Back to Qwen Code

External Context extension

integrations/external-context/README.md

0.21.821.1 KB
Original Source

External Context extension

This private Qwen Code integration connects one interactive CLI process to one administrator-bound external context corpus without changing Qwen Core. It has three managed deployment variants:

  • Read-only on-demand: version 1 configuration and the MCP context_search({ query }) tool.
  • Mem0 write: version 1 Mem0 configuration, the on-demand search tool, and the optional context_remember({ content }) tool with a separate content-visible confirmation Hook.
  • Auto-recall: version 2 configuration and an administrator-installed UserPromptSubmit Hook, with no external-context MCP server.

The built-in adapters support Mem0 Platform V3 search and a small Generic HTTP Search V1 contract for existing knowledge or RAG services. Only Mem0 has an optional write path. There is no generic ingestion protocol, personal memory, trusted user identity, per-document ACL, or tamper-resistant audit.

Use the governed Gateway/Orchestrator Profile described in #7449 when those controls are required.

Trust boundary

In the read-only on-demand profile, the model can provide only the search query. In the Mem0 write variant, it can additionally provide only the exact content to store. In the auto-recall profile, the query is derived only from Qwen's optional submitted_prompt provenance, captured before model-bound expansions. Provider type, endpoint, credential, Mem0 app_id, and all other corpus selectors are fixed before either process starts.

The actual corpus-isolation boundary is the provider-side credential, project, index, or corpus. A Mem0 app_id or any other client-supplied filter is classification, not authorization. Read-only deployments should use a read-only credential where the provider supports one. The write variant requires a credential that can add memories and therefore grants same-UID code access to that capability outside Qwen as well.

For on-demand retrieval, the extension manifest alone is not a managed binding. Qwen merges MCP servers by name, and a settings, project, or command-line server named external-context can replace the manifest contribution while retaining the same permission-rule name. A managed deployment must therefore start Qwen with an administrator-owned --mcp-config based on examples/managed-mcp.json. The Phase 1 launcher must construct the complete Qwen argument vector itself and must not pass through arbitrary caller arguments. This command-line tier overrides user, project, workspace, and system MCP settings. The documented permission rule is safe to deploy only inside that pinned process.

The launcher must also construct an administrator-approved environment rather than inherit caller-controlled values. Qwen can subsequently load values from the repository's .env and .qwen/.env files, so the Direct Profile requires the repository, those files, and same-UID code to be trusted. The source pin prevents same-name MCP configuration collisions; it is not a process sandbox. Use the governed profile when those inputs may be hostile or when credentials and process execution must be isolated.

The search tool omits MCP readOnlyHint because a provider search may record access metadata or otherwise have provider-side read effects. It exposes no explicit mutation in the default extension manifest. The managed Mem0 write variant exposes an explicit add operation. Both variants pass content to an external service, so this integration is not a DLP boundary. Credentials inherited by Qwen may also be visible to same-UID processes and tools. Use the governed profile when the credential, write authorization, or outbound-content policy must be isolated from the CLI user.

Retrieved context remains untrusted reference data if the model proposes storing it back into the shared corpus. A user approval does not upgrade that trust label. Review the complete write carefully because persisting retrieved or injected text can expose it to later users and model turns.

The managed-settings example allows Qwen to invoke search without per-call confirmation. It is on-demand rather than prompt-triggered, but it is not necessarily initiated manually by the user. In interactive non-YOLO mode, placing the tool under permissions.ask requests confirmation. YOLO mode auto-approves ordinary tools despite ask, and users can change approval mode during a session. Phase 1 does not provide non-bypassable per-call confirmation; use the governed profile when that is required.

Configure

On-demand profile

  1. Give the repository its own provider-side project, index, or corpus and a credential restricted to it. Verify that the credential cannot access or select another corpus.

  2. Copy the applicable provider configuration from examples/ to an administrator-owned location outside the repository that the CLI user cannot modify. Configure apiKeyEnv or tokenEnv if needed, then set the referenced environment variable to the credential. timeoutMs defaults to 5000 and may be between 1 and 30000 milliseconds.

  3. Have the managed launcher set QWEN_EXTERNAL_CONTEXT_CONFIG to the absolute configuration path.

  4. From the Qwen Code checkout, install dependencies and build this workspace:

    bash
    npm install
    npm run build --workspace @qwen-code/external-context
    

    Phase 1 is a private monorepo workspace. Copying the directory or its npm tarball without packaging its runtime dependencies is not a supported deployment.

  5. Copy examples/managed-mcp.json to an administrator-owned location and replace every placeholder with an absolute path. The command, args, and cwd must identify an administrator-controlled Node executable, reviewed checkout, and dependency tree that the CLI user cannot modify. The managed launcher must accept no arbitrary Qwen arguments. It must construct a clean, administrator-approved environment, inject the provider configuration and credential, change to the intended repository, and invoke:

    bash
    qwen --mcp-config /administrator/path/external-context-mcp.json
    

    The MCP subprocess honors HTTP_PROXY, HTTPS_PROXY, and NO_PROXY through an environment-aware dispatcher. If the provider requires an egress proxy, the launcher must include the administrator-approved proxy variables in that clean environment. Include localhost, 127.0.0.1, and [::1] in NO_PROXY when using the loopback Generic HTTP provider.

  6. Point QWEN_CODE_SYSTEM_SETTINGS_PATH at an administrator-controlled copy of examples/managed-settings.json only inside this managed launcher; do not install its automatic allow rule for unrelated Qwen sessions. It disables /cd to reduce accidental workspace/corpus mismatch and allows the pinned search tool. Neither setting is an authorization boundary: the provider credential is still the corpus boundary, and a new Qwen process is required to switch repositories.

For a local trusted trial, the built directory may instead be linked with qwen extensions link. The extension manifest contribution and workspace-scoped enablement are convenience mechanisms only; they do not provide the managed MCP source binding described above. Before starting Qwen, the trial environment must set QWEN_EXTERNAL_CONTEXT_CONFIG to the absolute configuration path and set the credential environment variable referenced by that file.

Each MCP subprocess reads configuration and credentials once when it starts. Qwen may restart that subprocess, so the configuration path, file contents, and credential-to-corpus binding must remain immutable for the whole Qwen session. Do not overwrite or reuse a configuration path for another corpus. Changing the working directory does not change the configured corpus. The managed settings disable Qwen's /cd command as an accidental-misuse guard, but cannot prevent every same-UID action. To switch corpora, terminate the old Qwen session and start a new one with a new managed configuration path.

Managed Mem0 write variant

This variant is an explicit opt-in for trusted collaborators who need to save repository-shared notes. It uses Mem0 Direct Import with infer: false, so the validated input is sent unchanged as one user message. Direct Import skips Mem0 inference and duplicate detection. Repeating an approved write may create another copy of the same content.

  1. Provision a dedicated Mem0 Project for the repository. Do not share its key with another security domain. Mem0 app_id remains classification within the Project, not authorization.
  2. Copy examples/mem0-write.json to an administrator-owned path and set the referenced MEM0_API_KEY. write is accepted only for version 1 Mem0 configurations. Generic HTTP and version 2 configurations reject it. The top-level timeoutMs applies to both search and write; choose a write budget appropriate for the deployment because a client timeout produces an ambiguous unknown result.
  3. Build the workspace so both dist/main.js and dist/write-confirmation.js exist.
  4. Start Qwen with an administrator-owned copy of examples/managed-mem0-write-mcp.json. The default extension manifest stays search-only and does not enable context_remember. Register this managed MCP server under exactly external-context; the Hook matcher uses mcp__external-context__context_remember, and the command ignores other tools if a matcher is accidentally broadened. Another server name would therefore prevent the content confirmation from running.
  5. Point QWEN_CODE_SYSTEM_SETTINGS_PATH at an administrator-controlled copy of examples/managed-mem0-write-system-settings.json. The provided system settings deliberately omit the Hook. Put the applicable POSIX or Windows managed write user-settings example in a dedicated administrator-controlled QWEN_HOME/settings.json. User-scope Hooks take precedence over legacy merged Hooks, but a system-scope Hook can still run when user Hooks are absent, so audit both scopes and avoid additional matchers for this command.
  6. Use the same fixed-path, no-user-arguments, environment-allowlist launcher requirements as the read-only profile. In addition, pin the Hook and dedicated QWEN_HOME; refuse headless, ACP, serve, resume/continue, and startup YOLO modes; and set QWEN_CODE_MEMORY_TEAM=0, QWEN_CODE_MEMORY_TEAM_SYNC=0, QWEN_TELEMETRY_ENABLED=0, QWEN_TELEMETRY_LOG_PROMPTS=0, QWEN_TELEMETRY_INCLUDE_SENSITIVE_SPAN_ATTRIBUTES=0, and QWEN_USAGE_STATISTICS_ENABLED=0. On Windows, the approved PATH must resolve powershell to the system executable, and PowerShell profiles must be absent or administrator-controlled.

In regular interactive mode, the permissions.ask rule presents Qwen's normal server/tool confirmation and the PreToolUse Hook then presents a second confirmation containing a reversible, safely escaped representation of the complete text. The Hook confirmation is rendered literally rather than as inline Markdown, so formatting markers and link targets remain visible. If the text exceeds the constrained terminal area, Qwen reports the hidden line count and Ctrl-S reveals the remaining confirmation content. YOLO bypasses the ordinary permission rule, but a working Hook still asks once. Plan mode, unknown permission modes, malformed Hook input, and environments unable to confirm are denied. Extra tool arguments are ignored consistently by the Hook and MCP schema and never reach Mem0. Hook transport failure follows Qwen's existing fail-open semantics, so this is a user-experience safeguard, not an enforceable authorization boundary. Strict approval requires the governed profile.

The confirmation Hook code does not read the Provider configuration or credential, but Qwen command Hooks inherit ordinary third-party credentials from the parent environment. The Hook process may therefore receive the Mem0 key in its environment. The Direct Profile does not provide credential isolation from administrator-pinned child processes or same-UID code.

context_remember accepts at most 4000 Unicode code points, preserves leading and trailing whitespace, and rejects empty/invisible-only text and malformed UTF-16. It never accepts a model-selected tenant, Project, app_id, filter, or metadata. Each approval performs exactly one request; the integration does not pre-search, retry, poll, cache, or deduplicate.

Mem0 V3 Add is asynchronous in normal operation. PENDING with a valid event_id is therefore the expected successful response and returns accepted, not stored. A valid synchronous SUCCEEDED response returns stored as a defensive compatibility path. Explicit FAILED and HTTP 400, 401, 403, or 404 responses return a stable failed result. A timeout, cancellation, redirect, other HTTP status, broken response, unexpected status, or invalid operation identifier returns unknown and an MCP error warning the model not to retry automatically, because Mem0 may already have accepted the write. Cancellation therefore does not prove that no memory was created.

Auto-recall profile

Auto-recall sends a sanitized best-effort query to the external provider for each eligible ordinary interactive prompt. It requires a non-empty submitted_prompt captured by the supported interactive TUI before reminders, file and resource expansion, extension output, and vision expansion. It never falls back to the legacy model-bound prompt. Missing or invalid provenance fails closed before configuration or credentials are read. Common credential shapes are removed from the submitted text, but this is not DLP.

  1. Copy examples/auto-recall-mem0.json or examples/auto-recall-generic-http.json to an administrator-owned location. Set repositoryRoot to the one absolute repository bound to the Provider credential. The directory must exist and must not be a filesystem root.
  2. Build this workspace so dist/auto-recall.js exists.
  3. Put the applicable examples/managed-auto-recall-user-settings-posix.json or examples/managed-auto-recall-user-settings-windows.json content in the settings.json of a dedicated administrator-controlled QWEN_HOME. Replace all placeholders with fixed absolute Node and Hook paths.
  4. Point QWEN_CODE_SYSTEM_SETTINGS_PATH at an administrator-controlled copy of examples/managed-auto-recall-system-settings.json. Its system-level disableAllHooks: false prevents lower-precedence workspace settings from suppressing the required Hook.
  5. Use a launcher that accepts no arguments, verifies stdin and stdout are TTYs, starts in repositoryRoot, supplies only an administrator-approved environment, and fixes all Qwen, Node, Hook, configuration, settings, and credential paths. It must set QWEN_CODE_MEMORY_TEAM=0, QWEN_CODE_MEMORY_TEAM_SYNC=0, QWEN_TELEMETRY_ENABLED=0, QWEN_TELEMETRY_LOG_PROMPTS=0, QWEN_TELEMETRY_INCLUDE_SENSITIVE_SPAN_ATTRIBUTES=0, and QWEN_USAGE_STATISTICS_ENABLED=0. On Windows, the approved PATH must resolve powershell to the system executable, and PowerShell profiles must be absent or administrator-controlled. A user-controlled shell shim or profile is outside the Direct Profile trust model.

Do not link or enable the external-context Extension Manifest and do not install examples/managed-mcp.json in the auto-recall process; either action would add the on-demand MCP surface and permit duplicate retrieval. The shared configuration loader accepts v1 and v2, but the MCP process entry point requires v1 and the Hook requires v2. Supplying the same v2 configuration to the MCP therefore fails startup. The managed Auto Profile must still omit the extension and MCP configuration because a separately configured v1 MCP process would permit duplicate retrieval. In v2, autoRecall.timeoutMs is the only request timeout the Hook reads; the top-level timeoutMs remains for schema compatibility and has no current runtime consumer. The profile supports only a fresh interactive TTY session. The managed system settings explicitly disable speculative execution because accepted speculation can bypass the normal UserPromptSubmit path. It does not support -p, stream-json, ACP, serve, YOLO, --continue, --resume, arbitrary launcher arguments, or switching repositories in one process. Mid-turn steering messages do not fire UserPromptSubmit and therefore do not trigger recall; only ordinary submitted turns are eligible.

The Hook reads at most 1 MiB from stdin, emits at most 4000 code units of structured untrusted_external_context, performs no retries or caching, and fails open as {} after the Node entry point starts. Failure to spawn the pinned Node process and a Qwen outer command timeout retain Qwen's blocking command-Hook semantics. The Provider timeout defaults to 1500ms and is capped at 5000ms; the internal Hook wall-clock budget is 6500ms and the managed Qwen command timeout is 8000ms. Each Hook invocation destroys its own proxy dispatcher after the attempted retrieval so stalled proxy connections cannot retain the child process; the long-running MCP process keeps its dispatcher.

Retrieved results are sent to the model provider as user-layer additional context. The managed profile disables Qwen chat recording, native memory, usage statistics, and telemetry by default. Provider-side access logs remain outside this integration's control.

For Mem0 auto-recall, verify that Memory Decay is disabled for the bound Project. If that cannot be verified, use the on-demand profile.

The integration emits no local per-request audit record. It does not write queries, results, credentials, provider errors, or operation metadata to stderr. The on-demand MCP entry point may write a sanitized startup configuration error once before it exits; unexpected startup failures remain opaque. The auto-recall Hook emits only {} on failure. Operators who need access records may use provider-side logs, but those are outside this integration and are not a tamper-resistant compliance audit.

Generic HTTP Search V1

The configured baseUrl must be an origin with no path, query, credentials, or fragment. That origin receives a request at the fixed path /v1/context/search:

http
POST /v1/context/search
Authorization: Bearer <credential>
Accept: application/json
Content-Type: application/json

{"query":"normalized query","limit":5}

The response is:

json
{
  "items": [
    {
      "id": "opaque-id",
      "content": "retrieved text",
      "title": "optional title",
      "uri": "optional provenance URI",
      "score": 0.82,
      "updated_at": "2026-07-23T00:00:00Z"
    }
  ]
}

The fixed endpoint and the credential's effective capabilities must together restrict access to one corpus. A bearer credential that can access another corpus through another endpoint or selector does not meet the Direct Profile boundary. The request contains no client-selected tenant, repository, namespace, or filter. HTTPS is required except for explicit loopback HTTP used in local development.

Mem0 Platform V3

The adapter calls POST /v3/memories/search/ with the configured app_id, top_k: 5, threshold: 0.1, and rerank: false. The API key's effective Mem0 Project must already be restricted to the intended corpus; a different app_id in the same broadly accessible Project does not establish isolation. The default extension manifest and read-only examples do not enable writes.

The managed write variant additionally calls POST /v3/memories/add/ with one exact user message, the configured app_id, and infer: false. It exposes no get-all, update, delete, delete-all, entity, event, or Project-management tool. Mem0 Direct Import does not perform duplicate detection, and the asynchronous Add API can leave the client unable to determine whether a write was accepted. The integration therefore never retries an ambiguous write.

Mem0 Memory Decay is opt-in and off by default. If enabled, search reinforces returned memories, updates access history, and can affect later ranking. Keep it disabled when search must have no semantic provider-side state change. Provider audit or access logs may still be retained. See Mem0 Memory Decay.

Rollout and rollback

Start with the pinned read-only on-demand MCP for one workspace and validate search quality and provenance. For writes, first run the repository's interactive fake-Mem0 test harness, then progress through an isolated temporary Mem0 Project, one trusted repository, and a small team. The shipped Mem0 configuration always targets Mem0 Platform; only the test harness injects a local endpoint. Enable auto-recall only after the administrator accepts automatic query forwarding. Do not run auto-recall and an on-demand MCP in one process.

Removing the pinned MCP or auto-recall Hook from the managed launcher rolls back the Qwen integration; local on-demand trials can instead disable or remove the extension. Restore a preserved v1 configuration before rolling back to a binary that does not understand v2. To roll back the write variant, remove its MCP configuration, Hook, and credential, restore the read-only v1 configuration, and restart Qwen. Rollback does not delete memories already accepted by Mem0 or remove provider-side search logs and access metadata; administrators must handle existing records at the Provider.