The twin of the spec-probes skill on the other side of the code. That one probes a requirement before anything runs. This one probes the running system after. Reach for it when nothing has been reported and the job is to find out what is wrong.

Two failure modes shape the design. An unbounded probe run burns budget and returns noise. A promoted finding no caller can reach wastes a maintainer's day. Pass 1 bounds the search. Pass 2 bounds what escapes it.

Name the scope, the surface, and the stop predicate, in that order.

Scope. A PR diff, one feature, or a whole product. For a diff, take the change map from the blast-radius skill and probe what it names rather than re-deriving it.

Surface. Probes run against the real thing through the project's driver, which the create-verification-skill skill builds when the project has none. A probe that never reached the running surface is , never a pass.

Stop predicate. A budget and a floor, both written down before the first probe. "Two hours or four hundred probes, whichever comes first, and never fewer than ten probes per entry point" is the shape. The budget ends a run that is finding nothing. The floor stops a run ending early because it got lucky on probe three. When a loop-governance suite is installed, hand it the outer loop and keep this skill as the method that loop runs; otherwise this predicate bounds the run.

Open a through the show-me-your-work skill, one row per probe: id, entry point, category, probe, observed, reproduces, state, reason. A probe that produced no finding is , , or . Otherwise the row carries that finding's state from Pass 2. carries the dismissal reason, the guard name on a gap, or the open question on an escalation. Keep it out of the tree so it survives reverts.

Inventory the entry points, classify each by shape, then raise only the categories whose shapes intersect it. A read-only query is never asked about interrupted sequences. The taxonomy is closed on purpose. Eight categories cleared beat forty nobody finishes.

input

input

resource

sequence

state

sequence

state

state

boundary

boundary

resource

concurrency

state

Fan the probes out with the swarm skill, partitioned by entry point. That skill resolves its own worker model and effort, so do not restate the choice here. Workers report , , or with evidence, which is swarm's own contract. Do not mint a second one.

An entry point with no matching shape gets one soft "unclassified, probe by hand" row, not silence.

Every worker finding is a hypothesis. Judge each on your configured model, with its configured effort per the pstack-harness skill. A finding promotes only when all three hold.

Triage every finding against the / / rubric in , whose three verdicts are tagged on the states below. Record a learned dismissal pattern there in its own format instead of re-deriving it next run.

Each finding ends the run in exactly one state.

The probe harness is the artifact, so commit it. A reviewer reruns it.

Each promoted finding becomes one failing test staged before its fix per the poteto-tdd skill, and the fix itself runs the Bug fix playbook (), one finding at a time. Do not batch fixes into one commit.

A promoted test asserts an invariant, not a transcript. "Rejects a quantity below zero" survives a refactor. "Returns this exact error string for this exact byte sequence" does not. Put every promoted test through the verify-commands skill before it lands, because a generated test that passes green while measuring nothing is worse than no test at all.

Reply: scope and surface, the stop predicate and whether it was met, probes run per entry point, each promoted finding on one line with its test, the dismissed count, every gap with its named guard, every escalated finding with the question it needs answered, and the path.