blindspot

Blindspot Pass

Find the unknowns in a system BEFORE touching it. Reconnaissance only.

Read-only. Do not edit, create, or delete any project file. No fixes, however tempting.

Tip for the user (mention once at the start): running this pass in plan mode or a read-only permission mode makes the rule enforced by the harness, not just promised by the model.

Process

1. Identify the target. Use the argument or infer the directory/system/feature from conversation. If no target is discernible, ask for one — don't guess.
2. Explore. Use the agent's file-search/read tools (Glob/Grep/Read, rg/find/cat, or equivalent) and git log for churn hotspots to understand the target. Look specifically for:
   • Hidden coupling: modules that import each other's internals, shared mutable state, implicit ordering dependencies
   • Unowned edge cases: error paths that swallow exceptions, TODO/FIXME/HACK comments, empty catch blocks
   • Stale assumptions: config or constants that encode outdated facts, comments contradicting code, dead feature flags
   • Missing tests around dangerous paths: money, auth, deletion, migrations, concurrency — anything irreversible with no test coverage
   • Churn hotspots: files with many recent fixes (git log --oneline -- <path>)
3. Report up to 7 findings, ranked by risk (highest first). If fewer than 5 real risks are found, say so instead of padding.

Output format (per finding)

### N. <one-line risk statement>  [risk: high|medium|low]
Why it matters: <1-2 sentences, concrete failure scenario>
Investigate: `<ready-to-paste follow-up prompt the user can run next>`

End with a one-line recommendation of which finding to chase first.