ur-agent 1.59.0 → 1.60.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md
CHANGED
|
@@ -1,5 +1,47 @@
|
|
|
1
1
|
# Changelog
|
|
2
2
|
|
|
3
|
+
## 1.60.1
|
|
4
|
+
|
|
5
|
+
- `ur memory-integrity verify` no longer reports an empty or missing store as
|
|
6
|
+
verified. It printed "verified — 0 file(s) match the recorded digests" for an
|
|
7
|
+
empty directory, and that identical reassurance would have appeared for a
|
|
8
|
+
mistyped path or a store an attacker had just emptied. Zero files checked is
|
|
9
|
+
not evidence of integrity, so empty and nonexistent stores now say exactly
|
|
10
|
+
that and name the path.
|
|
11
|
+
- The exit code tracks tampering rather than validity, so an empty or
|
|
12
|
+
unbaselined store still exits 0. Failing on it would fire for every fresh
|
|
13
|
+
install and train the user to ignore the exit code.
|
|
14
|
+
- `ur agent-inspect --costs` names the directory it searched. "No subagent
|
|
15
|
+
transcripts found for this session" was indistinguishable from having
|
|
16
|
+
resolved the wrong path.
|
|
17
|
+
|
|
18
|
+
## 1.60.0
|
|
19
|
+
|
|
20
|
+
- Added `ur memory-integrity`, tamper-evidence for the file-backed memory
|
|
21
|
+
stores. Project task memory was hash-chained and could prove tampering; the
|
|
22
|
+
auto-memory and team-memory directories had nothing, and their contents are
|
|
23
|
+
injected straight into context. A chain over append-only lines does not
|
|
24
|
+
describe a mutable file tree, so this is a digest manifest: it detects files
|
|
25
|
+
modified, deleted outside UR, and — the case that matters most — dropped in
|
|
26
|
+
by something else, which is an injection vector with a direct path to the
|
|
27
|
+
model. `quarantine` moves suspect files aside rather than deleting them, and
|
|
28
|
+
`verify` exits non-zero so it can gate.
|
|
29
|
+
- Deletion is now provable: removing an entry rewrites the manifest, so a
|
|
30
|
+
previously-deleted memory that reappears is reported as untracked instead of
|
|
31
|
+
being quietly reloaded.
|
|
32
|
+
- The release gate now grades recorded eval trajectories and writes versioned
|
|
33
|
+
per-category scores to `dist-release/trajectory-scores.json`. A run that
|
|
34
|
+
edited files without verifying, issued a destructive command, or looped on an
|
|
35
|
+
identical failure fails the gate even when its conclusion was correct.
|
|
36
|
+
- Added `test/settingsDocCoverage.test.ts`: every `SettingsSchema` key must
|
|
37
|
+
appear in `technical/06-configuration.md`. It found six undocumented settings
|
|
38
|
+
(`worktree`, `channelsEnabled`, `allowedChannelPlugins`, `urMdExcludes`,
|
|
39
|
+
`pluginTrustMessage`, `$schema`), now documented. Command coverage was
|
|
40
|
+
already enforced; settings coverage was not, which is how they slipped.
|
|
41
|
+
- Refreshed `ur agent-trends`, which still listed the claim-to-source ledger,
|
|
42
|
+
trajectory grading and multimodal capability warnings as future work after
|
|
43
|
+
they shipped — leaving UR misreporting its own coverage.
|
|
44
|
+
|
|
3
45
|
## 1.59.0
|
|
4
46
|
|
|
5
47
|
- Added `ur sources`, a claim-to-source ledger. `wrapUntrusted` already stamped
|