Agent Teams

Skill: /agent-teams

Model preference: #midcost (per horizon_aios_model_prefs.md; overridable by a prompt directive).

Tooling to define and override Agent Teams without hand-editing markdown blind. Agent Teams are named, reusable multi-agent workflows (ordered roles, each on a model group, optionally with loops) that the user invokes in-session ("send an investigate-and-fix agent team to look at problems 2, 3, and 4").

Shipped OS defaults live in $HORIZON_ROOT/agent_teams.md (tracked — do NOT edit it here). User definitions and overrides live in local.agent_teams.md (machine-local, gitignored, never clobbered by sync), @-imported by that scope's agents.md. This skill always writes to a local.agent_teams.md, never to the shipped file. Authoritative spec: agent_teams.md and file_structure_invariants.md §13.

When to invoke

The user types /agent-teams ..., or asks to add/define/edit an agent team, to set up team definitions for a specific project/brain/subfolder scope, or to add/define a custom role flag.

Dispatch. A bare /agent-teams with NO arguments → run Default (list loaded teams) below, then stop — do not start creating or editing anything unless the user asks. A request to add/define a custom role flag → go to Step 5 — Adding a custom flag. Any request to add/define/edit/override/scope a team → go to Step 1 and the flow that follows.

Default (bare /agent-teams, no arguments) — list loaded team sources

D.1 Run the resolver utility — it does the discovery; do NOT hand-glob or rely on memory of what is "in context":

python $HORIZON_BIN/resolve_agent_teams.py <cwd>          # human-readable
python $HORIZON_BIN/resolve_agent_teams.py <cwd> --json   # structured

It walks the scope cascade from the path up to the OS root, reads the shipped $HORIZON_ROOT/agent_teams.md plus every local.agent_teams.md override (excluding any file merely named agent_teams.md, e.g. the doc under documentation/), and reports each source with its team names — and, in --json, each team's roles, model groups, and loop flag — plus the resolved set (most-specific scope wins; same-name overrides shipped, new names unioned in).

D.2 Present its output to the user: the sources (path → team names, stubs noted as "no custom teams") and the resolved teams in effect. Keep it tight.

D.3 Then ask whether they want to add a custom team, modify/override an existing one, or set up a project/brain scope — and stop. Proceed to Step 1 and the flow below only if they ask you to.

Step 1 — Resolve the scope (which local.agent_teams.md)

1.1 Default scope is the OS root: $HORIZON_ROOT/local.agent_teams.md.

1.2 If the user names a different scope (a project, a brain folder, "here"/cwd, or a subfolder), target that directory's local.agent_teams.md instead. Teams cascade OS → project → brain → subfolder, most-specific wins (same semantics as model-prefs Scope Precedence).

1.3 Confirm the resolved absolute path with the user before writing.

Step 2 — Ensure the override seam exists at that scope

2.1 The scope's directory must have an agents.md that @-imports its sibling ./local.agent_teams.md (per §13 / §12 override-file anchoring). Check for it.

2.2 If the directory already ships the seam (OS root, .claude/, a brain .claude/), nothing to wire — proceed.

2.3 If creating a NEW scope (e.g. a project subfolder) that has no agents.md:

1. Create agents.md in that directory with the override line @./local.agent_teams.md (and, to inherit OS defaults, @<relative-path>/agents.md up to the nearest parent agents.md). Keep it a real agents.md, not a CLAUDE.md (§12: overrides anchor on agents.md).
2. If the directory also needs a CLAUDE.md, make it a thin pointer importing ONLY ./agents.md (§12.3).

2.4 If local.agent_teams.md does not yet exist at the scope, materialize it: copy the nearest local.agent_teams.md.template if present, else create a short stub header. It is gitignored — never commit it.

Step 3 — Define or edit the team

3.1 Gather, with the user:

1. Team name — the phrase they'll invoke ("Investigate & Fix").
2. Roles, in order — for each: a label, a model group (#midcost, #lowcost, #highcap, #investigate, … from model-prefs), an optional condition (if needed = model skips it unless it adds value; if asked = runs only when the user asks — see agent_teams.md → "Conditional roles"), and a 1–2 line charter (what it does + what it hands to the next role).
3. Loops (optional) — see Step 4.

3.2 Verify each referenced model group exists in horizon_aios_model_prefs.md (or its extend file). If a group is missing, tell the user and offer to add it via /model-prefs.

3.3 Write the team into the scope's local.agent_teams.md, matching the format in agent_teams.md (a ### <Team Name> heading then a numbered role list). A same-named team overrides the shipped one; a new name is unioned in. Keep entries tight — the file loads into context every session.

3.4 To EDIT an existing team, read the scope's local.agent_teams.md (and the shipped agent_teams.md for reference), then modify the matching block in the local file only.

Step 4 — Loops (retry until pass)

4.1 A role may loop back to an earlier role with feedback until a pass condition or an iteration cap. Express it inline on the looping role, per agent_teams.md → "Loops":

Loop: on <condition>, return feedback to <role> and re-run from step <N>; repeat until <pass condition> or <max> iterations, then <action at cap>.

4.2 Always set an iteration cap (bounds cost, prevents infinite loops) and a cap action (typically: stop and report outstanding failures). Example (Full Team): "Validator → on fail, feedback to Implementer, re-run from step 4; repeat until clean or 3 iterations, then report."

Step 5 — Adding a custom flag

Role flags (the markers beyond a role's label + model group) are cataloged in $HORIZON_ETC/agent_team_flags.md (shipped) and local.agent_team_flags.md (machine-local, gitignored). The resolver parses flags generically, so a new flag works as soon as it is used — the registry just gives it meaning, validation, and a --flags listing entry. To add one:

5.1 Confirm the shipped catalog first: python $HORIZON_BIN/resolve_agent_teams.py --flags. If the flag already exists, there is nothing to add.

5.2 Gather: the flag name (the literal token, e.g. if blocked, or an annotation name like Gate), its form (inline = a token in the role's (#group, …) parenthetical; annotation = a **Name:** … line under the role), and a one-line meaning (what the acting model should do when it sees it).

5.3 Append a row to the table in $HORIZON_ETC/local.agent_team_flags.md (NOT the shipped agent_team_flags.md):

| <flag> | inline\|annotation | <meaning> |

If the file does not exist, materialize it from local.agent_team_flags.md.template first. It is gitignored — never commit it.

5.4 Verify: re-run --flags; the new flag should appear. Using it in a team will no longer raise the resolver's "not in the registry" warning.

Notes for the executing agent

1. Never edit the shipped $HORIZON_ROOT/agent_teams.md from this skill — it holds OS defaults overwritable by sync. All user changes go in a local.agent_teams.md.
2. local.agent_teams.md files are gitignored and machine-local — do not commit them or echo their full contents back; confirm the change in one or two lines.
3. Keep team definitions short and durable. They load every session at their scope; bloat costs tokens each time.
4. Model groups must resolve in model-prefs. Do not invent group names — reuse the shipped set or have the user define one via /model-prefs.
5. Honor the §12/§13 invariants: overrides anchor on agents.md, never CLAUDE.md; CLAUDE.md stays a thin pointer to its sibling agents.md.