Tail remote MCP logs before downloading

[xeophon]

Overview

Bound MCP server failure-log collection at the runtime boundary so diagnostics no longer transfer and decode an entire remote log just to keep its suffix.

Why

log_tail previously called runtime.read(log) and only applied limit after the complete file reached the host and was decoded. A large server log could therefore consume hundreds of MiB and block the event loop while the launcher was already trying to report a startup failure. In remote runtimes, the unnecessary file transfer adds another unbounded cost.

Change

• Run tail -c <limit> <log> inside the runtime with separate, argv-safe arguments.
• Return the bounded stdout when tail succeeds while preserving missing/unreadable log behavior as an empty diagnostic.
• Return immediately for nonpositive limits, avoiding the old [-0:] behavior that returned the entire stripped log.
• Keep the optimization localized to log_tail; successful server launches and their control flow are unchanged.

Performance impact

A three-round median benchmark over a 256 MiB log measured:

Metric	Before	After	Saved
Wall time	26.337 ms	3.087 ms	23.250 ms (88.3%)
Maximum event-loop stall	12.441 ms	1.180 ms	11.261 ms (90.5%)
Peak traced host allocation	512.01 MiB	0.27 MiB	511.74 MiB (>99.9%)
Runtime-to-host log payload	268,435,456 bytes	at most 2,000 bytes	268,433,456 bytes (>99.999%)

Wall time used perf_counter, loop stall used a 1 ms asyncio heartbeat, and host allocation used tracemalloc. The new measurement includes process startup for tail.

Tradeoff

Tiny failure logs may pay a remote command-startup cost. This path only runs while enriching launch failures, and bounding the transferred data prevents error reporting from downloading an arbitrarily large log.

---

Note

Low Risk
Change is confined to error-enrichment on failed server launch and assumes standard tail in runtimes; no auth or rollout success-path behavior changes.

Overview
log_tail no longer pulls the full remote log into the host and slices it locally. It runs tail -c <limit> inside the runtime so failure diagnostics stay bounded (default 2000 bytes), which avoids huge transfers and decode stalls when MCP startup fails on large logs.

Nonpositive limit now returns "" immediately instead of relying on Python slicing that could return the entire stripped log for limit == 0. Successful launches are unchanged; this only affects log snippets attached to ToolsetError paths in serve_in_runtime.

^{Reviewed by Cursor Bugbot for commit 65c4f52. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Use tail -c to read last N bytes of remote MCP logs before downloading

Updates log_tail in launch.py to invoke tail -c <limit> via runtime.run instead of reading and slicing the full log file in memory. Returns an empty string for non-positive limits or non-zero exit codes, and no longer strips whitespace or applies error-replacement decoding.

^{Macroscope summarized 65c4f52.}

[macroscopeapp]

Approvability

Verdict: Approved

Small performance optimization to the log_tail utility function that avoids downloading large remote logs into host memory by running tail -c remotely instead. The function is used only for error enrichment, maintains proper error handling, and the change is well-documented.

^{You can customize Macroscope's approvability policy. Learn more.}