Dogfood the goal-directed task graph and decide default behavior

[spences10]

Parent

• Epic/RFC: goal-directed task graph for my-pi #476

Depends on

• Package and register @spences10/pi-goal as an opt-in extension #483

Scope

Evaluate the extension on representative repository work before enabling it by default or expanding its orchestration scope.

Cases

• A small deterministic task with one node.
• Parallel-ready nodes with a shared downstream dependency.
• Invalid and cyclic graph proposals.
• Missing or false completion evidence.
• A blocked task and deadlock recovery.
• Pause, resume, branch, abort, and session restart.
• No-progress and budget exhaustion under auto-continuation.
• A harness-backed ready task without duplicate validation state.

Acceptance criteria

• Runs complete without manual state repair or duplicate continuation.
• Invalid graphs never partially persist.
• Completion audits reject unsupported claims.
• Operator effort and outcome quality are compared with plain Pi plus a prompt.
• Cost and continuation behavior are recorded.
• End with an explicit decision: keep opt-in, enable by default, revise, or retire.
• Worker spawning, Team Mode dispatch, and multi-goal support remain separate decisions.

Tool-activation comparison

For representative task-completion cases, compare the existing plain-Pi baseline with the goal feature under fully active and on-demand tool exposure. Reuse the planned cases rather than adding another benchmark framework; see #471.

• Match model/provider/thinking, task inputs, dependencies, working backends, resource configuration and budgets.
• Record actual active tools, schema/prompt size, activation events, input/output/cache tokens, cost, operator actions and stop reasons.
• Score verified artifacts separately from normal completion, budget exhaustion and process settlement; include failures and discovery/continuation overhead.
• Use representative tracked-repository work, document cache effects, and repeat comparable runs before changing defaults.
• Treat Track upstream AgentHarness v2 and telemetry convergence #471's synthetic pilot as motivation, not a passed dogfood gate.

Existing decision gates and orchestration exclusions remain unchanged.