@lmzhen/dsh-tool-skill-manage 0.32.4 → 0.32.5
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 +11 -114
- package/package.json +8 -8
package/README.md
CHANGED
|
@@ -1,126 +1,23 @@
|
|
|
1
1
|
# @lmzhen/dsh-tool-skill-manage
|
|
2
2
|
|
|
3
|
-
|
|
4
|
-
routes them through the evolution approval seam when one is mounted; it owns neither the
|
|
5
|
-
library nor its lifecycle rules.
|
|
3
|
+
> Agent 写技能的工具(创建/修补/归档/合并入口)。
|
|
6
4
|
|
|
7
|
-
##
|
|
5
|
+
## 你会看到
|
|
8
6
|
|
|
9
|
-
|
|
10
|
-
- **Prompt prefix / KV cache:** the tool schema is prefix-stable, and skill writes do not alter the current request prompt; catalog invalidation affects the next request; family rules: `packages/README.md` §"Model-visible prompt prefix and the KV cache".
|
|
11
|
-
- **Mount it?** yes — the `tool-skill-manage` row, carried by the `evolution-all` and one-click `evolution-preset` bundles and the Evolution agent preset delta.
|
|
7
|
+
用户看到的是 Agent 创建/修改技能;改错可看历史/撤销。
|
|
12
8
|
|
|
13
|
-
##
|
|
9
|
+
## 什么时候会用到
|
|
14
10
|
|
|
15
|
-
|
|
11
|
+
你让 Agent 把做法写成技能,或它自己改进时。
|
|
16
12
|
|
|
17
|
-
|
|
18
|
-
scalar argument shape, the required arguments per action, the policy's protected list,
|
|
19
|
-
read-before-write, and the operator confirmation. The
|
|
20
|
-
ADMISSION point runs the sequence before the approval seam, so a write that is refused is never
|
|
21
|
-
staged for approval and never spends the operator's attention; the EXECUTION point (`executeCore`)
|
|
22
|
-
re-runs the gates a replay or a direct write can still fail, because a replay's stored arguments
|
|
23
|
-
passed no schema and the protected list can change while a record sits pending. Both points read the
|
|
24
|
-
same table, so they cannot diverge.
|
|
13
|
+
## 默认 / 前提
|
|
25
14
|
|
|
26
|
-
|
|
27
|
-
handles: the `/evolution` and `/graph` command face writes what the operator typed (the operator is
|
|
28
|
-
the authority there), the review's plan executor writes a plan its own filter already screened
|
|
29
|
-
(`filterUnreadSkillOps`), and a record staged outside this tool replays into the execution point
|
|
30
|
-
without re-deciding the admission-only gates — there is no writing session to read.
|
|
15
|
+
all/分层预设才有;host 没有。
|
|
31
16
|
|
|
32
|
-
|
|
33
|
-
write whose target the session never read is refused with `E-318`, and only a read that did not fail
|
|
34
|
-
counts. A session log the tool cannot read proceeds with one warning rather than blocking every
|
|
35
|
-
autonomous write.
|
|
17
|
+
## 边界
|
|
36
18
|
|
|
37
|
-
|
|
38
|
-
before writing; a `delete` carrying `absorbed_into` is the merge protocol and does not ask. WHAT
|
|
39
|
-
that confirmation does is the `skillWriteConfirm` parameter (`/evolution params`, and the
|
|
40
|
-
技能写入规则 card):
|
|
19
|
+
写入有保护;删除是归档。
|
|
41
20
|
|
|
42
|
-
|
|
43
|
-
|---|---|
|
|
44
|
-
| `auto` (**the default**) | the write proceeds with no prompt — an unattended run (a background pass, a scheduled review, a headless session) must not park a tool call on a question nobody will answer |
|
|
45
|
-
| `ask` | the question waits for as long as it takes; the answer is `Create`/`Delete` to proceed, `Cancel` (or a dismissed card) to refuse with `E-317` |
|
|
46
|
-
| `timeout` | the question is asked with a deadline of `skillWriteConfirmTimeoutSeconds` (default 120 s); an unanswered prompt cancels the write with `E-319`, which names the three knobs |
|
|
21
|
+
## 维护者参考:参数别名(G0/S0.3)
|
|
47
22
|
|
|
48
|
-
|
|
49
|
-
this write (a background-origin write, or a deployment with `stageForeground`), the confirmation is
|
|
50
|
-
SKIPPED — the approval window is the stronger gate (it holds the bytes until someone decides and replays
|
|
51
|
-
them), and asking first as well asked the same question twice for one foreground write. The confirmation
|
|
52
|
-
still fires exactly once when there is no such seam. A session whose approval policy is `never` has
|
|
53
|
-
opted out of interrupts altogether; the seam allows the write outright there, which is that instruction
|
|
54
|
-
rather than two gates collapsing into none.
|
|
55
|
-
|
|
56
|
-
The deadline is enforced by the gate itself, not only by the abort signal it hands the question
|
|
57
|
-
service: a prompt that outlives it returns anyway, and the timer is unref'd and cleared, so an
|
|
58
|
-
answered prompt leaves neither a parked call nor a live timer behind. This gate is UX, not a
|
|
59
|
-
security door: it is admission-only (a replayed record already carries the human release that staged
|
|
60
|
-
it), and — in the two modes that ask at all — an unmounted question service, a caller that is not
|
|
61
|
-
the registry's exact live root agent, or a failing ask all PROCEED with one warning. The shipped
|
|
62
|
-
`auto` mode asks nothing, so it never reaches those fallbacks; the operator's own session stays the
|
|
63
|
-
authority that asked for the write.
|
|
64
|
-
|
|
65
|
-
### Approval seam
|
|
66
|
-
|
|
67
|
-
Mutations (create/edit/update/patch/delete/write_file/remove_file/restructure) pass through the evolution approval seam when `evolution-approval` is mounted; approved/staged writes are replayed by the registered runner with the library origin preserved. The admission gates run BEFORE that boundary, so a write that cannot execute is refused, never staged for approval.
|
|
68
|
-
|
|
69
|
-
### pin/unpin: explicit exception
|
|
70
|
-
|
|
71
|
-
`pin` and `unpin` are deliberately **outside** the approval seam. Pinning only lifts/restores the curator-lifecycle freeze (a lifecycle flag, never content) and is fully reversible by the same tool. Routing it through `policy:'ask'` would let a staged-but-never-approved request hold the library in a pinned state invisibly. Tradeoff accepted: no approval on a lifecycle-flag flip; if product policy changes, pin/unpin should be wired into the same staging path as `patch`.
|
|
72
|
-
|
|
73
|
-
## Configuration
|
|
74
|
-
|
|
75
|
-
- `maxSkillContentChars` / `maxSkillFileBytes` bound SKILL.md reads and support-file writes on this row.
|
|
76
|
-
- Deprecated name (G0/S0.3): `maxSkillContentChars` is the legacy spelling of the
|
|
77
|
-
policy row's `skillContentChars` (canonical id). The two are not auto-synchronised with it: this row's value is what
|
|
78
|
-
the write paths use, and since 0.6.0 the user settings layer can override it
|
|
79
|
-
per user. Reading the legacy name still works; writing it is refused. Removal was planned for 0.7.0 and is deferred:
|
|
80
|
-
the alias still ships, so no later version is claimed here.
|
|
81
|
-
|
|
82
|
-
### Description length: the three numbers
|
|
83
|
-
|
|
84
|
-
A skill description is measured against three different bounds. They are not interchangeable, and
|
|
85
|
-
only the second one refuses a write:
|
|
86
|
-
|
|
87
|
-
| Bound | Value | Source | What it does |
|
|
88
|
-
|---|---|---|---|
|
|
89
|
-
| Authoring bar | 60 chars | `AUTHORING_DESCRIPTION_BAR` (the upstream 60-char rule) | advisory feedback on every create/edit/update; `descriptionStrict: true` turns it into a refusal |
|
|
90
|
-
| Family storage ceiling | 1024 chars | `MAX_DESCRIPTION_LENGTH`, lowered per deployment by this row's `maxDescriptionLength` | the hard validation limit — the ONLY refusal threshold of the three |
|
|
91
|
-
| Platform catalog view | 500 chars by default | `catalogDescriptionMaxLength` on the platform's `tool-skill` row (`PLATFORM_CATALOG_DESCRIPTION_DEFAULT`) | truncates the description in the skill catalog the model reads; the family's own rows set that field to 60 (`evolution-host`, `evolution-all` and `evolution-preset` each carry it, and the composer injects it onto a generated preset row too), so 500 governs only a composition that overrides none of them; `packages/README.md` states which of those rows takes effect per install form |
|
|
92
|
-
|
|
93
|
-
The third row is a view, not a rule: a description the family accepts can still be cut in a catalog
|
|
94
|
-
viewer, which is why the authoring feedback names the bar rather than treating the cut as a limit.
|
|
95
|
-
|
|
96
|
-
### The create-time duplicate hint
|
|
97
|
-
|
|
98
|
-
A `create` compares its candidate against the library listing before it writes, over the name +
|
|
99
|
-
description projection (core `nearDuplicateSummaries`, at `SUMMARY_DUPLICATE_HINT_THRESHOLD`), and
|
|
100
|
-
names the closest existing skills in the same "Authoring check" block. Three properties are
|
|
101
|
-
deliberate: it adds no read of its own (`library.list()` already reads every SKILL.md to parse
|
|
102
|
-
its frontmatter — that listing is what publishes `description` — and the hint never calls
|
|
103
|
-
`library.read`, never holds a body and never compares content); it is a HINT (the write proceeds
|
|
104
|
-
— only the model's next decision changes); and a listing failure degrades to a named line instead of
|
|
105
|
-
silence, because "nothing similar was checked" and "nothing similar exists" must not look alike. The
|
|
106
|
-
library-wide BODY scan stays where it was, in the `review` action.
|
|
107
|
-
|
|
108
|
-
### The retention line
|
|
109
|
-
|
|
110
|
-
A whole-body replacement that keeps less than half of the previous SKILL.md says so in the same result
|
|
111
|
-
message (`Content kept 26% of the previous body (2104 of 8123 characters); the replaced version is
|
|
112
|
-
preserved in this skill's history.`). It is computed where the write path still holds both bodies — the
|
|
113
|
-
library's own in-lock read, core `contentRetentionFeedback` — because a caller re-reading the file
|
|
114
|
-
would race the writers the library serializes. Like the duplicate hint it is FEEDBACK, not a gate: no ratio
|
|
115
|
-
refuses a write, small skills stay quiet (a 40-character skill cut in half is not news), and patches,
|
|
116
|
-
support-file writes, creates and archives have no ratio to report.
|
|
117
|
-
|
|
118
|
-
## Known limitations
|
|
119
|
-
|
|
120
|
-
- No known durable consumer gaps at this time. Runtime contracts are covered by package and boundary tests.
|
|
121
|
-
|
|
122
|
-
**Runtime invariant:** No companion is published. The platform auto-assembles nothing and the family mounts no `<pkg>/invariant` cordis row, so a companion here would never execute (v37 S2.1 / I-3).
|
|
123
|
-
|
|
124
|
-
## Notes and history
|
|
125
|
-
|
|
126
|
-
- pin/unpin: explicit exception (0.3.18, E-70)
|
|
23
|
+
`maxSkillContentChars` → `skillContentChars`。旧名仍可读,写入会拒绝。
|
package/package.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@lmzhen/dsh-tool-skill-manage",
|
|
3
3
|
"description": "Model-facing skill_manage tool (community build)",
|
|
4
|
-
"version": "0.32.
|
|
4
|
+
"version": "0.32.5",
|
|
5
5
|
"publishConfig": {
|
|
6
6
|
"access": "public"
|
|
7
7
|
},
|
|
@@ -27,8 +27,8 @@
|
|
|
27
27
|
"license": "MIT",
|
|
28
28
|
"dependencies": {
|
|
29
29
|
"@deepseek-ai/schemastery": "^3.18.5",
|
|
30
|
-
"@lmzhen/dsh-evolution-approval": "^0.32.
|
|
31
|
-
"@lmzhen/dsh-evolution-core": "^0.32.
|
|
30
|
+
"@lmzhen/dsh-evolution-approval": "^0.32.5",
|
|
31
|
+
"@lmzhen/dsh-evolution-core": "^0.32.5"
|
|
32
32
|
},
|
|
33
33
|
"peerDependencies": {
|
|
34
34
|
"@deepseek-ai/cordis": "^4.0.5",
|
|
@@ -36,17 +36,17 @@
|
|
|
36
36
|
"@deepseek-ai/dsh-system-prompt": "^0.2.1-rc.2",
|
|
37
37
|
"@deepseek-ai/dsh-tools": "^0.2.1-rc.2",
|
|
38
38
|
"@deepseek-ai/dsh-agent": "^0.2.1-rc.2",
|
|
39
|
-
"@lmzhen/dsh-evolution-io": "^0.32.
|
|
40
|
-
"@lmzhen/dsh-skill-usage": "^0.32.
|
|
39
|
+
"@lmzhen/dsh-evolution-io": "^0.32.5",
|
|
40
|
+
"@lmzhen/dsh-skill-usage": "^0.32.5"
|
|
41
41
|
},
|
|
42
42
|
"devDependencies": {
|
|
43
43
|
"@deepseek-ai/dsh-agent-loop-testkit": "^0.2.1-rc.2",
|
|
44
44
|
"@deepseek-ai/dsh-skill": "^0.2.1-rc.2",
|
|
45
45
|
"@deepseek-ai/dsh-system-prompt": "^0.2.1-rc.2",
|
|
46
46
|
"@deepseek-ai/dsh-tools": "^0.2.1-rc.2",
|
|
47
|
-
"@lmzhen/dsh-evolution-core": "^0.32.
|
|
48
|
-
"@lmzhen/dsh-evolution-io": "^0.32.
|
|
49
|
-
"@lmzhen/dsh-skill-usage": "^0.32.
|
|
47
|
+
"@lmzhen/dsh-evolution-core": "^0.32.5",
|
|
48
|
+
"@lmzhen/dsh-evolution-io": "^0.32.5",
|
|
49
|
+
"@lmzhen/dsh-skill-usage": "^0.32.5"
|
|
50
50
|
},
|
|
51
51
|
"engines": {
|
|
52
52
|
"dsh": ">=0.2.1-rc.2 <0.2.2-0"
|