Strix (STRIX) — tools

docs / modules / tools

← all docs

Strix (STRIX) — tools

flowchart LR IN["IN: tool call from an agent\n(name + JSON args)"] --> ACT["ACT: perform the action\n(shell cmd, HTTP via proxy,\nreport write, mailbox send...)"] ACT --> VAL["VAL: per-tool —\ne.g. CVSS field validation,\ndedup check for reports"] VAL --> OUT["OUT: bounded JSON/text result\nback to the agent"]

What this is

strix/tools is the entire toolbox an agent can reach for — everything from "run a shell command" to "spawn another agent" to "file a validated vulnerability report." If strix/agents decides which tools an agent gets, this package is where each tool's actual behavior lives.

What flows through it

flowchart TD Proxy["proxy/\nCaido traffic (list_requests,\nrepeat_request, ...)"] --> Caido[("Caido proxy\n(sandbox)")] Graph["agents_graph/\ncreate_agent, wait_for_agents,\nagent_finish, send_message_to_agent"] --> Coord[("AgentCoordinator")] Reporting["reporting/\ncreate_vulnerability_report"] --> Dedupe["report.dedupe\nLLM-judged check"] Dedupe --> State[("ReportState")] Finish["finish/\nfinish_scan"] --> State Other["notes/, todo/, web_search/,\nthinking/, load_skill/,\noutput_store.py"]

Technical details

flowchart LR IN["IN: tool call from an agent\n(name + JSON args)"] --> ACT["ACT: perform the action\n(shell cmd, HTTP via proxy,\nreport write, mailbox send...)"] ACT --> VAL["VAL: per-tool —\ne.g. CVSS field validation,\ndedup check for reports"] VAL --> OUT["OUT: bounded JSON/text result\nback to the agent"]

Package: strix/tools/ Depends on: strix.core (coordinator/context access for agent-graph tools), strix.report (reporting tools write into ReportState). Depended on by: strix.agents (factory.py imports every tool module to build _BASE_TOOLS).

Files at a glance

Subpackage/file What it does
proxy/ Caido traffic inspection: list_requests, view_request, repeat_request (replay/modify — "Repeater"), list_sitemap/view_sitemap_entry, scope_rules.
agents_graph/tools.py The Graph-of-Agents primitives: view_agent_graph, send_message_to_agent, wait_for_agents, create_agent, agent_finish, stop_agent.
reporting/tool.py create_vulnerability_report, create_dependency_report, list_reports, get_report — heavy validation (CVSS metric enums, CVE/CWE checks, code-location schema) + dedup call before accepting a finding.
finish/tool.py finish_scan — root-only lifecycle tool ending a scan.
notes/tools.py create_note/list_notes/get_note/update_note/delete_note — a cross-agent scratchpad persisted to disk, independent of the mailbox.
todo/tools.py Per-agent todo CRUD, persisted to disk, for self-tracking coverage.
load_skill/tool.py load_skill(skills: list[str]) — fetches skill markdown on demand mid-run (vs. always-preloaded skills baked into the prompt).
web_search/tool.py web_search(query) — Perplexity-backed search with a security-research system prompt; requires PERPLEXITY_API_KEY.
thinking/tool.py think(thought) — a no-op scratch tool for private chain-of-thought, keeping planning text out of user-visible output.
output_store.py bound_and_store/bound_text — caps tool output size, spilling the full text to disk via a configurable SpillWriter.
apply_patch/, shell/, view_image/ No Python tool files — these are SDK-native (ApplyPatchTool, exec_command/write_stdin, ViewImageTool) configured in agents/factory.py, backed by the live sandbox session.
agent_browser/ No Python tool file — a sandboxed CLI (agent-browser npm package) invoked via exec_command, not a first-class SDK tool.

Files & functions, explained

agents_graph/tools.py — the concrete implementation of "Graph of Agents": - create_agent(name, task, inherit_context=True, skills=None) — validates requested skills, optionally seeds parent_history, calls the injected context["spawn_child_agent"] closure (ultimately core.execution.spawn_child_agent). - wait_for_agents(reason, timeout_seconds) — the blocking synchronization primitive; guards against duplicate waits within one turn, drains any already-pending mailbox immediately, otherwise parks (wait_kind="agents") until a message arrives or it times out. - agent_finish(result_summary, findings, success, report_to_parent, final_recommendations) — subagent-only termination; renders and sends a completion report to the parent.

reporting/tool.py — the validation gate before any finding is persisted: required narrative fields, CVSS v3.1 metric-by-metric enum validation feeding _calculate_cvss(), CVE/CWE extraction/validation, code-location schema checks, then report.dedupe.check_duplicate() before report_state.add_vulnerability_report().

Related: STRIX — Architecture (HLD), core.md, report.md, ../design/data-flow.md.