You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The one-machine file bridge: @path in, outputFile out
Control plane stays embedded (D28); the chat's workbench sandbox
becomes the agent's filesystem, bridged by the sim_cli handler in both
directions. @-shaped argv tokens are pre-read from the session sandbox
into the embed context, and the CLI's OWN argument resolver consults
that map — so only genuinely file-accepting flags get file semantics
(a literal --text @channel stays literal) and the server's filesystem
stays unreadable from model argv (the earlier refusal remains as
defense in depth, now worded as this machine's not-found). outputFile
lands command stdout on the machine as a file and returns a short ack,
so exports and traces never transit the model's context. Find-only by
design: a cold machine degrades to inline with boot guidance.
Live e2e of the full loop: trace pulled to file (ack only), run_code
computed on it, @env.json fed a manual run byte-exact.
Claude-Session: https://claude.ai/code/session_01CgaxNAaeD3taGdghbXn17w
result.stdout=`[stdout written to ${outputFile} on your machine: ${result.stdout.length} chars. Read or process it with run_code, or pass it back as @${outputFile}.]`
124
+
}else{
125
+
result.stdout+=
126
+
'\n[outputFile not written: your machine is not booted yet — run any run_code first. Output returned inline instead]'
0 commit comments