v0.9.0: Always-on stateful agent identity across workrooms, repos, and fleet runs

[Hmbown]

Problem

CodeWhale's current agent state is still too tied to a turn, a TUI session, or a single worker run. The desired product shape needs an always-on, stateful agent that can live across workrooms, repos, GitHub issues, mobile/chat surfaces, and Fleet/WhaleFlow runs.

This is related to Fleet, but not identical. Fleet gives durable worker execution. The always-on agent identity is the durable operator/persona that watches work, wakes up on events, remembers commitments, routes tasks to workers, and can be resumed or steered from any authorized surface.

Product direction

An always-on CodeWhale agent should be able to:

• Own an inbox across workrooms, mentions, approvals, failures, and GitHub events.
• Remember active commitments, watched repos/branches, blocked decisions, and follow-up timers.
• Wake on durable events rather than relying on a user keeping a TUI turn alive.
• Delegate work to Fleet/WhaleFlow workers with explicit role/model/tool policies.
• Be steered from TUI, Runtime API, mobile, or a chat bridge without losing state.
• Explain what it is watching, what it is doing, what is blocked, and what needs human approval.
• Survive process restart, laptop sleep, host reboot, or chat bridge reconnect.

Proposed abstraction

Add a typed AgentProfile / AgentIdentity layer:

• Stable id, display name, persona, owner, allowed workspaces/repos, trust level, and auth boundaries.
• Provider/model policy: explicit route, inherit, or route-effective auto policy from v0.8.61: Add route-effective model inventory and auto fleet model selector #3205.
• Tool/fleet role policy: scout, implementer, reviewer, verifier, operator, summarizer.
• Subscriptions: workroom mentions, GitHub issues/PRs/checks, fleet run barriers, failed receipts, timers, approvals, and manually pinned threads.
• Durable state: active commitments, next checks, blockers, watched refs, recent receipts, memory refs, and compact summary.
• Lifecycle: active, sleeping, paused, needs-human, degraded, failed, disabled.
• Controls: pause, resume, sleep, wake now, handoff, clear subscription, export state.

Runtime requirements

• State is persisted outside a model context window and has schema versions/migrations.
• The agent reads compact state summaries and artifact refs, not raw full transcripts by default.
• Event-driven wakeups are debounced, idempotent, and replay-safe after restart.
• Wakeups respect approvals, sandbox/trust level, provider budgets, and route health.
• Agent identity and worker execution remain separate: the always-on agent may schedule fleet workers, but each worker still has its own receipt and artifact boundary.
• No secrets are stored in agent state; secrets are referenced through configured secret refs or existing auth stores.

Acceptance criteria

• A design doc/RFC defines AgentProfile, AgentState, subscriptions, lifecycle states, and how they connect to workrooms and Fleet.
• Runtime API can list agent identities, read bounded state, pause/resume an agent, and trigger a manual wakeup.
• CLI/TUI exposes at least a minimal status surface: active agents, what each is watching, current state, and blocked/needs-human count.
• A CI-safe test proves an agent subscription wakeup is persisted, consumed once, and not duplicated after manager/runtime restart.
• A smoke test proves a paused agent does not schedule new workers, then resumes and processes a queued wakeup exactly once.
• A fleet-backed task can report a receipt to the owning always-on agent, and the agent state records the receipt ref rather than copying the full transcript.
• Docs explain how this differs from normal sub-agents, Fleet workers, and one-off codewhale exec runs.

Non-goals for the first version

• No autonomous destructive actions without approval.
• No hidden hosted memory service.
• No always-on behavior enabled by default for every install.
• No model-weight training loop.

Related

• v0.9.0 EPIC: Chat-native CodeWhale workrooms for threaded, shareable agent work #3209 chat-native workrooms.
• v0.8.61 EPIC: Agent Fleet control plane for always-running verifiable work #3154 Agent Fleet control plane.
• v0.8.61: Split sub-agents into a headless worker runtime with lightweight TUI projections #3096 headless worker runtime.
• v0.8.61: Model the Agent Fleet org chart, roles, and delegation policy #3167 fleet roles and delegation policy.
• v0.8.61: Add route-effective model inventory and auto fleet model selector #3205 route-effective model inventory and auto fleet selector.
• EPIC: v0.9.0 WhaleFlow branch/leaf workflow mode (tracker, replaces deleted #2667) #2981 WhaleFlow branch/leaf workflow mode.
• fix(core): capacity-memory checkpoint and cross-session recovery #1062 capacity-memory checkpoint and cross-session recovery.