Buff Validate

buff-validate v2

Project-wide quality engine with cross-file coherence and trending. Performs high-sensitivity static analysis across five dimensions: correctness, security, complexity, style, and coherence. Reports findings with severity and actionable fixes. Never a blunt "rewrite everything" — precise, targeted, ranked.

Analysis dimensions

1. Correctness

Bugs, logic errors, and runtime risks:

• Null/undefined access without guards
• Off-by-one errors in loops and array access
• Unhandled async errors (missing await, unhandled Promise rejection)
• Incorrect comparison operators (== vs ===, is vs == in Python)
• Type mismatches at function boundaries
• Missing return statements in non-void functions
• Race conditions in async code

2. Security

OWASP-aligned security checks:

• SQL injection via string concatenation
• XSS via unsanitized innerHTML or dangerouslySetInnerHTML
• Hardcoded secrets, API keys, passwords in source
• Unvalidated user input reaching filesystem, shell, or database
• Insecure deserialization
• Missing authentication/authorization checks on routes
• Sensitive data logged to console

3. Complexity

Maintainability signals:

• Functions > 50 lines (split candidates)
• Cyclomatic complexity > 10 (too many branches)
• Nesting depth > 4 levels (extract logic)
• Files > 300 lines (split candidates)
• Duplicated logic (3+ similar blocks = extract function)
• Parameter lists > 5 arguments (use object/struct)

4. Style

Consistency and clarity:

• Naming violations (detected vs. project convention from Counterfeit)
• Dead code (unreachable branches, unused variables/imports)
• Magic numbers/strings without named constants
• Inconsistent error handling patterns within the same file

5. Coherence

Cross-file analysis — structural integrity across the project:

Scope resolution

Flags combine: /buff:validate src/ --fix --high-only scans directory, fixes only high findings.

--quick mode

Graph-guided smart scan. Uses graph.json to prioritize what to scan:

1. Changed files — files changed since last validate (git diff)
2. Dependents of changed files — from graph.json imported_by fields
3. Files with prior HIGH findings — from learnings.json past scan results
4. High-complexity files — from graph.json lines count

Skip unchanged, no-dependent, recently-clean files. Cuts scan time 60-80% while catching the findings that matter. Ideal for mid-session checks.

Counterfeit pre-read

Before scanning, read these files to inform analysis:

• ~/.buff/memory/patterns.json — known issue patterns (avoid false positives, catch known-bad patterns)
• ~/.buff/memory/corrections.json — past mistakes (what was flagged incorrectly before, what was missed)
• .buff/memory/graph.json — structural data for coherence checks (imports, exports, dependency chains)

If any file is missing or empty, skip it silently — never fail on missing counterfeit data.

Workflow

1. Pre-read counterfeit — load patterns, corrections, graph data
2. Determine scope — project-wide (default), file, directory, --session, or --quick
3. Read each file in scope
4. Run analysis across all five dimensions (coherence only runs when graph.json exists)
5. Rank findings by severity
6. Report findings table
7. Write scan results to counterfeit (learnings.json)
8. If --fix flag: apply fixes for high-severity findings automatically

Quality trending

After each scan, write a scan record to .buff/memory/learnings.json:

{
  "type": "validate_scan",
  "date": "2026-03-21",
  "scope": "project-wide",
  "findings": { "high": 0, "med": 3, "low": 7 },
  "coherence": { "broken_imports": 0, "circular_deps": 1, "orphaned_exports": 4 },
  "trend": "improving"
}

Trend calculation: Compare current scan to the last 3 scans in learnings.json:

• improving — high findings decreased or total findings decreased by 20%+
• stable — findings within 10% of previous average
• degrading — high findings increased or total findings increased by 20%+

If no prior scans exist, trend = "baseline".

Output format

## Validation — [scope] · [N files, N lines analyzed]

| File | Line | Dimension | Severity | Finding | Fix |
|---|---|---|---|---|---|
| auth.ts | L42 | Security | HIGH | Raw SQL string concat | Use parameterized query |
| auth.ts | L87 | Correctness | HIGH | Missing await on async call | Add await |
| router.ts | L12 | Complexity | MED | Function is 73 lines | Extract validatePermissions() |
| — | — | Coherence | MED | Circular dep: auth→db→auth | Break cycle via interface |
| utils.ts | L5 | Style | LOW | Unused import 'path' | Remove import |
| — | — | Coherence | LOW | 4 orphaned exports | Remove or document intent |

Summary: 2 high · 1 medium · 3 low
Trend: improving (was 3 high · 2 med · 5 low last scan)
Coherence: 1 circular deps · 4 orphaned exports
Auto-fixable: 1 of 2 high findings

Always sort: HIGH first, then MED, then LOW. Within each level, sort by file then line. Coherence findings (no specific line) go at the end of their severity group.

Auto-fix behavior (--fix)

Apply fixes only for:

• Unused imports — remove
• Dead code — remove
• Missing await on clearly async calls — add
• Hardcoded magic numbers with obvious names — extract to const

Never auto-fix:

• Security findings (require understanding of data flow — show fix, don't apply)
• Complex refactors
• Anything that changes behavior, not just style
• Coherence findings (require architectural decisions)

For every auto-fix applied, print:

Fixed: [file]:[line] — [what was fixed]

Integration with buff loop

When called from /buff:execute (validate phase):

• Scope = all files in .buff/session.log modified this session
• High findings block the execute loop — must be resolved before commit
• Med/low findings are reported but do not block
• Results appended to .buff/session.log

Severity definitions

Writes to counterfeit

After scanning, write results to counterfeit memory:

• Scan results → .buff/memory/learnings.json — full scan record with findings counts, coherence stats, and trend
• Recurring patterns → ~/.buff/memory/corrections.json — if a pattern appears 3+ times across scans, record it as a generalizable correction
• File stats → .buff/memory/graph.json — update file-level stats (lines, complexity, last-validated timestamp)

Correction detection

Watch for the user disagreeing with findings (rejection, replacement, redirect). Follow the correction capture mechanism defined in counterfeit SKILL.md — detect signals, record to learnings.json + corrections.json, promote to patterns.json at 2+ occurrences. Use skill: "validate" when writing correction entries. Learn from corrections to reduce false positives over time.