Add a --json flag to `secrets list` that emits a structured object for
tooling/CI instead of the human table — feeds the EGB-671 install scripts,
which need to enumerate a cloned store programmatically.
Contract: {"store", "projects":[{"name","entries":[...]}]}, each entry
self-describing via a type discriminator — {type:dotenv,path} or
{type:external,subtype:properties|file,path}. cmd_list_json mirrors the same
recursive store walk as the human list (nested <project>/<relpath>.age +
external/<slug>.age); jq assembles the JSON so paths escape correctly and
stdout stays pure JSON (the non-default-store hint is suppressed; jq is a
hard dep only in --json mode).
Tests: 7 new bats cases (dotenv, nested relpath, external properties + file
subtypes, empty store, pure-stdout-under-notice, store path). Full suite
261 pass / 0 fail.
VERSION 0.7.1.0 -> 0.7.2.0; CHANGELOG/README/CLAUDE.md updated.
Completes EGB-700 — the store-format line (format: vN) shipped in v0.6.0.0;
this adds the manifest schema version to the `which` manifest header
(version N, validated == MANIFEST_VERSION by _check_manifest_file). Now a
single `secrets which` surfaces both the store format and the manifest
version for the dual-format debugging window. 1 bats test. Suite 237/237.
Pre-landing review (1 critical, 4 informational):
- CRITICAL: per-project `secrets verify` during the migration window flagged
the freshly-written v2 twin as a spurious orphan and exited 1 (store still
reads v1, so verify's expected set only held the .gradle-properties.age name).
Breaks the documented migrate->verify->finalize workflow and CI. Fix: verify's
orphan set now accounts for BOTH suffix forms of a properties external, so the
twin is never a false orphan mid-migration. Regression test added.
- Copy-forward now pushes the twins (mirrors push/rekey) so a --finalize on
another machine sees them; previously twins were local-only until finalize,
a multi-machine footgun.
- Dropped the undocumented `--force` alias (keep `--yes`).
- Clarified the EGB-700 comment (which-format line, folded into EGB-703).
Deferred to EGB-701: the two finalize find-walks over *.gradle-properties.age
could collapse to one pass. Full suite 236/236.
Coverage audit found 4 untested branches (all single-test fills): finalize on
a v1 dotenv/file-only store (v1count==0 stamp path — a real user scenario),
the migrate positional-arg die, finalize on an already-v2 store, and the
_store_format strict-parse defense (garbage marker reads as v1). Full suite 235/235.
Stage 2 of EGB-677. Makes the store self-describing and unifies the legacy
external `properties` blob suffix, via a copy-forward migration that never
destroys data until an explicit, gated finalize.
Scope decision (see eureka): the EGB-677 CEO plan's "flatten dotenv blobs to
basename" was dropped as LOSSY — it discards the restore relpath that makes
the store self-describing and adds basename collisions. Engineering analysis
(4 parallel design agents) showed the store is already relpath-self-describing;
the only real v1→v2 delta is the `properties` suffix. This implements the
minimal, safe v2 that achieves the epic's self-describing goal.
What's added:
- `.secrets-format` marker (committed, one line `2`). Absence ⇒ v1 (every
pre-EGB-703 store). `_store_format()` reads it; `init` stamps fresh stores
born-v2. `secrets which` prints `format: vN` (EGB-700 folded in).
- `_external_blob_suffix(type)` — single source of truth for the external
suffix (v2: gradle-properties → properties; file unchanged). push/pull/verify
all route through it, so v1 and v2 stores never disagree on blob location.
- `secrets migrate` — per-project copy-forward (writes `.properties.age` twins
beside v1 blobs; idempotent; needs the project manifest), `--dry-run`
(reports old→new, writes nothing), `--finalize` (store-wide, the only
destructive step: gates on `verify --all` green + every v1 blob twinned,
cuts a `pre-v2-migrate-<sha>` recovery tag, stamps the marker, then drops v1
blobs; refuses without `--yes`/operator confirmation).
rekey and verify --all stay format-agnostic (recursive find walk) — no change.
21 new bats tests (test/migrate.bats): marker/born-v2, format-aware suffix,
v1 back-compat, dry-run, copy-forward idempotency, no-manifest die, finalize
gates (verify-not-green refusal, untwinned refusal, recovery tag, confirmation),
and full v1→window→finalize round-trip. Updated 4 existing tests for the
born-v2 suffix. Full suite 231/231, bash 3.2 clean.
Pre-landing review (0 critical, all informational) auto-fixes:
- Tighten external-corrupt test to assert the decrypt-fail message, not any
external finding (was *"external"*, now *"does not decrypt"*).
- Pin the verified-count in the nested+external happy-path test so a silent
under-count (exit 0 while skipping a blob) is caught.
- Account for an unsafe dotenv entry in `expected` so a matching stray blob
isn't double-reported as both unsafe and orphan.
- Discard the unused external `keys` read field (read -r etype epath _).
- Document the optional [project] positional in the README verify row.
Deferred to EGB-701 (stage-2 dedup): the external blob-path literal and the
find-walk overlap with cmd_rekey/cmd_list. Full suite 210/210.
Coverage audit found 5 untested branches in cmd_verify (all single-test
fills, no logic defects): external blob decrypt-failure (only the missing
case was covered), the rail-skip finding for an unsafe dotenv path in the
manifest, the unknown-flag die, a malformed manifest through the verify
entry point, and the empty-store 'verify --all' no-op. Full suite 210/210.
Read-only integrity check, the safety net for the stage-2 store migration.
Default mode (current project) cross-checks $PWD/.secrets.json against the
store both ways — declared-but-missing blobs and orphaned blobs (no manifest
entry) — and decrypt-tests every dotenv + external blob with the current key,
streaming plaintext to /dev/null so nothing is ever written to disk. `verify
--all` decrypt-tests every blob in every project (integrity only; the store
carries no manifests, so consistency can't be checked store-wide). Both
recurse the whole project tree (find -type f), the same walk rekey/list use,
so nested manifest blobs are covered. Exits non-zero on any finding so it can
gate `migrate --finalize` and CI.
12 bats tests (clean, nested+external, missing blob, decrypt failure, orphan,
missing external, no-manifest die, symlink refusal, --all clean/corrupt/orphan,
nested decrypt failure). Full suite 205/205. bash 3.2 clean.
Add .ship-policy.json (opts out AI adversarial/red-team/security-specialist
review; requires local operator sign-off) and test/run-security.sh (the
operator-local security regression subset). Document the policy in CLAUDE.md
and README, fix stale test counts (manifest.bats 41->58, total 174->191), and
update the storage-recursion note to reflect rekey/list now walking the full
project tree.
Pre-landing review (testing + checklist specialists, reproduced) caught a
data-loss bug: cmd_rekey's decrypt/re-encrypt globs were non-recursive and
only special-cased external/. Nested manifest dotenv blobs
(<project>/<relpath>.age, new this branch) were never visited, so after a key
rotation they stayed encrypted under the discarded old key = permanently
undecryptable. cmd_list had the same blind spot (cosmetic: nested entries
invisible in listings).
Both now walk the entire project tree with `find -type f` (bash 3.2 safe,
includes dotfiles natively), unifying top-level / nested / external blobs into
one recursive pass and dropping the now-redundant external/ special-casing.
Regression tests: nested-blob rekey round-trip (survives rotation) + list
shows nested entry. Full suite 193/193.
- pull with .secrets.json restores exactly the declared entries (nested
paths get mkdir -p); stray store blobs are not restored
- dotenv rail re-runs at restore time: unsafe entries warn+skip (pull
never dies on one bad entry), missing blobs warn with a directed hint
- empty manifest = warn no-op instead of a confusing 'not found' death
- manifest-less projects keep the legacy glob pull verbatim
- .secrets.json external[] drives push/pull: type 'properties' (alias of
gradle-properties; blob suffix stays legacy-compatible in stage 1) and
type 'file'; same charset rails as the legacy parser
- push absorbs uncovered .secrets-files entries into the manifest
(idempotent, gradle-properties → properties) with a delete hint
- pull: manifest wins entirely; a coexisting .secrets-files warns as
superseded instead of being silently ignored
- basename rail generalized: properties targets must end '.properties'
(was exact 'gradle.properties') — rc files/gitconfig still blocked;
EGB-531 wrong-basename test updated for the sanctioned change
- push syncs FROM the manifest; v1 store layout unchanged (nested
entries land at <project>/<relpath>.age, same shape -w always used)
- discovery (root globs + quiet package.json workspace re-scan when a
manifest exists) feeds the manifest as a generator; new files auto-add
with ==> notice + undo guidance
- options.autoAdd committed toggle (default ON when absent); explicit
false warns on undeclared files instead of enrolling them
- push --frozen: declared-only for one invocation; push --dry-run:
reports would-add/would-sync, touches nothing
- bootstrap ordering: manifest written only after >=1 blob encrypts
- declared-but-missing warns and continues; unsafe manifest path dies
- jq // falsy gotcha: explicit autoAdd:false compared directly
bats on macOS runs under system bash 3.2, where a failing [[ ]] compound
command mid-test does NOT trip the ERR trap — every mid-test
[[ "$output" == *...* ]] assertion in the suite was decorative.
Append '|| false' to all 123 standalone [[ ]] assertion lines so failures
become plain-command failures, which bats catches. Full suite still green
(133/133) — no latent false positives were hiding.
Built for the Beacon Android upload keystore: binary files outside the
project root can now ride the encrypted store.
- manifest: 'file <path>' (no keys; keys present = rejected loudly)
- push: encrypts the file verbatim (age is binary-safe)
- pull: restores next to target (atomic same-fs mv), mode 600, TOCTOU
symlink recheck, divergent existing target backed up to .secrets-bak
- _validate_external_target_path parameterized by type (basename
restriction stays gradle-properties-only; $HOME/../symlink rails apply
to both)
- README/help/CLAUDE docs + 7 bats tests (133/133 pass)
Red-team + specialist findings from /ship pre-landing review:
- CRITICAL: ensure_store_protections ran BEFORE `git pull --ff-only` in
commit_and_push_secrets; a pull bringing history without .gitignore
left the post-pull window unprotected and `git add -A` committed AND
pushed key.txt (reproduced end-to-end). Heal now runs after the pull,
immediately before staging.
- CRITICAL: .gitignore can't untrack an already-tracked key.txt. Heal
now defensively `git rm --cached`s a tracked key with a history-scrub
warning.
- Content-aware heal: a present .gitignore missing the key.txt line is
rewritten, not skipped (security specialist).
- Init guard now renders the real sanitized clone URL from
.secrets-store when available, mirroring EGB-282's check_initialized.
- Tests: positive heal-message assertions, non-empty key guard in the
init test, 4 new regression tests (126 total).
Three fixes, found from a real Ubuntu second-machine setup failure:
1. cmd_init: if key.txt exists but the store has no .git, die BEFORE
git init with directed guidance to clone the existing secrets repo
instead. Previously init ran git init, then age-keygen died on the
existing key, leaving a half-initialized store (.git but no
.gitignore or hook).
2. ensure_store_protections: push/pull/rekey now restore a missing
store .gitignore (and pre-commit hook) before any `git add -A`.
Without the .gitignore, add -A would commit key.txt to the remote.
3. cmd_rekey: the re-encrypt loop used a bare "$dir"* glob, which never
matches dotfiles — .env blobs were decrypted to the tmpdir but never
re-encrypted, leaving them on the OLD key (undecryptable) while the
new key overwrote key.txt. Glob now matches the decrypt loop
("$dir"* "$dir".*). Exposed by the self-heal test: with nothing
staged, the empty commit failed mid-rekey.
Tests: 5 new bats tests (118 total) — init guard, .gitignore self-heal
on push/pull/rekey, and a rekey round-trip that survives key rotation.
Adversarial review of the diff surfaced 6 non-critical findings, all fixed:
- pull now warns + reports when a merge fails (read-only $HOME, full disk)
instead of silently exiting 0 as if it synced
- push skips a multi-line (continuation) managed value with a warning rather
than writing a dangling backslash that would corrupt the target
- absent-key pull check escapes regex '.' so systemProp.foo can't false-match
- fix garbled 'Merged 0\n0 key(s)' count when a blob has no '=' lines
- slug gets a cksum suffix so paths that clean to the same string (a/b vs a_b)
don't overwrite each other's blob
- docs: backup happens before each merge, not just the first
Tests: 112 -> 113 (continuation-value skip).
Add a committed .secrets-files manifest that lets secrets track designated
keys from files outside the project root (motivating case:
~/.gradle/gradle.properties for Android Clerk publishable keys, which
Android Studio GUI builds read but terminal env vars can't reach).
- push extracts only the named keys, encrypts under <project>/external/
- pull MERGES them into the target, preserving unrelated keys/comments/order
- pure-bash merge (no sed/regex): exact-string key match, opaque values
- path validator: basename gradle.properties, within $HOME, no symlink/..
- external/ subdir keeps blobs out of the dotenv *.age globs; rekey + list
recurse explicitly
- which reads back the manifest; list shows [external]; pre-commit blocks
plaintext gradle.properties
Also fixes two latent bugs in 'secrets rekey' (never completed before, no
prior test): age-keygen refusing to overwrite key.txt, and an EXIT trap
referencing an out-of-scope local under set -u.
Tests: 80 -> 104.
Reviewed via /autoplan (CEO/Eng/DX). EGB-531.
* chore: ignore .gstack/ (per-project local state)
* feat: optional remote URL in .secrets-store (EGB-282)
A second whitespace-separated token after the store name in .secrets-store
is treated as the store's git remote URL. When a teammate clones a project
bound to a store they don't have locally yet, the directed missing-store
error now fills in `git clone <url> <path>` so they can copy-paste instead
of asking the original setter for the URL.
Backward compatible: single-token .secrets-store files (the v0.1.0.x
format) continue to work and produce the existing `<their-store-remote>`
placeholder.
Security hardening (caught by adversarial review during /ship):
- The rendered git clone line is meant to be copy-pasted by a teammate.
Without sanitization, `work evil.git;rm -rf ~` would render verbatim
and execute `rm -rf ~` on paste. The parser now rejects URLs containing
shell metacharacters (;&|<>$`(){}*?!"'\\), control characters (incl.
ANSI escape sequences that could spoof terminal output), and embedded
whitespace. Rejected URLs are dropped with a stderr warning; the error
falls back to the safe placeholder.
- Switched from `set -- $line` to `read -r spec rest` so the URL field
isn't glob-expanded or word-split — important so `work *` from a
populated directory doesn't leak filenames into the URL field.
Tests 72 → 80. New: backward compat, SSH+HTTPS+~/-prefix URL forms,
comment-and-URL form, four named injection vectors (shell metachar,
backtick, $(), ANSI escape), multi-token URL, glob char, and a positive
test asserting standard git URL chars round-trip unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v0.1.1.0)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
_find_secrets_store_file used to set _LAST_FOUND_AT inside the $(...)
subshell that resolve_store invoked it from, so the parent shell never
received the value and `secrets which` reported `source: .secrets-store
file ()` with empty parens. Function now returns a tab-separated
<dir>\t<source-path> tuple; resolve_store splits it in the parent shell.
Caught by /land-and-deploy post-merge fresh-clone verification — the
existing test grepped for the substring ".secrets-store file" which
matched the broken truncated form. Tightened to assert the full file
path appears in the parenthetical.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: multi-store support via .secrets-store + --store flag
Layer four-rule store resolution on top of the existing SECRETS_DIR primitive
so users can manage multiple isolated encrypted stores (work vs personal,
per-client, etc.) without giving up the tool's small-bash-script pitch.
Resolution order (highest first):
1. --store <dir> flag (parsed in main pre-pass)
2. .secrets-store file in cwd or any ancestor up to $HOME
3. SECRETS_DIR env var (legacy escape hatch)
4. ~/.secrets default
resolve_store() updates both SECRETS_DIR and KEY_FILE so existing single-store
codepaths just work. New cmd_which / where / status report the active store.
cmd_init, push, pull, push_workspaces, pull_workspaces, list, rm, rekey, run,
which all call resolve_store at entry.
Hardening from the EGB-281 adversarial review:
- F1: cmd_run EXIT trap is now a named function (not string-interpolated),
so paths with apostrophes still get plaintext cleaned up
- F2: symlinked .secrets-store files are skipped, never read
- F3/F4: --store flag rejects flag-shaped values and empty --store=
- F5: HOME unset is detected up-front with a directed error
- F11: check_initialized / check_key give context-aware errors that name
both recovery paths (git clone vs secrets init) when a teammate clones
a project bound to a non-existent store on their machine
Tests: 37 → 66 (29 new). HOME=\$TEST_TMPDIR added to test setup so the
walk-up logic stays bounded inside fixtures.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: add Multiple stores section to README
Five subsections walk users through: how store resolution works, how to
set up a second store on a machine, how to bind a project, how teammates
join a bound project, and how to undo or change a binding. SECRETS_DIR
table entry now points readers at the new --store flag and .secrets-store
file as the preferred mechanisms.
* chore: bump version and changelog (v0.1.0.0)
First formal release. EGB-281 adds multi-store support; this commit
seeds the VERSION file (4-digit MAJOR.MINOR.PATCH.MICRO) and the
CHANGELOG.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Updated file tracking to include `.dev.vars` alongside `.env` and `.env.*`.
- Improved pre-commit hook to block plaintext secret files, including `.dev.vars`.
- Added `clear` command to remove plaintext secret files from the current directory and workspaces.
- Enhanced tests to cover new functionality for `.dev.vars` and the `clear` command.
- Updated documentation to reflect changes in tracked files and command usage.
- Introduced `--workspaces` flag for `push` and `pull` commands to handle environment files in monorepos.
- Updated README and CLAUDE.md to reflect new workspace functionality and installation instructions.
- Enhanced test suite with cases for workspace operations, ensuring proper encryption and decryption of environment files.
- Improved error handling for missing package.json and workspaces field.
- Increased test coverage from 20 to 25 tests.
