@intentius/chant-lexicon-fountain

fountain lexicon plugin for chant.

fountain runs coding agents in sandboxed VMs. This lexicon declares its workload layer as typed chant resources: Environment (sandbox baseline), Vault (env-var overrides), Agent (a runnable agent config). Conversations are runs, not resources — start them with the fountainRun op.

chant build serializes to fountain's own manifest YAML (fountain apply -f accepts it verbatim) — the same YAML fountainApply sends to fountain's bulk POST /api/apply endpoint in one request.

import { Environment, Agent } from "@intentius/chant-lexicon-fountain";

export const env = new Environment({
  name: "team-env",
  networking_type: "limited",                    // FTN010 requires explicit intent
  networking_config: { allowed_hosts: ["github.com"] },
  metadata: { "managed-by": "chant" },           // enables owned-only reconcile/prune
});

export const helper = new Agent({
  name: "helper",
  model: "anthropic/claude-sonnet-4-6",
  runtime: "claude",
  environment: env,                              // typed ref — dangling name = build error
});

For agents handling anything sensitive, ConciergeStack bundles the locked-down defaults (deny-all egress, no vault overrides, ownership marker on both resources) so loosening any of them is a visible, reviewable parameter.

Development

just generate    # regenerate types from the upstream OpenAPI spec
just validate    # check the generated artifacts
just docs        # build and serve the docs site

Generation is pinned to an upstream release artifact — FOUNTAIN_SPEC_VERSION in src/spec/fetch.ts. The same pin produces the same types on any machine, so a regeneration that changes the surface means upstream moved, not that someone's laptop was online.

Without a network it reads the committed snapshot (src/spec/fountain-openapi.snapshot.json) instead. Either way it prints which one it used and at what version.

chant dev coverage compares the generated surface against that spec and reports properties upstream has added, plus the request schemas deliberately left unmodeled.

Project structure

• src/plugin.ts — LexiconPlugin with all lifecycle methods
• src/serializer.ts — manifest YAML output
• src/codegen/, src/spec/ — generation pipeline and spec fetch/parse
• src/lint/rules/, src/lint/post-synth/ — FTN001 (AST) and FTN010–017
• src/op/activities/ — fountainApply (reconciler) and fountainRun
• src/composites/ — ConciergeStack
• src/skills/ — agent skills for authoring, secrets, and locked sandboxes
• src/import/, src/describe-resources.ts — adoption and live observation
• src/lsp/ — completions and hover
• src/generated/ — generated artifacts (do not edit)