Repository navigation
Tags
Tags
jit 2.3.3 Scan reads `export KEY=value` lines in env files. A .env of exported secrets counted "0 plaintext variables" and, since 2.3.2, was not reported at all (#186). `jit review FILE[:LINE]` marks a scan finding checked as not live, a real-looking key in a test file or a README; later scans leave it out and count it as reviewed. A mark matches what was found, so a changed value, or any change to a reviewed env file, is reported again. `jit unreview` removes marks (#187). AI jobs: a secret removed from the vault no longer stops a job, which runs without it; editing a profile manifest in the job's folder no longer stops it either. A rotated secret still stops the job (#188). `jit vault list --users` finds every profile that uses a secret, in any project folder; `name_looks_secret` says which names are credentials, and `jit migrate settings` never moves those out (#189). `jit decoys expect PROGRAM [--file F]` marks a program's decoy reads as expected, such as an editor indexing a folder; `jit audit` tags them. Interpreters and shells cannot be marked (#190).
jit 2.3.2
An env file with no active variable is no longer a Protect finding. An
empty .env was listed as "0 plaintext variables" with jit migrate as its
fix, and migrate then refused it ("has no active KEY=value lines to
migrate"). A file with no variables is now no finding; one whose
variables are all commented out stays a finding, as a manual one, since
jit has nothing to move.
jit 2.3.1 Only secrets go into the vault. Protecting a .env sends to the vault what the scan counts as a secret, plus names that look like one; every other line (URLs, IDs, file names, flags) stays a plain setting beside the vault, and the file keeps working. The live mount serves settings to every reader and decoys only for secrets. `jit migrate --secret` and `--setting` override a line; `jit migrate preview` shows the split first. `jit vault move-out` and `move-in` move one value either way, each its own Touch ID naming the value. `jit vault settings` lists the settings. `jit migrate settings` moves the settings out of profiles protected before, with one Touch ID. An AI job's settings are part of its fingerprint, so a changed setting stops the job. `jit vault export` carries settings; an export that holds any is version 3. Also: an MCP config whose --env-file is already protected is done, not a failure, and a project block that could not be parsed is still named. Leak Guard checks every push to main, and fails when it cannot list the commits.
jit 2.3.0 AI jobs: an AI tool runs an approved script with your secrets, and gets its output back with every secret value hidden. `jit job allow` approves a folder and its program; `jit mcp install` connects Claude Desktop and Cursor. A job's fingerprint covers its folder and what its program loads from outside it (the interpreter's standard library, site-packages, native libraries); any change stops the job and names the file. The vault key can live in this Mac's Secure Enclave: `jit vault rekey --wrapper secure-enclave` (or Settings › Protection in JitPass), after a recovery file. Only JitPass's signed helper can use it, after Touch ID or your password; grants and AI jobs follow with no re-approval. `--wrapper keychain` moves it back. Rotating an enclave-held key comes in a later version. Also: `jit mcp install` backups are found by scan and migrate; a jit outside JitPass.app no longer takes over the service on an enclave vault; status and doctor report an unfinished move, a pending restore and a key left in the keychain.
jit 2.2.8 A seatbelt profile's `(deny network-outbound)` covers unix-domain sockets, so every jit command run from a sandboxed shell was refused the connect to agent.sock — and a refused dial looked exactly like an absent listener. jit reported a service that was running the whole time as "installed but not running", sent the reader to `jit service restart`, and told a sandboxed agent it had no grants, which is wrong precisely when one is active. The kernel does distinguish the two, so jit does now. The audit trail had the same shape of problem one layer down: a sandboxed caller cannot write the config directory either, so the invocations most worth recording were the ones being lost. The record is handed to the service, which is already authenticated and already writing there, with the direct write kept as the fallback. - agent: a refused socket is a sandboxed caller, not a stopped service (#154) - agent: audit_append, so a sandboxed caller's invocations still reach the trail (#154) This is jit 2.2.7 plus #154; everything in 2.2.7 is included.
PreviousNext