examples

Marvel example manifests

Every manifest here ships in both formats. name.toml and name.yaml are the same team written twice, so pick whichever your project already uses:

just build
./bin/marvel daemon &
./bin/marvel work examples/<name>.toml     # or .yaml
./bin/marvel get sessions
./bin/marvel stop --teardown               # when you are done

Each file opens with a comment explaining what it demonstrates and how to watch it work. This index says which one to open.

Start with these

Manifest	What it shows	Needs
generic-agent	The smallest thing that works: marvel managing arbitrary commands through the generic adapter.	nothing
claude	Real Claude Code agents beside a simulator supervisor, with adapter selection, permission injection and identity injection.	claude on PATH, model auth
review-team	A heterogeneous team: several roles, per-role replicas and restart policy.	claude on PATH, model auth

generic-agent is the honest first run, because it needs no model auth and still exercises the whole loop: manifest, reconcile, pane, session, teardown.

Harnesses and adapters

Manifest	What it shows
claude	The claude adapter: flags, permission mode, injected session identity.
claude-headless	Headless Claude Code, the mode marvel can read a usage stream from.
mixed-adapters	The harness matrix: three BYOA harnesses in one workspace.
generic-agent	The fallback adapter managing any command.
forestage-team	The five-primitive taxonomy (persona, identity, role) on the frozen forestage reference adapter.

Context metering and shifts

Manifest	What it shows
context-feed	The cooperative statusline feed that gives an interactive session a real CTX%.
context-feed-off	The same team without the feed, so CTX% stays absent. The "before" half of the pair.
auto-shift	Shifting a role automatically on context pressure, triggered on a token remainder rather than a percentage.
shift-demo	An operator-driven rolling shift with heartbeats.

Read context-feed-off and context-feed together. They are the same team and the diff between them is the whole feature.

Health, restart and loss

Manifest	What it shows
demo-act1-health	Health-driven terminal cases: restart policy against a failing healthcheck.
demo-act1-recovery	Recovery after pane loss.
demo-act1-roles	A team before a role is removed.
demo-act1-roles-removed	The same team after, so the reconciler's response is visible.
demo-act1-fleet-loss	Losing every replica of a role in one tick, which is charged to the role once rather than once per replica.

demo-act1-roles and demo-act1-roles-removed are a before and after pair, as are context-feed-off and context-feed.

Policy projection

Manifest	What it shows
policy-projection	A named settings fragment projected into a per-session file the harness reads.
policy-projection-v2	The same policy at version 2, to watch a re-projection rewrite running agents' settings files. A permissions change takes effect on next spawn (marvel#313).

Admission and budgets

Manifest	What it shows
demo-act4-budget	Admission refusing over-budget work against a team-declared budget.

Simulator demos

Manifest	What it shows
demo	Simulator agents with a chaos supervisor. No model auth needed.
demo2	Five simulator agents with healthchecks.

The simulator exists so the control plane can be exercised without spending model tokens. docs/demo.md walks the acts these manifests belong to.