Forge task execution and catchup personas

Each plan step can carry :execution. plan-to-kanban preserves it in the kanban card's :context, alongside :plan-step-id and :files. Step tags remain card tags, including the computed wave tag.

{:id "inspect" :title "Inspect the session control boundary"
 :files ["src/session_control.clj"]
 :tags ["saa" "session-control"]
 :execution {:presets ["ling" "mcp-first" "saa"]
             :persona {:priority-tags ["session-control"]
                       :caps {:decisions 30 :conventions 10 :snippets 0}}}}

{:id "implement" :title "Implement the inspected boundary"
 :depends-on ["inspect"]
 :execution {:presets ["ling" "mcp-first" "saa"]
             :persona {:priority-tags ["session-control"]
                       :caps {:decisions 5 :conventions 25 :snippets 20}}}}

Optional :spawn-mode selects a registered terminal/headless strategy. It must name an available backend. An explicit provider incompatible with a backend's declared providers fails before spawn (for example OpenAI on Claude-only). :headless resolves the configured default and is not a provider selector. The :hive-agent mode carries provider/model through its in-process API router; it is distinct from native Codex/Claude subscription CLI runtimes. Native subscription run-control ports are not swarm spawn modes.

Optional :provider and :model are strings passed to the existing provider registry. Use configured provider/model IDs; omitted fields use strike defaults and existing provider resolution. No new provider or model is selected by these examples. Per-step presets replace the default preset vector.

Mixed mode sends every card to lings, honoring explicit execution settings. Orchestrator mode rejects these settings instead of ignoring them: per-task settings do not route orchestrator subagents. The former drone mode is removed and refused with :execution/unsupported-mode.

:persona is a catchup lens map, not a named profile. Before starting the CLI, forge registers it under the generated agent ID through hive-agent's :agent/register-persona-lens extension. Catchup resolves that exact caller ID through :catchup/persona-lens and :catchup/bundle-profile. A missing persona registration extension fails the spawn explicitly.

Caps select memory buckets before content hydration. Supported adjustable buckets are :principles, :priority-principles, :priority-conventions, :sessions, :recent-wraps, :decisions, :conventions, :snippets, and :expiring. Zero omits a bucket; valid counts are integers from 0 to 1000. Omitted or invalid counts retain the normal limit. Axiom and nomination limits remain unchanged even when a persona requests zero. The knowledge addon also uses :atlas for its separate atlas lane.

Personalized bundles bypass the shared project bundle cache, preventing one caller's selection from changing another caller's memories. Persona registrations are process-local. Native run continuations restore them from completion checkpoints as described below; ordinary existing sessions still need their registration restored after a coordinator restart. Presets and configured provider backends retain their existing availability requirements.

Native continuation

Native subscription submission accepts an optional :run/opts containing only :persona. The runtime requires a checkpoint port when a persona is supplied:

{:run/adapter :codex
 :run/agent-id "exploration-session"
 :run/task "Inspect the session boundary."
 :run/opts {:persona {:priority-tags ["session-control"]
                     :caps {:decisions 30 :snippets 0}}}}

Completion persists the normalized persona beside the native session checkpoint before terminal publication. A continuation with a fresh :run/agent-id inherits that persona unless it explicitly supplies another. Restoring a checkpoint through a new runtime control works after the in-memory registry is lost. Registration happens before worker execution under both the transcript agent ID and the native driver run UUID.

The host-side restore path has real-file checkpoint tests. Actual native CLI MCP caller-ID alignment remains to be verified; no live native catchup proof is implied by those tests.