--- summary: "Bounded portable research and explicitly enabled provider-managed research" title: "Deep Research" read_when: - Researching a question across sources or supplied documents - Configuring Google managed research --- Deep Research is a bundled, opt-in plugin with a lazy `deep_research` tool. Enable `plugins.entries.deep-research.enabled` and allow `deep_research` through tool policy. Startup scans bounded local job state and reconstructs due wakes. An empty startup does not resolve research credentials or make research network requests. Existing plugin allowlists and agent tool policy still apply. Disable it with `plugins.entries.deep-research.enabled: false`. ## Choose an engine | Engine | Execution | Evidence | Configuration | | -------------------- | ----------------------------------------------------------- | ------------------------------------------------------------------ | --------------------------------------------------------- | | `portable` (default) | Ordinary API completions plus configured search/fetch tools | Retrieved text, exact quotations, fallible model support audit | Existing model and web-tool configuration | | `google` | Google's managed Deep Research Interactions service | Provider-attributed report citations; not locally fetched passages | Explicit research enablement and a pinned API-key profile | The portable engine does not require Codex, native function calling, or a vendor-specific agent. Foreground and recovered completions use the originating agent's configured model and normal credential resolution. Model overrides obey existing plugin LLM policy. Recovery does not restore unrecorded session-specific model or auth-profile overrides. `requiredAuthMode: "oauth"` requires that mode through the existing auth owner; it never silently falls back to an API key. This feature does not implement a Codex research RPC or a private ChatGPT OAuth client. Codex API-key, Codex OAuth, and ChatGPT plan-usage/token-sharing qualification are separate. ## Start and inspect research Call the tool with a stable idempotency key for one request: ```json { "action": "start", "idempotencyKey": "river-report-1", "request": { "question": "Compare the reported river restoration totals.", "engine": "portable", "asOf": "2026", "output": "A concise comparison with unresolved differences", "sources": { "urls": ["https://example.org/report"] }, "budget": { "modelCalls": 8, "searches": 6, "fetches": 12, "parallelism": 2 } } } ``` Keep the returned `jobId`. Subsequent calls use `{ "action": "status" | "result" | "resume" | "cancel", "jobId": "..." }`. `{ "action": "capabilities" }` reports the portable route and the explicitly activated managed adapter; availability is not a live entitlement check. A repeated start with the same key returns the existing job without resubmitting. Changing the request under that key is rejected. Results are returned as tool responses; the plugin never sends a separate channel message. Repeated result reads return the same saved report, so there is no independently retried delivery. ## Portable evidence and budgets All exploration branches share the model/search/fetch counters. Defaults are 8 model calls, 6 searches, 12 fetches, 2 parallel fetches, and a 180-second global deadline starting when an execution slot is admitted. Maximums are 30, 20, 40, 4, and 900 seconds. Each model response requests at most 4,096 output tokens. Token totals are observed, not a hard token or dollar cap. Search and fetch use the originating session's normal tool policy. Use `sources.allowDomains` / `denyDomains` to select accepted evidence. Denied domains take precedence. These are evidence filters, not a network firewall: search and redirect transports retain their own network policies. Duplicate source text counts once. The retained corpus is limited to 60,000 UTF-8 bytes; omitted or unreadable material is reported as a gap. Supply `sources.documents: [{ "title": "...", "text": "..." }]` for local-text research. Documents disable public search by default; explicitly set `sources.web: true` for mixed research. Explicit URLs may still be fetched. Findings require substantive exact quotations and a separate model support/ contradiction check. Derived values additionally require mechanically checked inputs, units, dates, and arithmetic. These checks do not establish truth or completeness. Missing evidence, malformed model responses, inaccessible sources, and exhausted budgets produce partial results with explicit gaps. ## Google managed research Configure an existing Google API-key auth profile. No new auth store is used. ```json5 { plugins: { entries: { google: { config: { research: { enabled: true, profileId: "google:research" }, }, }, }, }, } ``` Then start with `request.engine: "google"`. The provider uses the documented Gemini Interactions API and `deep-research-preview-04-2026`, with background execution, storage enabled, web search and URL context. API billing and provider storage apply. The provider binds each interaction to its originating agent, conversation incarnation, and selected API-key profile/account. Foreign sessions, reset or deleted incarnations, and changed credentials are rejected before native read or cancellation. The binding survives process restart; it does not grant authority to whoever possesses the handle. Google research rejects supplied documents/URLs, local source restrictions, model/auth overrides, and local duration/call/search/token budgets. Use portable research when those controls are required. Each managed invocation performs one bounded reconciliation step, then returns a running remote job as paused. The service owns the next authorized poll and its persisted backoff. There is no foreground polling loop. Managed research rejects `durationMs` because it cannot control the provider's runtime or spend. ## Recovery and cancellation One feature service owns job, wake, cancellation, and report-publication state in the existing plugin state store. Jobs and operation receipts expire after seven days; each conversation retains at most 32 jobs. A bounded index holds up to 4,096 lightweight wake references, not report bodies. Startup reconstructs persisted due work. Overlapping wakes coalesce, and stored terminal results are not republished. Obsolete wake references are removed atomically without deleting a newer job or wake. Non-retryable poll errors defer automatic reconciliation immediately. Temporary host-authority refresh, unexpected execution/transport failures, initial adapter-preparation refusals, and retryable, definitively not-sent starts share persisted backoff and defer after five failures over the job lifetime; each retry acquires fresh authority. Retryable poll failures use persisted exponential backoff and defer after five consecutive failures; a successful poll clears the failure count. The native handle remains available for inspection or an explicitly authorized resume. If the plugin state store itself cannot be read or updated, the persisted wake survives but recovery may require restarting the service after storage is restored. Before accepting a job, the host records its original admitted source separately from plugin data. Each wake revalidates that source's current identity, grants, model restrictions, conversation incarnation, sharing, and tool policy. Saved targets and native handles are not authorization. Sources without recoverable authority, revoked access, or changed session policy leave a precise unavailable or deferred state; they do not acquire system privileges or switch accounts. Known managed receipts are polled automatically, including pending cancellation. Portable work reuses completed operation receipts and the original deadline; after expiry it renders available evidence into a partial report without new model/search/fetch calls. Queued jobs have a `jobId` before an execution slot opens. Reset, deletion, service retirement, and provider retirement fence late effects and report publication. `status`, `result`, and an explicit authorized `resume` remain available; ordinary recoverable jobs do not require another tool invocation. An interrupted operation without a recorded receipt is not automatically repeated. A managed submission with no recoverable remote handle is `indeterminate`; inspect the provider account instead of starting it again. Do not create a new idempotency key as an automatic retry for such a job. A still-live service can retain proof that dispatch was never attempted or that the provider/router definitively refused a start. Temporary session refresh, startup, or pairing refusals then use bounded automatic backoff; permanent policy refusals defer for an explicit authorized resume. This bookkeeping retains no report content or execution permission. Unclassified RPC failures, timeouts, and crashes without that durable proof do not establish that a submission was never sent. Cancellation intent is persisted before stopping local work. Once a managed start is admitted, cancellation waits for its receipt and then cancels that handle; it does not discard the receipt by aborting the local request. If only the feature's parent tool signal stops and the host transport still returns a receipt, the feature records it while service and conversation authority remain valid, then schedules reconciliation under fresh host authority. Host-scoped cancellation is not bypassed: it can abort the RPC and lose the receipt even when the feature's own signal argument remains active. Such transport loss, service shutdown, or actual authority retirement can leave an indeterminate submission. If the provider owner retires or the state write fails before the interaction binding is committed, a returned interaction ID alone cannot authorize recovery or cancellation through the tool. Do not resubmit; inspect the original provider account directly. This receipt-before-binding window is not solved by automatic recovery. Managed cancellation is confirmed only by a terminal provider response; an HTTP success or aborted local request does not prove the remote job stopped. Pending cancellation survives service stop/start and is reconciled by the next authorized wake. Provider completion can win the cancellation race; the observed final state is returned honestly. ## Qualification Deterministic tests cover the controller, evidence ledger, registered provider transport, recovery, and cancellation. They do not establish live provider access, cancellation latency, billing, retention compliance, or research quality. No native-service or OAuth live qualification is implied. See [Research adapters](/plugins/research-adapters) for the versioned provider contract.