Justin Bounded Investigation
Keep a multi-file investigation bounded by delegating broad discovery and recording the evidence needed to verify the conclusion. Targeted reads and narrow commands are assumed; they are not what this skill adds.
When to Use
Invoke this when the work is investigation and the relevant code is not yet located, such as:
- a stack trace or error with no obvious owning file
- a "why does X happen" question naming a behavior but no source
- an audit or archaeology pass that will span many files
Skip it once the relevant files are identified, or for ordinary edits to known code.
Delegate Fan-Out Searches
Delegate a sweep to a read-only search agent when the harness provides one, delegation is authorized, and both conditions hold:
- the location is unknown — many candidate files, uncertain naming, or a whole-repository code search
- you need the conclusion, not the surrounding code
Run the search yourself when the target is specific enough that the output is small, or when judging the surrounding code matters more than finding it. Keep synthesis and verification in the primary agent: inspect the cited source before relying on a delegated conclusion.
Keep an Evidence Ledger
Keep the ledger in your working notes or a scratch file for the session; it is not committed. Record one line per finding: the source, the observation, and what it rules in or out.
src/Session/Store.swift:88 — writes cache off the main actor — explains the
crash, rules out the decoder path
Record ruled-out branches and sources that were checked and added nothing. Without them, the same file gets read again later in the same investigation.
The ledger is the raw material for a handoff: its lines become the Decisions
and Touched sections of jww-handoff, so keeping it makes that handoff nearly
free to write.
Report
Lead with the conclusion, then the evidence that supports it and the caveats that affect trusting it. Cite by file and line so the reasoning can be rechecked without reproducing the search.