deep-dive

Invoke the deep-dive mode with a scope header specifying the target area, such as an auth module or payment flow. The protocol maps the scope, dispatches parallel explorer lanes, and returns a verified findings report without modifying any source code.

Five profiles are available: standard, security, ux, architecture, and full. Each profile activates a specific set of audit lanes, such as SECURITY_TRUST for security or UX_FLOW for user-facing behavior, with full enabling all eight lanes.

Explorers only emit candidate findings. Two parallel reviewers verify or reject each candidate, then HIGH and CRITICAL findings pass two sequential critic challenges covering false positives, root cause, and severity accuracy.

No. The protocol is strictly read-only: it never mutates source code, never delegates to a coder, and never creates files outside the .swarm/ directory. All analysis uses read-only repository tools.

Failed lanes trigger a coverage gate: the lane is retried up to two times with different parameters, then replaced with an equivalent dispatch if retries fail. If no equivalent exists, the audit reports INCOMPLETE rather than skipping coverage.

The scope map is capped at 50 files, each explorer mission handles at most 8 files and roughly 3500 lines, and reviewer shards hold at most 10 candidates each. Very large codebases should be split into multiple scoped audits.