Decision recorded: keep both Fleet and agent profiles — the confusion is naming and duplicated fields, not the model

[Hmbown]

The question

Founder, 2026-09-10: "i'm also confused by us having both 'fleets' and 'agents' and labels for it — i think that's the huge issue here", then: "we should probably just rip out fleets altogether and rebuild it just using agents?", then, after seeing what Fleet does: "but then we still use the term /fleet - i like it - and there is stuff that it lets us do."

The conclusion moved twice. Recording the evidence so it does not move a third time.

Both layers are load-bearing

Fleet is 34,372 lines of runtime, not a config format:

manager.rs 5,255   worker_runtime.rs 4,937   ledger.rs 4,314   exact.rs 2,705
executor.rs 2,267  host.rs 2,202   profile.rs 1,562   store.rs 1,483
task_spec.rs 1,275 members.rs 1,192  roster.rs 1,184  control.rs 1,050
scheduler.rs 1,039 identity.rs 911

Dispatch, scheduling, worker lifecycle, a usage ledger. Operate mode runs on it. Deleting it means rebuilding a scheduler and a ledger that work today.

Agent profiles are not redundant either. AgentProfile (fleet/profile.rs:71) carries description, requires (closed capability requirements), profile: FleetProfile and origin — behaviour and authority. The fleet shortlist carries none of that; FleetModel (fleet/members.rs:25) is only provider, model, roles, fleet.

The intended split is stated in members.rs:1-8: "A person builds their fleet by adding models from the providers they have configured; the operator model later picks sub-agent routes from that list only. Shortlist rows carry no role; the roles a model fills are the executable member rows that pin it."

So: Fleet = which routes are available. Agent profile = what a role does. That is also the local-config half of #6028's capability / Fleet / policy separation.

What is actually wrong

Three small things, none of which justify a rewrite:

1. The same field has two names. agents/*.toml calls it role_hint; fleet members call it role. Pick one.
2. A member duplicates profile fields rather than referencing the profile. On a real machine scout exists in both with identical id, model, provider and instructions, free to drift. FleetMember already supports the fix — omit model/provider and it inherits — but the code paths that write members do not use it (see the sibling over-pinning issue).
3. The roster merges three layers — [fleet.profiles] in config.toml, personal $CODEWHALE_HOME/agents/*.toml, workspace .codewhale/agents/*.toml (fleet/roster.rs:10-12). That is reasonable, but nothing surfaces which layer won, so a project profile silently shadowing a personal one is invisible. doctor_operate_fleet_report_json already computes the losing paths (Fleet config has one layer too many — and silent shadowing between the rest #5098) — surface it in the fleet view too.

Decision

Keep both. Keep the /fleet name. Work the three items above rather than a rewrite.

Rejected, with reasons

• Delete Fleet, keep agents — throws away 34k lines of working orchestration to fix a field-duplication bug.
• Delete agents, keep Fleet — loses description, requires and profile authority, which the shortlist has no place for.
• Merge into one store — re-creates the layering problem the roster already solves, and breaks the capability/Fleet separation Account link: one model config, auth as a reversible transform, policy separate #6028 depends on.