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 / configuration-and-integrations

Configuration and external integrations

Strix turns user settings and external service access into one execution environment for a security scan. Configuration selects the model route, controls request behavior, chooses the sandbox runtime, and connects the agent to optional services such as ChatGPT subscriptions, MCP servers, web search, and Caido.

Configuration model

Strix uses one resolved settings model so scan startup and later services read the same values. The model groups language-model, deduplication, context, runtime, telemetry, integration, and viewer settings instead of making each part of the system read configuration on its own.

Settings sources

The settings loader reads environment variables first. It then reads the persisted file at ~/…/cli-config.json, unless a caller selects another file, and uses field defaults for values that remain unset. The resolved result is cached for the process, while an explicit file change clears that cache (strix/…/loader.py• L36-60).

The persisted file stores an env object rather than a second settings format. This keeps aliases such as LLM_API_KEY and OPENAI_API_KEY consistent across environment and file sources. Environment values win even when the file uses another alias, and persisted settings are written with restricted file permissions. When the shell changes one part of the model connection, such as the model, key, or base URL, the loader drops the stored linked connection values instead of combining old and new providers (strix/…/loader.py• L30-34, strix/…/loader.py• L63-90, strix/…/loader.py• L106-154).

For the field groups and provider-specific values, see Configuration and provider settings.

References: strix/…/settings.py• L22-81 strix/…/settings.py• L201-238 strix/…/loader.py• L36-60 strix/…/loader.py• L63-154

Execution environment settings

Runtime settings define the sandbox image, the selected backend, and the number of image outputs that can remain in the agent's live context. The default sandbox image is ghcr.io/usestrix/strix-sandbox:1.4.1, and the default backend is Docker. The backend registry rejects unknown names instead of silently selecting a different runtime (strix/…/settings.py• L163-173, strix/…/backends.py• L53-75).

The runtime boundary lets the rest of the scan request a sandbox without depending on one deployment method. A registered backend can support local bind mounts or receive source files through another transfer path, so the selected backend must match where the scan's files can be reached (strix/…/backends.py• L78-101).

Context settings also protect the model request from oversized history and tool output. They control automatic compaction, retained history, summary size, per-tool output limits, and live image retention. These limits keep the execution environment usable as a scan produces more data (strix/…/settings.py• L143-160).

References: strix/…/settings.py• L143-173 strix/…/backends.py• L17-55 strix/…/backends.py• L60-101

Model and provider routing

The model name is both a user choice and a routing hint. Strix uses its prefix and the configured model catalog to choose the SDK's native OpenAI path, a LiteLLM-compatible provider path, or the ChatGPT subscription path.

Model routes

The primary model setting is STRIX_LLM. OpenAI-style routes can use the SDK's native behavior. Other provider prefixes are routed through LiteLLM while preserving the model name that users supply. An ollama/ route is adapted to the LiteLLM Ollama form (strix/…/models.py• L563-611).

The API route is selected explicitly when STRIX_API_TYPE or STRIX_FORCE_API is set. Otherwise, Strix checks the model catalog and uses the Responses route only when the model lists /v1/responses; other models use chat completions. A custom endpoint does not by itself force the main model onto one route, while a separately bound custom endpoint uses chat-completions behavior by default (strix/…/settings.py• L32-36, strix/…/models.py• L573-587, strix/…/models.py• L728-749).

A model with a chatgpt/ prefix instead uses the subscription backend and does not use the API-key route. The subscription route always streams, so the general non-streaming setting does not change that path (strix/…/models.py• L613-635).

Strix can bind a secondary model route to its own API key, base URL, headers, request body, and cost rates. This keeps a provider used for a supporting task separate from the process-wide defaults used by the main model (strix/…/settings.py• L103-135, strix/…/models.py• L541-595).

Provider compatibility

Strix adjusts the model request to fit the selected provider. It detects whether a route supports reasoning and strict tool schemas, and it avoids sending features that a provider cannot accept. Parallel tool calls are opt-in through STRIX_PARALLEL_TOOL_CALLS; when enabled, the setting reaches model requests that have tools, while the default remains single-call behavior (strix/…/settings.py• L63-80, strix/…/inputs.py• L245-265, strix/…/runner.py• L419-428). It also repairs blank, placeholder, or repeated tool-call IDs so the conversation remains valid when a provider uses weak or per-turn identifiers. This allows the same agent tool surface to work across providers with different request limits (strix/…/models.py• L976-1007, strix/…/tool_call_ids.py• L1-15, strix/…/tool_call_ids.py• L48-89).

The model layer also applies shared request safeguards. It can use a non-streaming request for gateways that do not support reliable event streams, limit tool calls in one response, repair invalid tool-call identifiers in streamed events, and stop a stream that waits too long for its first event, between events, or for the complete response. These safeguards apply around the provider request, while retry rules handle selected network, rate-limit, and server failures (strix/…/settings.py• L80-108, strix/…/models.py• L171-182, strix/…/models.py• L250-387, strix/…/models.py• L420-499, strix/…/models.py• L687-701).

Provider setup also carries required headers and cost signals. Extra headers apply to both supported routing paths. OpenRouter and Vercel AI Gateway receive Strix app attribution headers, while OpenRouter also receives its category header. Streamed cost data is retained for usage reporting. Configured input, output, and cached-input rates provide cost values when a provider does not report them. Sticky OpenRouter sessions are opt-in, so a scan does not pin its calls to one upstream provider unless STRIX_OPENROUTER_STICKY_SESSIONS is enabled (strix/…/settings.py• L53-92, strix/…/models.py• L704-725, strix/…/models.py• L777-876, strix/…/models.py• L887-920).

Prompt caching is enabled by default for routes that support it. The cache setting controls whether Strix preserves cache points in the system prompt, while the cache block size setting helps align the prompt with provider token blocks (strix/…/settings.py• L81-92, strix/…/models.py• L393-411, strix/…/models.py• L1066-1096).

References: strix/…/settings.py• L29-135 strix/…/models.py• L541-635 strix/…/models.py• L704-920 strix/…/models.py• L976-1096 strix/…/tool_call_ids.py• L1-156 strix/…/inputs.py• L245-265 strix/…/runner.py• L419-428

Authentication paths

Strix supports credentials that belong to the selected external service. API-key settings serve metered model APIs, while ChatGPT subscription credentials use a separate sign-in store and are refreshed during long scans.

API key authentication

API-key authentication reads LLM_API_KEY or an accepted OpenAI-compatible alias. The key can be paired with a custom API base, extra request headers, and extra request data. Strix applies those values to the provider route selected by STRIX_LLM, including provider environment keys when the provider requires a named key (strix/…/settings.py• L32-62, strix/…/models.py• L704-725, strix/…/models.py• L758-775, strix/…/models.py• L914-953).

Other external tools have their own integration keys and provider choices. Perplexity, Exa, and Postman keys are kept in the integration settings group. Web search can select an automatic or named provider, and Exa settings control search depth and result count (strix/…/settings.py• L186-217).

References: strix/…/settings.py• L29-62 strix/…/settings.py• L182-217 strix/…/models.py• L704-775 strix/…/models.py• L914-953

ChatGPT subscription authentication

The strix auth command signs in to ChatGPT through an OAuth flow with a local callback or a manual redirect paste. The resulting access token, refresh token, and account identity are stored separately in ~/…/subscription-auth.json, so subscription secrets do not enter the general configuration file (strix/…/auth_cli.py• L44-179, strix/…/codex.py• L57-105).

A scan uses the subscription only when STRIX_LLM starts with chatgpt/. The client checks the stored token before use and refreshes it near expiry under a process and file lock. A refresh response can keep the prior refresh token when the service does not rotate it. If another Strix process refreshes the single-use token first, the client can use that process's newly stored record (strix/…/codex.py• L241-289, strix/…/codex.py• L318-348). Each request stamps a current bearer token and account identifier, which lets a long scan continue after token expiry (strix/…/codex.py• L351-381).

The subscription path also treats a provider content guardrail as a terminal model choice error. Retrying the same request does not clear that block, so the scan must use a model that the service permits (strix/…/codex.py• L130-160, strix/…/models.py• L76-83).

Viewer relay authentication

The local viewer keeps email verification state in ~/…/viewer-auth.json and uses STRIX_APP_URL as the base for relay requests. The browser does not call the relay directly. The local server sends verification requests and encrypted reports instead (strix/…/settings.py• L220-227, strix/…/auth.py• L1-9, strix/…/auth.py• L129-154).

The report password stays local. Only the encrypted report and delivery metadata reach the relay, and an expired or invalid relay token requires verification again (strix/…/auth.py• L101-120, strix/…/auth.py• L220-248).

References: strix/…/settings.py• L220-227 strix/…/auth.py• L1-9 strix/…/auth.py• L101-120 strix/…/auth.py• L220-248

MCP integration

Model Context Protocol (MCP) gives the agent an optional way to reach external tool servers during a scan. Strix loads named connection definitions before the run and makes their tools available through a per-run registry. The agent discovers tools on demand instead of receiving every remote tool schema in its base tool set.

MCP connection configuration

Users define MCP connections in ~/…/mcp-servers.json, or in a path selected by STRIX_MCP_CONFIG. Each connection has one stable name and uses either streamable HTTP or a local standard-input/output process. HTTP connections require a URL. Local connections require a launch command (strix/…/config.py• L27-56).

Authentication is optional and currently uses a bearer token for HTTP access. A connection can pass process arguments and environment values to a local server. An allowed_tools list limits the tools exposed from that server. Without it, all listed tools are available. Connection settings also define request, stream, session, and concurrency limits so a slow or busy service does not control the whole run. Notes provide one purpose line for the connection in the agent's MCP inventory (strix/…/config.py• L15-24, strix/…/config.py• L49-81).

References: strix/…/config.py• L15-24 strix/…/config.py• L27-89 strix/…/client.py• L122-182

MCP run selection

The MCP file is optional. A missing or unreadable file gives the run no MCP connections, while an invalid entry is skipped so valid entries can still load. Duplicate names keep the first connection because the name namespaces that server's tools (strix/…/loader.py• L1-12, strix/…/loader.py• L47-66, strix/…/loader.py• L101-133).

Per-run selection can keep only named connections or remove named connections. The include selection is applied before the exclude selection. Names that are not configured are reported and ignored (strix/…/loader.py• L69-98). For the complete connection and session flow, see MCP integration.

References: strix/…/loader.py• L69-98 strix/…/loader.py• L101-133 strix/…/client.py• L282-338

MCP tool access

The agent carries three generic MCP tools rather than one model tool for every remote tool. This keeps the base request small while preserving access to the full set of tools allowed for the run (strix/…/agent_tools.py• L1-19).

The access sequence moves from connection discovery to schema discovery and then to one tool call. The connection name is the stable address for both discovery and dispatch. The tool allowlist is applied before the agent can describe or call a tool (strix/…/agent_tools.py• L59-69, strix/…/agent_tools.py• L89-116, strix/…/client.py• L122-180).

This diagram shows how the agent reaches an MCP tool without adding that tool to the base agent surface.

sequenceDiagram
    participant A as Agent
    participant R as Per-run MCP registry
    participant S as MCP connection
    A->>R: list_mcps()
    R-->>A: Connection names and tool counts
    A->>R: describe_mcp(connection)
    R->>S: List allowed tools
    S-->>A: Tool names and input schemas
    A->>R: call_mcp(connection, tool, arguments)
    R->>S: Dispatch one allowed tool
    S-->>A: Tool result or failed-tool result

References: strix/…/agent_tools.py• L39-181 strix/…/client.py• L212-240

MCP connection sessions

Each connected server has a session that owns its connection for the full run. A connection that fails during startup is skipped, so one unavailable service does not stop the scan. Later transport failures stay contained to that connection and do not cancel the agent or other MCP connections (strix/…/client.py• L282-338, strix/…/session.py• L141-158, strix/…/session.py• L426-489).

The session treats a failure according to what failed. It rebuilds and retries transient transport, timeout, rate-limit, and server failures. A permission or protocol failure that belongs to one tool call returns a failed tool result while the connection remains usable. Repeated transient failures quarantine the connection for a cooldown and can later retire it. Authentication failures retire it at once (strix/…/failures.py• L16-28, strix/…/session.py• L491-647).

MCP credentials stay in the live in-memory connection configuration. Failure reports keep only a safe class, status, reason, and request location. They do not retain request data or bearer tokens (strix/…/session.py• L219-283, strix/…/failures.py• L33-47, strix/…/failures.py• L119-150).

References: strix/…/client.py• L282-338 strix/…/session.py• L141-158 strix/…/session.py• L287-375 strix/…/session.py• L426-489 strix/…/session.py• L491-715 strix/…/failures.py• L16-28 strix/…/failures.py• L119-150

Proxy and Caido integration

Caido supplies the scan with a proxy view of captured traffic and a way to replay requests. The integration crosses the sandbox boundary: Caido runs with the sandbox, while the host-side client and agent tools coordinate access to its GraphQL API.

Caido resource boundary

During sandbox startup, Strix waits for the in-container Caido sidecar to accept a guest login. It then connects through the exposed host URL, creates a temporary project, and selects that project for the scan (strix/…/caido_bootstrap.py• L1-8, strix/…/caido_bootstrap.py• L32-78, strix/…/caido_bootstrap.py• L81-115).

Caido setup can finish after the sandbox is ready. A bootstrap handle lets scan startup continue while proxy consumers wait for the client at first use. If setup fails, proxy tools return an unavailable result instead of making the whole scan fail at startup (strix/…/caido_handle.py• L1-7, strix/…/caido_handle.py• L25-60, strix/…/tools.py• L51-67).

References: strix/…/caido_bootstrap.py• L1-8 strix/…/caido_bootstrap.py• L32-115 strix/…/caido_handle.py• L25-60 strix/…/tools.py• L51-67

Proxy analysis and replay

The agent-facing proxy tools turn Caido data into scan actions. The agent can list and inspect captured requests, browse the discovered sitemap, manage allow and deny scopes, and repeat a request with selected URL, parameter, header, cookie, or body changes (strix/…/tools.py• L108-165, strix/…/tools.py• L238-317, strix/…/tools.py• L359-389, strix/…/tools.py• L439-564).

The proxy path keeps request inspection bounded and structured. Lists use filters and cursors, request bodies can be viewed by page or searched with a pattern, and replay reconstructs HTTP framing from the modified body before sending it. Shared calls are serialized because the Caido transport does not support concurrent use (strix/…/caido_api.py• L92-127, strix/…/caido_api.py• L130-167, strix/…/caido_api.py• L333-380).

For the full proxy tool contract and Caido lifecycle, see Proxy and Caido integration.

References: strix/…/tools.py• L108-165 strix/…/tools.py• L238-317 strix/…/tools.py• L359-389 strix/…/tools.py• L439-564 strix/…/caido_api.py• L92-127 strix/…/caido_api.py• L333-380

Configuration flow

The configuration flow connects user choices and credentials to the runtime and its external services. The same resolved settings guide model access, sandbox setup, and optional integrations.

Execution configuration flow

This diagram shows how settings become an active scan environment and how optional external services join that environment.

flowchart LR
    A[Environment variables] --> C[Resolved settings]
    B[Persisted config file] --> C
    C --> D[Model and provider route]
    C --> E[Sandbox backend and image]
    F[API key or subscription auth] --> D
    E --> G[Running scan environment]
    D --> G
    G --> H[Agent model turns]
    G --> I[Caido proxy client]
    I --> J[Captured traffic and replay]
    C --> K[Viewer relay settings]
    K --> L[Encrypted report delivery]
    M[MCP server definitions] --> N[Per-run MCP registry]
    N --> O[On-demand tool discovery and dispatch]
    O --> G
    C --> P[Web search provider]
    P --> G

References: strix/…/loader.py• L36-52 strix/…/settings.py• L229-238 strix/…/models.py• L613-667 strix/…/backends.py• L20-55 strix/…/caido_bootstrap.py• L81-115 strix/…/loader.py• L101-133 strix/…/client.py• L341-374 strix/…/agent_tools.py• L59-181 strix/…/auth.py• L132-154

Change checklist

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

  • When a settings field or alias changes, update both environment resolution and persisted-file remapping. A field can appear to work from the shell while failing to load from cli-config.json.
  • When a linked model, API key, or API base changes, preserve the rule that the stored connection is discarded as a unit rather than mixed with live shell values.
  • When model cost-rate settings change, keep their names and non-negative validation aligned with cost fallback and usage reporting. Provider-reported cost must still take precedence when it exists.
  • When a model prefix, API type, or provider route changes, review model catalog lookup, tool-schema selection, reasoning support, streaming behavior, extra headers, app attribution for OpenRouter and Vercel AI Gateway, request body values, cost handling, and provider credential mapping as one contract.
  • When parallel tool-call settings change, preserve the opt-in default and pass the setting only when the model request has tools. Review provider support for parallel calls with the request contract.
  • When prompt cache settings change, keep cache-point handling aligned with the provider route and its supported cache block size.
  • When provider tool-call ID handling changes, preserve unique non-empty IDs across history and keep one replacement stable across streamed item events and the completed response. For placeholder IDs, preserve output-position pairing for parallel calls.
  • When OpenRouter session behavior changes, preserve the explicit opt-in default and keep session IDs scoped to the agent whose prompt cache they are meant to preserve.
  • When subscription token storage or refresh changes, preserve the separation between subscription-auth.json and cli-config.json, keep the account identifier aligned with each request, and preserve the fallback for non-rotated or concurrently refreshed tokens.
  • When subscription request handling changes, keep content guardrail errors terminal and keep token refresh active for each request.
  • When stream timeout settings or wrappers change, keep first-event, inter-event, and whole-stream limits distinct. Preserve the rule that non-streaming requests use the normal request timeout instead of stream limits.
  • When the runtime backend or context limits change, verify the sandbox data path and the amount of history, tool output, and image output that can reach the model.
  • When web search provider settings change, keep provider selection, provider credentials, Exa search mode, and result bounds aligned.
  • When Caido startup or URL handling changes, check the in-container listener, exposed host URL, temporary project selection, delayed bootstrap, and teardown as one lifecycle.
  • When proxy replay fields change, keep captured-request parsing, modified request construction, framing headers, and the tool result shape aligned.
  • When MCP connection fields or names change, keep transport validation, bearer auth, tool allowlists, connection names, and per-run registry entries aligned.
  • When the MCP config path or file shape changes, update both the default and override paths and preserve the fail-open behavior for optional connections.
  • When MCP dispatch or session failure handling changes, keep on-demand tool discovery, result error signals, retry and quarantine states, and cleanup ownership aligned.

Important failure modes

TriggerConsequenceGuard
An unknown STRIX_RUNTIME_BACKEND valueScan startup stops instead of selecting a usable sandboxBackend lookup raises on unsupported names (strix/…/backends.py• L60-75)
A persisted setting uses an alias that is also present in the environmentThe stale file value could override the user's live valueFile loading skips a field when any accepted alias is set in the environment (strix/…/loader.py• L112-133)
The shell changes one linked LLM connection value while old linked values remain in the fileThe run could use a mixed provider connectionPersistence drops all linked model, key, and base values when one changes (strix/…/loader.py• L30-34, strix/…/loader.py• L140-154)
An explicit API route conflicts with the provider endpointThe provider can reject the request or return an incompatible response shapeAPI type is resolved from the explicit setting first, then from the model's supported endpoint catalog (strix/…/models.py• L738-749)
A provider receives unsupported tool schemas or reasoning settingsThe provider rejects the model requestRoute checks select compatible request features (strix/…/models.py• L976-1007)
Parallel tool calls are enabled for a provider that does not support themThe provider can reject the model requestThe feature is disabled by default and is passed only for requests with tools (strix/…/settings.py• L77-80, strix/…/inputs.py• L245-265)
A provider returns blank, placeholder, or repeated tool-call IDsStrict providers reject the conversation history, and retries repeat the malformed historyHistory and turn guards assign stable fresh IDs before the turn is recorded (strix/…/models.py• L307-324, strix/…/models.py• L340-407, strix/…/tool_call_ids.py• L48-156)
A gateway cannot reliably stream while streaming remains enabledA model turn can fail or wait for a stalled streamLLM_DISABLE_STREAMING switches compatible routes to one completed request (strix/…/settings.py• L93-95, strix/…/models.py• L171-182, strix/…/models.py• L613-667)
A model emits no first event within LLM_STREAM_FIRST_EVENT_TIMEOUTThe turn fails instead of waiting without bound for the response to startThe first-event limit closes the stream and lets the retry path handle the turn (strix/…/settings.py• L99-103, strix/…/models.py• L420-451)
A model stream emits no event within the idle limitThe turn fails instead of waiting without bound between eventsLLM_STREAM_IDLE_TIMEOUT closes the stream and lets the retry path handle the turn (strix/…/settings.py• L99-103, strix/…/models.py• L420-451)
A model stream exceeds LLM_STREAM_TOTAL_TIMEOUTA slow turn fails instead of consuming the run indefinitelyThe total limit is measured from stream start and closes the stream (strix/…/settings.py• L99-103, strix/…/models.py• L420-451)
OpenRouter sticky sessions are expected without enabling their settingCalls can move between upstream providers, so prompt-cache reuse is not pinnedSTRIX_OPENROUTER_STICKY_SESSIONS is disabled by default and adds a per-agent session only when enabled (strix/…/settings.py• L89-92, strix/…/models.py• L863-871)
ChatGPT access token nears expiry during a scanLater model requests lose authorizationPer-request token refresh uses a shared refresh guard and accepts a peer process's fresh record (strix/…/codex.py• L325-348, strix/…/codex.py• L363-366)
ChatGPT blocks a request with a content guardrailRepeating the same request fails and the scan cannot use that modelThe error is marked terminal and tells the user to select another model (strix/…/codex.py• L136-147, strix/…/models.py• L76-83)
A configured model has no provider rates and the provider returns no costUsage reporting cannot calculate a provider-specific fallback chargeCost settings can supply non-negative input, output, and cached-input rates (strix/…/settings.py• L63-71)
An MCP connection has the wrong transport fieldsThat connection is rejected and is not loadedValidation requires a URL for HTTP and a command for stdio (strix/…/config.py• L83-89)
The MCP file is malformed or an entry is invalidMCP connections are absent or that entry is unavailable, while the scan can continueFile and entry errors return or skip entries instead of stopping the run (strix/…/loader.py• L101-133)
A configured MCP server cannot connect at startupThat connection is omitted while the scan continues with the remaining servicesEach connection starts independently and initial failures are skipped (strix/…/client.py• L304-323, strix/…/session.py• L426-450)
An MCP transport fails during a runThe affected connection is retried, quarantined, or retired without cancelling the whole scanSessions classify failures and contain the transport lifetime on the connection owner (strix/…/failures.py• L65-128, strix/…/session.py• L459-489, strix/…/session.py• L600-647)
A tool call is denied or has invalid protocol dataThe agent receives a failed-tool result, while a call-specific rejection does not make a healthy connection unavailableCall failures are distinguished from session failures and carry a top-level failure signal (strix/…/session.py• L361-396, strix/…/client.py• L243-265)
Caido is not ready or its bootstrap failsProxy tools cannot inspect or replay trafficBootstrap retries readiness and tools degrade to a client-unavailable result (strix/…/caido_bootstrap.py• L32-78, strix/…/tools.py• L91-105)
Two proxy calls use the shared GraphQL transport at the same timeCaido reports a transport connection errorClient creation and calls are serialized (strix/…/caido_api.py• L92-127)
A replay keeps the captured request's old body framingThe target may truncate or misread the modified requestReplay removes stale framing headers and recomputes Content-Length (strix/…/caido_api.py• L173-214)

Key files

FileWhy to read it
strix/…/settings.pyDefines the resolved settings groups, provider controls, model cost-rate overrides, parallel tool-call opt-in, request extras, context limits, prompt-cache controls, stream limits, runtime values, web search choices, telemetry, and environment aliases.
strix/…/loader.pyShows source precedence, linked connection invalidation, config-file overrides, caching, and safe persistence.
strix/…/models.pyExplains the high-level model route, provider compatibility decisions, request safeguards, authentication defaults, app and provider attribution, prompt-cache handling, stream limits, and provider cost handling.
strix/…/codex.pyDefines ChatGPT subscription selection, token storage, refresh coordination, content-guardrail handling, and client access.
strix/…/backends.pyDefines the sandbox backend boundary and backend selection rules.
strix/…/caido_bootstrap.pyShows how a sandbox-side Caido service becomes a host-side client.
strix/…/caido_handle.pyShows the delayed Caido bootstrap and cleanup boundary.
strix/…/config.pyDefines the connection contract for remote and local MCP servers.
strix/…/loader.pyShows MCP file selection, validation, filtering, and fail-open loading.
strix/…/agent_tools.pyShows the generic discovery and dispatch surface used by the agent.
strix/…/client.pyShows transport setup, allowlist enforcement, result conversion, and per-run attachment.
strix/…/session.pyShows connection ownership, failure isolation, retry, quarantine, and cleanup.
strix/…/failures.pyShows the safe failure classes used to decide whether a connection can recover.
strix/…/tools.pyShows the agent-facing traffic inspection, scope, sitemap, and replay capabilities.
strix/…/caido_api.pyShows the shared Caido client boundary and request replay contract.
strix/…/auth_cli.pyShows the user-facing ChatGPT sign-in, status, and sign-out flow.
strix/…/auth.pyShows local viewer verification and encrypted report delivery through the relay.