Skill library and loading
Strix skills are focused knowledge packages that guide an agent during an authorized security scan. The repository provides a bundled library for security testing and repository-level workflows for common scan uses. See Security testing skills for the wider knowledge model.
Skill sources
Bundled runtime library
The bundled library is the runtime source for focused security guidance. It groups Markdown packages under strix/skills so an agent can combine a weakness class with the target's framework, protocol, technology, discovery method, or scan strategy.
Each package can declare a name and short description in YAML front matter. The loader uses this metadata to present available choices. It uses the Markdown body as the guidance that enters the agent context. Packages marked as templates can adapt their guidance to scan capabilities before the body is returned. The browser package is one such template. It adjusts image-related guidance to the agent's capabilities while also covering isolated browser sessions and safe interaction with target content.
References: strix/…/README.md• L4-24 strix/…/__init__.py• L16-27 strix/…/__init__.py• L169-185 strix/…/__init__.py• L287-329 strix/…/agent_browser.md• L1-20 strix/…/agent_browser.md• L533-540
Repository-level workflows
The top-level skills/ directory contains repository-facing SKILL.md workflows. These packages give an Agent Skills-compatible tool user-facing paths for running a general Strix pentest, adding CI security gates, or performing an API-focused assessment. The pentest workflow presents local self-hosted and managed cloud paths. The CI workflow covers change-scoped security gates. The API workflow adds schema, credential, authorization, and API-specific coverage guidance.
These workflows are a separate layer from the bundled runtime library. The internal loader searches the bundled root and roots added through skill-directory registration. A workflow under the top-level skills/ directory is not part of that runtime search path only because it exists there. It must be exposed through a registered root that follows the runtime layout before the loader can use it.
References: skills/…/SKILL.md• L10-35 skills/…/SKILL.md• L10-17 skills/…/SKILL.md• L10-15 strix/…/__init__.py• L38-65
Knowledge categories
Testing domains
The selectable library categories describe the context that changes how an agent tests a target. They work together rather than acting as separate scan engines.
| Category | Purpose | Example focus |
|---|---|---|
vulnerabilities | Test a weakness class and confirm its impact. | Injection, object access control, CSRF, credentials, unsafe output, and agent authority. |
tooling | Use a security tool through a focused workflow. | Port scans, HTTP probes, browser automation, crawling, fuzzing, and source triage with Opengrep. |
frameworks | Adapt tests to application framework behavior. | Routes, middleware, validation, and rendering. |
protocols | Interpret a communication model and its attack surface. | GraphQL operations and OAuth redirect and token flows. |
technologies | Test a named service or platform. | Identity systems, cloud services, observability stacks, and LLM applications. |
reconnaissance | Discover and map assets before focused testing. | Asset enumeration and infrastructure lifecycle checks. |
cloud | Assess provider and cluster boundaries. | AWS identity and resource paths and Kubernetes boundaries. |
custom | Add a focused workflow for a special testing need. | Source-aware static triage, dependency CVE review, and other non-standard testing needs. |
These domains provide different views of the same target. A vulnerability skill explains what weakness to prove. A framework or protocol skill explains how that weakness may appear in the target. A technology skill adds platform context. A tooling skill turns a tool into a repeatable security task. A vulnerability workflow can also narrow its method: IDOR follows object identity across users and services, while CSRF checks whether browser authority can trigger a state change across request methods and transports. NoSQL injection distinguishes database operators from scalar values that may become another query language. This split lets an agent load only the knowledge that matches its next task.
References: strix/…/README.md• L28-56 strix/…/opengrep.md• L11-18 strix/…/agent_browser.md• L8-20 strix/…/django.md• L6-8 strix/…/graphql.md• L1-8 strix/…/oauth.md• L1-8 strix/…/supabase.md• L1-8 strix/…/csrf.md• L1-8 strix/…/xss.md• L1-8 strix/…/idor.md• L6-32 strix/…/nosql_injection.md• L10-26 strix/…/query_language_injection.md• L8-18 strix/…/aws.md• L1-8 strix/…/dependency_cve_scanning.md• L6-23
Scan modes
Scan-mode packages set the assessment posture. They tell the agent how to balance speed, coverage, and depth while the other categories provide knowledge about what to test.
| Mode | Assessment aim | Guidance shape |
|---|---|---|
quick | Find high-impact issues quickly. | Time-boxed checks of changed or high-value surfaces with minimal proof. |
standard | Balance effort and coverage. | Systematic attack-surface review with bounded source and dynamic analysis. |
deep | Maximize coverage and attack-path depth. | Broad discovery, detailed workflow review, source-aware triage, and chaining. |
diff | Review a pull request, commit, or branch change set. | Focus on problems introduced, re-introduced, or newly reached by the change. |
For source-aware scans, the modes now make static triage part of the starting strategy. Quick mode narrows Opengrep, AST, secret, and dependency checks to changed or high-risk areas. Standard mode uses those checks to map routes, sinks, and trust boundaries. Deep mode expands the same source-aware stack before detailed review and exploit chaining.
Diff mode reviews a change set rather than the whole repository. It can include an old weakness that the change newly exposes or a control that the change weakens. It does not lower the evidence bar, and unrelated pre-existing issues remain outside the report.
The selected mode is added to every agent's starting prompt. It therefore supplies a shared work strategy even when child agents receive different vulnerability or technology skills.
References: strix/…/quick.md• L1-24 strix/…/standard.md• L1-26 strix/…/deep.md• L1-34 strix/…/diff.md• L7-36 strix/…/diff.md• L62-94 strix/…/prompt.py• L41-45
Internal guidance
scan_modes, coordination, and analysis are bundled support domains rather than ordinary user-selected testing topics. Scan modes control assessment depth. Coordination packages define agent roles and handoffs. Analysis packages define review disciplines such as source-aware discovery and counterevidence-based closure.
The library hides these domains from the normal available-skill list. Prompt composition adds the package required by the active scan mode or agent role instead. The root-agent guidance makes the root agent coordinate delegated work instead of testing the target itself. Source-aware white-box guidance adds static triage and dynamic validation when source is available. It favors Opengrep, structural AST mapping, secret checks, and dependency checks as inputs to focused testing. The dependency CVE workflow treats a lockfile advisory and usage analysis as reportable evidence even when a dynamic proof is not available. Counterevidence requires every candidate to end as confirmed, ruled_out, or open_proof_gap; missing evidence is not treated as proof of safety.
The source-aware packages complement each other. source_aware_whitebox defines the coordination goal and coverage bundle. source_aware_sast turns that goal into a repeatable static workflow. dependency_cve_scanning handles known vulnerable packages and their reachability evidence. semantic_confusion handles cases where different components give the same input different meanings. query_language_injection handles scalar fields that may become executable query expressions, including scalar sinks found during NoSQL review. These packages shape the next testing action, but static signals remain hypotheses until validated, except for the dedicated dependency-report path.
References: strix/…/__init__.py• L27-28 strix/…/__init__.py• L68-91 strix/…/source_aware_whitebox.md• L8-47 strix/…/source_aware_sast.md• L6-8 strix/…/source_aware_sast.md• L18-80 strix/…/dependency_cve_scanning.md• L15-23 strix/…/dependency_cve_scanning.md• L115-148 strix/…/semantic_confusion.md• L7-19 strix/…/query_language_injection.md• L10-36 strix/…/nosql_injection.md• L12-18 strix/…/source_aware_discovery.md• L7-42 strix/…/counterevidence.md• L7-49 strix/…/root_agent.md• L7-19 strix/…/prompt.py• L42-49
Skill discovery
Skill roots
Skill discovery searches a sequence of roots and keeps the first matching package. Registered roots have higher priority than the bundled root. A later registration has higher priority than an earlier one.
This precedence supports two related uses. A registered root can add a package for a scan. It can also replace a bundled package at the same category and name without changing the packaged library.
References: strix/…/__init__.py• L38-65 strix/…/__init__.py• L128-166
Skill names
An agent can request a skill by its bare name or by its category-qualified name. A bare name is convenient when only one package has that name. A qualified name such as vulnerabilities/sql_injection is required when more than one category exposes the same bare name.
Files at a skill root can also be addressed as root skills. Documentation files and names that start with __ are excluded from normal selection. Internal categories are also excluded from the user-selectable list so support guidance stays attached to the scan flow that needs it.
References: strix/…/__init__.py• L68-91 strix/…/__init__.py• L94-108 strix/…/__init__.py• L225-249
Skill metadata
Metadata supports discovery while the body supplies the actual knowledge. The available-skill view groups names by category and normalizes descriptions for display.
The loader removes front matter before it returns a body. A malformed front matter block does not discard the body, but it can leave the description empty. A missing or unreadable package is skipped and recorded in logs rather than treated as loaded guidance. A package can opt into template rendering when its guidance depends on scan capabilities. For example, the browser package can include image-specific guidance only when image support is available.
References: strix/…/__init__.py• L169-222 strix/…/__init__.py• L312-326 strix/…/agent_browser.md• L1-5 strix/…/agent_browser.md• L262-309
Load tracking
The loader records the distinct skill names loaded during the current process. Bundled skills keep their names. Any skill loaded from a registered root is recorded as custom, which gives telemetry a stable signal without exposing custom path details.
This tracking is process-local and does not change skill resolution or the guidance returned to the agent. The loader updates the set when a body loads and returns the sorted names when callers request the record.
References: strix/…/__init__.py• L252-264 strix/…/__init__.py• L312-328
Agent loading
Prompt preload
Prompt preload gives an agent its specialist context before the first scan action. The prompt builder combines caller-selected skills with guidance required by the scan mode and the agent's role, then places the loaded bodies in the system prompt's specialized-knowledge section.
The resolver keeps this order:
- Caller-selected skills.
- The selected
scan_modes/<mode>package. - Browser and Python tooling guidance.
- Root-agent coordination guidance when the agent is the root.
- Source-aware guidance for a white-box scan.
Duplicate requests are removed while the first occurrence keeps its position. This lets a caller add focused knowledge without removing the baseline guidance needed by the agent's mode or role. Browser tooling guidance also gives the agent a common model for session ownership, page changes, and safe handling of content returned by a target.
The following flow shows how scan settings and requested knowledge become the agent's initial context.
flowchart TD
A[Scan settings and requested skills] --> B[Resolve ordered skill names]
B --> C[Search registered roots and bundled root]
C --> D[Read metadata and Markdown bodies]
D --> E[Render system prompt]
E --> F[Start scanning agent]
References: strix/…/prompt.py• L20-57 strix/…/prompt.py• L60-109 strix/…/system_prompt.jinja• L487-506 strix/…/agent_browser.md• L35-80 strix/…/agent_browser.md• L533-540
On-demand reference
On-demand loading adds guidance to the current conversation when the agent discovers a need that was not known at scan start. The load_skill tool returns the skill body as an inline result. It does not change the agent's permanent system prompt.
This path is useful for exact technology, browser, payload, or tool guidance before the next action. For example, an agent can load tooling/agent_browser for browser session and interaction guidance, load tooling/opengrep for exact offline rule and output guidance, or load semantic_confusion when a source map shows a representation mismatch. A scalar query field found during a NoSQL review can load query_language_injection before it is tested. If the focus must last for a specialist task, the parent passes the skill list to create_agent instead. The child then starts with that assignment.
References: strix/…/tool.py• L10-35 strix/…/system_prompt.jinja• L155-162 strix/…/agent_browser.md• L26-80 strix/…/opengrep.md• L11-18 strix/…/semantic_confusion.md• L21-44 strix/…/nosql_injection.md• L14-18 strix/…/query_language_injection.md• L10-36
Specialist assignment
Skill assignment follows the agent tree. A parent selects a small group of related packages for a child. The runner carries that list into the child's agent construction so the child can work in parallel with a focused context.
Each request allows at most five skills. One to three related skills are preferred. This limit protects prompt space and keeps the link between a child task and its knowledge clear.
References: strix/…/tools.py• L407-453 strix/…/tools.py• L476-493 strix/…/execution.py• L303-332
Loading sequence
During a scan, skill use follows a simple sequence. Preloaded packages set the starting frame. The agent then uses the available-skill list to identify guidance it may need later.
- Scan settings determine the mode and agent role.
- The prompt builder resolves and loads the starting packages.
- The rendered system prompt lists loaded knowledge and available on-demand choices.
- The agent uses
load_skillfor a missing focused reference or assigns skills to a child specialist. - The child builder repeats prompt composition with the child's selected skills.
This separation keeps stable guidance in the system prompt and temporary guidance in the conversation. It also keeps specialist knowledge close to the agent that performs the focused work.
References: strix/…/prompt.py• L81-97 strix/…/system_prompt.jinja• L487-506 strix/…/tool.py• L11-35 strix/…/execution.py• L303-356
Selection boundaries
Request validation
Skill requests are checked before loading or child creation. The request must name an available package, stay within the five-skill limit, and use a category-qualified name when a bare name is ambiguous.
The validation result is written for the model. This makes a bad name or an unclear choice visible before a specialist starts work, rather than silently giving it the wrong guidance.
References: strix/…/__init__.py• L225-249 strix/…/tool.py• L27-34 strix/…/tools.py• L476-483
Focus boundaries
Skills provide knowledge and direction. They do not create a new scanner or grant an agent a new permission. The agent's tools, scope, runtime, and role still define what work it can perform.
Focused selection is therefore part of scan design. A vulnerability skill can be paired with a framework or protocol skill when their concerns overlap. For example, CSRF work can add OAuth or GraphQL guidance when state changes cross those protocol boundaries. A source-aware workflow can pair tooling/opengrep with frameworks/django or vulnerabilities/semantic_confusion when the target and the question require it. A NoSQL workflow can add vulnerabilities/query_language_injection when a scalar reaches expression construction. Broad unrelated assignments are avoided so each agent can spend its context on one clear task.
References: strix/…/README.md• L60-72 strix/…/graphql.md• L62-101 strix/…/oauth.md• L45-118 strix/…/csrf.md• L59-101 strix/…/opengrep.md• L50-66 strix/…/django.md• L206-215 strix/…/semantic_confusion.md• L165-187 strix/…/nosql_injection.md• L14-18 strix/…/query_language_injection.md• L38-50 strix/…/factory.py• L592-676 strix/…/root_agent.md• L45-68
Change checklist
When these files or sections are changed, remember to consider these:
- Keep category names aligned across the directory layout, available-skill display, and category-qualified requests.
- Keep the registered-root precedence rule aligned with any mechanism that exposes repository-level skill packages to the runtime loader.
- Keep scan-mode names aligned across scan configuration, prompt resolution, and all mode packages, including
diff. - Preserve the distinction between front matter used for discovery and Markdown bodies used for agent guidance. Check template capability flags when a body uses conditional guidance.
- Keep internal coordination and analysis packages out of ordinary selection while ensuring prompt composition adds the package required by the scan.
- Keep the preload and on-demand paths distinct so temporary reference material does not become a permanent prompt assignment.
- Keep the five-skill validation limit aligned with child-agent specialization guidance.
- Keep loaded-skill telemetry aligned with the process-local distinction between bundled skill names and registered-root content.
- Keep tooling guidance aligned with the tools available in the sandbox, including browser session isolation, proxy capture, and trust-boundary rules.
- Keep source-aware mode guidance aligned with the available static tools, offline rule locations, bounded AST targets, and the requirement for dynamic validation.
- Keep framework guidance aligned with the tool names and rule paths that its white-box workflow recommends.
- Keep protocol guidance aligned with the request boundaries used by GraphQL and OAuth skills, and with vulnerability handoffs such as CSRF to those protocols.
- Keep semantic-confusion guidance aligned with the representation, lifecycle, and namespace cases that can trigger a focused follow-up skill.
- Keep query-language guidance aligned with the scalar-sink matrix and the NoSQL handoff for fields that may become expressions.
- Keep dependency-CVE guidance aligned with the dedicated report contract, per-advisory reachability verdict, and lockfile attribution.
- Keep diff-mode scope and closure rules aligned with the source-aware and counterevidence guidance it relies on.
- Keep repository-level pentest, CI, and API workflow guidance aligned with the runtime skill names and supported scan paths they direct users to.
Important failure modes
| Trigger | Consequence | Guard |
|---|---|---|
| A requested bare name exists in more than one category. | Loading is rejected or skipped, so the agent lacks the intended guidance. | Use a category-qualified name. |
| A registered root contains a package with the same relative path as a bundled package. | The registered content silently replaces the bundled guidance. | Review registered roots and treat precedence as part of scan configuration. |
A skill is placed only in the top-level repository skills/ directory. | The internal runtime loader does not see it as a bundled skill. | Expose it through the runtime's registered-root layout when runtime loading is required. |
| A source-aware scan uses a network rule shorthand or omits an explicit Opengrep rule path. | The scan may fail or use rules outside the intended offline baseline. | Use tooling/opengrep guidance and an explicit /opt/opengrep-rules or local rule path. |
| A source-aware scan treats Opengrep, AST, secret, or dependency output as a finding. | The report contains unvalidated static hypotheses. | Use scanner output to rank work, then prove impact with dynamic validation. |
| A known dependency CVE is dropped because the vulnerable path has no dynamic proof. | A pinned vulnerable package is absent from the report even though the advisory and lockfile prove exposure of the package. | Use dependency_cve_scanning, give each CVE a reachability verdict, and use the dedicated dependency report path. |
| A scalar NoSQL search or filter field is tested only with object operators. | Query-language injection in a scalar sink can be missed or closed without proof. | Load query_language_injection and complete its balanced true/false matrix for each scalar sink. |
| A diff review treats unrelated old bugs as findings or treats the diff as a whole-repository scan. | The report loses its change link and scope. | Report only introduced, re-introduced, or newly reachable problems. |
| A root agent tests the target directly instead of delegating. | Work breaks the coordination boundary and may duplicate specialist coverage. | Load the root-agent guidance and delegate all target-touching work. |
| A candidate lacks a named control but is marked safe. | A real issue can disappear as a false clean result. | Keep it as open_proof_gap until evidence supports ruled_out. |
| A source review merges separate reachable call sites into one candidate. | Fix scope and coverage become unclear. | Preserve each distinct source, control, sink, and impact path. |
| A skill exceeds five requested packages or mixes unrelated topics. | The request is rejected or the agent's context loses focus. | Prefer one to three related packages and never exceed five. |
| A package has invalid metadata or cannot be read. | Its description may be empty or its guidance is skipped. | Keep a valid Markdown body and readable front matter for discoverable packages. |
| A required technology or tooling skill is not preloaded and the agent does not request it on demand. | The agent may use weak or incorrect focused guidance. | Use the available-skill list and call load_skill before the focused action. |
| A browser page changes after a snapshot is taken. | The agent uses stale references and may fail to interact with the intended control. | Re-snapshot after navigation, form submission, or dynamic page changes. |
| Browser work uses a shared session or treats page content as instructions. | Another agent can change the working state, or untrusted target content can redirect the workflow. | Use a named session and treat browser content as untrusted data. |
| A semantic-confusion candidate has only different errors or log values. | The agent reports a mismatch without proving a changed security outcome. | Show the representation seen by the control and the different representation seen by the final consumer. |
Key files
| File | Why to read it |
|---|---|
| strix/…/__init__.py | Defines skill roots, names, precedence, metadata, validation, body loading, templates, and loaded-skill tracking. |
| strix/…/README.md | Summarizes the purpose, domains, and expected content of bundled skills. |
| strix/…/agent_browser.md | Defines browser automation guidance, session boundaries, page interaction, recovery, and safe handling of target content. |
| strix/…/opengrep.md | Defines the offline source-triage tool, rule scopes, safe output forms, and scan recovery guidance. |
| strix/…/source_aware_whitebox.md | Defines the source-aware coordination goal, baseline coverage, and validation boundary. |
| strix/…/source_aware_sast.md | Shows how static and structural signals guide focused dynamic testing. |
| strix/…/dependency_cve_scanning.md | Defines lockfile CVE review, dependency reachability evidence, and the dedicated report contract. |
| strix/…/django.md | Illustrates how framework-specific guidance maps application behavior to security checks and source tools. |
| strix/…/graphql.md | Illustrates protocol guidance for schema mapping, resolver authorization, transport parity, and federation boundaries. |
| strix/…/oauth.md | Illustrates protocol guidance for redirect binding, PKCE, state, token exchange, and client boundaries. |
| strix/…/csrf.md | Illustrates weakness guidance for browser authority, state-changing requests, origin checks, and protocol handoffs. |
| strix/…/idor.md | Illustrates object-level authorization guidance across identifiers, tenants, services, and transport types. |
| strix/…/nosql_injection.md | Illustrates database operator injection and the handoff from scalar sinks to query-language guidance. |
| strix/…/query_language_injection.md | Defines the first-pass matrix for scalar inputs that may become query or expression syntax. |
| strix/…/semantic_confusion.md | Defines cross-component representation mismatches and the evidence needed to prove them. |
| strix/…/quick.md | Defines the time-boxed scan posture and its source-aware priorities. |
| strix/…/standard.md | Defines the balanced scan posture and its systematic coverage. |
| strix/…/deep.md | Defines the broad scan posture, source-aware reconnaissance, and vulnerability chaining. |
| strix/…/prompt.py | Defines the order in which requested, mode, role, and source-aware skills enter an agent prompt. |
strix/agents/prompts/system_prompt.jinja | Shows how loaded knowledge and available on-demand skills are presented to the model. |
| strix/…/tool.py | Defines temporary in-conversation skill loading. |
| strix/…/tools.py | Validates skill assignments made while creating specialist agents. |
| strix/…/execution.py | Carries child skill assignments into child-agent construction and execution. |
| strix/…/root_agent.md | Defines delegation, threat-model setup, agent roles, and handoffs for the root agent. |
| strix/…/source_aware_discovery.md | Defines how source-aware review keeps distinct candidates and family sweeps visible. |
| strix/…/counterevidence.md | Defines closure states and the evidence needed to rule out a candidate. |
| strix/…/diff.md | Defines the change-set scope and evidence rules for diff review. |
| skills/…/SKILL.md | Shows the repository-level workflow for choosing local self-hosted or managed scans. |
| skills/…/SKILL.md | Shows the repository-level workflow for adding security gates to CI. |
| skills/…/SKILL.md | Shows the repository-level workflow for API targets, schemas, credentials, and API checks. |