The orchestrator now knows about prp-spike (engine row in both routing lists, verdict:* terminal status, spike → Phase 5 gate → downstream). Two other skills should be able to reach for a spike, and currently cannot.
The case
A planner that hits a genuine unknown mid-plan has nowhere to put it. Today it either guesses, or writes a plan whose feasibility nobody has established — which is exactly the failure prp-spike was built for. The useful shape is: planner hits an unknown → dispatches a spike → gates the plan on the verdict. That is the same spike → gate → downstream composition the orchestrator now uses, so one pattern covers both.
Same for prp-prd: a product spec resting on an unvalidated technical assumption should be able to say so and get it settled, rather than shipping the assumption as a requirement.
Scope
prp-plan — a route for "this step rests on something unproven": frame it, dispatch a spike, gate the plan on the verdict. A CONDITIONAL verdict is the interesting case, since it turns a plan step into a priced decision about changing a constraint.prp-prd — the same, for assumptions the spec depends on.- Decide whether either dispatches a spike itself or hands back to the operator/orchestrator. They are not orchestrators and should probably not grow into ones.
Related, decide alongside
A spike-runner subagent. The open design question is the unit: one agent per hypothesis, versus splitting by evidence-gathering vs judgment. Evidence from the first two real runs (helm #320, #322) favours the second — sub-claims of one mechanism share an expensive setup, so per-hypothesis agents redo it N times, and verdict judgment needs cross-claim sight that a single-claim agent does not have. A planner wanting a verdict it can gate on reinforces that: judgment is the part that should not be delegated to a cheap parallel worker.
Prior art in the pack: prp-review --agents for the fan-out shape; the advisory agents are all report-only by prompt, which is the right precedent for "a spike subagent returns evidence, not verdicts."
Not in scope
The orchestrator integration, which has landed. This issue is only the planning-side skills.
The orchestrator now knows about
prp-spike(engine row in both routing lists,verdict:*terminal status, spike → Phase 5 gate → downstream). Two other skills should be able to reach for a spike, and currently cannot.The case
A planner that hits a genuine unknown mid-plan has nowhere to put it. Today it either guesses, or writes a plan whose feasibility nobody has established — which is exactly the failure
prp-spikewas built for. The useful shape is: planner hits an unknown → dispatches a spike → gates the plan on the verdict. That is the samespike → gate → downstreamcomposition the orchestrator now uses, so one pattern covers both.Same for
prp-prd: a product spec resting on an unvalidated technical assumption should be able to say so and get it settled, rather than shipping the assumption as a requirement.Scope
prp-plan— a route for "this step rests on something unproven": frame it, dispatch a spike, gate the plan on the verdict. ACONDITIONALverdict is the interesting case, since it turns a plan step into a priced decision about changing a constraint.prp-prd— the same, for assumptions the spec depends on.Related, decide alongside
A
spike-runnersubagent. The open design question is the unit: one agent per hypothesis, versus splitting by evidence-gathering vs judgment. Evidence from the first two real runs (helm #320, #322) favours the second — sub-claims of one mechanism share an expensive setup, so per-hypothesis agents redo it N times, and verdict judgment needs cross-claim sight that a single-claim agent does not have. A planner wanting a verdict it can gate on reinforces that: judgment is the part that should not be delegated to a cheap parallel worker.Prior art in the pack:
prp-review --agentsfor the fan-out shape; the advisory agents are all report-only by prompt, which is the right precedent for "a spike subagent returns evidence, not verdicts."Not in scope
The orchestrator integration, which has landed. This issue is only the planning-side skills.