Closes 10 audit findings from 4-agent wave (critic + security +
architect + validator) on v0.21.0.
CRITIC HIGH (5):
H1 s3_cloud::commit() was listing with delimiter='/' — nested
writes silently dropped from manifest hash. Added
list_recursive() (no delimiter), filter manifest-*.json from
hash input.
H2 S3Cfg access_key_env + secret_key_env were advertised in TOML
but never read. Wired via resolve_explicit_creds() with
aws-credential-types. Partial-set or empty-resolve → error.
H3 display::sanitize_display missing in detach.rs + mount.rs
(regression of v0.19.2 L9 ANSI injection fix). Applied at 8
print sites. 2 new integration tests.
H4 adapters/jsonmcp.rs RESTORED (was lost in earlier merge).
107 LOC shared module: load_json_or_empty / upsert_under_key /
remove_under_key / persist. claude_code 163→105, cursor 165→106,
zed 178→114. Unified error handling via ConfigParseError.
H5 ENV_LOCK shared across kei-store tests. New test_env.rs (24 LOC)
exposed under cfg(any(test, feature='s3')). github.rs +
s3_cloud/tests.rs + s3_smoke.rs all use shared mutex. Fixes
parallel-test race on KEI_STORE_S3_ENDPOINT.
SECURITY HIGH (2):
SEC-H1 scripts/install-actionlint.sh — added sha256 verify
(shasum/sha256sum) before extract. ACTIONLINT_SHA256_OVERRIDE
env var for CI injection. Per-platform constants marked
[UNVERIFIED: SKIP] pending live checksums.txt fetch (agent had
no WebFetch this session — user follow-up: paste from
https://github.com/rhysd/actionlint/releases/download/v1.7.12/checksums.txt).
SEC-H2 S3 SSRF/IMDS guard. validate_endpoint() rejects:
loopback (127/8, ::1, localhost), link-local (169.254/16,
fe80::/10), metadata hostnames (google/azure). Override via
KEI_STORE_S3_ALLOW_INTERNAL=1. HTTP rejected unless
KEI_STORE_S3_ALLOW_INSECURE=1. Custom endpoint now REQUIRES
explicit creds (no IMDS chain leak via third-party endpoint).
4 reject + 3 accept tests pass.
POLISH (3):
D1 docs/USB-BRAIN-GUIDE.md — ⚠️ WARNING block under Prerequisites:
exFAT/FAT32 NOT safe for multi-client attach (SQLite WAL needs
shared-mem mmap). Use ONE client at a time on those FSes.
New Troubleshooting entry 'SQLite corruption on mount-attach'.
D2 '~5 MB release binary growth' now labelled [estimate, E5 —
not yet measured] in CHANGELOG.md + s3_cloud/mod.rs header.
D3 scripts/validate-workflow-shas.sh exits 2 (not 0) when
UNVERIFIED_COUNT > 0 and GITHUB_TOKEN absent. Distinguishes
'network denied' from 'all good'.
REAL VERIFICATION (pasted by agent):
cargo check -p keisei -p kei-store: Finished (clean)
cargo test -p keisei --release: 30 passed 0 failed
cargo test -p kei-store --release: 10 + 9 passed (default features)
cargo test -p kei-store --features s3 --release:
31 + 9 + 6 = 46 passed (with s3)
bash -n scripts/*.sh: OK
regen-counts.sh --check: no drift
Constructor Pattern: largest new src 200 LOC (s3_cloud/mod.rs, at
limit). jsonmcp.rs 107 LOC. test_env.rs 24 LOC.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
11 KiB
USB Exobrain — Step-by-Step Test Recipe
Goal: put a KeiSeiKit "brain" on a physical USB drive or flash card, mount it into Claude Code (and Cursor / Continue / Zed if installed), verify MCP tools are live, unplug — verify clean detach.
Target audience: someone trying the v0.21 exobrain feature for the first time on macOS (adapt paths for Linux).
0. Prerequisites
On the host (your laptop):
- KeiSeiKit installed:
./install.sh --profile=fullran successfully,cargo test -p keiseishows 28/28 pass. keiseiCLI on PATH:ls ~/.claude/agents/_primitives/_rust/target/release/keiseireturns the binary. Optionallyln -sf ~/.claude/agents/_primitives/_rust/target/release/keisei ~/.local/bin/keiseiso you can just typekeisei.- Claude Code installed (or Cursor / Continue / Zed — the CLI auto-detects all four).
- A USB stick or flash card mounted on macOS — expected path
/Volumes/<NAME>(e.g./Volumes/EXOBRAIN).
On the USB drive:
- Filesystem: exFAT or APFS (not HFS+ if you want cross-platform). FAT32 works but has 4 GB per-file limit — fine for brain dir (< 200 MB total even with 5 platform binaries).
- Free space: ~500 MB recommended (5 mcp-server binaries × ~90 MB each = ~450 MB, plus room for memory/artifacts SQLite).
⚠️ WARNING — exFAT / FAT32 are NOT safe for multi-client attach.
SQLite WAL mode (used by
kei-memory,kei-artifact,kei-social-storeinside a brain) requires a filesystem with reliable shared-memory mmap. exFAT and FAT32 do NOT provide this, andkeisei mount(multi-client fan-out in step 7 below) WILL corrupt the memory DBs if the brain lives on one of those filesystems.
- On exFAT / FAT32, use the brain with ONE client at a time (single
keisei attach --scope=user). Do not runkeisei mount.- For reliable multi-client use, put the brain on APFS (macOS), ext4 (Linux), or NTFS (Windows). exFAT is fine for single-client or read-only cross-platform transport.
- If you've already mounted a brain from exFAT and suspect corruption, see "SQLite corruption on mount-attach" in Troubleshooting.
1. Create the brain directory
Replace EXOBRAIN with your actual USB volume name.
BRAIN=/Volumes/EXOBRAIN/my-brain
mkdir -p "$BRAIN"/{bin,memory,artifacts,manifests}
Structure after step 1:
/Volumes/EXOBRAIN/my-brain/
├── bin/
├── memory/
├── artifacts/
└── manifests/
2. Download MCP server binaries for every platform
Grab the 5 single-binary compiles from the latest GitHub release. For v0.21.0:
BASE=https://github.com/KeiSei84/KeiSeiKit/releases/download/v0.21.0
cd "$BRAIN/bin"
# macOS Apple Silicon
curl -fL -O "$BASE/kei-mcp-server-darwin-arm64"
curl -fL -O "$BASE/kei-mcp-server-darwin-arm64.sha256"
# macOS Intel
curl -fL -O "$BASE/kei-mcp-server-darwin-x64"
curl -fL -O "$BASE/kei-mcp-server-darwin-x64.sha256"
# Linux x86_64
curl -fL -O "$BASE/kei-mcp-server-linux-x64"
curl -fL -O "$BASE/kei-mcp-server-linux-x64.sha256"
# Linux arm64 (may be unavailable on older releases — continue-on-error in CI)
curl -fL -O "$BASE/kei-mcp-server-linux-arm64" 2>/dev/null || echo "skipped linux-arm64"
# Windows x86_64
curl -fL -O "$BASE/kei-mcp-server-windows-x64.exe"
curl -fL -O "$BASE/kei-mcp-server-windows-x64.exe.sha256"
# Verify every downloaded binary against its .sha256
for f in kei-mcp-server-*.sha256; do
shasum -a 256 -c "$f" || echo "FAIL: $f"
done
# Strip macOS Gatekeeper quarantine + chmod +x on Unix binaries
chmod +x kei-mcp-server-darwin-* kei-mcp-server-linux-* 2>/dev/null || true
xattr -d com.apple.quarantine kei-mcp-server-darwin-* 2>/dev/null || true
Expected: every shasum -c prints OK.
3. Write manifest.toml (schema v2, per-platform)
cat > "$BRAIN/manifest.toml" <<'EOF'
[brain]
schema_version = 2
name = "my-brain"
created = "2026-04-22T00:00:00Z"
[paths]
memory = "memory/"
artifacts = "artifacts/"
manifests = "manifests/"
[paths.mcp_server]
darwin-arm64 = "bin/kei-mcp-server-darwin-arm64"
darwin-x64 = "bin/kei-mcp-server-darwin-x64"
linux-x64 = "bin/kei-mcp-server-linux-x64"
linux-arm64 = "bin/kei-mcp-server-linux-arm64"
windows-x64 = "bin/kei-mcp-server-windows-x64.exe"
EOF
Validation: keisei rejects the brain if name contains anything outside ^[a-z][a-z0-9_-]{0,63}$, if any path is absolute, if any path contains .., or if the root is a symlink.
4. Verify the brain loads before attaching
keisei list-adapters
Expected: 2-column table showing every adapter, detected: yes for at least claude-code (the others depend on whether Cursor / Continue / Zed are installed on this host).
If you want a dry-run of the brain itself without touching any client config:
# Loads manifest, validates path confinement + schema — then errors because
# nothing attached yet. Use the error to confirm load-path is clean.
keisei status
# Expected: "no brain attached"
5. Attach the brain to Claude Code
Single-client, user scope (the safe default):
keisei attach "$BRAIN" --scope=user
Expected output (literally):
attached brain 'my-brain' to claude-code (user scope)
brain path: /Volumes/EXOBRAIN/my-brain
mcp server: /Volumes/EXOBRAIN/my-brain/bin/kei-mcp-server-darwin-arm64
client cfg: /Users/you/.claude/settings.json
marker: /Users/you/.keisei/attached.toml
run /help in Claude Code to verify the MCP server is reachable
The last line is the client-specific post-attach hint. Each adapter emits its own.
6. Verify in Claude Code
Close and reopen Claude Code (or run /help → "MCP servers" section). You should see my-brain listed with the /Volumes/EXOBRAIN/my-brain/bin/kei-mcp-server-darwin-arm64 command.
# On the host, inspect what just got written:
cat ~/.claude/settings.json | jq '.mcpServers["my-brain"]'
# Expected:
# {
# "command": "/Volumes/EXOBRAIN/my-brain/bin/kei-mcp-server-darwin-arm64",
# "env": {
# "KEISEI_BRAIN_ROOT": "/Volumes/EXOBRAIN/my-brain"
# }
# }
Run a Claude Code command that uses a tool from kei-mcp-server (any of the 25+ MCP tool endpoints). If Claude Code reports success, the brain is live.
7. Multi-client mount (all detected clients at once)
keisei mount "$BRAIN"
Fan-out attach to every client detected on this host. For each, writes to user-scope config (~/.claude/settings.json, ~/.cursor/mcp.json, ~/.continue/config.json, ~/Library/Application Support/Zed/settings.json).
Expected summary:
mounted brain 'my-brain' to:
claude-code ✓
cursor ✓
continue ✗ (not detected)
zed ✓
8. Project-scope attach (Claude Code / Cursor only)
Useful for per-project brains that stay with the repo:
cd ~/path/to/your-repo
keisei attach "$BRAIN" --scope=project
Writes to ./.claude/settings.json (inside the repo) instead of ~/.claude/. Continue and Zed reject this scope with a clear error — they don't have per-project MCP config.
9. Verify brain health
keisei status
Expected:
attached: my-brain
brain path: /Volumes/EXOBRAIN/my-brain
attached at: 2026-04-22T17:30:00Z
clients:
- claude-code (user scope) ~/.claude/settings.json
- cursor (user scope) ~/.cursor/mcp.json
- zed (user scope) ~/Library/Application Support/Zed/settings.json
health: OK (manifest readable, mcp_server binary exists)
If you unplug the USB, re-run keisei status:
health: WARN — mcp_server binary at <path> is not reachable
(brain was mounted at /Volumes/EXOBRAIN/my-brain, which no longer exists)
10. Detach — clean up when done
keisei detach
Strips the mcpServers.my-brain entry from every client's config (preserving any other MCP servers you have configured), then deletes the marker.
Expected:
detached 'my-brain' from:
claude-code ✓
cursor ✓
zed ✓
marker removed: ~/.keisei/attached.toml
Confirm:
keisei status # → "no brain attached"
jq '.mcpServers' ~/.claude/settings.json # → your own entries only
11. Eject the USB safely
# macOS:
diskutil eject /Volumes/EXOBRAIN
# Linux:
umount /media/$USER/EXOBRAIN
Physically unplug.
Troubleshooting
"BrainNotFound" on attach
- Check
/Volumes/EXOBRAIN/my-brain/manifest.tomlexists - Check the path to
keisei attachis absolute (/Volumes/...), not relative
"PathEscape" on attach
- Every path under
[paths]must be relative and resolve inside the brain root. No../, no absolute paths.
"BrainIsSymlink" on attach
/Volumes/EXOBRAIN/my-brainis a symlink to elsewhere. Pass the real resolved path instead —keiseirefuses symlink roots to prevent accidental host-dir pivot via crafted USB.
"NoPlatformBinary" on Claude Code first-use
- Your platform's binary isn't in
bin/. Checkstd::env::consts::{OS, ARCH}on your host — the expected filename iskei-mcp-server-<os-renamed>-<arch-renamed>where macos→darwin, x86_64→x64, aarch64→arm64.
Claude Code can't spawn the MCP server
- Ensure
chmod +xapplied to the binary - On macOS:
xattr -d com.apple.quarantine <binary>to clear Gatekeeper - Check the binary actually runs standalone:
/Volumes/EXOBRAIN/my-brain/bin/kei-mcp-server-darwin-arm64 --help
"NameConflict" on attach
- Another MCP server with the same name already exists in the client's config. Either rename your brain (
namein manifest.toml) or remove the existing entry manually.
SQLite corruption on mount-attach
kei-memory/kei-artifact/kei-social-storedatabases show "disk image is malformed" or "database is locked" errors after akeisei mounton a USB drive.- Root cause: the brain lives on exFAT or FAT32, which do not reliably support SQLite WAL-mode shared-memory mmap. Multi-client attach corrupts the DB.
- Fix:
- Copy the brain dir to an APFS (macOS) / ext4 (Linux) volume, either internal disk or a reformatted USB.
- Re-attach via
keisei attach <new-path>orkeisei mount <new-path>. - For single-client use on the same exFAT drive, restrict to
keisei attach --scope=useronly — do NOT usekeisei mount.
- Prevention: re-read step 0 → Prerequisites → "⚠️ WARNING" above.
What this tests end-to-end
- Portable brain — works from read-only USB, no installer needed on the brain itself
- Per-platform dispatch — schema v2 picks the right binary based on host OS+arch
- Multi-client fan-out —
keisei mountattaches to every detected client in one call - Clean detach — zero residue in host configs, preserves unrelated MCP servers
- Safety gates — path confinement, name regex, symlink rejection, size bound (64 KiB manifest cap)
- Schema v1 compat — drop in an older v1 brain with a single-string mcp_server, still works
If all 11 steps above pass, the v0.21 exobrain is production-ready for single-user workflows.
For shared-brain scenarios (team members all mounting the same brain over git / S3) see the kei-store backend docs — S3 backend via keisei attach s3://my-bucket/brain/ (requires --features s3 at install).