Skip to content
Release candidate documentation. This version is qualified source, not yet an independently verified registry release. Use stable 1.0.0

Develop with Codex and Claude Code

Minco 1.1 packages the same focused application workflows for Codex and Claude Code. The skills select Minco commands and source authorities; they do not add a hosted agent runtime or replace the application graph.

Start with a read-only plan

From a generated or existing Minco application:

bash
cargo minco agent plan --target all --json

Review the fixed project-relative destinations and retain the returned plan_digest. Planning writes nothing. It classifies files as creates, unchanged managed files, safe managed updates, conflicts, or manual actions.

Install only the exact reviewed plan:

bash
cargo minco agent sync \
  --target all \
  --expect-plan-digest <sha256> \
  --json

Sync rechecks the project and destination identities before every write. A stale digest, edited managed file, unowned destination, symlink, non-regular file, or path race fails closed. Neighboring files are never deleted.

Client projections

ClientManaged project paths
Codex.agents/skills/minco-*/
Claude Code.claude/skills/minco-*/
Both.minco/agent-manifest.json ownership receipt

For Claude, Minco may also create a three-line CLAUDE.md that imports @AGENTS.md. It does so only when AGENTS.md is a regular file and the Claude file is absent or already Minco-managed. An existing user-owned CLAUDE.md is preserved byte-for-byte and reported as a manual action.

Minco does not parse or rewrite .mcp.json, .codex/config.toml, global client settings, or non-Minco skills.

Choose a focused workflow

The package contains eight skills:

  • web-application routing;
  • OpenAPI-first operations;
  • statically linked plugins;
  • explicit application lifecycle;
  • bounded diagnosis;
  • evidence-aware review;
  • framework contribution through the owning task and JJ workspace; and
  • explicitly requested release preparation.

Generated applications receive application guidance. They do not inherit the Minco framework repository's internal task, JJ, or release policy.

Inspect bounded context

Use the project model instead of asking an agent to infer hidden structure:

bash
cargo minco agent context --json
cargo minco agent context --operation placeOrder --json
cargo minco agent context --task M0-T01 --json

The response is schema-versioned and capped at 64 KiB. It reuses the bounded ProjectView allowlist and reports zero writes, child commands, network requests, and arbitrary file reads. Unknown IDs return found: false with a stable diagnostic; Minco does not guess.

Diagnose and evaluate

bash
cargo minco agent doctor --target all --json
cargo minco agent eval --target all --json

Doctor reports discovery, exact version, ownership, drift, projection parity, and manual MCP configuration state without modifying the project. Eval checks portable skill format, exact installed bytes, Codex/Claude parity, and 16 positive/negative workflow contracts. It performs no writes, commands, network requests, or model invocations.

A passing evaluation is local projection evidence. It is not model quality, hosted CI, review, deployment, runtime, or production proof.

Authority remains explicit

Skills can explain commands, but they never authorize commit, push, merge, release, registry publication, deployment, database mutation, provider access, promotion, or production feature enablement. Those actions keep their ordinary Minco evidence and user-approval gates.

For the underlying read model, see ProjectView, MCP, and the workbench. For exact command syntax, see the CLI reference.

Minimal cost, maximum capability.