Agent spawning: create scoped agents with distinct harnesses

[Jacobobber]

The orchestrator can now create agents, not just dispatch to ones that already exist, each with a different harness and a different task, all coordinating over the channels. And the reason to build it on hallpass:

Each spawned agent is a scoped identity, not a trusted one.

Team.spawn mints each agent a token carrying only its harness scopes, then launches it through a pluggable Spawner, passing the identity, task, and channel by environment:

_, mint = dev_app()   # or your IdP's client-credentials flow in production
team = Team(mint=lambda name, scopes: mint(name, scopes),
            spawner=SubprocessSpawner(["python", "agent.py"]), channel="work")

team.spawn(AgentSpec("reviewer",  scopes=frozenset({"github:read"}), task="review PR 42"))
team.spawn(AgentSpec("messenger", scopes=frozenset({"slack:write"}), task="post to #eng"))

# inside agent.py
ctx = AgentContext.from_env()   # ctx.name, ctx.token (scoped), ctx.task, ctx.channel

• The reviewer's token carries github:read and nothing else; the messenger's carries slack:write and nothing else. A confused or compromised agent can only reach the tools its harness grants, enforced at call time and audited. Isolation between agents is the auth layer.
• hallpass stays model-agnostic: it provisions the identity and harness and launches the process; what thinks inside is yours. SubprocessSpawner is the default; swap it for any Spawner.
• Team tracks handles (alive / shutdown).

7 new tests: AgentContext.from_env round-trip + missing-env, Team.spawn mints a scope-limited token and builds the env, different agents get different harnesses, alive/shutdown, and two real-subprocess tests (env passing end to end, and lifecycle/terminate). Runnable two-process demo: examples/spawn_agents.py + examples/spawned_agent.py. Additive over 1.x. 268 tests, mypy --strict, ruff (whole repo), catalog-freshness green. v1.2.0.