xdog-skills

Common medium-level skills with isolated domain support and extension code. The hub ships goal and plan. Flow and Eyes keep their independent engines and packaged skills; they are not copied here.

Coding and Claw include xdog-skills by default. Their skills extra remains compatible. Other consumers can install the hub directly. Namespace discovery lists the skill metadata without copying files or starting a model, host, store or observer. Full skill bodies load on demand. There is no Goal enable switch; availability is not user delegation.

Goal: records and reminders, not an agent runner

The current main Agent knows the conversation and owns planning, optional worker assignment using the host's existing tools, task reconciliation and completion. The goal skill explains that workflow. A host without subagents can do everything on its main Agent.

The trusted installed native Goal extension provides:

/goal create <outcome>
/goal status
/goal pause [reason]
/goal resume
/goal clear [reason]
/goal history

The goal model tool provides create/status/summary/tasks/history/update/plan/report/task_update/wait/ pause/resume/complete/clear. plan saves the main Agent's plan, not another model call. complete records its own evidence-based judgment, not a forced independent review. There are no Goal worker/reviewer runners, role tools or attempt/report acceptance protocol. Existing permissions and approvals still apply.

Each stable main-Agent owner has at most one current Goal, including paused or waiting Goals. Completed/cleared Goals release the binding but keep history. Clear does not delete task files, cancel unrelated Eyes observers or undo accepted work.

Data and events

GoalStore owns a leased xdog.goal.registry.v2 registry with owner binding, Goal revision history and immutable reminder outbox. Task state has ONE authority: <workspace>/.xdog/goals/<goal-id>/tasks/<task-id>.json. Workers update their assigned files; SDK mutations use expected versions and atomic no-follow writes. The local python -m xdog.skills.goal.task_cli ... list|update helper shares this CAS protocol under the caller's existing tool permissions. The plan is plan.md; errors are not empty results.

Before every continuation, the main Agent compares actual results with its previous report and calls goal(action="report", report={...}) BEFORE doing more work. progress.md is the sole report authority, with human-readable and structured history in one atomic, version-checked Markdown file. Baseline/advanced/unchanged/ regressed are Agent judgments, not script verdicts. With repeated no progress and no viable new approach, the Agent should report blockers and pause. No fixed round threshold or automatic progress judge exists.

/goal status and the model status default to counts, bounded previews, latest report and distinct execution/observation/delivery facts. Use tasks for filtered or single-task reads, and plan without content for reading. Snapshot byte pages retain query and content hashes. SDK status(detail=True) keeps full history reads.

Generic host settled/available events trigger a check. When continuation is appropriate, the runtime sends a stable-ID automatic message through ACP enqueue to the same main Agent and original Inbox/controller. ACP delivers; the skill instructs; the main Agent decides. Both settled and task-file signals enter one durable coalescer. An accepted pending/running reminder prevents another send; new changes remain dirty until consumed. Recovery inspects original receipts, not idle time, and known-not-sent retries retain original ID/content. New waits support only user or tasks; historical host waits stay readable and require explicit user conversion. A user wait, historical host wait or user pause cannot be washed into an automatic task wait and resumed. Execution/delivery failures are separate durable runtime facts, never automatic Goal state changes. Interrupted/unknown effects are not replayed; unresolved delivery uncertainty remains fenced even across pause/resume. No fixed polling loop, hidden tool approval or second Agent is created.

The runtime generates eyes.json and uses the installed Eyes CLI to own a daemon registration observing only this Goal's task JSON directory. It imports no Eyes SDK, runtime, loader, or daemon protocol. Events arrive through an owned events --follow JSONL process; only a durable Goal commit permits a separate CLI ACK. Configuration/state files are excluded. Pausing stops both proactive continuation and this owned registration; only explicit resume restores them and reconciles changed task files. Pause does not cancel accepted inputs, tools, children or other observers. Clear stops only the owned registration. Host retirement/close only detaches its CLI stream (suspend=False), allowing daemon-owned offline observation to remain; Goal business stops use suspend=True. Missing Eyes or an unavailable actual owner endpoint is reported, never silently installed or retargeted. Goal depends directly only on Agent/stdlib; optional Eyes operations cross the real CLI process boundary, not a Python SDK.

Old journal.sqlite3 and goals.json stay untouched, with their locations reported in status/history; no old worker/reviewer is restored. They are not silently converted into a new current Goal.

The optional packaged goal_ui worker remains readonly and exposes /goals via explicit global extension selection. Default /goal commands/model tools use the native installed capability, not this display worker, so a failed UI cannot own the business lifetime. Native /goal explicitly overlaps only its skill command; the skill itself remains discoverable/loadable. Builtins and other extension owners cannot be shadowed.

See docs/goal-host-faults.md for the current no-host-wait/fault/receipt contract, docs/goal-optimization-design.md, docs/goal-optimization-requirements.json, docs/goal-simple-design.md, docs/goal-simple-requirements.json and the current verification record. Earlier Goal kernel/adapter/reviewer records are historical.