OTEL_EXPORTER_OTLP_HEADERS and GH_AW_OTLP_ENDPOINTS reach the agent sandbox through awf --env-all #67156

Summary

With observability.otlp configured, the compiler puts OTEL_EXPORTER_OTLP_ENDPOINT, OTEL_EXPORTER_OTLP_HEADERS and GH_AW_OTLP_ENDPOINTS in the workflow-level env:. The agent (and the threat-detection run) executes under awf ... --env-all, and the --exclude-env list does not cover them.

Unpatched v0.91.6 output, agent step:

--env-all --exclude-env ACTIONS_ID_TOKEN_REQUEST_TOKEN --exclude-env ACTIONS_ID_TOKEN_REQUEST_URL \
  --exclude-env COPILOT_GITHUB_TOKEN --exclude-env GITHUB_MCP_SERVER_TOKEN --exclude-env MCP_GATEWAY_AGENT_ID

Detection step: the first three only.

Impact

1. Credential exposure. OTEL_EXPORTER_OTLP_HEADERS carries the collector's auth (for example a Langfuse basic-auth key from secrets.*). It is readable by the agent process and by anything the agent runs.
2. Duplicate telemetry. Copilot CLI auto-exports its own OTel spans when it sees the standard OTEL_EXPORTER_OTLP_* variables. Every LLM call was recorded twice: once by the api-proxy in the attributed gh-aw trace, and once by the CLI in a separate unattributed trace, doubling cost and token dashboards. That also shows the variables do reach the agent process.

The api-proxy and the MCP gateway receive their OTLP configuration host-side, so excluding the variables from the agent container does not break gh-aw's own export.

Expected

Whenever OTLP is configured, append

--exclude-env OTEL_EXPORTER_OTLP_ENDPOINT --exclude-env OTEL_EXPORTER_OTLP_HEADERS --exclude-env GH_AW_OTLP_ENDPOINTS

to every awf --env-all invocation (agent and detection). Our patch notes record that v0.82.x emitted the first two natively; they are absent from current output.

Workaround today

Post-compile rewrite of every awf command line to add the three flags (two fleets, roughly 70 lock files).