# Coding-agent and provider-side telemetry

Point your developers' coding agents, and your model provider's own spans, at one Bucker project.
Everything here is OTLP/HTTP into the intake described in [otlp.md](otlp.md); nothing needs a new
exporter. What the data becomes (sessions, spend, detectors) is in [coding-agents.md](coding-agents.md).

Distributing the settings is your administrator's act. Bucker generates the text from your project
and never pushes it anywhere.

## Where the snippet comes from

`sdk_inventory_get` with `include: [setup]` returns a `codingAgents` section beside the install plan,
and the dashboard's setup page shows the same section. It holds:

- managed-settings snippets for Claude Code, Codex, Gemini CLI and VS Code Copilot;
- the provider-side span destination below;
- the producers (`service.name` and instrumentation scope) the inventory recognises as coding agents.

## Credentials are variables, not values

Every snippet refers to two variables instead of holding a DSN:

| Variable | Holds |
| --- | --- |
| `BUCKER_OTLP_HEADERS` | `Authorization=DSN <your project's DSN>` |
| `BUCKER_OTLP_DSN` | the bare DSN, for tools that take the header apart |

Put the values in your MDM profile or secret store. An agent's read of `sdk_inventory_get` never
returns the DSN; the dashboard setup page shows the project's own DSN beside the variables for a
person to copy. A DSN of another project is refused when the section is built, and the snippets of
one project never contain another's key.

## Claude Code managed settings

```json
{
  "env": {
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_METRICS_EXPORTER": "otlp",
    "OTEL_LOGS_EXPORTER": "otlp",
    "OTEL_EXPORTER_OTLP_PROTOCOL": "http/json",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "https://us.ingest.bucker.io",
    "OTEL_EXPORTER_OTLP_HEADERS": "${BUCKER_OTLP_HEADERS}",
    "OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE": "delta",
    "OTEL_RESOURCE_ATTRIBUTES": "team.id=platform,cost_center=cc-1",
    "OTEL_LOG_USER_PROMPTS": "0",
    "OTEL_LOG_ASSISTANT_RESPONSES": "0",
    "OTEL_LOG_TOOL_CONTENT": "0",
    "OTEL_LOG_RAW_API_BODIES": "0",
    "OTEL_LOG_TOOL_DETAILS": "0"
  }
}
```

Install it as the managed settings file (`/Library/Application Support/ClaudeCode/managed-settings.json`
on macOS, `/etc/claude-code/managed-settings.json` on Linux) and substitute the variable when your MDM
renders the profile. Your region's endpoint replaces the example host.

### Content flags, and what each means

All of these are off in every snippet, and no option turns them on.

| Flag | If you turned it on |
| --- | --- |
| `OTEL_LOG_USER_PROMPTS` | developer prompt text would be exported |
| `OTEL_LOG_ASSISTANT_RESPONSES` | model replies would be exported |
| `OTEL_LOG_TOOL_CONTENT` | tool inputs and outputs would be exported |
| `OTEL_LOG_RAW_API_BODIES` | whole API request and response bodies would be exported |
| `OTEL_LOG_TOOL_DETAILS` | commands and file paths would be exported (the one optional flag) |

Exported content is untrusted data: it is scrubbed and scanned before any model reads it. A developer
whose local settings turn a flag on, overriding the managed file, shows up as a `content_flag_on`
finding in the SDK inventory.

## Other agents

Codex (`[otel]`, `log_user_prompt = false`), Gemini CLI (`logPrompts: false`) and VS Code Copilot
(`captureContent: false`) are generated the same way. Their key names are marked `unverified`: check
them against the vendor's reference before rollout.

## Provider-side spans

Model providers that can export their own spans to an administrator-added destination send them here:

| Setting | Value |
| --- | --- |
| Protocol | OTLP/HTTP, protobuf |
| Traces URL | `https://us.ingest.bucker.io/api/<projectId>/otlp/v1/traces` |
| Logs and metrics | `.../otlp/v1/logs`, `.../otlp/v1/metrics` |
| Header | `${BUCKER_OTLP_HEADERS}` |
| Content capture | off |

The project is in the URL and the key must belong to it. Use one destination per project. The GenAI
attribute names the providers use are reconciled on intake, so the spans join your application's own
LLM spans by trace id.

## Checking it worked

Within a few minutes of applying the snippet, the SDK inventory lists the producer (for example
`claude-code`) and the coding-agents project shows its first session. A producer that is silent after
that is a setup problem: check the endpoint, the header variable and that the machine reaches the
ingest host.
