Tailrocks Improve Deep

Improve Deep

Own exhaustive non-security repository audit. Apply [runtime-trust.md](references/runtime-trust.md) and [repository-audit-lanes.md](references/repository-audit-lanes.md). This owner is read-only and returns one report; it never plans, implements, reconciles, or seeds delivery work. Resolve every relative link in this file against the directory containing this SKILL.md, never the plugin skills root.

No category means the whole-repository deep route. One accepted category means the focused deep route. --batch may accompany either form. The caller does not pass a redundant --deep flag: selecting this owner already selects deep audit.

Audit

1. Bind root, revision, dirty state, whole-repository or one category scope, and

every package in scope. The only focused categories accepted here are correctness, perf, tests, tech-debt, dependencies, dx, docs, direction, and agent-legibility. Refuse quick, security, platform-design categories (ux, tui, liquid-glass), branch review, unknown categories, and multiple primary categories. Name tailrocks-improve-security for security, tailrocks-web-design-audit for ux, tailrocks-tui-design-audit for tui, and tailrocks-macos-design-review for liquid-glass; report the route and stop, never dispatch it. --batch only makes this owner's selection non-interactive; it changes neither exhaustive coverage nor authority.

1. Map every applicable common lane over every package. No hotspot sampling,

leverage cutoff, silent package omission, or inferred execution authority. A target command runs only with explicit authority in an enforceably read-only tree using frozen inputs, scrubbed secrets, disabled network, owner-only external cache/output, bounded time/output/processes, TERM-then-KILL cleanup, and before/after hashes; otherwise record NOT_RUN.

1. Dispatch one bounded read-only investigator per independent lane/package

slice. Every lane returns candidates or an explicit skip with evidence.

1. The orchestrator reopens every citation. Then a fresh-context independent

verifier receives only the claim, target identity, and cited locations—not the raising lane's reasoning—and returns CONFIRMED or REFUTED with cause.

1. Report only twice-confirmed candidates. Rank by correctness, consistency,

goal fit, severity, confidence, and fix risk; effort stays metadata. Separate direction options from defects.

Output and final gate

Return one exhaustive report containing package/lane coverage, skipped scopes, both verification receipts per surviving finding, rejected claims, and next owners. No file, comment, plan, roadmap artifact, commit, or external action.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tailrocks
• Source: tailrocks/tailrocks-skills
• License: Apache-2.0
• Homepage: https://skills.tailrocks.com

Install and usage instructions live in the source repository linked above.