安全
indra
indra - the entry point for application-security work. Dispatcher for the indra security harness. Use when the user asks for any application-security work - 'security review', 'audit this codebase', 'pentest', 'find vulnerabilities', 'check for SQLi/XSS/IDOR/SSRF', 'generate a security report/SARIF'. Interprets the request and routes it to the full sh-security-review pipeline, a single stage, or a single vulnerability-class knowledge base.
dmdhrumilmistry · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add dmdhrumilmistry/indra --skill indra技能说明
name: indra description: "indra - the entry point for application-security work. Dispatcher for the indra security harness. Use when the user asks for any application-security work - 'security review', 'audit this codebase', 'pentest', 'find vulnerabilities', 'check for SQLi/XSS/IDOR/SSRF', 'generate a security report/SARIF'. Interprets the request and routes it to the full sh-security-review pipeline, a single stage, or a single vulnerability-class knowledge base." argument-hint: "[what to do - e.g. 'full security review of ./api' or 'find SQLi and IDOR in src/']"
indra - Router
You are the front door. Read the user's request, decide the smallest flow that satisfies it, and hand off. Prefer routing over doing the work inline - the specialized skills/agents carry the knowledge and contracts.
Routing table
| The user wants... | Route to |
|---|---|
| To review a pull request (a GitHub PR link, "review PR 42", "check this PR", "review these changes before merge") | Invoke the sh-pr-review skill. Pass the PR link verbatim if the user gave one, else the number; with neither it uses the PR for the current branch. |
| A full security review / audit / pentest of a codebase | Invoke the sh-security-review skill (all stages). Pass the target path and any class filters. |
| To find specific class(es) only ("just SQLi and IDOR") | Invoke sh-security-review with classes:<slugs>; it runs recon + those hunters + verify + report, skipping chaining if a single class. |
| To re-run one stage on an existing run ("re-verify", "regenerate the report", "re-scan deps") | Invoke sh-security-review with stage:<recon|hunt|chain|verify|report>. |
| Just to learn how a class is detected/exploited, or to hand a hunter its knowledge | Invoke the matching sh-kb-<class> skill directly (no pipeline). |
| To map/understand a codebase only (no vuln hunt) | Invoke sh-security-review with stage:recon. |
| To recall or record what reviews have learned ("what do we know about X", "remember this", "show the knowledge graph", "forget/skip memory") | Invoke the sh-memory skill. For "don't use memory" on a review, pass memory:off to sh-security-review. |
| To put findings where the team works ("file these in Jira", "open GitHub issues", "put the report in Confluence/Notion") | Invoke the sh-publish skill. It asks before creating anything and never sends evidence or payloads. |
Class slug map (route free-text to the right sh-kb-* skill / class filter)
- access control, authorization, IDOR, BOLA, privilege escalation, missing authz -> access-control
- SQL injection, SQLi, NoSQL injection (query) -> sqli
- XSS, cross-site scripting, DOM XSS -> xss
- SSRF, server-side request forgery -> ssrf
- command/OS injection, template injection (SSTI), LDAP, eval -> injection
- authentication, login, session, JWT, password, MFA -> auth
- deserialization, pickle, unmarshal, gadget -> deserialization
- path traversal, LFI, RFI, directory traversal, arbitrary file read -> path-traversal
- secrets, hardcoded credentials, API keys, tokens in code -> secrets
- CSRF, cross-site request forgery -> csrf
- XXE, XML external entity -> xxe
- open redirect, unvalidated redirect -> open-redirect
- crypto, weak hashing, insecure random, encryption misuse -> crypto
- race condition, TOCTOU, concurrency bug -> race-conditions
- file upload, unrestricted upload, malicious file -> file-upload
- outdated/vulnerable dependency, CVE, SBOM -> handled in recon (stage:recon or full run)
Steps
- Is this about a pull request? If the request contains a GitHub PR link, names a PR number, says "this PR", "these changes",
"before merge", or the user is on a branch with an open PR and asks for a review, route to
sh-pr-reviewand stop. A PR review is scoped to the diff, posts back to the PR, and sets a commit status; a codebase review is not and does neither. Do not substitute one for the other. - Identify the target path (default: current working directory) and whether the user restricted the scope to specific classes or a single stage.
- Confirm authorization context briefly if the request implies attacking systems the user may not own (this harness does static analysis only; decline live attacks on third-party targets).
- Route per the table. When ambiguous between "full review" and "specific classes", default to a full review but state the assumption so the user can narrow it.
- Hand off by invoking the chosen skill with the parsed arguments. Pass through any
classes:,stage:,depth:, andmodels:tokens verbatim so the pipeline honors them (see the model-override options insh-security-review). Do not duplicate its work here.
Notes
- Model / cost control:
sh-security-reviewruns each stage on a cost-appropriate model by default (recon=sonnet, hunt=tiered haiku/sonnet/opus per class, chain/verify=opus, report=haiku). Users can override withmodels:(e.g.models:max,models:cheap,models:verify=opus,hunt=sonnet) - pass these through unchanged. - The pipeline writes everything under
<target>/.security-harness/<run-id>/; point the user there. - If the user asks something outside security review (general coding), say this harness is scoped to security work and let the normal assistant handle it.