Introducing a new way to review your PRs.

Read the announcement

usestrix / strix

Open-source AI penetration testing tool to find and fix your app’s vulnerabilities.

View on GitHub

26 documents · Updated Nov 20, 2026

docs / agent-tools-and-workflows

Agent tools and workflows

Strix gives each scanning agent a shared tool surface for inspection, research, planning, coordination, and report delivery. Root and child agents use the same core capabilities, while their finish actions reflect their different roles in the scan; see Security agent orchestration, Agent factory and prompts, and Skill library and loading for the wider agent model.

Tool surface

The tool surface connects model decisions to the sandbox, the target application, captured proxy traffic, shared scan state, and the final report. The factory adds the common tools to both root and child agents, then adds the role-specific terminal action.

AreaPurposeMain boundary
Shell and browserRun commands, inspect targets, and drive web applicationsSandbox and target traffic
Proxy trafficSearch, inspect, replay, and scope captured HTTP trafficShared Caido project
MCP connectionsDiscover and call tools exposed by configured external servicesPer-run MCP sessions
Web researchObtain current security facts and tool guidanceExternal search service
Reasoning and work statePlan work and retain useful scan knowledgeAgent or scan scope
CoordinationCreate specialists, exchange messages, and inspect progressAgent graph
ReportingFile confirmed findings and review themShared scan report
CompletionYield to a user or end an agent or scanInteractive session or scan lifecycle

Shell and browser actions

The shell is the main action path for security testing. exec_command runs commands in the sandbox, and write_stdin continues a live process when that process was started with a TTY. This supports normal command-line tools as well as long-running or interactive tools.

The agent-browser command gives the agent a browser workflow through the same shell surface. It can open pages, inspect an accessibility snapshot, interact with fresh element references, capture screenshots, and use separate sessions for parallel work. Browser traffic uses the scan proxy, so browser activity can be reviewed with the other captured target traffic.

The normal browser flow is:

  1. Open the target page in an agent-owned browser session.
  2. Take an interactive snapshot and act on its current references.
  3. Wait for the expected page state after a navigation or action.
  4. Re-snapshot before the next reference-based action.
  5. Use view_image only when visual pixels add information that the snapshot does not provide.
flowchart LR
    A[Agent] --> B[exec_command]
    B --> C[Shell tools]
    B --> D[agent-browser]
    D --> E[Target web application]
    C --> F[Sandbox tools and files]
    E --> G[Captured target traffic]

References: strix/…/README.md• L1-32 strix/…/README.md• L1-12 strix/…/agent_browser.md• L14-37 strix/…/README.md• L1-17 strix/…/factory.py• L656-675

Proxy traffic inspection

The proxy tool surface lets an agent use captured HTTP traffic as both evidence and test input. It can find requests with HTTPQL filters, inspect request or response content, replay a request with selected changes, browse the discovered sitemap, and restrict traffic with named allow and deny rules.

The tools share one proxy project for a scan. Host-side calls are serialized so parallel agents do not use the proxy connection at the same time. A typical traffic workflow is:

  1. Use list_requests or list_sitemap to find an endpoint or captured exchange.
  2. Use view_request or view_sitemap_entry to inspect its content and context.
  3. Use repeat_request to test a controlled change while retaining the original request context.
  4. Pass verified request IDs to a report when the exchange supports a confirmed finding.

Scopes provide the boundary for this work. An empty allowlist includes all domains, and a denylist takes precedence over an allowlist. Request lists use cursor pagination, while request content uses line pages or compact regular-expression hits.

flowchart LR
    A[Captured traffic] --> B[list_requests or list_sitemap]
    B --> C[Inspect request or sitemap entry]
    C --> D[view_request or view_sitemap_entry]
    D --> E[repeat_request]
    E --> F[New observed response]
    D --> G[Verified exchange ID]
    G --> H[Finding report evidence]

References: strix/…/tools.py• L47-69 strix/…/tools.py• L71-93 strix/…/tools.py• L135-192 strix/…/tools.py• L269-387 strix/…/tools.py• L390-467 strix/…/tools.py• L470-549 strix/…/tools.py• L552-679

MCP connections

The Model Context Protocol (MCP) tool surface lets an agent reach tools provided by configured external services through a per-run registry. The registry keeps live connections shared by the root and child agents, while the agent discovers only the connection and tool information needed for its current task instead of receiving every remote tool schema at startup.

An MCP connection can use a remote streamable HTTP endpoint or a local process. Its configuration can also limit the tools that the service exposes and identify a smaller set of active tools for search ranking. A connection may be prepared in the background, but the first discovery or call can also open it on demand. A failed connection is isolated from the rest of the run and is shown as unavailable.

The targeted discovery flow is:

  1. list_mcps returns connection IDs, purpose, tool counts, and connection state without tool schemas.
  2. search_mcp_tools returns a small ranked set of tool names and descriptions for one connection.
  3. get_mcp_tool_schema returns the exact input schema for the selected tool.
  4. call_mcp sends an object that matches that schema and returns the remote result.

describe_mcp remains a compatibility path that returns a full catalog when targeted search cannot identify a tool. Unknown connections, unknown tools, unavailable sessions, and invalid argument shapes return an explanatory tool result. Remote results use the common output bounds and disk-spill handling, so large responses remain recoverable without filling the model history.

This diagram shows how a run shares configured MCP connections while keeping remote tool discovery on demand.

flowchart LR
    A[Configured connection] --> B[Per-run MCP registry]
    B --> C[list_mcps]
    C --> D[search_mcp_tools]
    D --> E[get_mcp_tool_schema]
    E --> F[call_mcp]
    F --> G[External MCP service]
    B --> H[Shared root and child access]

References: strix/…/config.py• L15-92 strix/…/client.py• L123-183 strix/…/client.py• L213-241 strix/…/client.py• L283-379 strix/…/registry.py• L43-75 strix/…/registry.py• L79-240 strix/…/registry.py• L319-462 strix/…/agent_tools.py• L1-23 strix/…/agent_tools.py• L62-89 strix/…/agent_tools.py• L92-118 strix/…/agent_tools.py• L146-233 strix/…/agent_tools.py• L236-296

Web research

web_search provides current security research when the agent needs facts that may have changed or are not in its prompt. Its security-focused search path is suited to advisories, exact product versions, exploit research, tool use, and error diagnosis. The selected provider receives the agent's query as the search request, so the search service can match the target question rather than the tool's guidance text.

With the Exa provider, web_search returns ranked pages with security-focused summaries. Its summary step adds security context after Exa finds the pages. web_get_contents is the follow-up path for reading selected public pages in full; it accepts up to ten URLs per call and reports URLs that did not return content. Page fetching needs Exa, so a run pinned to Perplexity must use the synthesized search answer instead.

Search and page-fetch results are returned as information for the current conversation. Missing configuration, timeouts, rejected requests, network errors, and unexpected service responses return a usable error result so the agent can narrow the query or continue without search.

References: strix/…/tool.py• L19-41 strix/…/tool.py• L44-118 strix/…/tool.py• L119-161 strix/…/tool.py• L164-224 strix/…/tool.py• L266-316 strix/…/tool.py• L319-413

File edits and output handling

The agent edits sandbox files through the first-class patch action. This is the model-facing name for the SDK file-patching capability, and it gives code changes a clear boundary separate from command execution.

Tool results are bounded before they enter the model history. If a result is too large, the system keeps a head and tail preview and writes the full result to a sandbox output file when the spill path is available. The agent can then read the saved output through its normal file and shell tools without filling the conversation with raw command output.

References: strix/…/README.md• L1-10 strix/…/output_store.py• L1-7 strix/…/output_store.py• L131-166 strix/…/factory.py• L76-83 strix/…/factory.py• L120-144

Working state

The tools separate short-lived reasoning from durable work state. This lets an agent make a decision, track unfinished work, and share useful facts without treating every item as a user-facing report.

Reasoning and work tracking

think records a private reasoning note with no external side effect. It is useful when the agent must check evidence, scope, policy, or a multi-step attack plan before it acts. It is not a user response and does not add new information.

The todo tools track actionable work for one agent. A todo can be created, listed, updated, marked done or pending, and deleted. Each agent has its own list, while the list is mirrored to scan state so work can survive the active process and be inspected in a later run.

State toolRole
thinkPrivate decision and evidence note
create_todoAdd one or more planned actions
list_todosReview work by status or priority
update_todoChange task text, priority, or status
mark_todo_done / mark_todo_pendingMove tasks across the completion boundary

References: strix/…/tool.py• L10-37 strix/…/tools.py• L20-32 strix/…/tools.py• L41-105 strix/…/tools.py• L266-359 strix/…/tools.py• L362-496

Threat models

A threat model is the shared scan baseline for the target. It states who the attacker is, where trust boundaries sit, what inputs are attacker-controlled, which surfaces matter, and how severity should be judged. Agents read it before hunting so they use one view of the target instead of deriving conflicting assumptions.

The model is scoped to the current scan and keyed to a normalized target identity. One agent can write the base model. Other agents can add attributed amendments when new evidence corrects or extends that baseline. Replacing the base model is reserved for establishing or deliberately folding in the shared document; amendments preserve concurrent corrections until they are folded.

References: strix/…/tools.py• L1-14 strix/…/tools.py• L480-517 strix/…/tools.py• L520-588 strix/…/tools.py• L591-638

Shared notes and skills

Notes preserve observations that matter across agents in the same scan. They support categories such as findings, methodology, questions, plans, and repository maps; other agents can filter notes by category, tags, or text and can see who wrote them.

Skills provide focused reference material for a task. load_skill places selected skill content into the current conversation for immediate use, while a child created with a skill receives that specialization as part of its assigned context. Loading a skill informs the agent; it does not permanently rewrite the agent prompt.

References: strix/…/tools.py• L21-32 strix/…/tools.py• L53-106 strix/…/tools.py• L339-377 strix/…/tools.py• L388-444 strix/…/tool.py• L10-36

Coverage ledger

The coverage ledger records what an agent reviewed and how each surface closed. It supplies the negative space that findings alone cannot show, so the final report can distinguish tested areas from areas that were never examined or still need work.

Agents record coverage as they finish a review. Each entry names a surface and risk area and uses one outcome: reported, no_issue_found, ruled_out, not_applicable, or needs_follow_up. A surface and risk pair has one shared entry; later work updates that entry and keeps its prior conclusion as history. The root agent lists the ledger before completion and reconciles unresolved entries and missing risk coverage into the final report.

References: strix/…/tools.py• L1-14 strix/…/tools.py• L40-49 strix/…/tools.py• L377-535 strix/…/tool.py• L86-135 strix/…/tool.py• L200-207

Scan coordination

Graph tools turn a single scanning agent into a coordinated set of specialists. The coordinator keeps parent-child relationships, status, messages, waiting state, and resume information so work can proceed in parallel without losing ownership.

Agent graph coordination

view_agent_graph shows the current tree and status. The root can use it before creating work to avoid overlap and before completion to confirm that no active child remains. create_agent starts a focused child, validates its requested skills, and can pass the parent history as background or start a clean task with selected skills.

Messages are targeted to an agent inbox. A message can ask a question, give an instruction, or share information, and it wakes the recipient so the next cycle can use the new context. In an interactive run, a message can wake an agent that is completed, stopped, or failed. In a non-interactive run, a finished agent cannot be revived, so the coordinator's filed reports or a new child provide the recovery path.

wait_for_agents is one blocking wait for another agent's message. It is not a process poll, and it is not a user handoff; shell processes use their own polling path and users use wait_for_user. The tool allows one wait per model turn, returns as soon as a message arrives, and reports a timeout, stop, or no-active-agent result when it cannot wait usefully. If no active agent can send a message, the coordinator's filed reports or a new child provide the recovery path instead of an idle wait. stop_agent provides a controlled stop for active work. A cascading stop handles descendants, and the current turn completes before the stop takes effect.

The coordination flow is:

sequenceDiagram
    participant R as Root agent
    participant G as Agent graph
    participant C as Child agent
    R->>G: view_agent_graph
    R->>G: create_agent(task, skills)
    G-->>C: Start focused work
    C->>G: send_message_to_agent or agent_finish
    G-->>R: Inbox message or completion report
    R->>G: wait_for_agents or view_agent_graph

References: strix/…/tools.py• L117-166 strix/…/tools.py• L169-270 strix/…/tools.py• L293-491 strix/…/tools.py• L494-607 strix/…/tools.py• L610-752 strix/…/tools.py• L755-856 strix/…/agents.py• L44-65 strix/…/agents.py• L278-360

User handoff and scan completion

wait_for_user is the interactive handoff. The agent writes the user-visible reply as plain text, then parks until the user replies; it does not carry a second message body. It is only available when a user is attached, and it is a pause rather than a terminal action.

Completion is role-specific. A child calls agent_finish to mark its task complete and send a structured summary to its parent. The completion report includes the vulnerability reports actually filed by that child, while its narrative findings remain summary text. The root calls finish_scan once to persist the customer-facing narrative and end the scan. Successful lifecycle results are treated as final tool output; a failed validation leaves the agent running so it can correct the request.

Before root completion, the agent must inspect the graph, account for filed findings, consider attack chains, reconcile coverage, and ensure no child is still active. The completion response reports coverage counts, unresolved needs_follow_up surfaces, and risk classes assigned to agents that have no coverage entry. Completion fields become report content, so they must contain finished assessment prose rather than progress text or placeholders.

References: strix/…/tool.py• L17-102 strix/…/tools.py• L610-752 strix/…/tool.py• L86-135 strix/…/tool.py• L138-207 strix/…/tool.py• L314-353 strix/…/factory.py• L460-514

Finding reports

Reporting tools turn verified evidence into shared scan findings. create_vulnerability_report records a proof-backed dynamic issue, while create_dependency_report records a known-CVE dependency issue; update_vulnerability_report lets an agent correct or extend an existing finding without changing its identity. delete_vulnerability_report withdraws a finding that later testing disproves. list_reports and get_report let the root review what the scan has filed.

The two creation paths use different evidence contracts:

Finding typeEvidence and rating basisMain record boundary
DynamicA working proof of unauthorized impact and an eight-metric CVSS breakdownEndpoint and PoC details
Dependency CVEA verified advisory, affected package version, repository-relative manifest path, usage evidence, and contextual CVSS reasoningPackage and dependency metadata

Dynamic reports require a concrete proof of unauthorized impact. They can also point to repository-relative code locations, including an applyable fix suggestion only when the report carries verification that the fix closes the issue, checks bypasses, preserves valid behavior, and passes the relevant repository checks. Dependency reports use a verified advisory and affected version instead of a live exploit, then record reachability as a prioritization signal. The dependency path requires evidence for its reachability level and a contextual rating with reasoning that a reviewer can check. Neither path treats reachability alone as proof of exploitability.

HTTP-backed findings can also carry proxy exchange IDs. The reporting tools validate that each ID is a bounded, numeric request ID and confirm that it exists in the current proxy project before storing it. If the proxy does not respond during verification, the IDs are kept with a warning so a slow proxy cannot strip evidence; consumers resolve each ID against recorded traffic before rendering it. A missing or invalid ID is rejected.

An update replaces only the fields supplied by the caller, keeps the report id and original filing identity, and records the reason in update history. A revision must stay within the finding class: dynamic findings cannot receive dependency-only metadata, and dependency findings cannot receive endpoint, method, or PoC fields. A revised CVSS breakdown is recalculated, while dependency re-rating keeps the published advisory score as its reference and stores the codebase-specific reasoning. A report that later proves false is withdrawn with delete_vulnerability_report; a real finding that is weaker than first reported is revised instead. Withdrawal removes it from the scan and records why it was disproved.

list_reports gives the root a compact scan-wide view for coverage and summary work. It supports severity, finding-class, target, and text filters, marks reports filed by the calling agent, and returns severity counts across the scan. get_report retrieves one finding in full. Reporting is separate from child completion: a narrative findings entry in agent_finish does not register a vulnerability.

This diagram shows how verified observations become shared reports and how later evidence updates or withdraws the same finding.

flowchart LR
    A[Observed behavior or advisory] --> B{Verified finding?}
    B -->|Dynamic proof| C[create_vulnerability_report]
    B -->|Known dependency CVE| D[create_dependency_report]
    C --> E[Shared report state]
    D --> E
    E --> F[update_vulnerability_report]
    E --> G[delete_vulnerability_report]
    E --> H[list_reports]
    E --> I[get_report]
    F --> E
    G --> J[Withdrawn finding]
    H --> K[Root final report]
    I --> K

References: strix/…/tool.py• L183-265 strix/…/tool.py• L335-355 strix/…/tool.py• L587-696 strix/…/tool.py• L699-777 strix/…/tool.py• L780-927 strix/…/tool.py• L1327-1379 strix/…/tool.py• L1463-1671 strix/…/tool.py• L1763-1961 strix/…/tool.py• L1964-2162 strix/…/tool.py• L2268-2448 strix/…/tool.py• L2485-2571 strix/…/tools.py• L71-93

Change checklist

When these files or sections are changed, remember to consider these:

  • Keep the root and child tool sets aligned with their distinct terminal actions. A lifecycle tool added in the wrong role can allow a child to end the whole scan or leave the root without a valid finish path.
  • Preserve the distinction between wait_for_user, wait_for_agents, and shell process polling. These actions pause for different actors and use different wake-up paths.
  • Keep graph status, message delivery, one-wait-per-turn behavior, timeout ceilings, and terminal notifications aligned. A change in one path can leave a parent waiting for a message that cannot arrive.
  • Keep child creation skill validation and inherited-history behavior aligned with the child prompt and skill loader. A child must receive the scope and specialization that the parent requested.
  • Keep child completion reports tied to the reports actually filed by that child. Narrative findings, authoritative report IDs, parent notification, and terminal status must not drift apart.
  • Keep graceful stop behavior aligned with active statuses, descendant cancellation, and parent notification. A stop must not act as self-completion or leave an ancestor waiting on a stopped descendant.
  • Keep the MCP discovery contract aligned across list_mcps, targeted search, schema lookup, compatibility catalog lookup, and dispatch. Connection IDs and exact tool names must remain stable across every step.
  • Keep MCP registry state, connection ownership, cleanup, and status reporting aligned. Root and child agents must reuse the same per-run connection without exposing secrets or orphaning a session.
  • Keep MCP transport validation, authentication, tool allowlists, active-tool ranking, and timeout settings aligned with the remote service contract. A valid connection must expose the same reachable catalog that the agent can discover and call.
  • Keep the web provider contract aligned with the query and page-fetch paths. Exa search must receive the agent's query directly and can add security-focused summaries afterward; Exa supports web_get_contents, while a run pinned to Perplexity must use web_search.
  • Keep threat-model target normalization aligned with the scan's target spelling. Separate identities can cause agents testing one target to derive different baselines.
  • Keep coverage outcomes, duplicate handling, and final coverage reconciliation aligned. A new or renamed outcome can silently turn an unexamined surface into an apparently clean result.
  • Keep report creation, revision, withdrawal, and finding-class rules aligned. A field accepted by one report path can be invalid when written onto the other class, and a disproved report must be removed rather than left as a false-positive note.
  • Keep dependency report contracts aligned across advisory scores, repository-relative manifest paths, reachability evidence, and contextual CVSS reasoning. Missing one part can prevent a finding from being rated or tied to the right source file.
  • Keep report code-location and fix-verification rules aligned. A suggested replacement without proof that it closes the issue can turn a report into an unsafe one-click change.
  • Keep proxy request IDs tied to the current proxy project. Report validation, proxy lookup, and the warning path for an unavailable proxy must agree so evidence is not silently stripped or rendered as unverified traffic.
  • Keep report type rules aligned with report deduplication and final-summary guidance. A dependency CVE and a proof-backed dynamic issue have different evidence contracts.
  • Keep tool output limits and spill-file access aligned. A changed preview contract must still leave the agent a way to recover omitted command or tool output.
  • Keep browser session and screenshot guidance aligned with the sandbox defaults. Shared sessions, stale snapshot references, or unreadable image paths can silently invalidate browser work.

Important failure modes

TriggerConsequenceGuard
write_stdin targets a process that was not started with a TTY or has already exitedInteractive input cannot reach the processStart interactive work with tty=true and poll only while the process is live
A browser action uses a reference after the page changesThe reference is stale and the action fails or targets the wrong stateWait for the new state and take a fresh snapshot
Multiple agents use the default browser sessionOne agent changes another agent's page, tabs, cookies, or referencesGive each concurrent agent its own named session
An oversized result is returned without a usable spill pathThe model sees only a bounded preview and may miss evidenceFollow the saved sandbox path when present, or narrow the command and collect focused output
Parallel agents issue host-side proxy calls without serializationA shared proxy transport can reject a call or return incomplete traffic workUse the proxy tool surface, which serializes calls for the shared project
A proxy request is replayed with stale or wrong request contentThe test does not represent the intended target behaviorSelect the request from the current project, inspect it, then apply only the intended modifications
A scope has an overly broad allowlist or a denylist that overrides the intended targetRequests are omitted or noisy out-of-scope traffic is includedReview scope rules and confirm the effective allow and deny patterns before testing
An agent calls an MCP tool with an unknown connection, unknown tool, or wrong argument shapeThe external action is rejected and the intended service work does not occurUse list_mcps, search the connection, read the exact schema, and pass an object that matches it
An MCP connection fails, becomes unavailable, or has a tool outside its allowlistThe requested remote capability is unavailable while the rest of the scan continuesCheck connection state and reachable tool results, then continue without that service or retry after it can reconnect
Web search is pinned to a provider without its key, or page fetch is attempted through PerplexityResearch or page extraction is unavailableCheck the provider contract and use the returned error to retry search or continue without it
Exa receives the tool's search guidance together with the agent querySearch results can match generic security terms instead of the target questionSend only the agent query to Exa and keep security guidance in the result-summary step
A target is named with an identity that does not match the scan scopeAgents miss the shared threat model and derive separate trust assumptionsUse the scan's target spelling or confirm the normalized target identity
Two agents record the same surface and risk area as separate coverage rowsThe report can show a stale conclusion beside its replacementUse update_coverage on the existing entry and retain the change as history
The root calls finish_scan while agents are running or waitingChild work is orphaned and the final report can be incompleteInspect view_agent_graph; finish, wait for, or stop active children first
The root finishes with missing or unresolved coverageThe report can imply clean coverage while assigned risk classes remain unexaminedReconcile list_coverage; record unresolved work as needs_follow_up or dispatch it
A child places a finding only in agent_finish.findingsThe narrative reaches the parent but no report record is createdFile the finding with the matching reporting tool before child completion
A report repeats an existing vulnerabilityDeduplication rejects the report and the duplicate must not be retriedReview shared reports before filing and use one report per distinct issue
A report stores a proxy request ID that was not verified in the current projectThe finding points to missing or unrelated traffic evidenceCopy IDs from captured proxy traffic and let report creation or update verify them
The proxy is unavailable while a report includes exchange IDsThe finding can be filed with the IDs and a warning, but rendering may omit exchanges until the proxy respondsRetry update_vulnerability_report with the IDs when the proxy responds
A dependency report omits its manifest path, reachability evidence, or contextual rating explanationValidation rejects the report or leaves the finding without a codebase-specific ratingRecord the exact repository-relative manifest and evidence-backed usage and CVSS context
A code location uses an absolute or parent-traversing path, or has invalid line boundsThe source location cannot safely identify the affected repository codeUse a repository-relative path and verify positive start and end lines
A code location includes fix_after without a complete verification statementThe report cannot safely present the change as an applyable fixProve security closure, bypass coverage, preserved behavior, and relevant checks before adding fix_after
A revision sends endpoint or PoC fields to a dependency finding, or dependency-only rating fields to a dynamic findingThe update is rejected instead of mixing two finding classesKeep revisions within the class of the existing report and file a separate dynamic issue for a proven exploit
A revision omits its reason or supplies no changed fieldThe existing report is not changedName the new evidence and pass at least one replacement field
A report is withdrawn without evidence that testing disproved itA valid finding disappears from the scan and cannot be restoredUse delete_vulnerability_report only after re-testing; revise real but weaker findings instead
A requested child skill is invalid or the child is spawned with an unclear taskChild creation fails or the child repeats work without useful scopeValidate skill names and give the child a concrete objective with the target and known shared state
A message targets the sender itself or a finished non-interactive agentThe message is rejected and no new work startsUse think for a private note; read filed reports or spawn a new child when a finished agent cannot resume
wait_for_agents is called twice in one turn, with no active sender, or for a shell processThe agent returns early or remains idle until timeout without useful workIssue one wait, then react to its outcome; poll commands with write_stdin and finish or spawn work when no sender remains
stop_agent targets a completed agent or the caller itselfThe stop request is rejected and the intended lifecycle action does not occurStop only a running or waiting target; use agent_finish or finish_scan for self-completion
A child finishes without filing its confirmed report or coverage entryThe parent receives a summary but shared scan state misses the finding or reviewed surfaceFile reports and record coverage before agent_finish
wait_for_user is used in an autonomous runNo user handoff occurs and the tool returns an errorContinue the task or use the role-specific finish tool

Key files

FileWhy to read it
strix/…/factory.pyDefines the common tool set, root/child role split, sandbox capabilities, output bounds, and lifecycle handling.
strix/…/README.mdExplains the shell command boundary and the TTY rule for live processes.
strix/…/agent_browser.mdDescribes the browser session model, snapshot workflow, waits, screenshots, and safety boundary.
strix/…/tools.pyDefines captured-traffic search, request inspection and replay, sitemap navigation, scope rules, proxy evidence lookup, and serialized access to the shared project.
strix/…/agent_tools.pyDefines connection discovery, targeted tool search, schema lookup, compatibility catalog access, and generic dispatch.
strix/…/registry.pyDefines the shared per-run connection registry, lazy connection and catalog state, cleanup, status reporting, and call attribution.
strix/…/client.pyDefines HTTP and local-process connection setup, tool filtering, result handling, bounded connection startup, and session ownership.
strix/…/config.pyDefines the stable connection contract for transport, authentication, tool reachability, active-tool hints, and timeouts.
strix/…/__init__.pyExposes the public MCP connection, registry, discovery, dispatch, and configuration surface.
strix/…/output_store.pyShows how large results become recoverable head-and-tail previews plus sandbox files.
strix/…/tools.pyDefines graph inspection, child creation, messages, one-shot waits, completion reports, and graceful stops.
strix/…/agents.pyExplains the shared coordinator state behind agent status, mailboxes, waiting, and snapshots.
strix/…/tool.pyDefines provider selection, direct query search, security-focused summaries, Exa page retrieval, and recoverable provider errors.
strix/…/tools.pyDefines the shared target baseline and its read, replace, and amendment workflow.
strix/…/tools.pyDefines the shared ledger for reviewed surfaces, closure outcomes, and unresolved work.
strix/…/tool.pyDefines dynamic and dependency finding creation, code-location and fix-verification rules, proxy evidence validation and fallback warnings, revision and withdrawal, evidence and severity validation, reachability context, deduplication, and report review.
strix/…/tool.pyDefines the root completion contract, coverage reconciliation, and final customer-facing report handoff.
strix/…/tool.pyDefines the interactive user pause and its separation from terminal completion.
strix/…/tools.pyDefines shared scan notes, authorship, categories, filters, and persistence.
strix/…/tools.pyDefines per-agent work lists and their status and priority model.