Agent Enhancer Utilities

Add Agent Enhancer to an existing agent workflow for planning, checkpoints, deduplication, recovery, and evidence.

Service status · Effectiveness · 24-module catalog · Beta partner guide · Video demo · Glama verified · MIT license

Current release: skills v1.7.2 and backend 0.7.0. A fail-closed checkpoint proves one owner before a duplicate-sensitive external write and blocks blind replay after an uncertain result. All 24 modules remain free and real USDC is disabled. The separate exploratory diagnostic passed checkpoint adherence 10/10 and low-risk abstention 5/5; see its limitations and raw report.

Agent Enhancer is a reliability sidecar, not a replacement for the agent or domain tools you already use. Connect it to an existing workflow and call the smallest relevant utility only when coordination or failure handling will materially help.

_{Methodology, raw results, and limitations.}

Skills

Skill	Use it for
guard-external-plugin-workflows	Add honest reliability profiles, verification, and recovery around repeated or high-consequence work performed by other plugins
coordinate-parallel-agents	Locks, deduplication, one-use batons, semaphores, shared rate gates, barriers, and freshness leases
test-http-failure-paths	Bounded status/failure sequences, webhook delivery attempts, Retry-After behavior, and synthetic fixtures
debug-x402-integrations	Evidence-linked x402 error diagnosis, requirement drift, and facilitator compatibility checks
review-mcp-tool-contracts	Tool schema, annotation, capability-handshake, elicitation, and edge-case review
guard-x402-retries	Detect changed payment requirements and bind one retry identity to one normalized quote
measure-webhook-delivery	Collect one-use, hash-only evidence about bounded webhook delivery attempts

Every skill is a small Markdown workflow. Inspect its SKILL.md before installation.

Sidecar model

Agent Enhancer does not automatically intercept or extend another plugin. Agents deliberately combine it with domain tools when a workflow is repeated, parallel, scheduled, quota-sensitive, freshness-sensitive, or vulnerable to harmful duplicates.

The sidecar coordinates the workflow; the domain plugin still performs and verifies its own reads, writes, sends, and analysis. See the guard-external-plugin-workflows skill and the vendor-neutral Reliability Sidecar Contract v1 for the standard lifecycle and guarantee labels.

Install

List all seven skills before installing:

npx skills add artiehinz/Agent-Enhancer-Utilities --list

GitHub CLI 2.90 or later can preview or install one skill:

gh skill preview artiehinz/Agent-Enhancer-Utilities coordinate-parallel-agents
gh skill install artiehinz/Agent-Enhancer-Utilities coordinate-parallel-agents

For the lowest-overhead reliability path, install only the workflow guard:

gh skill install artiehinz/Agent-Enhancer-Utilities guard-external-plugin-workflows

This skills-first path has no persistent MCP connection. It selects locally, abstains without network access for ordinary one-time low-risk work, and calls the public HTTP planner only for risk-bearing work. The release also includes agent-enhancer-on-demand-skill.zip, which expands directly into a Codex repository's .agents/skills/ folder and deliberately omits .mcp.json and MCP dependency metadata.

The same repository includes versioned package entry points:

Host	Entry point	Current evidence
ChatGPT and Codex	.codex-plugin/plugin.json	Package validator passed
Claude Code	.claude-plugin/plugin.json	Claude validator passed
GitHub Copilot CLI	plugin.json	Isolated install found all seven skills
Gemini CLI	gemini-extension.json	Extension prepared; clean host install remains
Generic MCP clients	.mcp.json	Live Streamable HTTP contract passed

See the exact, deliberately conservative compatibility evidence. A prepared manifest is not a claim of marketplace approval.

Agents can also connect without installing a skill:

Surface	Endpoint	Shape
Progressive MCP (recommended)	https://liberated.site/mcp?profile=core	Search, describe, and invoke all 24 modules through three host-visible tools
Progressive MCP (full control plane)	https://liberated.site/mcp	The core tools plus capability-request and catalog-management tools
Claude	https://liberated.site/mcp/claude	37 direct action-specific tools on backend 0.6.0 and later
ChatGPT	https://liberated.site/mcp/chatgpt	The same release-dependent direct surface
HTTP/OpenAPI	https://liberated.site/v1/openapi.json	24 generated module contracts on backend 0.6.0 and later

All public modules are currently free. Real USDC settlement is disabled.

Runnable examples

The examples use Python's standard library, synthetic identifiers, and the public no-auth MCP endpoint:

python -B examples/reliability-sidecar/run.py
python -B examples/multi-agent-checkpoint/run.py
python -B examples/sidecar-benchmark/test_benchmark.py
python -B examples/sidecar-agent-benchmark/test_benchmark.py
python -B examples/on-demand-sidecar/run.py both
python -B examples/on-demand-agent-benchmark/test_benchmark.py
goose recipe validate examples/goose/agent-enhancer-reliability-sidecar.yaml

• The general sidecar example combines the live guard planner with a mock domain agent and proves one searchable record survives a replay.
• The multi-agent checkpoint example races two workers, admits one owner, records one synthetic action, and keeps external_proof: false.
• The Goose recipe adds the no-auth sidecar to a Goose task, stops ambiguous retries, and requires destination read-back before recording caller verification.
• The paired sidecar benchmark executes five deterministic with/without scenarios, publishes 20 pairs per scenario, and exposes model-token fields as unavailable rather than inferring savings.
• The metered Codex tier freezes the same five scenario classes, a condition-blind evaluator, real host token and latency capture, five excluded validation pairs, and 20 later publishable pairs. Both completed validations are published rather than hidden. Backend 0.6.5 reduced harmful events from six to zero and improved verified completion from 80% to 90%, but failed both connection-only overhead gates, so the publication sample remains blocked. The harness rejects app/MCP contamination and keeps cost null when the host supplies no defensible dollar rate.
• The skills-first on-demand sidecar makes no remote call for a locally classified low-risk contract and calls the existing hosted planner once for risk-bearing work. Local/hosted plan drift, raw identifiers, private fields, and unknown tools fail closed. Its new preregistered measured-agent validation remains separate from all persistent MCP results.
• The on-demand metered benchmark freezes five paired Codex scenarios, installs only one repo-scoped skill in the with condition, rejects every MCP call, measures activation and abstention, and blocks publication unless the complete five-pair validation passes every fixed gate. The separate 200-run publication sample passed its preregistered gates. A post-hoc run-level reanalysis found that confirmed-harm runs across the four risk scenarios fell from 14/80 to 1/80 and scenario acceptance rose from 66/80 to 79/80, with no unresolved outcomes in either condition. The older pooled counters fell from 26 to 2 but are secondary because duplicate and conflict events can occur in the same run. The strongest result remained overlapping workers, where verified completion improved from 8/20 to 19/20. Scheduled refresh improved from 18/20 to 20/20; ambiguous create and shared-rate runs were already 20/20 without the sidecar. Guarded risk-bearing runs used substantially more tokens and time. All 20 low-risk skill runs correctly abstained with zero adapter or remote calls; their lower observed token and latency totals are treated as host variance, not product savings. See the case study and run-level report.

Remote MCP examples search, describe, and invoke through the progressive facade. The on-demand prototype uses the public generated HTTP contract only after local selection. No example copies private backend code or schemas.

Directory listings

• Official MCP Registry — 0.7.0 latest
• Glama — ownership verified
• Smithery

These listings point to the live service. The implementation is not published in this skills repository.

Trust

The service publishes:

• current per-module self-test results and dataset freshness at /status;
• privacy-safe methodology and usage evidence at /effectiveness;
• reproducible run-level sidecar benchmark evidence in examples/sidecar-benchmark/results/latest.json;
• preregistered metered Codex evidence in examples/on-demand-agent-benchmark/results/publication-latest.json;
• schemas, limits, side effects, retention, and typed errors in the catalog;
• privacy, acceptable-use, security, support, and pricing policies from the service homepage.

The skills contain no hidden install telemetry. Do not send credentials, wallet private keys, personal data, customer records, or conversation history to the service.

Repository scope

This public repository contains the reusable skills, their host metadata, and project-level validation/documentation. The hosted service implementation is operated separately. A skill must use progressive discovery and must not copy private schemas, credentials, or backend source.

See the architecture, CONTRIBUTING.md, SECURITY.md, and effectiveness methodology. The product direction and planned cross-plugin reliability work are described in docs/ROADMAP.md. The proposed read-only planning contract is in docs/WORKFLOW_GUARD_PLANNER.md, the bounded checkpoint contract is in docs/OPAQUE_WORKFLOW_CHECKPOINTS.md, and current live primitive evidence is recorded in docs/SIDECAR_RECIPE_TESTS.md. The exact open-source contribution sequence and module map are in docs/OPEN_SOURCE_INTEGRATION_PLAN.md.

License

MIT