Test Agent Teams

Skill: /test-agent-teams

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

End-to-end verification that Agent Teams actually resolve and spawn each role on its intended model. Every spawned role echoes a nonce, its role, what it was told to do (its charter, from the resolver), and the model it actually ran as — a nonce the orchestrator cannot fake, so a correct echo proves the role really spawned, the #model-group routed, and the chain executed. This SPAWNS real agents (one per role across all teams) — a deliberate, possibly costly integration test; run it to verify, not routinely.

When to invoke

/test-agent-teams --all-teams (or all) tests every defined team; /test-agent-teams <team-name> tests just that one. Also triggered by "test the agent teams" / "run the agent-team self-test". The skill mints its own nonce.

Dispatch. --all-teams / all → every resolved team. A team name (matched against the resolved set) → only that team. No argument → list the teams and ask which (or --all-teams); never default to a large spawn silently.

Step 1 — Nonce

1.1 GENERATE a fresh short random nonce (8 hex chars) and state it up front — the skill always mints its own; never ask the user for one. The SAME nonce is passed to every role spawned this run.

Step 2 — Enumerate the teams (use the tooling, do not hand-list)

2.1 Run python $HORIZON_BIN/resolve_agent_teams.py --json. It returns every resolved team with its roles, model groups, flags, charters, and loop target/cap. Use that.

2.2 Apply scope: --all-teams/all → keep all teams; a team name → keep only the team whose name matches (case-insensitive); otherwise ask which.

Step 3 — For each team, run an example flow

3.1 Compose a one-line natural-language example flow that exercises the team's roles and flags (mirror documentation/system/agent_teams.md §1.1 style). State it before spawning.

3.2 For each role, in order, spawn a sub-agent (Task/Agent tool) on that role's model group: resolve #group → a concrete model via horizon_aios_model_prefs.md (+ its local file); if a group has no runnable member, note it and fall back to the harness default. Record the model you spawned it as (the expected model). Honor flags for the test:

1. if needed / if asked — still spawn (note the flag) so every part is exercised.
2. parallel — spawn the adjacent parallel roles concurrently.
3. ask user — note it and continue; do not block the test.
4. Loop — spawn the looping role once (note the loop + cap + target); do NOT actually iterate.

3.3 Each role sub-agent gets this prompt (fill in the placeholders; <CHARTER> is the role's charter from the resolver JSON — its real job in the team):

You are the "" role of the "" agent team (model group <GROUP>). Your job in the team: . This is a wiring test — do NOT carry the job out. Reply with EXACTLY four lines, nothing else: nonce: role: told: model: <the model you are actually running as — your real model id/name>

Step 4 — Collect and verify

4.1 Gather each role's four-line reply. Build one report table per team:

| Team | Role | Told to do (charter) | Model group | Expected model | Reported model | Nonce OK |

4.2 Flag: a missing/incorrect nonce (the role did not really run, or the chain broke); a reported model whose tier does not match the group (a routing problem); or a role that failed to spawn.

Step 5 — Report

5.1 Print the table(s) and a PASS/FAIL summary: PASS when every spawned role echoed the nonce and ran on a model consistent with its #group. List every mismatch explicitly.

Notes for the executing agent

1. The nonce is the proof of execution — NEVER fill it in yourself on a role's behalf; it must come back from the spawned agent or the test fails for that role.
2. Reuse resolve_agent_teams.py for discovery; never hand-parse agent_teams.md.
3. This is a real, fan-out spawn (≈ one agent per role across all teams). Tell the user the spawn count before a large run if they have many custom teams.
4. The point is to exercise EVERY part: every role, every model group, and a note for every flag — so the report doubles as a coverage check of the resolved team set.