Modules
Your LLM, with MCP tools and agent skills built in — the top-level API.
A2A (agent-to-agent) — OUTBOUND: call remote A2A agents (SPEC §7A).
Provider adapters — turn the uniform tool list into each LLM's tool schema (SPEC §0.7).
Sub-agents — the Level-1 surface over the agent runtime (SPEC §7D).
An agent definition — a name plus its spec (see Toolnexus.Agents.agent/2).
The runtime's injectable clock (SPEC §7D — lifecycle & runtime obligations).
Context compaction (SPEC §7F) as a before_llm helper — no core-loop change.
A live agent — one GenServer per handle (SPEC §7D).
Agent home — personas over the §7D runtime (SPEC §7E).
Handle to a running persona (see Toolnexus.Agents.Home.start_agent/3).
The agent runtime (SPEC §7D) as OTP.
The runtime's transition trace — one append-only list shared by every handle of a
runtime. §0 conformance for §7D is judged on these lines: identical per-handle
transition sequences against the shared examples/subagent-* fixtures.
The global turn gate (SPEC §7D backpressure gate 3) — a counting semaphore wrapped around the LLM HTTP call ONLY, never a whole Run (holding it across a Run deadlocks a parent against its children).
A §10 answer: {id, ok, data?} — keys byte-identical across ports.
Built-in tool source (source: "builtin") — SPEC §4A.
Unified LLM client — the host loop (SPEC §8, §10).
Where ask/4 conversations are remembered — two callbacks (SPEC §8 Conversation memory).
Default conversation provider — keeps transcripts in memory for the client's lifetime.
Full outcome + telemetry of one run/ask (SPEC §8, §10).
Execution context passed to every tool (SPEC §1).
HTTP / REST tools — declare a remote endpoint as a uniform Toolnexus.Tool (SPEC §7).
Dynamic MCP source (SPEC §2, §0.2–0.4). Mirrors js/src/mcp.ts: parse an
mcp.json (path | raw JSON string | map; top-level mcpServers|servers|mcp
accepted), connect to every enabled server (local stdio + remote
streamable-HTTP) with per-server failure isolation, list tools (cursor
pagination), and expose each as a uniform Toolnexus.Tool named
sanitize(server)_sanitize(tool).
One GenServer per configured MCP server (SPEC §2). Owns the transport
(stdio port or streamable-HTTP state), correlates JSON-RPC request ids with
per-call timeouts (config timeout, default 30000 ms), performs the
initialize handshake, paginates tools/list, executes tools/call, and
bridges inbound elicitation/create onto the host wait_for (§2 + §10):
form → kind:"input", URL → kind:"authorization"; no wait_for ⇒ decline.
§2 Gap 6: list-only inventory — original unprefixed tool defs per server.
Pure JSON-RPC 2.0 codec for the in-house MCP client (SPEC §2).
A loaded MCP source: tools, per-server status, and its own supervisor.
Incremental server-sent-events parser: feed chunks, collect complete events.
MCP-style streams (event:/data: lines, no [DONE] sentinel).
Stdio transport for a local MCP server (SPEC §2 "local").
Streamable-HTTP transport for a remote MCP server (SPEC §2 "remote"), built
on Req. POSTs JSON-RPC to the configured url with the configured headers
(after ${ENV_VAR} expansion — values are never logged), accepts either a
JSON response or an SSE stream, carries the Mcp-Session-Id response header
onto subsequent requests, and DELETEs the session on close.
MCP — INBOUND: serve a toolkit as an MCP server (SPEC §7C).
Native tools — turn a plain function into a uniform Toolnexus.Tool (SPEC §6).
A §10 suspension request. JSON keys are byte-identical across ports:
{id, kind, prompt, url?, data?, expiresAt?}.
A2A (agent-to-agent) — INBOUND: serve a toolkit as a remote A2A agent
(SPEC §7B), with the MCP profile co-mounted at /mcp (SPEC §7C).
File-backed store — one JSON file per Task id, under dir.
Handle returned by serve — the base URL + port plus stop/close.
Default store — Tasks live only for the lifetime of the process.
Pluggable Task persistence: get(store, id) and save(store, task).
Dynamic agent-skill source (SPEC §3). Mirrors js/src/skill.ts:
discover **/SKILL.md, parse frontmatter, and expose ONE skill tool that
loads a skill's instructions + sampled resources on demand (progressive
disclosure).
One discovered skill (SPEC §3).
The built skill source: skills + the single skill tool + system-prompt catalog.
The uniform tool: a named, described, JSON-Schema'd callable (SPEC §1).
Uniform tool result (SPEC §1). metadata.pending carrying a Request = suspension (§10).
Toolkit — aggregates dynamic MCP tools + the skill tool + built-ins + remote
A2A agents + custom tools into a single uniform list, with provider adapters
and a single execute/4 router (SPEC §4).
Single-turn translation — the translator path (SPEC.md §11, ADR-0011).
An OpenAI-shaped single-turn result (§11).
One tool call the model asked for, in OpenAI shape (§11).