Skip to content

feat: add the engineering agent catalog and an interactive installer - #21

Merged
archae0pteryx merged 3 commits into
mainfrom
feat/engineering-catalog
Oct 28, 2026
Merged

archae0pteryx merged 3 commits into
mainfrom
feat/engineering-catalog

Conversation

@archae0pteryx

@archae0pteryx archae0pteryx commented Oct 28, 2026 •

Copy link
Copy Markdown
Contributor

What

  • Catalog. The 24 engineering agents from agency-agents (MIT, license copied to factory/catalog/LICENSE) now live in factory/catalog/, one source file each. catalog.py renders them per host at install time: Claude Code gets model, effort, tools and maxTurns; Codex gets model, model_reasoning_effort and, for agents that only read, sandbox_mode = "read-only"; Cursor gets readonly.
  • Installer. install.py with no flags asks for scope (user-wide or project), tools, a multi-select of agents and the catalog models, then confirms before writing. q or end of input writes nothing. A rerun preselects what is installed, and the prompts make the folders match the selection. Flags (--scope, --hosts, --agents, --model, --yes, --remove, --force, --prune) keep the skill and CI non-interactive. A flags run only adds to what is installed; --prune also removes copies the installer wrote that are not selected. Existing --dest behavior is unchanged.
  • File names. Installed files are named by slug (code-reviewer.md, code-reviewer.toml), with no prefix. The installer only treats a file as its own when it carries a marker and has a name it writes, so a user's file that mentions the marker is never deleted, and an existing same-named file is refused unless --force.
  • Size. The catalog went from about 315 KB to 195 KB. A script removed the persona, memory, metrics and advanced-capability sections and the emoji headings (253 KB), then the 8 largest agents were hand-trimmed of most sample code.
  • Plugin version 2.6.0 to 2.7.0.
  • The README is compressed with caveman-compress (original kept outside the repo).

Tool restrictions

Before, no agent set tools, so each inherited every tool including all MCP servers. Advisory agents now get no Edit/Write, and incident-response-commander and network-engineer get no Bash either.

Testing

  • scripts/build_agents.py --check, test_install.py, test_wrap_up.py and git diff --check pass.
  • New test_catalog.py (added to CI) covers source validity, rendering for all three hosts, user-wide and project installs, additive runs versus --prune, a user file that mentions the marker surviving a prune, the interactive flow including the empty-tools retry and cancel, refusing to overwrite files the installer did not write, and the lane agents' Codex model ids agreeing with the catalog tiers.
  • Rendered Claude frontmatter parses as YAML and every Codex TOML parses. The installer was only run against throwaway home folders and repositories.

Not verified

  • No Codex agent has been run with sandbox_mode set from an agent file. Codex 0.164's config schema accepts the key.
  • The trimmed agents have not been used in real work yet. Only the incident-response agent was read in full after trimming; the other seven were checked by script (frontmatter unchanged, fences balanced).
  • The catalog is now a trimmed fork, so upstream changes have to be merged by hand.

Follow-ups are listed under "Engineering catalog" in tasks/todo.md.

Bundle the 24 engineering agents from agency-agents (MIT) in
factory/catalog/, one source file each, and render them per host at
install time with model, effort, tools, maxTurns, and Codex sandbox_mode
filled in. Read-only agents lose Edit and Write in Claude Code and get a
read-only sandbox in Codex.

The sources drop the persona, memory, metrics, and advanced-capability
sections and the emoji headings, and the eight largest lose most sample
code, so the set shrinks from about 315 KB to under 200 KB.

install.py run without flags asks for the scope (user-wide or project),
the tools, a multi-select of agents, and the catalog models, then
confirms before writing. A rerun preselects what is installed. Flags
(--scope, --hosts, --agents, --model, --yes) keep the skill and CI
non-interactive.

Bump the plugin to 2.7.0.
…ering- file prefix

Flag runs now add to what is installed; --prune, --remove and the prompts
make the folders match the selection. ours() requires a known file name
beside the marker, so a user's file that mentions it is spared. Picker
applies a/n only when the whole line parses, the tty check counts --hosts,
--model and --force, and the Claude name is quoted when needed. Installed
files are named by slug, and test_catalog checks the Codex ids agree with
build_agents.py.
@archae0pteryx
archae0pteryx merged commit ae9d712 into main Oct 28, 2026
1 check passed
@archae0pteryx
archae0pteryx deleted the feat/engineering-catalog branch October 28, 2026 16:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant