rhachet-brains-anthropic

rhachet brain.atom and brain.repl adapter for anthropic

install

npm install rhachet-brains-anthropic

usage

import { genBrainAtom, genBrainRepl } from 'rhachet-brains-anthropic';
import { z } from 'zod';

// create a brain atom for direct model inference
const brainAtom = genBrainAtom({ slug: 'claude/sonnet' });

// simple string output
const { output: explanation } = await brainAtom.ask({
  role: { briefs: [] },
  prompt: 'explain this code',
  schema: { output: z.string() },
});

// structured object output
const { output: { summary, issues } } = await brainAtom.ask({
  role: { briefs: [] },
  prompt: 'analyze this code',
  schema: { output: z.object({ summary: z.string(), issues: z.array(z.string()) }) },
});

// create a brain repl for agentic tasks
const brainRepl = genBrainRepl({ slug: 'claude/code' });

// use ask() for read-only operations
const { output: { analysis } } = await brainRepl.ask({
  role: { briefs: [] },
  prompt: 'analyze this codebase',
  schema: { output: z.object({ analysis: z.string() }) },
});

// use act() for read+write operations
const { output: { proposal } } = await brainRepl.act({
  role: { briefs: [] },
  prompt: 'refactor this module',
  schema: { output: z.object({ proposal: z.string() }) },
});

available brains

atoms (via genBrainAtom)

stateless inference without tool use. uses anthropic messages api with structured outputs.

| slug | model | cost ($/MTok) | cutoff | description | | --- | --- | --- | --- | --- | | claude/haiku | claude-haiku-4-5-20251001 | $1 / $5 | 2025-04 | fastest and most cost-effective | | claude/haiku/v3.5 | claude-3-5-haiku-20241022 | $0.80 / $4 | 2024-04 | fast and cost-effective | | claude/haiku/v4.5 | claude-haiku-4-5-20251001 | $1 / $5 | 2025-04 | fastest and most cost-effective | | claude/sonnet | claude-sonnet-4-5-20250929 | $3 / $15 | 2025-04 | balanced performance and capability | | claude/sonnet/v4 | claude-sonnet-4-20250514 | $3 / $15 | 2025-04 | balanced performance and capability | | claude/sonnet/v4.5 | claude-sonnet-4-5-20250929 | $3 / $15 | 2025-04 | balanced performance and capability | | claude/opus | claude-opus-4-5-20251101 | $5 / $25 | 2025-05 | most capable for complex reasoning | | claude/opus/v4 | claude-opus-4-20250514 | $15 / $75 | 2025-04 | highly capable for complex reasoning | | claude/opus/v4.5 | claude-opus-4-5-20251101 | $5 / $25 | 2025-05 | most capable for complex reasoning |

repls (via genBrainRepl)

agentic code assistant with tool use via claude-agent-sdk. repl slugs map to atom configs.

| slug | atom | cost ($/MTok) | cutoff | description | | --- | --- | --- | --- | --- | | claude/code | claude/sonnet | $3 / $15 | 2025-04 | balanced agentic capability | | claude/code/haiku | claude/haiku | $1 / $5 | 2025-04 | fast and cost-effective agent | | claude/code/haiku/v4.5 | claude/haiku/v4.5 | $1 / $5 | 2025-04 | fast and cost-effective agent | | claude/code/sonnet | claude/sonnet | $3 / $15 | 2025-04 | balanced agentic capability | | claude/code/sonnet/v4 | claude/sonnet/v4 | $3 / $15 | 2025-04 | balanced agentic capability | | claude/code/sonnet/v4.5 | claude/sonnet/v4.5 | $3 / $15 | 2025-04 | balanced agentic capability | | claude/code/opus | claude/opus | $5 / $25 | 2025-05 | most capable agent | | claude/code/opus/v4.5 | claude/opus/v4.5 | $5 / $25 | 2025-05 | most capable agent |

episode continuation

rhachet supports multi-turn conversations via episode continuation. each brain output includes an episode that can be passed back to continue the conversation.

atoms (supported)

atoms support episode continuation for cross-supplier workflows. prior exchanges are injected as actual user/assistant messages.

const brainAtom = genBrainAtom({ slug: 'claude/sonnet' });

// first turn
const resultFirst = await brainAtom.ask({
  role: {},
  prompt: 'remember this code: MANGO77',
  schema: { output: z.object({ content: z.string() }) },
});

// continue the conversation
const resultSecond = await brainAtom.ask({
  on: { episode: resultFirst.episode },
  role: {},
  prompt: 'what was the code i told you to remember?',
  schema: { output: z.object({ content: z.string() }) },
});
// resultSecond.output.content contains "MANGO77"

limitations:

• haiku does not support continuation with structured outputs (will throw BadRequestError)
• use sonnet or opus for continuation workflows

repls (not supported)

repls do not support episode continuation due to claude-agent-sdk limitations:

1. session resumption doesn't work with structured outputs (returns plain text)
2. cross-supplier continuation requires message injection which the sdk doesn't support

const brainRepl = genBrainRepl({ slug: 'claude/code' });

// this will throw BadRequestError
await brainRepl.ask({
  on: { episode: someEpisode }, // ❌ not supported
  role: {},
  prompt: 'continue...',
  schema: { output: z.object({ content: z.string() }) },
});

repls still export episode/series data for tracking and audit purposes. the episode.exid contains session info in the format anthropic/claude-agent-sdk/{machineHash}/{sessionId}.

summary

| brain | continuation | notes | | --- | --- | --- | | atom (sonnet, opus) | ✅ supported | use on: { episode } to continue | | atom (haiku) | ❌ not supported | throws BadRequestError | | repl (all) | ❌ not supported | throws BadRequestError |

for workflows requiring continuation, use genBrainAtom with sonnet or opus.

sources

• anthropic api documentation
• claude-agent-sdk documentation