@uzysjung/agent-harness 26.185.0 → 26.186.0

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/dist/index.js CHANGED
@@ -150,7 +150,7 @@ import {
150
150
  withoutAppleDouble,
151
151
  wrapHarnessBlock,
152
152
  writeInstallLog
153
- } from "./chunk-3OFLAQWT.js";
153
+ } from "./chunk-Q7JZK7RQ.js";
154
154
 
155
155
  // node_modules/sisteransi/src/index.js
156
156
  var require_src = __commonJS({
@@ -17,7 +17,7 @@ import {
17
17
  init_esm_shims,
18
18
  residentCost,
19
19
  resolveBundleRoot
20
- } from "./chunk-3OFLAQWT.js";
20
+ } from "./chunk-Q7JZK7RQ.js";
21
21
 
22
22
  // src/trust-tier-drift.ts
23
23
  init_esm_shims();
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@uzysjung/agent-harness",
3
- "version": "26.185.0",
3
+ "version": "26.186.0",
4
4
  "description": "Curate vetted AI-coding skills & plugins by your tech stack — install only what you need, across Claude Code, Codex, OpenCode & Antigravity",
5
5
  "type": "module",
6
6
  "publishConfig": {
@@ -12,7 +12,9 @@ and explicit authority take precedence over preferred methods.
12
12
  Start from the intended user's goal, the core usage flow, and what makes the
13
13
  outcome worth choosing over existing alternatives. Reuse relevant context and
14
14
  ground consequential decisions in project evidence or applicable authoritative
15
- sources. Make assumptions visible when they affect the result.
15
+ sources. Make assumptions visible when they affect the result. When an input
16
+ the task depends on is missing, report what is missing and what it blocks
17
+ instead of producing stand-in results that could pass for real ones.
16
18
 
17
19
  Define success by the outcome users need and the failures they must be protected
18
20
  from. Distinguish essential readiness from optional polish, using the intended
@@ -66,11 +66,60 @@ design. The session integrates either result; neither arrangement requires an
66
66
  extra review by default. Compare separately developed alternatives when their
67
67
  different approaches justify the additional work.
68
68
 
69
+ Route by how open the work is. When what to do, or the answer itself, is still
70
+ open (setting direction, product planning, scoping, building a plan), use the
71
+ strongest reasoning model available. When a judgment stalls mid-task or a problem
72
+ needs a fresh angle, first measure what you can (run it, check the data, read the
73
+ code) and bring in the strongest model only if the judgment still splits. Design,
74
+ implementation, and verification inside a scope that a spec, the project's
75
+ direction, or a benchmark has already fixed go to a capable general model or the
76
+ session itself. Work with no judgment at all, such as link checks or confirming
77
+ that a rebased change matches the reviewed one, goes to tools, direct execution,
78
+ or the implementer.
79
+
80
+ Creative work with no single right answer (a name, a brand or product concept, a
81
+ visual direction, copy) gains more from breadth than from depth. Send the same
82
+ brief separately to two or three models with different tendencies and collect
83
+ their candidates. Leave your own examples and hypotheses out of that brief: one
84
+ example pulls every candidate toward it. Let the user choose from the candidates
85
+ side by side; simulated personas are not the user's judgment. Refining the chosen
86
+ direction is then work inside a settled scope.
87
+
69
88
  Choose effort by actual difficulty, interacting decisions, remaining uncertainty,
70
89
  likely rework, and the benefit of further reasoning. High stakes call for stronger
71
90
  assurance, not automatically maximum effort: inspecting code, contracts, tests,
72
91
  or external facts may resolve a risk that more reasoning alone cannot.
73
92
 
93
+ For open work, start at high effort. When open work is also hard to reverse
94
+ (permission boundaries, data deletion, migrations, money), start at a higher
95
+ effort; a high run followed by a rerun pays twice. Being hard to reverse alone
96
+ does not make work open: a migration whose answer is already settled goes to a
97
+ capable builder with a stronger reviewer, not to the strongest model. Count the
98
+ answer as settled only when a spec, a benchmark, or an established pattern
99
+ settles it, and send the work back up when the builder uncovers an open choice. Treat the maximum effort level as an
100
+ exception to raise with the user.
101
+
102
+ Judge every delegated result, from a planner, an implementer, or a reviewer, by
103
+ evidence rather than tone: a shallow report reads as confidently as a sound one.
104
+ An implementation or verification report needs the commands run with their exit
105
+ codes, and for a new guard, a record of the test failing when the guard is
106
+ reverted. A review of a consequential area (permissions, approvals, personal
107
+ data, exports, audit, migrations) needs a record of reverting the key lines and
108
+ watching the tests fail, the counterexample inputs actually run with their
109
+ results, and boundary-value checks; a claim of having tried a counterexample
110
+ without the input and the result counts as missing. A design or decision needs
111
+ the location of each current fact, what breaks if each option is wrong, the
112
+ conditions that would change the recommendation, and the assumptions it could
113
+ not confirm. A calculation needs its steps and raw data. Rerun only the part that
114
+ lacks this, with the doubt written into the request, instead of rerunning the
115
+ whole at a higher effort. When the doubt is about a calculation or a behavior,
116
+ measure it first and raise the effort only if the judgment still splits.
117
+ Narrow confirmation whose
118
+ answer is fixed (that a fix now passes the cited counterexample, that a document
119
+ points to the right decision, that issue or pull request status matches the
120
+ documents) needs only moderate effort, and moves up when it turns into finding
121
+ new counterexamples or making a judgment.
122
+
74
123
  ## 3. Make additional contexts earn their cost
75
124
 
76
125
  Add a worker for a concrete benefit: needed expertise or tools, useful context or
@@ -106,6 +155,24 @@ references, unresolved question, allowed changes, and completion criteria. Suppl
106
155
  enough context to decide correctly, including material failed approaches; favor
107
156
  accessible source references over replaying the conversation. Keep critical
108
157
  constraints explicit rather than relying on a lossy summary or hidden context.
158
+ A worker asked to plan or design from facts it cannot see will design from
159
+ assumption, so name where it can confirm the current facts (files, records,
160
+ queries) and give it read access to them. Where access itself would add a
161
+ boundary to check, such as production data, measure the few values it needs
162
+ yourself and pass them instead; when it cannot yet know which values it needs,
163
+ give read access to a copy or test environment rather than production.
164
+
165
+ How much to specify depends on how open the work is. For work whose scope is
166
+ already fixed, spell out the files, interfaces, expected values, and checks. For
167
+ open work handed to a stronger model, give direction instead of a procedure: the
168
+ background and context (who it is for, what has happened, the current facts and
169
+ where to confirm them), the goal to pursue and how the result will be judged, and
170
+ the guardrails (scope, constraints, budget, what must not happen, where to put
171
+ the result). Leave the decomposition to that model: which questions to ask, which
172
+ options to compare, and how to structure the answer. A list of sub-questions or a
173
+ prescribed outline fixes the shape of the answer before the stronger model has
174
+ looked at the problem. If you hold a hypothesis, pass it as one labelled input
175
+ the worker may reject.
109
176
 
110
177
  Ask for the result needed for integration, its evidence, and remaining gaps.
111
178
  Short results can return directly. Use durable files or artifacts for large,
@@ -155,7 +222,10 @@ context makes correction efficient; reassess only the affected evidence.
155
222
 
156
223
  Resolve model availability, effort controls, inheritance, and tool behavior from
157
224
  the actual environment or applicable documentation when needed; reuse current
158
- evidence. Distinguish requested settings from observed execution. Do not claim
225
+ evidence. Distinguish requested settings from observed execution: a requested effort or
226
+ model can be silently not applied. Where a per-call setting can be dropped, fix it
227
+ in the worker's definition, and still confirm the actual value from the
228
+ execution record before relying on it. Do not claim
159
229
  a model, effort level, or independent review that was not established.
160
230
 
161
231
  Apply the same scope, data protections, and approval boundaries to every worker,
@@ -109,6 +109,25 @@ description: >-
109
109
 
110
110
  직접 호출하지 않은 완료 보고는 **무엇이 달라졌는지 → 무엇을 실제로 확인했는지 → 남은 제한이나 필요한 행동**을 중심으로 전달한다. 해당하지 않는 항목은 생략한다.
111
111
 
112
+ ### 표현 수단
113
+
114
+ 글과 표가 기본이다. 글로는 한 번에 들어오지 않는 내용이면 수단을 바꾼다.
115
+
116
+ - 구성 요소의 관계 · 흐름 · 순서 · 상태 변화는 도식으로 보여준다. 읽는 화면이 그려 주는 형식이 있으면 그것을, 없으면 코드 블록의 텍스트 도식을 쓴다.
117
+ - 항목이 많아 훑어보며 골라 읽어야 하거나 눌러 봐야 이해되는 내용은 한 파일로 열리는 HTML 페이지가 낫다. 쓰고 버리는 산출물이므로 미리보기 화면이나 프로젝트 소스 밖의 임시 위치에 두고, 무엇을 어디에 만들었는지 알린다.
118
+ - 영상은 사용자가 요청할 때, 설치된 영상 도구로 만든다.
119
+
120
+ 만드는 부담이 이해에 주는 이점보다 크면 글로 답하고, 더 나은 수단이 있어 보이면 한 줄로 제안한다.
121
+
122
+ ### 쉬운 문장
123
+
124
+ 처음 읽는 사람이 한 번에 따라오게 쓴다.
125
+
126
+ - 한 문장에는 한 가지만 말한다. 절차는 한 문장에 동작 하나다.
127
+ - 누가 무엇을 하는지 드러나게 쓴다.
128
+ - 한 개념에는 끝까지 한 용어를 쓰고, 처음 나올 때 뜻을 밝힌다.
129
+ - 일상어로 바꿀 수 있는 전문어 · 약어는 바꾸고, 꼭 필요한 것만 남긴다.
130
+
112
131
  ### 화면·동작의 전후 비교
113
132
 
114
133
  UI/UX 변경에서 배치·문구·동작의 차이를 말만으로 판단하기 어렵거나 사용자가 시각적인 비교를 요청하면, 변경 전후를 보여준다.