Blindspot Pass

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

<HARD-RULE>Read-only. Do not edit, create, or delete any project file. No fixes, however tempting.</HARD-RULE>

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.

1. Name the domain, derive its dangerous paths. Before scanning, state in

one line what kind of system the target is, and list the 2–3 failure classes that domain actually bleeds from. The generic list below is a backend-shaped floor, not a ceiling — a native UI app bleeds from accessibility gaps, lifecycle/identity bugs, and state-restoration loss; a data pipeline from loss, duplication, and ordering; a CLI from argument edge cases and exit codes. Findings should come from the derived list first.

1. 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 for the domain-derived risks from step 2, plus:

• 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• Stale assumptions: config or constants that encode outdated facts, comments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 --no-merges -- <path>, ignoring generated files: lockfiles, *.pbxproj, build outputs — their churn is noise)

1. 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.