Secrets are harder to reach and worth less if reached

After this entry you'll know how the platform now protects the credentials a run works with. Three layers of hardening shipped together, each reducing what a misbehaving or manipulated run could expose.

What changes for you

• Redaction at the source. Anything a run persists or displays (session transcripts, the run timeline, Slack replies) is now scrubbed of known secret values before it leaves the agent, so a command that happens to print a credential no longer stores it anywhere.
• A minimal shell environment. The agent's shell now sees only the variables it actually needs to do its job, instead of everything the container was launched with. Less in the environment means less for anything, accidental or hostile, to read.
• Repo-scoped GitHub tokens. The short-lived GitHub token each run receives is now valid only for the single repository that run works on, not for every repository the app is installed on. Child bees no longer inherit their parent's token; each gets its own, scoped to its own repo.

Nothing changes in how you trigger or review runs. The effect is purely blast-radius: if something ever does leak, it reaches one repo for a short time instead of everything.