clipboard and used to be a dead end
everywhere but macOS. Verified the round trip through real pbcopy/pbpaste.
**Home directories.** `/home/<user>` was hardcoded — wrong on the platform
this was written on. A named user is now looked for beside the current user's
home first, which is right wherever homes live together whatever that
directory is called, then in /home and /Users, and the failure names every
path tried instead of feeding a nonexistent one to readdir. For the current
user, `HOME` still wins, with `os.userInfo()` behind it: `process.env.HOME ||
''` made an unset HOME fatal, which it is not in a cron job or a container.
**Keys that are not named id_*.** A key called `deploy_ed25519` was absent
from every menu with nothing said. It still is — the vault stores
`<name minus id_>/id_<name>.age` and decrypt rebuilds the filename from the
directory, so relaxing discovery means changing the on-disk layout, which the
plan sizes as its largest single item and is not folded in here. What it does
do is say so: any file whose first line carries a private key header and whose
name lacks the prefix is now reported, per directory, with the reason. A
bounded 64-byte read, because classifying a key is no reason to load one.
**Plaintext hygiene.** A "Clear decrypted keys" entry, defaulting to no and
listing what it would delete first, and a vault `.gitignore` written on first
run covering the age identity and the tmp directory — which the README asked
the user to do by hand. Never overwritten, and silent about a configured
directory that sits outside the vault, since a .gitignore cannot speak for a
path above itself and pretending otherwise reads as protection that is absent.
Three things about .keymanrc.json.
`z.object` strips a key it does not know, so `{"vaultroot": "…"}` was
indistinguishable from an empty file: the vault stayed at the default and
nothing said why. Now warned per file, listing the known keys, because for a
casing slip naming the alternatives is most of the help. Warned rather than
fatal — this module degrades to defaults throughout — and warned inside the
per-file loop, the only place the filename exists: z.strictObject on the
merged result cannot say which file said it. The known-key list is derived
from the schema shape, so it cannot drift.
`--print-config` now includes `configFiles`, in merge order. That was the one
question it could not answer, and it existed only as unstructured stderr from
loadConfig — the wrong half of the output for it. Assembled in
describeConfig() rather than in cli.ts, which is excluded from coverage.
And the `resolution` machinery is gone: roughly 45 lines that could not change
an outcome, because every schema property is a string and both strategies
return the child's value for primitives. Its one test passed either way.
mergeConfigs is now a spread. The divergence from nopy, where the same
machinery is load-bearing, is recorded in the comment above it.
Generate prompted for the passphrase itself and passed it as `-N <value>`,
so it sat in this process's argv — readable by any user on the box through
`ps` for the length of the spawn — and in keyman's memory before that.
Verified that omitting `-N` makes ssh-keygen prompt *and* confirm, so the
prompt and the flag are both gone and the spawn inherits stdio. keyman no
longer learns the passphrase, which is strictly better than handling it more
carefully, and it deletes code.
The other half is the missing `.pub`. The selection list is built from
private keys, so an orphan is offered like any other, and copyFileSync
discovered the absent sibling only *after* age had written the encrypted
key: a vault entry with no public key, and an exception that took the rest
of the batch with it. It is now derived with `ssh-keygen -y -f`, before the
vault directory is created. Verified against real binaries that the derived
key matches the original byte for byte, that an encrypted key prompts (on
stderr — hence stdout piped, stdin and stderr inherited), and that a refused
derivation degrades to storing the private key alone rather than failing.
encrypt's loop now isolates per key and reports which ones did not make it,
except for ToolNotFoundError: age missing is not a per-key problem and nine
more identical errors help nobody.
storeInVault is the shared write path both callers had a copy of. It also
undoes its own mess: age has to write into a directory that already exists,
so a failure could leave an empty directory or a truncated .age — which list
counts as a vault entry and decrypt offers. The .age is removed because we
named it, the directory only while empty, since one holding an earlier key
is not ours to delete.
encryptKeys and decryptKeys each took `vaultDir` and rebuilt `<vault>/keys`
and `<vault>/tmp` from it, so `keysDir` and `tmpDir` in .keymanrc.json were
honoured by main and list and silently ignored by the two operations that
write. main was also passing vaultRoot where encrypt expected the keys
directory, which put encrypted keys one level above where list looks for
them: with any config at all, a key encrypted a second ago was invisible.
Both now take keysDir and tmpDir explicitly. The decrypt location prompt
names the real directories instead of the hardcoded `vault/tmp` and
`~/.ssh`, which meant its labels were also its values — hence LOCAL_MODE.
tests/vault-layout.test.ts is the regression: encrypt then list, driven
through keyman() with only age and the prompts mocked, against a config
using keysDir `encrypted` and tmpDir `plain`. Every unit suite passed
through this bug because each was told which directory to use; the seam
between them was untested. Verified it fails when main is reverted to pass
vaultRoot.
Verified before the fix: `age -d -o <existing>` overwrites without a word
("PRECIOUS EXISTING KEY" became "secret"), and the old `cp` for the public
key did the same. Decrypting a vault entry on top of a newer working key
in ~/.ssh destroyed it with no prompt, no backup and no mention. It is the
only finding in the audit that loses data the user never asked to touch.
Every collision — private and public, both output modes — is now settled
before anything is written, so the questions are asked about files that
still exist. Default is to keep what is there.
cp and chmod are gone. Three spawns per key become one, it works where
those binaries do not, and the chmod happens in-process immediately after
age returns: age creates its output 0644 regardless of umask, so a
plaintext private key was world-readable for the length of two spawns and
stayed 0644 whenever the chmod itself failed. ~/.ssh is created 0700 when
absent rather than assumed.
decrypt.test.ts stops asserting on which binaries were spawned. The age
stand-in now writes its -o file at 0644 the way age does, and the tests
assert the bytes and the mode on disk — the outcome rather than the
mechanism.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
main.ts asserted the recipient non-null twice — extractAgePublicKey(...)!
— and the type already said null was possible. With no age.key the vault
encrypted to the string "null": execa stringifies it, age exits 1, and on
the generate path that happens *after* ssh-keygen has written a plaintext
private key into tmpDir, so the user is told the operation failed and left
with a key on disk. Now the recipient is resolved once, remembered on
success, and a null prints the remedy (age-keygen -o <path>) and returns
to the menu. list, copy and decrypt still work without one.
extractAgePublicKey now derives the public key with `age-keygen -y`
instead of scraping the `# public key:` comment. The comment is ordinary
text nothing re-checks; verified that rewriting it does not change what
-y reports, so a stale or forged comment silently encrypted the vault to
a recipient nobody holds the private half of.
The comment survives as a fallback for a machine with no age-keygen,
behind a warning that it is unverified — but not when age-keygen runs and
refuses the file. That means age cannot read the identity, and trusting
the comment there would encrypt to a recipient the vault could never
decrypt with.
runTool throws ToolNotFoundError for ENOENT so the two cases can be told
apart. Its own tests move to tool.test.ts, which keeps real processes;
utils.test.ts mocks execa, since the gate cannot require age installed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
encrypt read ~/.ssh and the tmp directory, and decrypt read <vault>/keys,
with no existsSync between them. main.ts created vaultRoot and tmpDir but
never keysDir, so decrypt on a fresh vault threw ENOENT instead of
printing the "no encrypted keys" message it already had — the message was
unreachable until something else created the directory.
Both functions now fall through to their warning. main.ts creates all
three directories, 0700: the vault holds the age identity and tmp holds
plaintext private keys.
age spawns go through runTool, which separates "not installed" (ENOENT,
whose message is `spawn age ENOENT`) from "age refused" (whose reason is
on stderr and nowhere in the thrown message). Tested against real
processes, not a mocked execa — the shape of the failure is the point.
list.ts kept statSync rather than switching to withFileTypes as planned:
withFileTypes reports a symlinked key directory as a link and would have
silently dropped it. `throwIfNoEntry: false` fixes the dangling-symlink
throw and keeps following the good ones. Both cases now have a test.
Also deletes the three debug logs (encrypt.ts printed both key arrays,
decrypt.ts printed every candidate path from inside a filter).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
docs/AUDIT.md and docs/PLAN.md record the review and the ten phases it
turns into. This commit is phase 1.
keyman.cli.ts fell through to an interactive session for --help, ignored
unknown flags, and called keyman() unawaited — so Ctrl-C at any prompt,
and any rejection inside the menu loop, became an unhandled-rejection
stack trace. flagValue() also read `--channel --force` as the channel
"--force", which reached the dist-tag lookup as a key that cannot exist
and reported an unreachable registry.
New keyman.args.ts owns the parse: both --flag value and --flag=value, a
UsageError for an unknown flag or command, --channel validated against
the three real channels, and self-update-only flags rejected rather than
silently ignored. It is a separate module because cli.ts is excluded from
coverage and these are rules, not wiring. --help short-circuits before
tokenising, so it answers a line the parser would otherwise reject.
Usage errors exit 2; ExitPromptError is caught by name (@inquirer/core is
transitive here and does not resolve) and prints Goodbye.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Publishing infrastructure
- Three Gitea workflows: ci.yml (PRs, non-main pushes), publish-snapshot.yml
(main -> Gitea under dist-tag @main) and release.yml (tags -> Gitea + npmjs)
- Tag-driven releases as <package-dir>-v<version>; the manifest stays the
source of truth and release.yml refuses to run if tag and manifest disagree
- Every publish is idempotent: each step checks the registry first, so a run
that fails on the second registry can simply be re-run
- Hard coverage gate (85% branches) shared by CI, the pre-push hook and local
runs, since the thresholds live in vitest.config.ts rather than a CI flag
- README.PUBLISH.md documents the whole mechanism
Toolchain
- TypeScript 7 native compiler; drop tsgo and ts-node, use tsx for dev runs
- Biome 1.9 -> 2.x, Vitest 1 -> 4, zod 3 -> 4, inquirer 8 -> 14, pnpm 11.17.0
- Replace inquirer-checkbox-plus-prompt, which is peer-capped at inquirer <9,
with enquirer's AutoComplete; the CubeSelection contract is unchanged
- Stand in for zod 4's removed z.AnyZodObject with a local AnyObjectSchema
Repo hygiene
- Stop tracking dist/; ignore coverage/, *.tsbuildinfo, .npmrc* and release.json
- Drop package-lock.json in favour of pnpm-lock.yaml
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Benjamin Diedrichsen