This feature is in early access and behind a feature flag. Contact your Augment Solutions Architect to enable it for your tenant
At a glance
Content levels
Every span includes metadata, which can’t be turned off. An admin can turn on each content level separately in Settings → Trace Export → Content. These settings apply to the whole organization and normally take effect within about 30 seconds, on or off. Content already queued or being sent can still arrive during that time. If Cosmos briefly can’t read your settings, it keeps using the last settings it read for up to five minutes.
Choose levels by what you need:
- Read what users asked and what the agent answered: Conversation. This is the smallest content level.
- Debug why the model did something, or build evaluations from real traffic: Model inputs. To keep spans small, each model request sends only what changed since the previous one. See Content attributes.
- Debug a failing tool or command: Tool inputs and outputs. This is usually the largest level, and the most likely to contain source code.
Send less content to one destination
A destination can receive fewer content levels than the organization setting, never more. The Trace Export page doesn’t have a general control for this yet. Today a destination is narrowed in one case only: when it rejects spans as too large, its problem banner offers Turn off tool outputs, which stops sending Tool inputs and outputs to that destination alone. Its details then show This destination receives less content than your organization setting, and Use organization setting removes the narrowing. To keep content away from a platform, turn the level off for the whole organization, or send to your own OpenTelemetry Collector and remove the content attributes there before forwarding.Names and hashes in metadata
Some metadata is derived from your configuration, so treat it as you would your expert and model settings:- Expert names name run spans (
invoke_agent PR Fixer) and are sent asgen_ai.agent.name. Model picker names, including custom Prism titles, name model request spans. A name that contains something that looks like a credential isn’t sent. cosmos.system_prompt.hashis the SHA-256 of a model request’s system instructions, sent at every content level so you can see when a prompt changed without sending the prompt. A hash can’t be reversed, but anyone who can guess the exact system instructions can confirm the guess. If your system instructions are short and predictable and that matters, tell your platform’s admins.- MCP server names name MCP setup spans (
initialize linear) and are sent ascosmos.mcp.server.name.
Before content leaves Cosmos
-
Credentials in content are redacted. These are replaced with
[REDACTED], including inside nested JSON:- private keys (
-----BEGIN … PRIVATE KEY-----), bearer tokens, JWTs, andAuthorization,CookieandX-API-Keyheader values; - GitHub and GitLab tokens, AWS access key IDs, Slack tokens,
sk-…API keys and Google API keys; - assignments such as
api_key=…,secret: …orPASSWORD=…; - values under JSON keys named like credentials (for example
password,api_key,authorization,cookie, or a name ending intoken), andvaluefields next to such a name, as in lists of headers or environment variables; - user names and passwords in URLs (
https://user:pass@host); - the value of any environment variable in the agent’s environment whose name looks like a secret (contains
TOKEN,SECRET,PASSWORD,API_KEY,CREDENTIAL,PRIVATE_KEYorAUTH) and that is at least 8 characters long, wherever it appears.
env, is replaced as a whole. - private keys (
- Free-text metadata is scrubbed further. Error and status text also has email addresses, file paths, home directories, UUIDs and long hex or token-like strings replaced. Free-text attributes are cut to 256 characters and status messages to 1 KB.
- Private reasoning is removed. Model thinking and reasoning blocks are never sent.
- Metadata is checked against an allowlist. Only the attributes in the Export reference are sent, each in a fixed format.
- MCP connection failures use fixed messages. A failed MCP server connection reports one of 17 fixed messages, never the server’s own error text. This covers connecting only: an MCP tool call’s result, including an error result, is ordinary content and is sent when Tool inputs and outputs or Model inputs is on.
Content limits
Content is limited in two steps:- Capture limit, about 120 KB per span. When a span’s content is larger, the agent shortens it before upload: it first cuts every long text value (to 32 KB, then smaller as needed), then keeps fewer items in long lists. Images appear as
[image], and JSON nested more than 32 levels deep as[truncated]. This limit can’t be raised. - Destination limit, Max content size. Each destination can lower the limit to between 1 and 128 KB. Leave it empty to send everything that was captured. A new Honeycomb destination starts at 50 KB, below Honeycomb’s attribute limit.
- Drops tool definitions and system instructions.
- Drops the earliest input messages.
- Shortens long text from its end.
- Drops whole attributes, as a last resort.
- Your platform, or a proxy in front of it, rejects large attributes or requests, for example with HTTP 413 or by truncating fields: lower it.
- Long prompts or tool results are cut short and your platform accepts larger values: leave it empty.
The capture limit is lower than some other agent platforms’ defaults. For example, the Claude API sends up to 0.75 MB of content per span by default. Very long model inputs are shortened, so use Model inputs to see what changed between requests rather than to archive full prompts.
User identity
Include user identity is an organization setting in the Content section, off by default. When it’s on, spans, metrics and log events carryuser.id and user.email for the user each session belongs to. Use it for per-user dashboards and chargeback.
Turning it off stops sending both attributes within about 30 seconds and restarts metric totals, so per-user series end cleanly.
Which sessions are exported
By default, every session in the organization is exported. To export only some, choose Only selected experts under Sessions to export. See Choose which sessions to export.Destination credentials
- Header values and the URL’s path and query are write-only. After you save, Cosmos shows only the URL’s scheme and host, and
••••for each header value. - Saved header values are sent only to the host they were entered for. Changing the URL’s scheme, host or port requires entering them again.
- Destinations must be public HTTPS endpoints. Cosmos won’t send to private, loopback, link-local or reserved addresses, and checks this again on every connection.
- Resource attributes aren’t secret. Keys named like credentials are rejected.