Coding agent

A coding agent uses the same thread runtime as a support bot. The difference is its workspace, mode, and tool policy.

Define lazy tools

import {
  createFlueLazyTools,
  createFlueRequestUserInputTool,
  defineFlaryAgent,
} from "flary/flue";

export default defineFlaryAgent<Env>({
  resolveContext: ({ env, id }) => env.RUN_BINDINGS.read(id),

  resolveAgent: ({ trusted }) => ({
    agentId: "coding",
    revisionId: trusted.revisionId,
    instructions:
      "Inspect the bound workspace. Make small changes, run checks, " +
      "and report the exact files changed.",
    model: {
      provider: "openai",
      model: "gpt-5.6-codex",
      cacheRetention: "short",
    },
    thinkingLevel: "high",
    mode: "build",
  }),

  resolveModel: ({ env, agent, trusted }) =>
    env.MODELS.resolve(trusted, agent.model),

  resolveTools: ({ env, trusted }) => [
    ...createFlueLazyTools(env.TOOLS.lazyRuntime(trusted)),
    createFlueRequestUserInputTool(env.INPUT.forThread(trusted)),
  ],
});

The model sees a small fixed surface:

• tool_search
• tool_describe
• tool_call
• tool_batch
• request_user_input

Tool schemas load only after search or describe. tool_batch runs independent reads in parallel. Writes to the same resource run in order.

Register application tools

import { InMemoryToolCatalog } from "flary/tools";

export function createWorkspaceTools(workspace: Workspace) {
  const catalog = new InMemoryToolCatalog();

  catalog.register({
    definition: {
      id: "workspace.read_file",
      name: "Read file",
      description: "Read one file from the bound workspace.",
      kind: "native",
      operation: "read",
      capabilities: ["file.read"],
      tags: ["files", "read"],
      inputSchema: {
        type: "object",
        properties: { path: { type: "string" } },
        required: ["path"],
        additionalProperties: false,
      },
    },
    execute: ({ path }: { path: string }) => workspace.read(path),
  });

  catalog.register({
    definition: {
      id: "workspace.apply_patch",
      name: "Apply patch",
      description: "Apply one checked patch to the bound branch.",
      kind: "native",
      operation: "write",
      capabilities: ["file.write"],
      tags: ["files", "write"],
      requiresApproval: true,
    },
    resourceKey: "workspace",
    execute: (input: PatchInput) => workspace.applyPatch(input),
  });

  return catalog;
}

The host can register native tools and use createMcpTools() for authorized remote MCP servers. Raw credentials stay in trusted host code.

Modes

Flary includes ask, plan, build, and review modes. Modes are Zod-backed permission profiles, not separate runtimes.

import { resolveAgentMode } from "flary";

const plan = resolveAgentMode("plan");
const build = resolveAgentMode("build");

Plan Mode can write only plan artifacts. Build Mode can write the bound workspace, but risky operations can still require approval. A model cannot grant itself a stronger mode. See Modes and permissions for the complete profile shape, custom modes, thread storage, and recovery rules.