AI Agent Linter Ecosystem: Quality, Cost and Safety Control

Tue, 24 Mar 2026 00:00:00 +0000

The Problem

AI agents write code fast, but speed comes at a cost. Literally - tokens cost money. And figuratively - code quality degrades without oversight.

I hit three problems at once:

Architecture degrades - agents write working code that violates layer boundaries, creates circular dependencies and god classes
Costs grow - complex prompts go to Opus at $15/1M tokens when Haiku at $0.80 would suffice
Content isn’t filtered - production prompts contain things that shouldn’t be there

Each problem has its own tool. But when the tools work together, they amplify each other.

The Pipeline

Every request passes through a chain of linters, each responsible for its domain:

seclint - checks content safety (6+/12+/16+/18+ ratings)
promptlint - scores complexity, picks the model
Agent - executes the task on the selected model
archlint - validates the result for architecture violations
costlint - records cost, tracks cache hit rate

If archlint rejects the result, the task escalates to a more powerful model. costlint records the escalation cost.

The Tools

archlint - Architecture Linter

Scans Go projects for structural violations.

1

archlint scan --format json ./project/

What it finds:

Layer violations (handler calls repository directly)
God classes (>20 methods or >15 dependencies)
Circular dependencies
Interface Segregation violations (interfaces >5 methods)

Metrics: fan-out, coupling (Ca/Ce), component and link counts.

GitHub: mshogin/archlint

promptlint - Complexity-Based Router

Scores prompt complexity and picks the model. No LLM, pure metrics, <10ms.

1
2
3
4
5


echo "Fix typo in README" | promptlint analyze --output-model
# haiku

echo "Design microservices with CQRS" | promptlint analyze --output-model
# opus

Signals: length, sentence count, domain keywords, action type (fix/create/refactor), code presence.

Integrates with ccproxy for real Claude Code request routing through a proxy.

GitHub: mikeshogin/promptlint

costlint - Token Cost Analysis

Tracks spending, analyzes caching, runs A/B tests between models.

1

costlint report --source telemetry.jsonl

1
2
3
4
5
6
7


Cost Report:
 Total requests: 342
 By model:
 opus: 48 requests, ~$12.60
 sonnet: 195 requests, ~$6.50
 haiku: 99 requests, ~$0.44
 With optimal routing: ~$8.20 (savings: 58%)

Cache metrics: hit rate, block reuse, content entropy, Jaccard similarity. A/B testing: 30/30/40 traffic split, per-group cost and quality metrics.

GitHub: mikeshogin/costlint

seclint / promptsec - Content Filter

Age ratings for prompts: 6+, 12+, 16+, 18+.

1
2
3
4
5


echo "Help with math homework" | seclint rate
# {"rating": "6+", "safe": true}

echo "Explain SQL injection for security course" | seclint rate
# {"rating": "16+", "safe": true, "flags": ["security_tools"]}

Considers educational context - explaining SQL injection for a security course gets 16+ instead of 18+.

GitHub: mikeshogin/seclint / mikeshogin/promptsec

Shared Principles

All tools follow the same rules:

Go - single stack, single build
No LLM - pure metrics, regex, keyword matching. <10ms per request
CLI + HTTP - each tool works as a command and as a server
JSONL telemetry - unified log format for analysis
Pipeline-friendly - exit codes, stdout, pipes

Numbers

On test workload (342 requests over a week):

Metric	Before	After
Token spending	$19.54	$8.20
Architecture violations	63	12
Requests to expensive model	100% opus	14% opus
Routing latency	-	<10ms

58% cost savings without quality loss - simple tasks go to Haiku, architecture tasks stay on Opus.

Orchestration

The tools work autonomously, but deliver maximum impact together. For orchestration we use myhome - daemon-based AI agent management with workflow stages and scheduled tasks.

The stack:

myhome launches agents
promptlint picks the model (pre-route hook)
archlint validates output (quality gate stage)
costlint tracks cost (telemetry consumer)
seclint filters incoming prompts (pre-filter)

Get Started

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15


# Install all tools
go install github.com/mshogin/archlint@latest
go install github.com/mikeshogin/promptlint/cmd/promptlint@latest
go install github.com/mikeshogin/costlint/cmd/costlint@latest
go install github.com/mikeshogin/seclint/cmd/seclint@latest

# Start routing
promptlint serve 8090 &
seclint serve 8091 &

# Check architecture
archlint scan --format json ./my-project/

# View costs
costlint report --source ~/.promptlint/telemetry.jsonl

Cost-Optimization on Mikhail Shogin