Back to skill

Security audit

网页重复操作录一遍以后自动跑

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this browser-workflow skill suspicious because scanners report weaknesses that could expose recorded data, exceed stated limits, or perform actions without the promised safeguards.

Consider avoiding installation until the reported data-capture, approval, browser cleanup, and execution-scope issues have been reviewed and addressed. Coverage is incomplete: Skillspector inspected 51 of 79 components, and its analysis was partial; the static analyzer also had degraded coverage. The suspicious verdict does not establish malicious intent.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (7)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/lib-cdp.mjs:117
Finding

Substring-based browser target selection exposes request data to unrelated pages ## Remediation Suggestions - Compare parsed, exact origins rather than matching substrings of complete URLs. - Bind execution to a specific CDP target ID and revalidate its origin immediately before evaluating requests. - Abort if the target navigates outside the authorized origin. - Avoid placing credentials in an untrusted page’s main-world JavaScript environment. - Add regression tests with attacker URLs containing the legitimate host in their query string, path, or hostname suffix.

Content
View full analysis

Vulnerability Details

File Location: scripts/lib-cdp.mjs:117–140; scripts/api_run.mjs:137–157; scripts/probe.mjs:56–79
Vulnerability Type: Incorrect origin validation and sensitive request exposure
Risk Level: High

Technical Analysis

Browser targets are selected by checking whether their entire URL contains the intended host string:

js
const list = await listTargets(port);
const pages = list.filter(t => t.type === 'page');
let hit = urlPart ? pages.find(p => (p.url || '').includes(urlPart)) : pages[0];

Both API execution and probing attach using this host substring:

js
const cdp = await Cdp.attach({ port, urlPart: host });
await cdp.enableSession();

The probe copies recorded headers and constructs a request that executes in the selected page:

js
const hdrs = { ...(target.headers || {}) };
delete hdrs['content-type'];

async function callUrl(url) {
  const expr = `(async () => {
    try {
      const opt = { method: ${JSON.stringify(target.method || 'GET')}, credentials: 'include'
        ${Object.keys(hdrs).length ? `, headers: ${JSON.stringify(hdrs)}` : ''}
        ${target.body ? `, body: ${JSON.stringify(target.body)}` : ''} };
      const r = await fetch(${JSON.stringify(url)}, opt);
      const t = await r.text();
      return JSON.stringify({ ok: true, status: r.status, ct: r.headers.get('content-type') || '', len: t.length, body: t.slice(0, 200000) });
    } catch (e) { return JSON.stringify({ ok: false, err: String(e && e.message || e) }); }
  })()`;

An unrelated page can include the expected host in its path or query string. If selected, its JavaScript environment receives the request arguments. An overridden window.fetch can inspect those arguments before any network request occurs; CORS does not prevent this interception.

Attack Path

  1. An attacker gets a page opened in the debugging browser with a URL containing the authorized service host, such as `https://attacker.example/?target=servic ...[truncated 852 chars]

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/evidence.mjs:143
Finding

Evidence reports interpret recorded webpage content as executable HTML ## Remediation Suggestions - Introduce a dedicated HTML encoder for all dynamic text, including locators, action details, notes, URLs, and response field names. - Use context-specific encoding for attributes. - Prefer DOM construction with `textContent` over HTML string concatenation. - Apply a restrictive Content Security Policy that blocks scripts and outbound connections in static evidence reports. - Test report generation using HTML payloads in page-controlled locator attributes.

Content
View full analysis

Vulnerability Details

File Location: scripts/evidence.mjs:58,143–170; input collection in assets/recorder.js:74–90
Vulnerability Type: Stored HTML and script injection
Risk Level: Medium

Technical Analysis

The report encoder handles Markdown separators and line breaks, but does not encode HTML metacharacters:

js
const esc = s => String(s ?? '').replace(/\|/g, '\\|').replace(/\r?\n/g, ' ⏎ ').slice(0, 160);

Recorded locator content is then inserted into the generated HTML:

js
const loc = a.locator?.best || a.locator?.L4_text || '';

The corresponding HTML sink is:

html
<td><code>${esc(loc)}</code></td>

The recorder derives locators from page-controlled attributes:

js
var tids = ['data-testid', 'data-test', 'data-qa', 'data-cy'];
for (var i = 0; i < tids.length; i++) {
  var v = el.getAttribute(tids[i]);
  if (v) { r.L2_testid = '[' + tids[i] + '="' + v + '"]'; break; }
}

Consequently, markup stored in an attribute or other recorded field is interpreted as HTML when the evidence report is opened.

Attack Path

  1. An external content provider controls a recorded element’s text or locator attribute.
  2. The provider places HTML containing an executable event handler in that value.
  3. The user interacts with the element during recording.
  4. evidence.mjs incorporates the recorded value into its HTML report without HTML encoding.
  5. Opening the report activates the injected markup.

This path uses independently controlled webpage content, not merely a user creating an HTML file for themselves.

Impact Assessment

Injected JavaScript can modify the evidence display, read content embedded in the report, and attempt outbound transmission of that content. It can mislead the person reviewing the recorded workflow.

Execution requires opening the HTML report. Browser-origin restrictions remain applicable; arbitrary local-file access or operating-system code execution was not established.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/analyze.py:385
Finding

Mapping-review pages allow script injection from recorded table metadata ## Remediation Suggestions - HTML-encode all mapping labels, table names, unresolved descriptions, and column headings. - Attribute-encode radio-button values. - Serialize JavaScript data safely rather than interpolating it into quoted source. - Prevent HTML parser termination sequences such as `</script>` when embedding serialized data. - Prefer external scripts and DOM APIs using `textContent`. - Test titles and column headings containing angle brackets, quotes, event handlers, and script-termination sequences.

Content
View full analysis

Vulnerability Details

File Location: scripts/analyze.py:150–160,173–188,385–424
Vulnerability Type: HTML, attribute, and inline JavaScript injection
Risk Level: Medium

Technical Analysis

The analyzer incorporates recorded document titles and copied table headings into mapping data. The review generator inserts mapping values directly into table cells without HTML encoding.

It also constructs radio-button attributes using unescaped candidate values:

python
key_opts = ''.join('<label><input type="radio" name="key" value="%s"%s> %s</label> '
                   % (c, ' checked' if c == chosen_key else '', c) for c in key_names) or '<input name="key" placeholder="...">'

The first expression is the vulnerable construction; the fallback placeholder is abbreviated here. Separately, the HTML template inserts chosen_key into a single-quoted inline JavaScript string:

js
var rows=[['KEY',document.querySelector('input[name=key]:checked')?document.querySelector('input[name=key]:checked').value:document.querySelector('input[name=key]').value||'%s']];

The same untrusted values therefore cross several interpretation contexts without context-specific encoding.

Attack Path

  1. An attacker controls a shared document’s title or table headings on a page used in the supported recording workflow.
  2. The user records copying or pasting the table.
  3. The analyzer derives mapping labels or key candidates from that external content.
  4. The generator places those values into the mapping-review HTML.
  5. The user opens the review page, as required by the documented UI workflow.
  6. Injected markup or script executes and can alter the review controls or exported confirmation data.

The reportable boundary is external page/table content becoming executable review code. A user’s self-authored local table alone does not establish that boundary.

Impact Assessment

The attacker can manipulate the mapping-approval interface and the confirmation C ...[truncated 231 chars]

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/record.mjs:60
Finding

Network recording ignores the visible start and pause consent boundary ## Remediation Suggestions - Enforce a host-side recording-state check before collecting each request. - Apply start, pause, and stop consistently across all attached sessions. - Require explicit target selection or an approved origin/target scope. - Discard responses belonging to requests outside the authorized interval. - Redact authentication material and sensitive request fields before persistence. - Add tests proving that idle and paused network activity never enters output artifacts.

Content
View full analysis

Vulnerability Details

File Location: scripts/record.mjs:60–82,131,157–175,208,258–263; scripts/lib-cdp.mjs:274–305
Vulnerability Type: Overcollection outside authorized recording state
Risk Level: Medium

Technical Analysis

Every eligible browser tab receives a network listener when attached:

js
const rec = { id: t.id, cdp: c, net: [], dl: null, url: t.url, title: t.title || '', injectedAt: 0 };
rec.net = watchNetwork(c);
rec.dl = watchDownloads(c, dlDir);

The network listener does not check whether recording has started or is paused:

js
cdp.on('Network.requestWillBeSent', p => {
  if (!['XHR', 'Fetch'].includes(p.type)) return;
  if (reqs.length >= MAX) { const old = reqs.shift(); byId.delete(old.id); }
  const r = {
    id: p.requestId, at: Date.now(), type: p.type,
    method: p.request.method,
    url: p.request.url.slice(0, 500),
    body: p.request.postData ? String(p.request.postData).slice(0, 1500) : '',
    headers: pickHeaders(p.request.headers),
    status: null, mime: '', size: 0, ms: null,
    respPreview: '', fields: null, rows: null
  };
  byId.set(p.requestId, r);
  reqs.push(r);
});

At completion, the collected requests are merged:

js
const netReqs = [...tabs.values()].flatMap(r => r.net || []);

They are included in the recording object and persisted:

js
fs.writeFileSync(outFile, JSON.stringify(out, null, 1), 'utf8');

The visible start and pause state controls page-event recording but not this host-side CDP collection.

Attack Path

  1. The user starts the recorder in a browser containing eligible tabs.
  2. The recorder attaches network listeners before the visible start action.
  3. A tab issues XHR/Fetch requests while recording is idle or paused.
  4. Those requests are retained regardless of the recording state.
  5. Finishing the session writes them into the recording artifact.

Independent page scripts can generate these requests. An attacker is not required for the consent violation ...[truncated 467 chars]

T09 · Insecure Skill Coding Practices

Warning
Location
assets/recorder.js:287
Finding

Password input and paste events are persisted as recording content ## Remediation Suggestions - Exclude password inputs before constructing any action. - Apply the same exclusion to paste, snapshot, and attribute collection. - Recognize other sensitive fields using explicit configuration and relevant autocomplete attributes. - Represent credentials as runtime-only parameters rather than recorded defaults. - Redact authentication bodies in network capture so field-level exclusions cannot be bypassed through request recording. - Add tests covering typed, pasted, and autofilled passwords.

Content
View full analysis

Vulnerability Details

File Location: assets/recorder.js:249–267,287–300; scripts/record.mjs:136–142,263
Vulnerability Type: Sensitive credential persistence contrary to the declared boundary
Risk Level: Medium

Technical Analysis

The change handler excludes checkbox, radio, and file inputs, but not password inputs:

js
if (tag === 'input' || tag === 'textarea') {
  if (/^(checkbox|radio)$/i.test(el.type)) { return; }
  if (el.type === 'file') { push('upload', el, { value: el.value || '(...)' }); return; }
  push('fill', el, { value: el.value });
}

The file-input fallback message is abbreviated; the password-relevant branches are unchanged.

Paste handling also stores the clipboard text without excluding sensitive targets:

js
push('paste', el, {
  value: text.slice(0, 8000),
  rows: lines.length,
  cols: cols,
  grid: gridContextOf(el)
});

Recorded actions are merged into the output and written to disk:

js
fs.writeFileSync(outFile, JSON.stringify(out, null, 1), 'utf8');

This contradicts the skill’s explicit promise that login passwords remain in memory and do not enter artifacts. The --no-copy-text option only affects copy events.

Attack Path

  1. Recording is active.
  2. The user types a password into a password input and commits the change, or pastes a credential into that field.
  3. The recorder stores the plaintext value as a fill or paste action.
  4. Session completion writes it into the recording JSON.
  5. Downstream evidence or analysis can create additional copies of the recorded value.

Impact Assessment

Credentials intended only for website authentication become durable automation data. Subsequent artifact handling can expose them beyond the intended authentication operation.

No actual secret was identified in the supplied fixtures, no unauthorized file reader was demonstrated, and no external transmission was established. The confirmed issue is unwanted credential persistence, not proven creden ...[truncated 11 chars]

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/probe.mjs:34
Finding

Live API probing bypasses the shared write-approval gate ## Remediation Suggestions - Reuse the common write gate and write audit for every live request path. - Require explicit approval of the method, target, body, and planned request count. - Disable minimization for mutating requests by default. - Permit write probing only in an isolated test environment or under an explicitly reviewed idempotency plan. - Provide a render-only mode that cannot attach to the browser or issue requests.

Content
View full analysis

Vulnerability Details

File Location: scripts/probe.mjs:34–38,66–79,104–133
Vulnerability Type: Inconsistent authorization for production writes
Risk Level: Medium

Technical Analysis

The probe checks only its own --go switch for methods outside GET, HEAD, and OPTIONS. It does not invoke the common writeGate, require --allow-write and an approval acknowledgement, or create the corresponding write audit record.

After that check, it replays the original request:

js
const opt = { method: ${JSON.stringify(target.method || 'GET')}, credentials: 'include'
  ${Object.keys(hdrs).length ? `, headers: ${JSON.stringify(hdrs)}` : ''}
  ${target.body ? `, body: ${JSON.stringify(target.body)}` : ''} };
const r = await fetch(${JSON.stringify(url)}, opt);

Parameter minimization repeats the request after deleting each query parameter:

js
for (const k of keys) {
  const test = new URL(url0); test.searchParams.delete(k);
  const r = await callUrl(test.toString());
  const s = summarize(r.json);
  const is2xx = r.ok && r.status >= 200 && r.status < 300;
  const stillOk = is2xx && (sum.rows != null ? (s.rows != null && s.rows > 0) : !!s.fields);
  console.log(`  ${k}: ${r.status ?? r.err}`);
  if (!stillOk) keep.push(k);
  await sleep(300);
}

The diagnostic logging is shortened above; the request repetition and result checks are preserved.

Attack Path

  1. A recording contains a production write request.
  2. An agent invokes the probe with --go to investigate it, without the shared limit-and-approver acknowledgement required elsewhere.
  3. The probe sends the authenticated request with its original method and body.
  4. If --minimize is enabled, it sends additional requests for individual query-parameter deletions.
  5. Non-idempotent endpoints can perform repeated production mutations.

Impact Assessment

The probe can execute production writes without the promised shared approval and audit boundary. Parameter minimization can mu ...[truncated 281 chars]

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/build_skill.mjs:134
Finding

Generated offline-preview instructions issue live authenticated requests ## Remediation Suggestions - Add `--dry` to every generated offline-preview command and related terminal hint. - Make render-only behavior the default. - Separate offline rendering, live reads, approved writes, and output-file creation into clearly named modes. - Ensure render-only mode cannot attach to or close a browser. - Add tests asserting that generated preview commands perform no CDP or service requests. The audit was static only: no project commands or tests were executed. The findings establish implementation defects and high-risk behavior, not malicious author intent.

Content
View full analysis

Vulnerability Details

File Location: scripts/build_skill.mjs:134–136,159; scripts/api_run.mjs:81–88,137–157,197–208
Vulnerability Type: Misleading execution mode that exceeds preview authorization
Risk Level: Medium

Technical Analysis

The builder labels this generated command as rendering the request without touching the browser:

sh
node scripts/api_run.mjs --config <configuration-directory>

However, the runner only exits before browser access when --dry is supplied:

js
if (flag('dry')) {
  console.log(JSON.stringify({ url, method, params, bodyTemplate: bodyTemplate || null, body }, null, 1));
  process.exit(0);
}

Without that flag, it attaches to a browser target and evaluates a real request:

js
const cdp = await Cdp.attach({ port, urlPart: host });
await cdp.enableSession();
js
const r = await fetch(${JSON.stringify(url)}, opt);
js
const raw = await cdp.evalJs(expr, { awaitPromise: true }).catch(e => JSON.stringify({ ok: false, err: String(e.message || e) }));

The later output condition does not prevent the request:

js
if (!flag('go') || !out) {

Thus, --go controls CSV persistence rather than whether browser interaction and network execution occur.

Attack Path

  1. The builder generates an API child skill and command reference.
  2. The user or agent selects the command described as an offline preview.
  3. That command omits --dry.
  4. The runner connects to the authenticated browser and sends the configured request.
  5. Only afterward does it select the output-preview branch.

No independent attacker is necessary for this high-risk authorization mismatch: the generated instructions themselves expand a preview operation into live execution.

Impact Assessment

An operation authorized as offline inspection instead interacts with the configured service using the browser session and processes live account data. The completion path can also invoke browser cleanup.

Standard writ ...[truncated 132 chars]

Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (85)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The delivery instructions require executing test files explicitly outside the declared script whitelist. Although testing is legitimate, these instructions encourage an agent to cross an execution-permission boundary; the test files' behavior is not established by this source.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 125)May include surrounding context.

md
| 登录态 | 复用你已登录的浏览器;要无人值守才用 `login.mjs` 走账密独立通道 | 未登录时录到的都是登录页,接口 401(运行时归因码 `LOGIN`) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 204)May include surrounding context.

md
| 10 | `login.mjs` | 独立登录通道(账密只进内存,不落盘) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 296)May include surrounding context.

md
3. **独立账密通道**:`login.mjs` 开专属 `user-data-dir` 登一次,之后 `api_run --profile … --bg` 脱人跑;密码只从 `SF_USER`/`SF_PASS` 读,不落盘。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 186)May include surrounding context.

md
- ❌ **不执行任意用户命令**:全库无 `eval` / `new Function`;子进程调用只有两处——关浏览器(端口号来自 CLI 参数,模板写死)和跑同包的 `diffrec.mjs`(`process.execPath`),没有任何一条会把用户输入拼成命令交给 shell。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 198)May include surrounding context.

md
| 4 | `diffrec.mjs` | 两遍录制对比,判出真正的变量 |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 264)May include surrounding context.

md
| ②.5 判参 | **首选**:同一操作再录一遍,`node scripts/diffrec.mjs --a 第二遍.rec.json --b 第一遍.rec.json`(`record.mjs` 发现同名上一份录制会**自动**跑这步;录完也会自动按端口关浏览器)<br>**只有一遍时**:`node scrip

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 189)May include surrounding context.

md
> 复核方式:`grep -rn "child_process\|eval(" scripts/` 只应命中 `lib-cdp.mjs`(关浏览器)与 `record.mjs`(跑同包 diffrec)。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 189)May include surrounding context.

md
> 复核方式:`grep -rn "child_process\|eval(" scripts/` 只应命中 `lib-cdp.mjs`(关浏览器)与 `record.mjs`(跑同包 diffrec)。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 195)May include surrounding context.

md
| 1 | `record.mjs` | 起独立 Chrome + 注入浮标,录点击/跳转/接口/备注(多标签整链录) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 262)May include surrounding context.

md
| ① 录 | `node scripts/record.mjs --url <起始页,可 about:blank> --name <链路名> [--port 18802]`<br>页面右上角自动出现**浮标**:状态(录制中/已暂停)+ 实时计数 + 操作流水 + **「说一句」备注框** + 暂停 / 结束按钮 |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 264)May include surrounding context.

md
| **首选**:同一操作再录一遍,`node scripts/diffrec.mjs --a 第二遍.rec.json --b 第一遍.rec.json`(`record.mjs` 发现同名上一份录制会**自动**跑这步;录完也会自动按端口关浏览器)<br>**只有一遍时**:`node scripts/infer_

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 276)May include surrounding context.

md
2. forge 全部脚本照常加 `--port <档位口>`:`record.mjs` / `api_run.mjs` / `run.mjs` 都行——`ensureChrome` 发现端口活着就直接挂上,不再另开 Chrome。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 307)May include surrounding context.

md
`action.note`,这步干什么 / 什么规则),没动作时进 `rec.json.notes` 当全局规则;点浮标日志某一行可把备注挂到**指定步**;`record.mjs` 收尾打步骤清单,`evidence` 时间线出「这步干什么/规则」列,`interview.mjs --rec` 自动带出来 | 备注不

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 196)May include surrounding context.

md
| 2 | `evidence.mjs` | 录制 → 证据包(时间线 + 接口候选打分) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 263)May include surrounding context.

md
| ② 证 + 探 | `node scripts/evidence.mjs --recording X.rec.json`<br>`node scripts/probe.mjs --recording X.rec.json --index N [--minimize]` | `证据包.md/html`(时间线 + 接

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 197)May include surrounding context.

md
| 3 | `probe.mjs` | 探查接口(构造请求、看响应、出配置草稿) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 263)May include surrounding context.

md
| ② 证 + 探 | `node scripts/evidence.mjs --recording X.rec.json`<br>`node scripts/probe.mjs --recording X.rec.json --index N [--minimize]` | `证据包.md/html`(时间线 + 接

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 199)May include surrounding context.

md
| 5 | `infer_params.mjs` | 单遍录制判参(提示,不是定论) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 264)May include surrounding context.

md
一遍.rec.json`(`record.mjs` 发现同名上一份录制会**自动**跑这步;录完也会自动按端口关浏览器)<br>**只有一遍时**:`node scripts/infer_params.mjs --rec X.rec.json [--out 参数判定.csv]` 先出变量提示(日期/单号/金额/接口带出

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 306)May include surrounding context.

md
| **单遍判参** | 只有「录两遍比 diff」才知道哪个是变量,成本翻倍 | `node scripts/infer_params.mjs --rec X.rec.json [--out 参数判定.csv]`:日期 / 长单号 / 金额 / URL 带出 / 接口响应里出现 → 判变量并给理由 | **提示,不是

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 200)May include surrounding context.

md
| 6 | `interview.mjs` | 录完问需求(什么时候跑/参数哪来/给谁),写 `触发.json` |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 265)May include surrounding context.

md
| ④ 问 | `node scripts/interview.mjs --link <名> [--rec X.rec.json] [--type api\|ui]`<br>人在场就一条条问;agent 已问过就用 `--answers answers.json`;**录制时浮标里说的原话会自动带出来**,不用再问一遍

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 201)May include surrounding context.

md
| 7 | `build_skill.mjs` | 生成子技能包(接口版 / UI 版 + 说明书) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 266)May include surrounding context.

md
| ⑤ 固 | `node scripts/build_skill.mjs --link <名> --config <探查目录> --type api --verdict-from <参数判定.json> --require-trigger`<br>或 `--type ui` | `produced/<链路>/` 自包