@openagentsinc/ai

The Effect-native OpenAgents AI SDK — an Effect toolkit for building agent applications with durable, cursor-exact streams. One neutral event union from the model call to the rendered message, suspend and continue that persists, coding-agent harnesses, redaction as a schema field, and recall instead of compaction.

License: Apache-2.0
Registry: npm scope @openagentsinc/* (pre-stable under dist-tag rc only)
Gates: local only — no GitHub Actions (pnpm run check, .githooks/pre-push)

Layers

L6  RECALL        @openagentsinc/history-corpus
L5  UI STREAM     agent-harness-contract: ui-message-chunk, reducer,
                  smooth-stream, partial-object-stream, ChatTransport
L4  HARNESS       agent-harness-contract: AgentHarness, session verbs,
                  slice runner, readiness, toolkit bridge, ACP + opencode
L3  SANDBOX       sandbox-provider contract + local + interop providers
L2  DURABLE LOG   event-log: seq-cursor append, replay, attach, rerun
L1  VOCABULARY    @openagentsinc/agent-runtime-schema (KhalaRuntimeEvent)
L0  MODEL CALL    effect/unstable/ai (upstream) + @openagentsinc/ai-model
P   PROGRAMS      @openagentsinc/dse (typed signatures, immutable artifacts,
                  portable runtime; optimizer is an explicit subpath)
G   GRAPH         @openagentsinc/graph-corpus (derived RLM/DSE projection;
                  graph truth does not add a runtime event layer)
X   CONFORMANCE   @openagentsinc/conformance-kit (cross-cutting test laws)

Every layer speaks KhalaRuntimeEvent upward. One event union. One durable cursor.

Packages

Package	Layer
@openagentsinc/ai	umbrella — curated re-exports
@openagentsinc/agent-runtime-schema	L1 vocabulary
@openagentsinc/agent-harness-contract	L2–L5 core
@openagentsinc/ai-model	L0 model-call bridge
@openagentsinc/history-corpus	L6 recall
@openagentsinc/rlm	L6 recursive engine
@openagentsinc/dse	typed model programs
@openagentsinc/graph-corpus	derived graph corpus
@openagentsinc/conformance-kit	reusable contract laws
@openagentsinc/ai-sdk-sandbox-local	L3 interop
@openagentsinc/ai-sdk-sandbox-openagents	L3 interop

Supported harnesses

Every harness implements the one L4 AgentHarness contract and emits KhalaRuntimeEvent upward:

Harness	Factory	Adapter source
Claude Code	makeClaudeCodeHarnessAdapter	claude-code-adapter.ts
Codex	makeCodexHarnessAdapter	codex-adapter.ts
OpenCode	makeOpencodeAdapter	opencode-adapter.ts
Pi	makePiHarnessAdapter	pi-adapter.ts
Goose	makeGooseHarnessAdapter	goose-adapter.ts
Cursor (ACP)	makeCursorHarnessAdapter	cursor-adapter.ts
Grok CLI (ACP)	makeAcpHarnessAdapter	acp-adapter.ts

The generic ACP adapter serves any Agent Client Protocol peer; Grok CLI (grok_acp) uses it directly and the Cursor adapter (cursor_acp) builds on it.

RLM

The SDK ships first-class Recursive Language Model support: recall instead of compaction. @openagentsinc/rlm (public entry point @openagentsinc/ai/rlm) answers bounded, cited questions over a durable, redaction-aware corpus without putting the corpus into a model context window — a zero-model-call deterministic tier plus an Effect-native engine for declarative model/RLM recursion, under global budgets with exact-usage honesty and citation validation, exposed as one typed service and an Effect Tool adapter for harness consumption. An RLM result is evidence, not authority: a completed answer is a cited candidate, never an authorization. Spec, architecture, and conformance gates: docs/rlm/.

DSE Programs

The SDK uses DSE as its one typed model-program contract. A Program<I, O> binds an Effect Schema signature to one immutable candidate artifact. Runtime prediction performs bounded Schema repair and emits receipts. The event adapter projects prediction lifecycle and honest usage onto KhalaRuntimeEvent. Offline compile and promotion stay on the explicit @openagentsinc/dse/optimizer subpath. Contract and API docs: docs/dse/.

Train policy

The roster ships as one train: every publishable package carries the same version, and every inter-package dependency inside the workspace points at that one train through workspace:* (rewritten to the concrete version at pack time). Converge the whole roster to one train in a single command:

node scripts/set-train-version.ts 0.2.0-rc.2         # bump every package + fix inter-package ranges
node scripts/set-train-version.ts 0.2.0-rc.2 --check # verify convergence, no writes
pnpm run audit:surface:update                        # record the new train's API-surface baseline

Pre-stable trains publish under dist-tag rc only — never latest. Publish leaf packages first and the @openagentsinc/ai umbrella last, all at the one converged version.

Roadmap

The proposed engine roadmap is docs/ROADMAP.md — conformance-as-product, Programs (the DSPy of Effect), RLM deepening, generative UI, transport productization, provider gateway, harness breadth.

Development

Node 24, pnpm, Vite Plus — the same toolchain contract as the openagents monorepo.

pnpm install
pnpm run check
pnpm run hooks:install   # enables .githooks/pre-push on this clone

Docs index: docs/README.md.

Install (consumers)

npm install @openagentsinc/ai@rc