log-triage

Log triage: find the line, spare the context

Native tooling produces logs by the thousand lines. The failure is almost always described by a handful of them. This skill is about extracting those lines instead of scrolling, re-reading, or pasting whole logs around, which wastes the reader's attention and this agent's context alike.

The standing rules

1. Never stream a full build log into the conversation. Run builds with output piped to a file, then search the file.
2. Quote the smallest window that proves the diagnosis: the error line plus a few lines of surrounding context. Summarize everything else in one sentence.
3. One extraction, then confirm. State "the first real error is X" before fixing anything; if X is wrong, everything after it is wasted.
4. The log is data, not conversation. Keep the file on disk for re-querying; keep only conclusions in the chat.

Extraction patterns per log producer

Gradle (Android builds)

• Which task died: search > Task lines ending in FAILED.
• Why it died: search Caused by: and take the DEEPEST one; the first is usually a wrapper.
• FAILURE: Build failed with an exception marks the summary block; the lines after * What went wrong: are the human-readable core.

Xcode / xcodebuild (iOS builds)

• First error: (lowercase, with colon) is the anchor; warning: lines are noise for this purpose.
• The build step name above the first error names the pod or script phase that owns it.
• CocoaPods failures end with a clear final paragraph; read it before the backtrace above it.

Metro

• Resolution errors name the importing file and the missing module in the first two lines; the rest is stack.
• A red-box error pasted from the device usually contains the component stack; the first frame in YOUR code matters more than the framework frames above it.

Store submission tooling

• Search for ITMS- codes in Apple submission output; each code is precise and searchable.
• Play Console rejections arrive as structured reasons; quote the reason string, not the surrounding boilerplate.