Steering Audit & Optimize

Audit agent configuration surfaces for issues, then apply R1-R7 rewrites to fix them.

Phase 1 -- Audit (find issues)

1. Lint scan: Run agnix --show-fixes on config directories. Categorize findings as real errors, false positives, or auto-fixable.
2. Hook efficiency: Count hooks per event type. Flag unconditional Bash hooks, duplicate references, prompt-type hooks doing pure string checks.
3. Duplication scan: Cross-reference AGENTS.md, rules, skills, and agent
   definitions. Flag policies stated in 2+ places, and rules that tools enforce
   mechanically. The native <domain-rules> listing already renders every rule
   name + description; always-apply indexes that duplicate it are waste.
4. Stale file detection: Empty rule files, unreferenced agent files, outdated
   memory entries, empty directories.
5. Token budget: Identify always-apply rules that should be rulebook or TTSR
   instead. Flag files exceeding 5KB. Suggest lazy-loading candidates.
6. Bootstrap leakage: Global/bootstrap skills that are present as project-local
   copies when they should remain global.
7. AGENTS.md minimality: Root AGENTS.md should stay a thin context file;
   scoped AGENTS.md files should be path-specific.
8. Rule buckets: Sort each rule as alwaysApply, rulebook (globs / on-demand),
   or TTSR (condition + scope). Do not dump large global content into alwaysApply.

Phase 2 -- Optimize (fix issues)

Apply R1-R7 rewrites to files flagged in Phase 1. LOAD skill://audit-steering/references/rules.md for the exact checks.

Rule	What
R1	description in YAML frontmatter on every file
R2	Imperative tone, no model names, frame as actions
R3	Tables for mappings, bullets for rules
R4	write-agentic's templates + lint.py own the per-kind format contract
R5	Relative paths for files, backticks for skill/agent names
R6	lint.py's per-kind line caps; split oversized files
R7	Index files as routing tables, detail in referenced files

Measure before and after: LOAD skill://audit-steering/references/measurement.md for token estimation and report format.

Output Format

• Summary line: X errors, Y warnings, Z suggestions
• Section per check with findings (file, severity, description, fix)
• Priority actions sorted by impact
• No prose filler

Steering

• Prefer enforceable automation for mechanical policy.
• Prefer short guidance for judgment-heavy policy.
• Focus findings on real drift, not stylistic preferences.

References

File	When to load
skill://audit-steering/references/checklist.md	Phase 1: full audit pass
skill://audit-steering/references/rules.md	Phase 2: R1-R7 checks
skill://audit-steering/references/measurement.md	Phase 2: token estimation and report format