Skip to main content
AI observability records each model call and tool call that your app makes, with its tokens, cost, latency, and status. You can see where AI spend goes, find slow or failing calls, and follow one chat or agent run through its steps.

​
Send AI data to Squasher

Prompt and output text are off by default in the Agent SDK. Squasher still records tokens, cost, model, and timing.

​
Find your AI data

Select AI observability under Telemetry in the sidebar. Choose a time range from Last 30 minutes to Last 30 days; the default is Last 24 hours. Select a request to open its record. Overview shows latency, cost, and input and output tokens with their cached and reasoning parts, and a bar that splits the tokens. Below that, it shows the conversation when you capture content:
  • The request settings, the system instructions, and the tools the model could call.
  • Each message, labeled User, Assistant, or Tool. The model’s reply is marked Response.
  • Thinking, closed by default. Select it to read it.
  • Each tool call with its arguments and the result it returned. Failed calls open by default.
  • Code in tool arguments and in replies, for example JavaScript or SQL, in color.
Search the messages, open or close every block with Expand all and Collapse all, or select JSON to see the payload exactly as recorded. Input and Output show each side at full width. Squasher reads OpenAI, Anthropic, OpenTelemetry GenAI, and Vercel AI SDK message formats. Under the conversation, Trace shows every step of the request’s trace on a waterfall, with this request selected. Select another step to open it. Select a session or user ID to open its page, or the copy button beside it to copy the ID. Each step shows its kind as an icon and a word: Agent session (a solid tile) for the run of an agent, Generation for a model call, Tool call for a tool, and Span for any other step. An agent session records no tokens of its own, so its page counts the model calls and tool calls below it and adds up their tokens and cost.

​
Find where a run failed

When a step failed, its page starts with Failed: the error type and the error message that the step recorded (exception.message), with a copy button. The trace waterfall follows, with the failed step selected. A step that did not fail but holds failed steps shows N failed inside beside its status, and a Failures list: each failed step in time order, with its kind, the step it ran under, how long into the run it failed, and its error message. Select a failure to open it. On the waterfall, a failed step has an error icon and an outlined bar. A step that holds failed steps shows a warning triangle and their number. Errors only keeps the failed steps and the steps above them. Next error opens and selects each failed step in turn. From a trace, you see its steps on a waterfall and its failures beside it. Select a step, or press j and k, to see its error, its conversation, or the arguments and result of a tool call, beside the waterfall.

​
Read a session

A session page shows the totals of one chat or agent run, its requests, and its traces. A session includes every call of each trace in which a call has the session ID, so model calls without a session ID still add their tokens and cost. Needs attention lists the checks for the session and the number of calls that failed each one: rate limits, context window, refusals, tool errors, timeouts, and provider errors. The session’s verdict is the first check that failed, or Completed when all pass. Squasher reads the finish reason (gen_ai.response.finish_reasons), the HTTP status, and the error type and message of each call.

​
How Squasher counts cost

  1. If the call reports its own cost, Squasher uses that number. The Agent SDK reads the cost that OpenRouter returns. In HTTP ingest, send costDetails.total.
  2. If not, Squasher computes the cost from the token counts and a public price list for each model. Cached input tokens use the cached price, and tokens written to the cache use the cache-write price. As in the OpenTelemetry GenAI conventions, the input tokens include the cache reads and writes. Squasher updates the prices every 24 hours.
  3. If Squasher does not know the model, the cost is 0.
When a model call is in a span that another model call span contains, such as a Claude SDK span over the Claude API span, Squasher counts the tokens and cost of the inner span only.

​
Record scores and feedback

Scores tell you whether an answer was good, not only fast and cheap. Send them with the Agent SDK: captureFeedback for thumbs and ratings from users, and captureScore for automated checks. They appear on the Scores tab, linked to the trace or call that they judge. See Record quality scores.

​
Check AI telemetry in CI

squasher agent-observations quality checks the last 24 hours for missing traces, prompts, outputs, models, tokens, and cost. It exits with an error when a check fails:
Add --min-scores or --min-scored-observations to check that scores still arrive, and --require-privacy-mode to check that every call was recorded in privacy mode.

​
Export data for evaluation

squasher agent-observations export writes model calls and their scores as JSON Lines, for offline evaluation or fine-tuning. Filter with --score-name, --min-score, --model, and other flags. The API key needs the ai:export scope. Session IDs, user IDs, and metadata are left out unless you add --include-identifiers or --include-metadata.

​
Read complete model and tool content

When your integration sends full content, the call detail includes the prompt, input messages, output, returned reasoning, and tool arguments and results. Sensitive fields are redacted before capture. Squasher can record only the reasoning that the model provider returns. For complete coverage, send canonical tool arguments and results along with the model call. If content is omitted or fails to serialize, mark that in the record. Use the telemetry quality check above to confirm the coverage of your integration.

​
Limits

  • The plans list an AI request allowance, but Squasher does not enforce or bill it yet. Squasher does not drop AI data when you pass the allowance. See Plans and limits.
  • AI observability records the calls that your app makes. It does not include Squasher’s own AI triage and fixes.

​
Next steps

Agent SDK

Record every Vercel AI SDK call with one wrapper.

OpenRouter

Record OpenRouter calls with their exact cost.

Agent Observations API

Query and export AI data over HTTP.

Send OpenTelemetry

Send gen_ai.* spans from any language.

​
Credential redaction

The JavaScript SDKs mask common credential formats, private-key blocks, URL passwords, and credential assignments in diagnostic text. Numeric token counts remain visible. Explicit key and path masks take priority. Add explicit masks for your own secret fields; pattern matching cannot identify every secret.