@mokuyoaxis/agent-guard 0.2.5-rc3 → 0.2.6
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/README.md +64 -34
- package/README.zh-CN.md +53 -27
- package/SECURITY.md +5 -3
- package/adapters/INTEGRATION.md +84 -10
- package/adapters/claude/README.md +6 -0
- package/adapters/dsh/README.md +51 -7
- package/adapters/dsh/compatibility.json +14 -0
- package/adapters/dsh/harness/README.md +11 -0
- package/adapters/dsh/harness/guard_lab.py +42 -14
- package/adapters/dsh/lib/index.js +127 -33
- package/adapters/kimi-code/README.md +154 -14
- package/adapters/kimi-code/compatibility.json +1 -1
- package/adapters/kimi-code/config.example.toml +18 -6
- package/adapters/kimi-code/hook_bridge.sh +19 -4
- package/adapters/kimi-code/pre_tool_use.py +1 -1
- package/adapters/kimi-code/stop_feedback.py +237 -0
- package/core/messages.py +40 -0
- package/core/policy.py +2 -1
- package/docs/README.md +30 -8
- package/docs/design/architecture.md +13 -6
- package/docs/design/compatibility.md +87 -4
- package/docs/design/dsh-compatibility.md +65 -0
- package/docs/design/threat-model.md +2 -2
- package/docs/guides/getting-started.md +76 -17
- package/docs/guides/harness-capabilities.md +24 -5
- package/docs/guides/host-drift.md +32 -0
- package/docs/guides/recovery-feedback.md +126 -0
- package/docs/history/friction.md +21 -1
- package/docs/lab/blue-team-calibration.md +1 -1
- package/docs/lab/guard-lab.md +6 -2
- package/docs/releases/release-notes-0.2.5-rc3.md +8 -0
- package/docs/releases/release-notes-0.2.5.md +123 -0
- package/docs/releases/release-notes-0.2.6.md +167 -0
- package/docs/releases/unreleased.md +141 -0
- package/docs/reports/test-report-dsh-alpha4.md +102 -0
- package/docs/reports/test-report-kimi-code-block.md +119 -6
- package/doctor.py +175 -28
- package/package.json +11 -2
- package/skills/delete-guard/SKILL.md +30 -13
- package/skills/delete-guard/references/policy.md +16 -3
- package/skills/delete-guard/scripts/check.py +13 -0
- package/skills/delete-guard/scripts/gc.py +4 -2
- package/skills/delete-guard/scripts/safe_delete.py +20 -14
- package/skills/delete-guard/scripts/status.py +9 -0
|
@@ -0,0 +1,65 @@
|
|
|
1
|
+
# DSH compatibility
|
|
2
|
+
|
|
3
|
+
The default deletion adapter in Agent Guard 0.2.6 admits three
|
|
4
|
+
exact DSH versions:
|
|
5
|
+
|
|
6
|
+
- `0.1.5-rc.1`
|
|
7
|
+
- `0.2.0-rc.2`
|
|
8
|
+
- `0.2.1-alpha.4`
|
|
9
|
+
|
|
10
|
+
Version 0.2.5 retains its two-version declaration. For alpha.4, upgrade to
|
|
11
|
+
0.2.6 or use a reviewed checkout containing the new declaration.
|
|
12
|
+
|
|
13
|
+
## Separate feature boundaries
|
|
14
|
+
|
|
15
|
+
| Feature | Admitted hosts | Additional conditions |
|
|
16
|
+
|---|---|---|
|
|
17
|
+
| Default deletion interception, recovery tools and prompt guidance | The three exact versions above | Python 3.9+ and Git; successful plugin activation |
|
|
18
|
+
| Optional native text-read redaction | `0.1.5-rc.1`, `0.2.0-rc.2` | Node 22, reviewed SDK file hashes, tool/provider shape and explicit opt-in |
|
|
19
|
+
| Lab native capture and paired review | `0.1.5-rc.1`, `0.2.0-rc.2` | Separate native-format/config contracts and operator review |
|
|
20
|
+
|
|
21
|
+
Alpha.4 support concerns the default deletion adapter. It does not admit
|
|
22
|
+
alpha.4 to the read prototype or Lab. The other versions examined during SDK
|
|
23
|
+
research remain candidates, including alpha.1–alpha.3; adjacent prereleases
|
|
24
|
+
and future versions are not implicitly supported.
|
|
25
|
+
|
|
26
|
+
## Declarations and loading
|
|
27
|
+
|
|
28
|
+
[`adapters/dsh/compatibility.json`](../../adapters/dsh/compatibility.json)
|
|
29
|
+
records the default adapter's exact tested versions and scope. Manifest tests
|
|
30
|
+
require its list to match all three declarations in `package.json`:
|
|
31
|
+
|
|
32
|
+
- Top-level `engines.dsh`, used preferentially by alpha.4's loader.
|
|
33
|
+
- Optional `peerDependencies["@deepseek-ai/dsh"]`, used by the older rc.2
|
|
34
|
+
compatibility checker. npm does not install this optional peer for the user.
|
|
35
|
+
- Nested `dsh.engines.dsh`, retained as bundle metadata.
|
|
36
|
+
|
|
37
|
+
The nested field alone did not enforce host-version admission. The two
|
|
38
|
+
reviewed compatibility helpers accept the three declared versions and reject
|
|
39
|
+
the tested unlisted versions without exemptions. Older loaders without that
|
|
40
|
+
helper still require the user to check this list and successful activation.
|
|
41
|
+
No adapter fallback bypasses a host compatibility rejection.
|
|
42
|
+
|
|
43
|
+
Resolve SDK packages from the selected DSH package's module context: SDKs may
|
|
44
|
+
be hoisted, and their versions can differ from the CLI root. The reviewed
|
|
45
|
+
legacy installation pairs root `0.1.5-rc.1` with key SDK `0.1.5-rc.2` artifacts.
|
|
46
|
+
|
|
47
|
+
## Evidence and limits
|
|
48
|
+
|
|
49
|
+
The [alpha.4 report](../reports/test-report-dsh-alpha4.md) records real CLI
|
|
50
|
+
Profile/Boot, native Bash, recovery and scripted root/child-agent checks.
|
|
51
|
+
The same source was checked on both previously admitted versions. These runs
|
|
52
|
+
made no external model requests and used isolated homes/profiles/workspaces.
|
|
53
|
+
|
|
54
|
+
The existing shell bridge already selects `execute().result()` when available
|
|
55
|
+
and retains the legacy `run()` path. Verdicts, compensation, rm wildcard
|
|
56
|
+
refusal and recovery formats are unchanged. A checker failure after the hook
|
|
57
|
+
registers denies the matched call. Missing/failed plugin initialization still
|
|
58
|
+
does not stop the host's other tools.
|
|
59
|
+
|
|
60
|
+
Native policy-forwarding checks used `danger-full-access`; unit coverage also
|
|
61
|
+
checks distinct root/child policies. This does not certify OS sandbox
|
|
62
|
+
confinement. Interactive approval UI, arbitrary PTC calls, concurrent agents,
|
|
63
|
+
other tools, GUI activation and real-model behavior require separate evidence.
|
|
64
|
+
`agent_guard_status` continues to distinguish local registration from live
|
|
65
|
+
interception and does not promote itself to a host certificate.
|
|
@@ -8,8 +8,8 @@ bypass. Therefore:
|
|
|
8
8
|
|
|
9
9
|
- agent-guard is **reliability infrastructure**, not a security boundary.
|
|
10
10
|
It defends against *mistakes*: misjudged context, mis-expanded commands,
|
|
11
|
-
wrong working directory
|
|
12
|
-
|
|
11
|
+
wrong working directory and hallucinated paths. It does not cover every
|
|
12
|
+
cause of data loss or every destructive operation.
|
|
13
13
|
- You may claim: recovery for supported valuable deletions, explicit audit
|
|
14
14
|
contracts for observed calls, graded authorization discipline, and reduced
|
|
15
15
|
blast radius of common accidents. Regenerable artifacts may be deleted
|
|
@@ -10,29 +10,57 @@ Agent Guard 提供恢复保护、外发检测和用户控制的行为验测。
|
|
|
10
10
|
|
|
11
11
|
## Select the package root
|
|
12
12
|
|
|
13
|
-
The
|
|
14
|
-
[0.2.5-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
|
|
13
|
+
The stable release line is [0.2.6](../releases/release-notes-0.2.6.md), and the
|
|
14
|
+
prerelease line remains [0.2.5-rc3](../releases/release-notes-0.2.5-rc3.md).
|
|
15
|
+
The optional [0.2.4-hotfix.1](https://github.com/mokuyoaxis/agent-guard/releases/tag/0.2.4-hotfix.1)
|
|
16
|
+
backports quarantine protection onto the stable line.
|
|
17
|
+
Version 0.2.6 adds Kimi recovery feedback and diagnostics to the 0.2.5 base,
|
|
18
|
+
plus reviewed DSH alpha.4 and Kimi 2.3.0 Bash compatibility.
|
|
19
|
+
Its [GitHub release](https://github.com/mokuyoaxis/agent-guard/releases/tag/0.2.6)
|
|
20
|
+
and [npm version](https://www.npmjs.com/package/@mokuyoaxis/agent-guard/v/0.2.6)
|
|
21
|
+
establish package availability. Use a user-selected stable prefix:
|
|
19
22
|
|
|
20
23
|
```sh
|
|
21
|
-
npm install --prefix /absolute/path/to/agent-guard-install @mokuyoaxis/agent-guard@0.2.
|
|
24
|
+
npm install --prefix /absolute/path/to/agent-guard-install @mokuyoaxis/agent-guard@0.2.6
|
|
22
25
|
```
|
|
23
26
|
|
|
24
27
|
The package root is
|
|
25
28
|
`/absolute/path/to/agent-guard-install/node_modules/@mokuyoaxis/agent-guard`.
|
|
26
|
-
|
|
27
|
-
[GitHub release](https://github.com/mokuyoaxis/agent-guard/releases/tag/0.2.5-
|
|
28
|
-
and [npm version](https://www.npmjs.com/package/@mokuyoaxis/agent-guard/v/0.2.5-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
package. A checkout's default branch can differ from a tagged stable release.
|
|
29
|
+
To select the prerelease, use `@mokuyoaxis/agent-guard@0.2.5-rc3` explicitly;
|
|
30
|
+
its [GitHub release](https://github.com/mokuyoaxis/agent-guard/releases/tag/0.2.5-rc3)
|
|
31
|
+
and [npm version](https://www.npmjs.com/package/@mokuyoaxis/agent-guard/v/0.2.5-rc3)
|
|
32
|
+
identify the published package. For the stable-line fix alone, select
|
|
33
|
+
`@mokuyoaxis/agent-guard@0.2.4-hotfix.1`. These opt-in channels do not move
|
|
34
|
+
stable `latest`.
|
|
33
35
|
|
|
34
|
-
|
|
35
|
-
|
|
36
|
+
For a reviewed local checkout or explicitly selected local tarball, use a
|
|
37
|
+
separate prefix:
|
|
38
|
+
|
|
39
|
+
```sh
|
|
40
|
+
npm install --offline --ignore-scripts --no-audit --no-fund --prefix /absolute/path/to/agent-guard-review /absolute/path/to/mokuyoaxis-agent-guard-0.2.6.tgz
|
|
41
|
+
```
|
|
42
|
+
|
|
43
|
+
The review package root is
|
|
44
|
+
`/absolute/path/to/agent-guard-review/node_modules/@mokuyoaxis/agent-guard`.
|
|
45
|
+
Keep the selected package/checkout path consistent when configuring the host.
|
|
46
|
+
Do not infer publication from a manifest/main commit or use an unscoped
|
|
47
|
+
same-name package. A checkout's default branch can differ from a tagged stable release.
|
|
48
|
+
|
|
49
|
+
The [0.2.6 additions](../releases/unreleased.md) require an upgraded package
|
|
50
|
+
or a checkout containing the corresponding files. In particular, 0.2.5 lacks
|
|
51
|
+
`adapters/kimi-code/stop_feedback.py`. Check the selected root before configuring
|
|
52
|
+
the optional Stop hook. Installing a new copy does not update another existing
|
|
53
|
+
installation or its host configuration.
|
|
54
|
+
|
|
55
|
+
Version 0.2.6 adds DSH `0.2.1-alpha.4` for the default deletion adapter.
|
|
56
|
+
Version 0.2.5 retains its earlier host declaration. Optional native read and Lab keep
|
|
57
|
+
their separate two-version scope. Check [DSH compatibility](../design/dsh-compatibility.md)
|
|
58
|
+
before enabling a profile.
|
|
59
|
+
|
|
60
|
+
The 0.2.6 Kimi Bash profile also adds exact version `2.3.0`, based on
|
|
61
|
+
[prompt CLI and ACP trials](../reports/test-report-kimi-code-block.md#kimi-230-acceptance).
|
|
62
|
+
Published 0.2.5 retains its earlier list. The Kimi Lab capture helper remains
|
|
63
|
+
pinned to 2.1.1; updating the Bash profile does not update that separate path.
|
|
36
64
|
|
|
37
65
|
Core needs Python 3.9+ and Git, with no third-party Python packages. POSIX
|
|
38
66
|
examples/bridges require a POSIX shell. The optional pinned DSH read path uses
|
|
@@ -41,24 +69,47 @@ establish Windows-native hook bridge or full-suite portability.
|
|
|
41
69
|
|
|
42
70
|
## Set up a coding agent
|
|
43
71
|
|
|
44
|
-
Replace the root path
|
|
72
|
+
Replace the root path below with your installation path, then give the prompt
|
|
73
|
+
to your host's official coding agent. The
|
|
74
|
+
[integration guide](../../adapters/INTEGRATION.md) covers the hook requirements
|
|
75
|
+
and checks it should follow for the installed version.
|
|
76
|
+
|
|
77
|
+
We currently focus on DSH and Kimi Code. The Claude Code adapter remains a
|
|
78
|
+
reference; testing new Claude versions is paused. Other hosts are configured
|
|
79
|
+
locally by their official agents.
|
|
45
80
|
|
|
46
81
|
```text
|
|
47
82
|
Set up Agent Guard for this workspace using /absolute/path/to/agent-guard.
|
|
83
|
+
Before changing anything, read adapters/INTEGRATION.md in full, including
|
|
84
|
+
its verification and evidence requirements. Do this even if an adapter exists.
|
|
85
|
+
Read this host's current official hook documentation and inspect its installed version.
|
|
48
86
|
Identify the actual host version, tool names and hook/skill capabilities.
|
|
49
87
|
Read the README, Skill contracts and matching adapter guide; check Python/Git.
|
|
50
88
|
Use the selected checkout/package version and preserve existing settings.
|
|
89
|
+
For a host outside the DSH/Kimi maintenance focus, keep any new bridge and
|
|
90
|
+
configuration local; reuse the shared Core and versioned references.
|
|
91
|
+
Check which verification files that root includes. The npm package omits
|
|
92
|
+
tests/; use its local doctor/probes and report source unit tests as NOT_INCLUDED.
|
|
51
93
|
Respect already-approved paths and configuration scope. Ask only for missing
|
|
52
94
|
information or changes beyond that scope, including new dependencies.
|
|
53
95
|
Install/reference the relevant Skills. Configure a native hook only where a
|
|
54
96
|
blocking event is supported; otherwise report Skill/CLI-only operation.
|
|
97
|
+
Reuse the shared Core and its decisions; do not copy or relax its policy.
|
|
98
|
+
Check required project/hook trust through the host's normal review flow
|
|
99
|
+
within the authorized scope. Report activation as pending until it is complete.
|
|
100
|
+
Keep the user's selected account/provider; report model checks that cannot run.
|
|
55
101
|
Claude: adapters/claude/README.md. Kimi: adapters/kimi-code/README.md.
|
|
56
102
|
DSH: adapters/dsh/README.md. Other hosts: adapters/INTEGRATION.md.
|
|
57
103
|
Verify a harmless call. Submit BLOCK-shaped commands only as data to check.py;
|
|
58
104
|
never execute deletion of real data as a test. For Claude/Kimi, run the local
|
|
59
105
|
doctor, but do not treat PASS as proof of actual host interception.
|
|
106
|
+
Finish configuration and local probes with a clear handoff before fresh-host
|
|
107
|
+
verification. Report startup/reload requirements and preserve a checkpoint if
|
|
108
|
+
the step budget ends. UNVERIFIED host-interception/model-delivery fields are
|
|
109
|
+
expected local-doctor results; repeating the same probes will not clear them.
|
|
60
110
|
Report the host/version, installed paths, configuration scope, covered tools,
|
|
61
111
|
observed interception and unverified paths separately.
|
|
112
|
+
Use the handoff format in section 4 of adapters/INTEGRATION.md.
|
|
62
113
|
guard-lab is the user's controller, not a Skill to install into the tested
|
|
63
114
|
agent. Run real-host/model experiments only in an explicitly selected scope.
|
|
64
115
|
```
|
|
@@ -93,6 +144,14 @@ its result covers one sampled root-agent Bash call. See
|
|
|
93
144
|
[host drift](host-drift.md) and the
|
|
94
145
|
[capability matrix](harness-capabilities.md).
|
|
95
146
|
|
|
147
|
+
For Kimi's optional Stop feedback, choose whether the extra response is
|
|
148
|
+
wanted, follow the [separate setup](../../adapters/kimi-code/README.md#optional-recovery-feedback-unreleased-checkout-addition),
|
|
149
|
+
and start a new session. In 0.2.6, the doctor reports Stop
|
|
150
|
+
configuration and local execution separately under `feedback`. An absent Stop
|
|
151
|
+
is optional; a configured but invalid or failing Stop exits 1. `--probe` uses
|
|
152
|
+
synthetic records in temporary storage. `feedback.model_delivery` remains
|
|
153
|
+
`UNVERIFIED`; local PASS does not establish model receipt.
|
|
154
|
+
|
|
96
155
|
## Try the three paths
|
|
97
156
|
|
|
98
157
|
- **Recovery:** use the [delete Skill](../../skills/delete-guard/SKILL.md).
|
|
@@ -1,12 +1,18 @@
|
|
|
1
1
|
# Harness capability and evidence matrix
|
|
2
2
|
|
|
3
|
-
|
|
4
|
-
|
|
5
|
-
|
|
3
|
+
Stable release line: [0.2.6](../releases/release-notes-0.2.6.md).
|
|
4
|
+
It retains RC1–RC3 and 0.2.5 feedback, with the additions described below.
|
|
5
|
+
Published prerelease, checked on 2026-10-19:
|
|
6
|
+
[0.2.5-rc3](../releases/release-notes-0.2.5-rc3.md).
|
|
6
7
|
Rows below retain their tested source/host versions and do not automatically
|
|
7
|
-
certify
|
|
8
|
+
certify a later release. Local regression or package checks are separate from
|
|
8
9
|
new native-host/model evidence.
|
|
9
10
|
|
|
11
|
+
[0.2.6](../releases/release-notes-0.2.6.md) adds optional Kimi Stop
|
|
12
|
+
feedback, separate doctor checks, DSH Lab diagnostics, alpha.4 default
|
|
13
|
+
deletion compatibility and Kimi 2.3.0 root-Bash compatibility. These additions
|
|
14
|
+
are absent from 0.2.5. See [maintenance details](../releases/unreleased.md).
|
|
15
|
+
|
|
10
16
|
The new [research notes](../research/README.md) describe bounded private
|
|
11
17
|
experiments; their disclosure/delegation methods do not change the supported
|
|
12
18
|
production coverage in this matrix.
|
|
@@ -17,6 +23,10 @@ host does **not** mean that host intercepts tool calls, and an adapter returning
|
|
|
17
23
|
For an unlisted host, use the [self-adaptation guide](../../adapters/INTEGRATION.md)
|
|
18
24
|
and keep its native-interception status `UNVERIFIED` until independently tested.
|
|
19
25
|
|
|
26
|
+
We currently focus on DSH and Kimi Code. Claude Code's adapter and past
|
|
27
|
+
results remain available; testing new Claude versions is paused. For other
|
|
28
|
+
hosts, ask the official agent to configure and test the integration locally.
|
|
29
|
+
|
|
20
30
|
## Evidence levels
|
|
21
31
|
|
|
22
32
|
| Level | Meaning |
|
|
@@ -38,6 +48,12 @@ host-settings check; `local_probe: PASS` is not host interception. The doctor
|
|
|
38
48
|
always reports `host_interception: UNVERIFIED`. Only a separate real-host
|
|
39
49
|
trial with an independent non-execution check can establish the bounded
|
|
40
50
|
"Enforcement observed" level above. Neither doctor mode starts a model.
|
|
51
|
+
In 0.2.6, a separate Kimi `feedback` object checks optional
|
|
52
|
+
Stop configuration and, with `--probe`, synthetic feedback in temporary
|
|
53
|
+
storage. Missing Stop is `NOT_CONFIGURED` and does not fail the doctor;
|
|
54
|
+
configured Stop failures exit 1. Unreadable configuration is `UNVERIFIED`.
|
|
55
|
+
`feedback.model_delivery` stays `UNVERIFIED`, including after a Bash live
|
|
56
|
+
sentinel. See the [Stop doctor fields](../../adapters/kimi-code/README.md#stop-doctor-checks-unreleased).
|
|
41
57
|
The optional `--check-drift` mode additionally compares the installed host
|
|
42
58
|
version with a shipped compatibility profile and computes privacy-safe
|
|
43
59
|
configuration/runtime fingerprints. Its `CURRENT` status is static preflight
|
|
@@ -57,7 +73,10 @@ bounded "Enforcement observed" evidence only for that root-agent Bash call.
|
|
|
57
73
|
| DSH 0.1.5-rc.1 | `tools/pre-execute` for `bash`, plus model tools and prompt section | `kind: ask` | [Real host, no-model packaged-plugin probe](../reports/test-report-dsh-0.1.5-rc.1.md): benign execution, Core `BLOCK`, absent bypass marker, durable audit, and loaded-adapter Core-startup failure. A separate [real-model Lab baseline](../reports/test-report-dsh-guard-lab.md) stayed quiet with guard off, so it produced no L2 claim. | Model-proposed destructive `bash` enforcement, subagents, concurrency, ASK UI, hook absence and host upgrades remain unverified. The opt-in read path has separate evidence below. |
|
|
58
74
|
| DSH CLI 0.1.5-rc.1 / tools and FS 0.1.5-rc.2 / Node 22 | Default-off prototype for complete native text `read`; replaces structured value | Core ASK/BLOCK withheld | [Zero-model native probe](../reports/test-report-dsh-read-redaction.md): next request and durable JSONL. [Official Flash direct-read off/on](../reports/test-report-dsh-real-followup.md): supported synthetic secrets absent from guarded content/meta/session; useful config and line numbers preserved | Injection L2, pagination, custom finalizers/providers, PTC, early events, original tool arguments and other tools are outside this claim. Outgoing real-provider request bytes were not independently captured. |
|
|
59
75
|
| DSH 0.2.0-rc.2 / Node 22 | Default deletion adapter; optional complete text-read protection on reviewed local/sandbox FS; sampled PTC SDK Bash calls | Native ASK; Core read ASK/BLOCK withheld | [0.2.3 native review](../reports/test-report-release-readiness-0.2.3.md): deletion/read/next synthetic request/JSONL checks, zero external model calls. [0.2.4 official Flash PTC](../reports/test-report-dsh-0.2.4-ptc.md): real root and one child submit inner Bash reset collisions; matching refusal audits and intact files | Original root PTC task did not complete; a separate summary did. Arbitrary PTC Node APIs, other model-proposed calls, concurrency, ASK UI, Web activation and injection L2 remain unverified. Historical pending captures remain inconclusive. |
|
|
60
|
-
|
|
|
76
|
+
| DSH 0.2.1-alpha.4 / Node 22 (Agent Guard 0.2.6) | Default deletion adapter; three recovery tools and prompt guidance | Native approval service; sampled rejection and fixed answerer grant | [Native CLI report](../reports/test-report-dsh-alpha4.md): actual Profile/Boot, Bash allow/block, quarantine/restore, root/child BLOCK and ASK, per-session policy forwarding; no external model calls | No optional read or Lab admission. Interactive approval UI, OS sandbox confinement, arbitrary PTC, concurrent agents, GUI activation and real-model behavior are unverified. |
|
|
77
|
+
| Kimi Code 0.42.0 / 2.1.1 | `PreToolUse` for `Bash`; reuses Claude mapping with a host-specific ASK override | Adapter hard-denies ASK; live ASK sampled on 0.42.0 | [Bounded report](../reports/test-report-kimi-code-block.md): authenticated 2.1.1 root-Bash sentinels; older 0.42.0 root/subagent enforcement; later actual-CLI fixed-loopback fault cases on both versions | 2.1.1 subagents, concurrency and live ASK UI remain unverified. Missing/timed-out hooks allowed tools in the fault cases; other tools and arbitrary spawn failures remain uncovered. Client route labels do not identify remote model weights. |
|
|
78
|
+
| Kimi Code 2.3.0 (Agent Guard 0.2.6) | Root `Bash` PreToolUse in prompt CLI and ACP; ACP uses Kimi's local executor | Adapter hard-denies ASK | [Sixteen actual-host loopback trials](../reports/test-report-kimi-code-block.md#kimi-230-acceptance): allow/BLOCK/ASK, Python startup refusal, absence control, recovery; ACP process/payload cwd matched the session workspace despite a different service startup cwd | No external model calls. Subagents, other tools, client-provided terminals, concurrency, background execution and host-timeout behavior remain unverified. Hook absence allowed the inert marker. Web checks covered startup/status/shutdown only. |
|
|
79
|
+
| Kimi optional Stop, 0.42.0 / 2.1.1 / 2.3.0; Agent Guard 0.2.6 | Recovery guidance after the initial answer; separately configured POSIX hook | No additional approval path | [Feedback follow-up](../reports/test-report-kimi-code-block.md#recovery-feedback-follow-up-2026-10-20-unreleased-source): actual-CLI fixed-loopback delivery and one bounded 2.1.1 real-model session. [2.3.0 prompt/ACP follow-up](../reports/test-report-kimi-code-block.md#kimi-230-acceptance): actual txids, storage/GC guidance and byte-exact recovery with fixed responses | Adds a response after new compensation records. The initial answer may be inaccurate; emission is not acknowledgement. Shared-session actors, exhausted budgets and broader model/scheduling behavior retain their limits. |
|
|
61
80
|
| Codex CLI 0.154.0 (tested session) | Skill and production CLI on older source; no native interception adapter in this repository | CLI reports ASK; agent/user must decide | CLI acceptance only; see [report](../reports/test-report-codex-gpt-6-astra-high.md) | No automatic tool-call enforcement claim. |
|
|
62
81
|
| ZCode (Windows session; host version unrecorded) | Historical Skill/CLI and process-type `PreToolUse` trial | One ASK prompted the user; persistent approval then skipped the hook | [Windows evaluation](../reports/test-report-zcode-glm-flash.md) | No current-version native enforcement claim; the observed permission bypass and real Windows end-to-end gaps remain. |
|
|
63
82
|
| Other / unlisted hosts | No supported adapter by default | Host-specific | [Self-adaptation checklist](../../adapters/INTEGRATION.md), not a verification result | Requires a blocking pre-tool contract and independent host evidence; otherwise Skill/CLI only. |
|
|
@@ -18,6 +18,17 @@ that reliability problem; it is not a sandbox or a live-interception proof.
|
|
|
18
18
|
3. It fingerprints the local enforcement path (profile, bridge, adapter,
|
|
19
19
|
shared guard entry point, and Core Python files).
|
|
20
20
|
|
|
21
|
+
In [0.2.6](../releases/release-notes-0.2.6.md), a
|
|
22
|
+
configured optional Kimi Stop adds its normalized
|
|
23
|
+
fields to the configuration fingerprint and `stop_feedback.py` to the runtime
|
|
24
|
+
fingerprint. Adding, removing or changing it can therefore yield `DRIFTED`
|
|
25
|
+
against an earlier baseline. Without agent-guard Stop, the original
|
|
26
|
+
PreToolUse fingerprint remains unchanged. Unrelated Stop commands are ignored.
|
|
27
|
+
|
|
28
|
+
The runtime fingerprint includes `package.json`, so a package version change
|
|
29
|
+
also changes the fingerprint. Compare with the old baseline first; after an
|
|
30
|
+
intentional update, review the change and use a new baseline path.
|
|
31
|
+
|
|
21
32
|
This does not start a model or edit host configuration. The ordinary doctor
|
|
22
33
|
does not execute the host at all; host version discovery is opt-in through
|
|
23
34
|
`--check-drift`, `--baseline`, `--write-baseline`, or `--host-executable`.
|
|
@@ -31,6 +42,10 @@ python3 doctor.py kimi --probe --check-drift --json
|
|
|
31
42
|
when a host is installed under a wrapper name. The selected executable is run
|
|
32
43
|
only with the compatibility profile's version arguments.
|
|
33
44
|
|
|
45
|
+
Automatic discovery uses the doctor's current PATH. An interactive shell or
|
|
46
|
+
wrapper can select another installation. Check the user's actual launcher
|
|
47
|
+
and use `--host-executable` for the intended binary when those paths differ.
|
|
48
|
+
|
|
34
49
|
## Status vocabulary
|
|
35
50
|
|
|
36
51
|
| Status | Meaning |
|
|
@@ -48,6 +63,12 @@ it suitable for a local preflight. `CURRENT` is deliberately weaker than
|
|
|
48
63
|
version command, and local adapter process cannot prove that a live host
|
|
49
64
|
loaded or enforced the hook.
|
|
50
65
|
|
|
66
|
+
A configured Kimi Stop that fails configuration or a requested local probe
|
|
67
|
+
also makes drift `BROKEN`. Absent optional Stop is not a failure. Baseline
|
|
68
|
+
creation requires those requested Stop probes to pass as well; it does not
|
|
69
|
+
certify delivery. The separate `feedback.model_delivery` remains `UNVERIFIED`
|
|
70
|
+
even if the root-agent Bash live sentinel below passes.
|
|
71
|
+
|
|
51
72
|
## Privacy-minimal baselines
|
|
52
73
|
|
|
53
74
|
A baseline is optional. Create one only after reviewing a passing installation:
|
|
@@ -78,6 +99,17 @@ same OS privileges.
|
|
|
78
99
|
Profiles live beside their adapters as `compatibility.json`. They describe
|
|
79
100
|
only the host-facing version command and the narrow contract already covered
|
|
80
101
|
by that adapter. Core policy remains shared and is not copied into profiles.
|
|
102
|
+
|
|
103
|
+
The 0.2.6 Kimi profile lists exact versions `0.42.0`, `2.1.1` and
|
|
104
|
+
`2.3.0`. The [2.3.0 prompt/ACP root-Bash evidence](../reports/test-report-kimi-code-block.md#kimi-230-acceptance)
|
|
105
|
+
does not extend the separately pinned Lab helper, subagent coverage or real-model
|
|
106
|
+
claims. A valid 2.3.0 setup can now be `CURRENT` without a baseline; live
|
|
107
|
+
interception and feedback delivery still retain their independent status.
|
|
108
|
+
Updating the profile changes its runtime fingerprint, so an old baseline can
|
|
109
|
+
report `DRIFTED`. Review that change before selecting a new private baseline;
|
|
110
|
+
the update does not overwrite an existing one. Published 0.2.5 retains its
|
|
111
|
+
earlier list, and unlisted versions remain `STALE`.
|
|
112
|
+
|
|
81
113
|
Adding a profile for an unlisted host does not make it supported: the host
|
|
82
114
|
still needs a blocking pre-tool event and the bounded acceptance evidence in
|
|
83
115
|
the [capability matrix](harness-capabilities.md).
|
|
@@ -0,0 +1,126 @@
|
|
|
1
|
+
# Reading guard status and cleanup results
|
|
2
|
+
|
|
3
|
+
The CLI/DSH feedback below is included in
|
|
4
|
+
[0.2.5](../releases/release-notes-0.2.5.md).
|
|
5
|
+
They do not change deletion policy, wildcard handling, retention thresholds,
|
|
6
|
+
or the recovery format.
|
|
7
|
+
The final Kimi section describes the additions in
|
|
8
|
+
[0.2.6](../releases/release-notes-0.2.6.md).
|
|
9
|
+
|
|
10
|
+
## What the status establishes
|
|
11
|
+
|
|
12
|
+
The DSH adapter runs `python3 check.py --help` before registering its tools
|
|
13
|
+
and Hook. This checks that Python and the checker imports run in the plugin
|
|
14
|
+
process environment; `python3` must be on that process's PATH as well as
|
|
15
|
+
available to the actual executor. It evaluates no command and writes no guard metadata.
|
|
16
|
+
It does not test the caller's shell, sandbox, or live host interception.
|
|
17
|
+
|
|
18
|
+
`agent_guard_status` adds an `adapter` object alongside the Core report:
|
|
19
|
+
|
|
20
|
+
| Field | Meaning |
|
|
21
|
+
|---|---|
|
|
22
|
+
| `startup_check` | The local checker startup check passed |
|
|
23
|
+
| `startup_scope` | `plugin_process`; the environment actually checked |
|
|
24
|
+
| `hook_registered` | This plugin registered its pre-execute listener |
|
|
25
|
+
| `last_checker_result` | `NOT_RUN`, `OK`, or the most recent checker infrastructure diagnostic; a policy refusal can still be an `OK` check |
|
|
26
|
+
| `host_interception` | `UNVERIFIED`; registration and local queries do not certify a live host path |
|
|
27
|
+
|
|
28
|
+
Missing shell service, missing checker, Python launch/import failure, a
|
|
29
|
+
10-second startup timeout, or excessive startup output now fails plugin initialization explicitly.
|
|
30
|
+
The default path reports `GUARD_STARTUP_FAILED`; the optional read guard
|
|
31
|
+
retains `READ_GUARD_STARTUP_REFUSED`. Check the host's plugin status for
|
|
32
|
+
successful activation. If the plugin is absent or inactive, its own status
|
|
33
|
+
tool may not exist. A running host or a CLI status result is insufficient:
|
|
34
|
+
the host can continue running other tools after this plugin fails.
|
|
35
|
+
|
|
36
|
+
If the status script fails after registration, the model tool returns
|
|
37
|
+
`ok: false` and `report: null` with its adapter facts. It omits raw process
|
|
38
|
+
exceptions and unusable process output. These diagnostics describe the status
|
|
39
|
+
query; they do not overwrite the last command-checker result.
|
|
40
|
+
|
|
41
|
+
| Diagnostic | Meaning |
|
|
42
|
+
|---|---|
|
|
43
|
+
| `STATUS_EXECUTION_FAILED` | The caller's shell could not start or finish the status process |
|
|
44
|
+
| `STATUS_TIMEOUT` | The status process timed out, even if it produced output |
|
|
45
|
+
| `STATUS_PROCESS_FAILED` | The status process exited unsuccessfully |
|
|
46
|
+
| `STATUS_INVALID_OUTPUT` | A zero exit status did not provide a usable recovery report |
|
|
47
|
+
|
|
48
|
+
Once a command reaches the checker, process failure, timeout and unusable
|
|
49
|
+
output remain denials. Static `CHECKER_*` diagnostics distinguish these
|
|
50
|
+
failures from a Core policy refusal without copying private command text or
|
|
51
|
+
raw process errors into adapter logs.
|
|
52
|
+
|
|
53
|
+
## What ALLOW_NOOP means
|
|
54
|
+
|
|
55
|
+
`ALLOW_NOOP` keeps its original decision and reason code. Its explanation is:
|
|
56
|
+
|
|
57
|
+
> No recovery action is needed for the recognized operations. Unrecognized command effects have not been checked.
|
|
58
|
+
|
|
59
|
+
For an unrecognized script or interpreter command, this is not a claim that
|
|
60
|
+
the operation is harmless. Scripts, overwrites and other uncovered paths
|
|
61
|
+
retain their existing limits. See the [threat model](../design/threat-model.md).
|
|
62
|
+
|
|
63
|
+
## Moving a file does not remove its recovery copy
|
|
64
|
+
|
|
65
|
+
A successful relocation now reports `storage_notice` and `gc` guidance in
|
|
66
|
+
CLI JSON and displays the same information in terminal output. Enforced
|
|
67
|
+
checks and `safe_delete` also report retained data after a partial relocation,
|
|
68
|
+
including a storage failure after some files moved. The failed decision stays
|
|
69
|
+
unchanged; both JSON and terminal output include the recovery transaction ID.
|
|
70
|
+
A storage failure before any move does not claim retained data. Direct artifact
|
|
71
|
+
deletion, a no-match result and a deletion dry-run do not claim a relocation.
|
|
72
|
+
The DSH model-facing preview puts adapter/storage feedback before long
|
|
73
|
+
target or audit arrays within the existing preview limit.
|
|
74
|
+
|
|
75
|
+
Quarantined data still uses storage. Cross-filesystem moves and actual free
|
|
76
|
+
blocks depend on the filesystem; no exact freed-space measurement is claimed.
|
|
77
|
+
Status uses the existing size accounting and eligibility plan. Regenerable
|
|
78
|
+
artifacts still require the existing ignored-path and artifact-pattern rules.
|
|
79
|
+
|
|
80
|
+
Run maintenance from the same working directory with the same
|
|
81
|
+
`AGENT_GUARD_WORKSPACE` and `AGENT_GUARD_TRASH` settings as the operation.
|
|
82
|
+
Replace the package path below with your installation; CLI results supply
|
|
83
|
+
POSIX-shell-quoted commands for that package and Python interpreter. Use a
|
|
84
|
+
POSIX shell for these command strings; their quoting is not PowerShell or CMD syntax.
|
|
85
|
+
|
|
86
|
+
```sh
|
|
87
|
+
# Inspect the plan; this removes nothing.
|
|
88
|
+
python3 /absolute/agent-guard/skills/delete-guard/scripts/gc.py
|
|
89
|
+
|
|
90
|
+
# Permanently remove the recovery copies selected by the plan.
|
|
91
|
+
python3 /absolute/agent-guard/skills/delete-guard/scripts/gc.py --execute --all-eligible
|
|
92
|
+
|
|
93
|
+
# Or explicitly select one transaction, including one not yet eligible.
|
|
94
|
+
python3 /absolute/agent-guard/skills/delete-guard/scripts/gc.py --execute --txid '<txid>'
|
|
95
|
+
```
|
|
96
|
+
|
|
97
|
+
Review the plan and selected transactions before choosing a purge. A manual
|
|
98
|
+
`--txid` selection is not limited to age/capacity eligibility. The 30-day and
|
|
99
|
+
5-GiB thresholds do not run GC automatically. `--execute` alone selects
|
|
100
|
+
nothing; GC does not prune Git stash snapshots.
|
|
101
|
+
|
|
102
|
+
## Kimi: tool success and recovery feedback
|
|
103
|
+
|
|
104
|
+
In the sampled Kimi 0.42.0, 2.1.1 and 2.3.0 paths, successful PreToolUse recovery
|
|
105
|
+
explanations did not enter the next model request. A successful Bash result
|
|
106
|
+
alone therefore does not tell the model that files were retained.
|
|
107
|
+
The [2.3.0 observations](../reports/test-report-kimi-code-block.md#kimi-230-acceptance)
|
|
108
|
+
used fixed responses through the actual prompt CLI and ACP host, with no
|
|
109
|
+
external model calls.
|
|
110
|
+
|
|
111
|
+
Version 0.2.6 provides a separate, opt-in POSIX Stop hook. It supplies
|
|
112
|
+
actual transaction IDs, state/restore commands, retained-storage guidance and
|
|
113
|
+
full GC commands, then asks for one extra response after the initial answer.
|
|
114
|
+
The initial answer may already be inaccurate, and the model can still omit
|
|
115
|
+
details. Emission cursors track notices, not model acknowledgement.
|
|
116
|
+
|
|
117
|
+
Restoring files returns them to their original location; it does not free
|
|
118
|
+
total disk space. GC permanently removes recovery copies and remains an
|
|
119
|
+
explicit operation. The Stop hook itself executes no restore or GC.
|
|
120
|
+
Published 0.2.5 lacks the Stop helper. See [setup and limits](../../adapters/kimi-code/README.md#optional-recovery-feedback-unreleased-checkout-addition)
|
|
121
|
+
and the [bounded real-host report](../reports/test-report-kimi-code-block.md).
|
|
122
|
+
In 0.2.6, the doctor separates PreToolUse from Stop
|
|
123
|
+
configuration and local feedback probes. `--probe` uses isolated synthetic
|
|
124
|
+
records and leaves existing recovery data alone. Stop's `model_delivery`
|
|
125
|
+
remains `UNVERIFIED`; local PASS does not establish model receipt. See the
|
|
126
|
+
[Stop doctor fields](../../adapters/kimi-code/README.md#stop-doctor-checks-unreleased).
|
package/docs/history/friction.md
CHANGED
|
@@ -9,9 +9,29 @@ F21–F23 changes subsequently shipped in 0.2.3; F24–F25 shipped in 0.2.4.
|
|
|
9
9
|
F26–F31 subsequently shipped in 0.2.5-rc1. RC2 adds shared audit metadata and
|
|
10
10
|
observer startup maintenance; publication is checked by its versioned receipts.
|
|
11
11
|
Current scope is in the
|
|
12
|
-
[candidate notes](../releases/release-notes-0.2.5
|
|
12
|
+
[candidate notes](../releases/release-notes-0.2.5.md); historical failures and
|
|
13
13
|
test counts below are not rewritten by a later success.
|
|
14
14
|
|
|
15
|
+
## Current feedback maintenance · status, explanations and retained storage
|
|
16
|
+
|
|
17
|
+
2026-10-19: this feedback work is included in 0.2.5.
|
|
18
|
+
The recording below describes the implementation work on 2026-10-18.
|
|
19
|
+
|
|
20
|
+
2026-10-18, unreleased 0.2.5 work: script existence was only a file check,
|
|
21
|
+
default DSH startup could return without registering protection, and generic
|
|
22
|
+
runtime errors did not explain what failed. A local Python/import probe and
|
|
23
|
+
explicit initialization failures now separate local readiness from listener
|
|
24
|
+
registration and unverified live-host interception. Runtime failures retain
|
|
25
|
+
deny behavior and report static diagnostics.
|
|
26
|
+
|
|
27
|
+
`ALLOW_NOOP` previously said “Nothing to delete” for commands whose effects
|
|
28
|
+
were not recognized. Its explanation now describes that limitation, without
|
|
29
|
+
changing the reason code or verdict. Relocation results/status/GC plans explain
|
|
30
|
+
retained storage and explicit plan/selection commands. They do not add
|
|
31
|
+
automatic purge or alter rm/glob policy. See the
|
|
32
|
+
[feedback guide](../guides/recovery-feedback.md) for current semantics;
|
|
33
|
+
historical test evidence below retains its original scope.
|
|
34
|
+
|
|
15
35
|
## F1 · Compound `cd X && rm y` resolves targets against the wrong base
|
|
16
36
|
|
|
17
37
|
The interceptor classifies against the tool call's declared workdir. A
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# 蓝队零模型校准 / Blue-team calibration
|
|
2
2
|
|
|
3
|
-
RC3
|
|
3
|
+
RC3 引入的便携示例保留在 0.2.5 中,将现有四种 Lab 控制依次运行并保存简短摘要。
|
|
4
4
|
它复用已交付的 Core,不需要 DSH、Kimi、模型账户、私有 home 或新的 Python 依赖。
|
|
5
5
|
需要 Python 3.9+,并允许本机 IPv4 回环监听;完整 Core 安装仍按[入门指南](../guides/getting-started.md)检查。
|
|
6
6
|
|
package/docs/lab/guard-lab.md
CHANGED
|
@@ -5,16 +5,20 @@ observable agent and harness behaviours. It creates disposable bait, watches
|
|
|
5
5
|
only declared channels, and produces a bounded report. Its built-in controls
|
|
6
6
|
call no model; its separate manual probe is designed for a user-run real
|
|
7
7
|
harness trial. It was introduced in `0.2.3-rc2` and shipped with `0.2.3`.
|
|
8
|
-
It remains available in stable `0.2.4` and published `0.2.5-
|
|
8
|
+
It remains available in stable `0.2.4` and published `0.2.5-rc3`;
|
|
9
9
|
RC2 introduced startup diagnosis and cancellation maintenance
|
|
10
10
|
without adding new model-trial evidence. See the [RC2 notes](../releases/release-notes-0.2.5-rc2.md).
|
|
11
11
|
|
|
12
|
-
The [RC3
|
|
12
|
+
The [RC3 prerelease](../releases/release-notes-0.2.5-rc3.md) includes
|
|
13
13
|
[blue-team research](../research/README.md), sanitized historical observations
|
|
14
14
|
and a [portable four-control example](blue-team-calibration.md).
|
|
15
15
|
Its private research methods do not expand this controller's supported
|
|
16
16
|
channels, evidence levels or native-host coverage.
|
|
17
17
|
|
|
18
|
+
Version [0.2.5](../releases/release-notes-0.2.5.md) retains
|
|
19
|
+
these controls and limits. Its status and storage explanations do not add
|
|
20
|
+
new model trials or production interception channels.
|
|
21
|
+
|
|
18
22
|
It is not a Skill for the tested agent, a security boundary, or a certificate
|
|
19
23
|
that a model or vendor is safe or malicious.
|
|
20
24
|
|
|
@@ -1,5 +1,13 @@
|
|
|
1
1
|
# Agent Guard 0.2.5-rc3
|
|
2
2
|
|
|
3
|
+
Publication status checked 2026-10-19: RC3 was published on 2026-10-14 as a
|
|
4
|
+
[GitHub prerelease](https://github.com/mokuyoaxis/agent-guard/releases/tag/0.2.5-rc3)
|
|
5
|
+
and [npm version](https://www.npmjs.com/package/@mokuyoaxis/agent-guard/v/0.2.5-rc3).
|
|
6
|
+
The next stable release line is [0.2.5](release-notes-0.2.5.md).
|
|
7
|
+
The following preparation snapshot retains its original results and limits.
|
|
8
|
+
|
|
9
|
+
## Preparation snapshot · 2026-10-13
|
|
10
|
+
|
|
3
11
|
Source candidate prepared 2026-10-13. Published baselines remain stable
|
|
4
12
|
[0.2.4](release-notes-0.2.4.md) and prerelease
|
|
5
13
|
[0.2.5-rc2](release-notes-0.2.5-rc2.md). The maintainer authorized the normal
|