flowstate-agents-llm-client Package Skill

Package: @epicdm/flowstate-agents-llm-client Repository: epic-flowstate-community Path: packages/flowstate-agents-llm-client Version: 1.0.7

Use this published Dojo skill as the package-level entrypoint. For detailed workflows, use the package-local skills listed below.

Primary Package Workflow

FlowState Agents LLM Client

Status: Active Purpose: Help agents use @epicdm/flowstate-agents-llm-client without trusting stale docs or generated feature lists. Scope: /Users/sthornock/code/epic/epic-flowstate-community/packages/flowstate-agents-llm-client Trigger: LLM provider setup, single-shot query calls, streaming output, memory/knowledge prompt composition, package maintenance, or Dojo/docs repair. Input: Provider/model config, prompt params, optional AMS/knowledge-store clients, package source/docs. Output: Correct LLMClient usage and source-backed package skill routing.

Overview

This package wraps the Vercel AI SDK behind LLMClient. The implemented center is query() and stream(). Some public config fields are reserved or only partially wired, so package skills must distinguish working behavior from typed-but-unimplemented surface area.

Route by Task

Minimum Happy Path

import { LLMClient } from '@epicdm/flowstate-agents-llm-client'

const client = new LLMClient({
  provider: 'anthropic',
  model: 'claude-3-5-sonnet-20241022',
  apiKey: process.env.ANTHROPIC_API_KEY,
})

const result = await client.query({
  prompt: 'Summarize this package skill review.',
  systemPrompt: 'Return concise engineering prose.',
})

if (!result.success) {
  throw result.error
}

console.log(result.content)

Source Authority

Prefer these files over generated Dojo content:

• src/llm-client.ts
• src/providers/*.ts
• src/types/*.ts
• src/utils/*.ts
• tests/memory-prompt-integration.test.ts
• .flowstate/docs/api/index.md

Critical Pitfalls

• createLLMProxy() is exported but throws.
• Budget fields are typed but not enforced.
• responseFormat and schema are typed but not implemented as structured output validation.
• Knowledge-store filters like knowledgeTypes and minImportance are typed but ignored.
• README still contains stale package scope and model examples.

Verification

yarn workspace @epicdm/flowstate-agents-llm-client test
yarn workspace @epicdm/flowstate-agents-llm-client typecheck
yarn workspace @epicdm/flowstate-agents-llm-client build

Done When

• The correct workflow skill has been used.
• Source behavior, not generated Dojo prose, backs the implementation.
• Reserved/partial config is not documented as operational.
• Downstream coordinator callers are checked for public behavior changes.

Package Skill Inventory

Package Skill Inventory: @epicdm/flowstate-agents-llm-client

Package Context

Package path: /Users/sthornock/code/epic/epic-flowstate-community/packages/flowstate-agents-llm-client

This package centers on LLMClient. Its skills should teach agents which parts of the config surface are implemented, how to call query() and stream(), how provider registry and cost calculation work, and how memory/knowledge composition behaves.

Existing .flowstate/dojo files are generated/prototype artifacts and should not be published as-is.

Skill Map

Anti-Sprawl Decision

Seven skills is appropriate. The package has one main class, but it has distinct operational tasks: provider setup, query, streaming, memory/knowledge composition, troubleshooting contract drift, and maintenance. Creating one skill per config field or provider would reproduce the bad generated pattern.

Feature Coverage

Composition Model

Start with flowstate-agents-llm-client. Use provider registry for config/model/pricing edits, query workflow for single-shot calls, streaming tools for async generator callers, memory/knowledge composition for stateful agent prompts, and troubleshooting before trusting README or generated Dojo claims. Consult flowstate-agents-knowledge-store and the future memory-client package skill when composing prompt context.

Deferred Skills

Review Questions

• Should includeAMSContext and saveToAMS become per-call controls or be removed from docs?
• Should knowledge filters pass into KnowledgeStoreClient.retrieve()?
• Should retry honor retryableErrors?
• Should query() return AI SDK tool calls in LLMResult.toolCalls?
• Should README be migrated from @epic-flow/* to @epicdm/* before Dojo publication?

Review Gate

Package Skills Review: @epicdm/flowstate-agents-llm-client

Verdict

Pass with fixes.

The new local skillset teaches actual package workflows and explicitly rejects generated Dojo claims around unimplemented budget, proxy, schema, and filter behavior. It should replace the generated feature-list content as the source for future Dojo artifacts.

Findings

Coverage

Drift and Contract Risks

Dojo Publication Decision

Do not publish existing generated Dojo files. The reviewed local skills are eligible to become Dojo source content after package verification commands pass and manifests are rebuilt.

Follow-Up Work

• Package verification passed:
  • yarn workspace @epicdm/flowstate-agents-llm-client test
  • yarn workspace @epicdm/flowstate-agents-llm-client typecheck
  • yarn workspace @epicdm/flowstate-agents-llm-client build
• Rebuild .flowstate/dojo from reviewed local skills.
• Update README/docs drift.
• Decide whether to implement or remove reserved config fields.

Start with this skill to understand the package purpose, then use the package-local focused skills for specific workflows. Keep this Dojo version aligned with packages/flowstate-agents-llm-client/package.json. Do not treat generated feature lists as lessons; use workflow evidence from CODE-REVIEW.md, SKILL-INVENTORY.md, and REVIEW.md.