Cut Agent API, subagents, and compression roles over to per-agent profile runtimes

[acoliver]

Parent: #2635

Objective

Cut the Agent API, subagents, compression, and reusable role runtimes over to the transactional profile architecture. Remove their independent profile stores/populators/activation paths and prove multiple concurrent execution owners are isolated without ambient runtime identity.

This task does not implement interactive OAuth coordination (#2562) or utility-model generation (#2629).

Scope

Agent API

Expose narrow public capabilities:

• dispatch one typed ProfileCommand protocol;
• immutable redacted profile state/document views, summaries, requested/effective diffs and typed health;
• factual agent+command+revision subscriptions;
• saved/v2-document/unconfigured initial state for createAgent/fromConfig/preflight;
• no runtime/provider/settings/credential/store escape hatches.

ProfilesControl, legacy activation APIs, and providerActivationExecutor are removed in a versioned change or retained only as thin adapters with no directory scan, alternate schema, private store, direct mutation, profile-as-CLI-overrides, or ambient access. Failed initial profile construction publishes no partial agent.

Subagents

• Every launch creates an independent controller/runtime/session; definitions are templates, not live shared active state.
• Pass qualified profile ID or explicit captured v2 document across launch boundaries, never foreground current state or runtime ID.
• Replace subagentSettingsPopulation, buildActivationCliOverrides, LB runWithRuntimeScope profile pinning, and parallel auth/keyfile population.
• Derive immutable runtime/tool/context snapshots from the shared resolver/factory.
• Foreground transitions never mutate running subagents.
• Parent delegation/definition/environment/profile policy intersections cannot expand authority.
• Parent turn awaiting subagent remains transition-blocked; independent subagents may continue.
• Missing auth escalates structured challenge to the host per Make lazy authentication cancellable and route subagent auth challenges through the interactive host #2562; subagent never starts hidden UI.

Compression and reusable roles

• Replace CompressionProfileResolver, CompressionLoadBalancingProvider population/auth/LB duplication, ProfileManager.applyLoadedProfile, and SettingsService.importFromProfile with shared graph resolution/runtime construction.
• Parent transition captures qualified role document + etag; no late repository lookup.
• Lazy role runtime is child of parent revision, cached/disposed with that revision, and owns its own provider/model/auth/LB operational state.
• Preserve characterized parent-overlay, nested-LB rejection, context-window, auth warning/failover, fallback/error, and disposal semantics.
• Apply role capability ceilings: compression can deny tools/shell even if referenced profile requests them.
• Establish narrow construction/runtime capability for future search/summarization roles without exposing interactive profile CRUD or adding role-specific fields to base profile schema.

Concurrency and routing

• Explicit Agent/Turn/ProviderCall/ToolInvocation contexts route all calls/events/messages.
• AsyncLocalStorage may carry tracing/request metadata only, never active execution identity.
• Multiple simultaneous subagents/roles/foreground agents may use the same named definition without sharing mutable state.
• Shared repositories/credential stores/catalogs/transports are borrowed through keyed ports; runtime disposal never closes shared infrastructure.

Coordination / non-overlap

• Make lazy authentication cancellable and route subagent auth challenges through the interactive host #2562 owns host interactive auth coordinator, challenge coalescing/cancellation, and waiter lifecycle.
• Utility-model capability: provider-routed auxiliary LLM calls + utilityModel setting (unblocks summarization and prompt completion) #2629 owns utility model setting and single-shot provider-routed generation; this task only supplies role-profile runtime construction.
• Allow parent agents to resume cancelled or failed subagents with full session context #2571 owns resumable cancelled/failed subagent sessions.
• Eliminate ambient runtime global state: explicit dependency injection, delete getCliRuntimeServices and the singleton maze #2616 owns broader ambient global deletion; this task removes agent/subagent/role profile identity use.
• Foreground CLI/Zed cutover is a sibling Transactional per-agent profiles: replace ambient provider/model state with declarative runtime ownership #2635 subissue.

Acceptance criteria

• Agent API exposes only typed commands/redacted immutable queries/factual subscriptions and explicit initial profile state.
• No Agent API/ProfilesControl/activation adapter owns a second store/schema/mutation path.
• Concurrent subagents have independent profile/provider/model/auth/policy/LB/session state and pass the Subagents fail intermittently: z.ai 400 error 1213 on context accumulation and Codex GPT-5.5 Null final message #2410 parallel isolation invariant without profile identity ALS.
• Subagent auth challenges route to the supervising host or fail fast noninteractively; no hidden prompt/instructions occur.
• Compression standard/LB/parent-overlay/reference capture/role-ceiling behavior runs through the shared resolver/factory with no late repository mutation.
• Role runtimes are parent-revision children with deterministic cleanup and no shared-infrastructure disposal.
• subagentSettingsPopulation, profile-derived CLI overrides, compression population/auth lists, applyLoadedProfile, importFromProfile, and independent ProfilesControl mutation have no production consumers.
• Agent API, createAgent, subagent concurrency, compression standard/LB, role lifecycle, auth escalation boundaries, and noninteractive behavioral tests pass with full test/lint/typecheck/format/build verification.