Agency Setup

Agency Setup

Configure (or refresh) this repo to use srid/agency. Each step below is idempotent — it inspects what's already on disk and acts only on what's missing or out of date. The skill works equally well as first-time bootstrap, full refresh, or partial-install upgrade (e.g. user already added srid/agency to apm.yml manually but never created workflow.instructions.md — the skill detects the gap and fills it without re-doing the parts that already exist).

When the repo already has srid/agency in apm.yml, this skill also refreshes it to the latest ref (via apm deps update srid/agency in step 6) — there's no separate "update" mode.

Don't commit anything — leave changes staged for the user to review.

Invariant: apm install and apm compile run after every file change

apm install regenerates the host folders (.claude/, .opencode/, .codex/) from apm.yml plus the contents of .apm/, and apm compile -t <subset> produces the project-root AGENTS.md for Codex / opencode from the same inputs. Any change to apm.yml or anything under .apm/ invalidates both outputs. So this skill makes all file changes first (steps 1–5) and runs apm install (and apm compile where needed) exactly once at the end (step 6). Don't run install or compile partway through — later steps may add or modify files that must land in the same regeneration. If you ever edit apm.yml or .apm/* outside the prescribed order, you must re-run both before reporting back.

1. Pick an apm invocation

apm does not need to be installed — run it through uvx. Try in order, stopping at the first that works:

1. uvx --from apm-cli apm --version
2. nix shell nixpkgs#uv -c uvx --from apm-cli apm --version

If neither works (no uvx and no nix), tell the user to install one of uv or nix and stop. Don't try to install package managers yourself.

Use whichever prefix succeeded as the apm invocation for every subsequent apm call in this run (e.g., uvx --from apm-cli apm install).

2. Detect the host targets

The host targets go into apm.yml (next step) — apm install reads them, no -t flag needed. Detect from what's already on disk and the host you're running in:

• .claude/ exists, or you're running in Claude Code → claude
• .opencode/ exists, or you're running in opencode → opencode
• .codex/ exists, or you're running in Codex → codex

Multiple matches are fine — declare all of them. If nothing matches and the host you're in isn't one of the three, use AskUserQuestion to confirm. Do not guess silently — installing for the wrong target wastes a round trip.

Single vs. multiple targets: apm has a bug where the plural targets: list breaks when only one entry is present. Use the singular target: <name> scalar for exactly one target, and the targets: list only when you have two or more.

3. Create or extend apm.yml

Before editing, note whether srid/agency is already listed under dependencies.apm: — step 6 needs that fact to decide whether to refresh the dep.

If apm.yml does not exist, write:

name: <repo-directory-name>
version: 1.0.0
type: hybrid

target: <detected-target>

dependencies:
  apm:
    - srid/agency#master

(If you detected two or more hosts, use the targets: list form instead — see the note in step 2.)

If apm.yml already exists, edit it idempotently:

• If dependencies.apm: is missing the srid/agency entry, append srid/agency#master. Preserve every existing entry. If the dependencies.apm: block itself is missing, add it.
• If neither target: nor targets: includes the detected host, add it. Don't remove existing targets. When adding a host pushes the count from one to two, convert target: <name> into a targets: list with both entries; when removing a host (not something this skill does, but worth knowing) drops the count back to one, convert the list back to the scalar form.

Don't touch unrelated entries. Refreshing an existing srid/agency pin is handled by apm deps update in step 6 — don't hand-edit the ref here.

4. Ensure .gitignore covers agency runtime artifacts

apm install (which runs in step 6) will add apm_modules/ for you, but do writes .do-results.json at the repo root during a workflow run and that should not be committed. Make sure both lines exist in .gitignore (create the file if missing), idempotently — don't duplicate entries that are already there:

• /.do-results.json
• /apm_modules/ (verify; apm install may already have added it as apm_modules/ — either form is fine)

5. Draft .apm/instructions/workflow.instructions.md

If this file already exists, leave it alone — the user has either already configured it or is intentionally hand-maintaining it. Skip to step 6.

If it's missing (whether this is a first-time setup or an upgrade where the user added srid/agency to apm.yml themselves but never wrote workflow instructions), create it now.

do runs autonomously but needs to know your project's check, format, test, and CI commands. Inspect the project to figure them out — look at:

• package.json scripts: (Node)
• justfile (just)
• Makefile
• Cargo.toml, flake.nix, pyproject.toml
• .github/workflows/ for CI hints

For each of the four sections (Check, Format, Test, CI), there are three possible outcomes:

• Found a clear command in the project → fill it in.
• Found a plausible command but you're not certain → use AskUserQuestion to confirm. Offer the candidate as one option and "skip this section" as another, with a free-form fallback for the user to type a different command.
• Found nothing → use AskUserQuestion to ask the user directly. Always include a "skip this section" option so they can explicitly discard it. Don't fabricate commands.

Sections the user discards are omitted from the generated file entirely — no # TODO placeholders. do already handles missing sections by skipping the corresponding step with a note, which is the right behavior for a section the user has consciously declined.

Final file uses this template, including only the sections the user kept:

---
description: Workflow commands for the do pipeline
applyTo: "**"
---

## Check command
<command>

## Format command
<command>

## Test command
<command>

## CI command
<command>

## Documentation
Keep `README.md` in sync with user-facing changes.

6. Refresh srid/agency (if already present), then run apm install (and apm compile for Codex / opencode)

If srid/agency was already listed in apm.yml at the start of this run (you noted this in step 3), first run <apm-invocation> deps update srid/agency from the directory containing apm.yml. apm install alone won't move an already-installed dependency to a newer ref — deps update is what pulls the latest commit on the pinned ref. Skip this sub-step on first-time setup, where step 3 just added the entry; apm install will fetch it fresh.

Then, regenerate the host configs in a single pass. Run <apm-invocation> install from the directory containing apm.yml. apm reads target: / targets: from the yml and generates the runtime-specific folders (.claude/ / .opencode/ / .codex/), plus adds apm_modules/ to .gitignore.

install does not generate the project-root AGENTS.md instruction file. Codex and opencode read AGENTS.md at session start; without it they will not see the workflow instructions, code-police rules, or any other .apm/instructions/ content. To produce it, also run:

<apm-invocation> compile -t <subset>

…where <subset> is the comma-separated list of codex and/or opencode from your declared targets (e.g., -t codex,opencode if both are declared, -t codex if only Codex). Skip the compile step entirely if claude is the only target — Claude Code reads .claude/ natively and doesn't need AGENTS.md (compiling CLAUDE.md is intentionally avoided).

If install or compile fails, surface the error verbatim and stop — don't paper over it.

If you discover after this step that you still need to touch apm.yml or anything under .apm/, run install (and compile if applicable) again before moving on. The invariant at the top is non-negotiable.

7. Report back

Summarize for the user, in this order:

1. Which apm invocation you used (so the user knows the exact command for ad-hoc apm calls later).

2. Which target(s) ended up in apm.yml (and which form — target: scalar or targets: list).

3. Which workflow sections were filled in (and from where) versus skipped at the user's request.

4. Files changed (staged, not committed). Tell them to review the diff before committing.

5. Optional instructions to consider adding — list whichever of these files do not yet exist under .apm/instructions/, and explain briefly what each is for. They're project-specific and can't be auto-generated, but the user should know they exist so they can layer them on:

   • code-police-rules.instructions.md — project-specific quality rules checked alongside the built-in code-police rules.
   • hickey-catalog.instructions.md — project-specific complecting patterns extending the Hickey Layer 4 catalog.
   • lowy-volatilities.instructions.md — project-declared areas of volatility used by the Lowy review pass.

   Point them at Kolu's .apm/instructions/ as a worked example. Skip files that already exist.

6. Restart the agent CLI (Claude Code, Codex, opencode, etc.) so it picks up the newly generated skills — without a restart, the new skills won't be available in the running session.

7. After restart, try talk or do to verify everything works. Tell the user the exact invocation syntax for the target(s) you installed for — don't make them guess:

   • Claude Code → /talk <question> and /do <task> (slash commands).
   • Codex → $talk <question> and $do <task> (dollar prefix).
   • opencode → invoke /skills and pick talk or do from the list.

   If you installed for multiple targets, list the syntax for each.