Spear Assistant

Spear Assistant investigates account data using RouteSpear’s tools. Open it from Analyze → Spear Assistant, or use Ask Spear beside a supported page. Start with a target, time window and specific question. Use skills for repeatable workflows and MCP connections to connect an external client.

Review proposed changes

Assistant writes pause at Review proposed change, showing the operation and its exact arguments. Approve change executes that stored proposal once, then starts read-only verification. Decline discards it. Typing “yes” or selecting Continue does not authorize a write.

Approvals expire after 15 minutes, belong to your user and conversation, and are superseded by a new message. Permissions are checked again when approving: changes require ai:execute and the affected resource permission. If execution is interrupted, inspect current configuration before requesting a new proposal; the consumed approval cannot be reused. Saving a change is not evidence that measurements have recovered.

A run pauses after reaching its step, tool-call or time budget, or after repeated calls or failures. Run paused explains the reason. Continue starts a new bounded run with the saved history; it does not retry automatically or expand read-only verification permissions.

For changes across several objects, Spear can propose an ordered change plan. Review each step’s purpose, exact arguments, recorded current configuration and recovery instructions, then approve the plan once. Spear checks the recorded configuration before writing and stops at the first failure. Completed changes remain applied; this is not an atomic transaction. Ask for the plan receipt to inspect completed, failed or uncertain steps before planning further work.

Return to ongoing work

Closing a tab or losing the connection does not stop a submitted run. Reopen the conversation to reconnect to its saved progress. Stop requests cancellation on the server. Runs retain the original account, user, page context and time window, and permission or session revocation stops further work.

After an API restart, Spear resumes from a saved analysis checkpoint without repeating approved changes. If a restart interrupts a write before a safe checkpoint, the run pauses for inspection. Ask Spear to inspect run progress and current configuration before requesting another change. Recovery stays within the original execution limits; it does not provide unlimited background work.

Explore and configure your account

Ask what Spear can do on a page or for an account feature. Its capability catalog and discovered tools reflect your permissions, with additional resource checks when an operation executes. Skills cover onboarding, monitoring, dashboards, access, alerts, baselines, logs, infrastructure and support workflows.

Credential entry, authentication and file uploads use links to the relevant secure settings screen. Spear can prepare the surrounding configuration and verify it afterward. A settings link does not mean that setup has completed.

The assistant model belongs to the account

Open Manage → AI & MCP → AI config → Assistant model. Users with ai:manage choose the provider, endpoint, available models and credentials for the tenant. The account shares this configuration and one starred default model. Personal client tokens and connection instructions are on the MCP config tab.

1. Open the Vendor dropdown and search for a provider, self-hosted runtime, or custom endpoint. Selecting a preset fills its protocol and base URL.
2. Open Model to add suggested models. For a custom model or Azure deployment, type its exact ID in the search field and choose Use custom model. Enter any required API key. You can add up to 32 models.
3. Star one model as the account default. Select a model chip to inspect its options; this does not change the default. Remove a model with its remove control.
4. Save the configuration, then use Test connection to test the saved settings.
5. Open Spear Assistant and begin with a specific target, time window and question. Select the compact model-name button inside the composer beside Send to open a popover containing Model, Reasoning effort and Context window controls. The button remains available when only one model is configured. The controls are disabled while a response streams.

Users with ai:read can inspect the configuration and use the assistant; changing it requires ai:manage. Role names are editable, so check the actual permissions in Users & roles. A successful connection test confirms provider connectivity; it does not establish that a monitoring target is healthy.

For a saved connection, the model picker also includes models available to its key when the provider supports listing them. List models refreshes that list. Suggestions and custom entry remain available if listing fails. Choose a model that supports tool calling.

Discovered models apply only to the saved connection. Changing the vendor, protocol, endpoint or key hides the old discovery results until the new connection is saved. Selecting a different vendor resets the model collection to a suggestion for that vendor.

Select a model chip to set Reasoning effort or Context window (tokens) when verified controls are available. Higher effort may use more time and tokens. Leave these at Provider default to let the provider decide. Custom IDs remain usable when their capabilities are unknown; their advanced controls stay provider-managed.

In the chat controls, Account default inherits the selected model’s saved reasoning effort or context window. Provider default lets the provider decide instead. Choose a supported reasoning effort or a context window from the dropdown to override that setting for subsequent messages without changing account defaults. Context choices stay within the model’s verified maximum. Unsupported controls are disabled and show Provider managed.

A smaller context budget reserves room for the reply and sends recent complete turns, leaving the full conversation history saved. RouteSpear estimates request size conservatively from serialized UTF-8 bytes rather than an exact model tokenizer. If the current turn cannot fit, the request fails instead of silently cutting that turn.

AI usage by feature and user

The AI usage tab shows provider-reported input and output tokens and model request counts for the last 24 hours, 7 days, 14 days or 30 days. Group by feature, user or model, and filter by user and feature. View users shows who used a feature; View features shows where a user’s tokens went.

AI managers (ai:manage) see account-wide usage; other AI readers see their own usage. Spear, specialist analysis, shared incident analysis, context summaries, conversation titles and connection tests are recorded separately. Background smart-alert work appears under Automation. A request is a provider attempt, so a conversation can include many requests and retries. Failed attempts retain any tokens the provider reported. Missing token counts make totals partial.

Tracking starts when the usage migration and updated API are deployed; earlier conversations are not backfilled. External MCP clients use their own models, whose token consumption RouteSpear cannot see. These are consumption metrics, not currency charges or a provider billing reconciliation.

Smart-alert levels

On AI config, choose a smart-alert level and select Save smart alerts:

• Off: no background AI assessments.
• Classify severity: an advisory severity with an explanation.
• Classify and group: also link alerts of the same type or likely incident.
• Classify, group and analyze: also summarize evidence, possible causes and suggested next checks.

An enabled, saved AI provider is required. New open alerts created after enabling are assessed in bounded batches using the saved default model. Changing levels reassesses eligible open alerts. Grouping uses the current batch and a bounded set of recently assessed open alerts; it is not an exhaustive correlation scan. Provider failures retry after five minutes and appear in the settings status.

Read the AI assessment and grouped alerts on incident details. Original rule severity, notifications, recorded evidence and recovery decisions remain intact. AI classification and grouping are suggestions, not proof of a shared cause.

Inspect the assistant’s thoughts

When the configured model shares reasoning, a collapsed Thinking… section appears while it works. Select it to read the available reasoning; it becomes Thought for Ns when that reasoning finishes. The duration measures time from the first shared reasoning text to the end of that segment, rounded to seconds. It does not include hidden work before that text arrives. Older entries without timing show Thought. The answer appears separately, and the thoughts and recorded duration remain available when you reopen the conversation.

Availability depends on the provider and model. Supported OpenAI models expose reasoning summaries, supported Claude models expose thinking content, and compatible endpoints may return their own reasoning text. If none is available, the assistant keeps its usual activity indicator. These sections contain only reasoning the provider makes available, which may be summarized.

For Sonnet 5, RouteSpear explicitly requests summarized thinking for this disclosure.

Read and copy code

Fenced code blocks in assistant responses use syntax highlighting when tagged with a supported language: MetricQL (metricql or mql), JavaScript, TypeScript, Python, Bash, PowerShell, JSON, YAML or SQL. MetricQL uses the same tokenizer as Metrics Explorer. Copy copies the code’s exact text without highlighting markup. Unknown or untagged languages remain plain text, as do blocks longer than 100,000 characters.

Follow tools and charts

Tool badges explicitly show Running, Complete, Failed, or Stopped. Open a badge to inspect its arguments and result; completed tools also show their reported duration.

When the assistant creates a chart, a Loading panel… frame appears at that point in the conversation and becomes the chart in place. Use its Panel tool details info button to inspect the arguments, result, and duration. Successful chart creation does not add a separate tool badge. A failed chart call shows a Failed badge; empty data or interrupted loading leaves an explanatory frame instead of an ongoing spinner.

The assistant can use all 13 metric visualizations, including stats, gauges, ranked bars, tables, histograms, heatmaps and maps. It also has specialized panels for paths, browser request waterfalls, screenshots, hosts, containers and custom metrics. Its tool guidance selects a presentation for the question: a current ranking, a distribution and a trend should have different views. You can request a specific presentation, such as “rank the sites by their latest latency” or “show the browser request waterfall.”

Summary widgets describe current account data, while screenshot and detail panels show the latest available capture or detail. These panels do not become historical views just because the conversation specifies an earlier time window. Histograms describe returned metric samples, which may already be aggregated; they do not establish a raw request distribution.

Chat panels save absolute start and end timestamps when they are created. Reopening a conversation keeps that window; older relative panels use the saved message timestamp as their anchor. Historical path inspection stays inside the saved time context. Empty metric series whose source retention has expired show Data is no longer retained, rather than silently switching to recent data.

Current-only summary widgets and latest-detail panels cannot replay an old observation. Existing chat panels of those types show Historical snapshot unavailable; new chat panels use time-aware charts and diagnostics instead. This does not change the panels available on live dashboards.

Panels in one assistant message share their zoom and measurement cursor. Drag to zoom or select a measurement time in one chart to update the other time-aware panels in that message. Reset all zoom restores their saved windows; clearing measurement time clears it for the message. Other messages remain independent.

The assistant can include analysis references such as Inspect the spike, Zoom to the outage, or See this panel. Selecting one scrolls to the referenced panel and applies its time or interval to the message’s panels. References use fixed timestamps and saved panel IDs; unavailable references do not navigate elsewhere.

MCP tokens belong to you

Under Your MCP connections, each user sees and manages their own tokens within the account. Tokens act as their creator, capped by the selected scope. Account administrators do not manage another user’s tokens through this personal settings panel.

Creating a token requires ai:manage; revoking your own token requires ai:read. See Connect an MCP client for setup, permissions and skill prompts.

1. Choose New token, give it a recognizable client name, select a scope and expiry.
2. Copy the secret when it is shown. It is stored hashed and cannot be retrieved later.
3. Use the endpoint and configuration snippet for your MCP client.
4. Replace the placeholder with your secret and keep that secret out of shared documents or source control.

A read token is for inspection. A write token also permits supported changes, subject to the caller’s permissions. The tool catalog identifies read, write and destructive operations.

Rotate or revoke access

Create a replacement token, update your client, then revoke the old token. Revocation stops clients using that token on their next call. If a token is lost, create another; its original secret cannot be recovered.

When a connection fails

• Assistant model: test the saved endpoint and model, then check required credentials and provider availability.
• MCP client: check the endpoint, bearer token, expiry and revocation status.
• Read works, changes fail: check the token scope and your current account role.
• Tools return no data: verify the target, placement and time window in the portal before changing client settings.

Continue with reading results or query examples.