lanterna-profiler

Lanterna produces agent-facing Node.js profiling reports. Your job is not to summarize the report; it is to drive an interactive performance investigation until the most likely cause, missing evidence, and next measurement are clear.

Core Rules

• The report.agent.md contract is the primary evidence source. Always capture in JSON, then render it: $LANTERNA report report.json --format agent --output report.agent.md (set $LANTERNA per the Lanterna Capture & Rerun Commands prefix). Keep report.json on disk for fields the agent format omits (full retainer paths, source-map failures, complete memory series). Never analyse from raw JSON, --format text, or --format markdown — JSON is a fallback for targeted field lookup, not the analysis surface.
• Treat profiling artifacts as sensitive: reports can expose commands, cwd, file paths, function names, source-map paths, allocation/heap details, and attach/plugin execution context. For production-like or sensitive targets, prefer staging, scrub reports before sharing, avoid unreviewed plugins, and call out that attach/async/plugin hooks can execute against the target process.
• Do not claim a root cause without report evidence and, when code is available, confirmation from the relevant source files.
• Every recommendation must cite a concrete report observation, code observation, or be explicitly labeled a hypothesis.
• If the signal is weak, mostly idle, blocked by caveats, or not representative, stop diagnosis and ask for a rerun with a suitable workload.
• If the codebase is accessible, inspect Files To Read First before recommending changes; otherwise ask for the exact files/functions needed.

Diagnostic Workflow