Spike — a question, not a change

Plugin root: ${CLAUDE_PLUGIN_ROOT} — resolve it to the path this skill was loaded from.

The output of a spike is an answer; anything built along the way is throwaway. Probing that only reads needs no nod from the user.

Procedure

1. Open no work item: no kf new, no .works/ file, no branch. The spike leaves nothing in the repository.
2. State in two sentences what you will probe, then start. Do not wait for a nod; read-only probing never waits for one.
3. Dispatch Agent(kf:researcher) for each question that research can answer (docs, APIs, versions, prior art, how an existing piece of code works), using the template in ${CLAUDE_PLUGIN_ROOT}/skills/flow/references/researcher-brief.md. Probe the rest yourself, as cheaply as correctness allows.
4. Write throwaway code only in a scratch dir under $CLAUDE_JOB_DIR/tmp when that variable is set, otherwise under the OS temp dir (mktemp -d). Never write it into the repository.
5. Wait for the user's explicit "go" before any irreversible step: a delete, a publish, a setting flipped on a shared system, a write to a remote. Name the step and its effect when you ask.
6. Report a recommendation, the sources it rests on, and the exact commands you ran, so the user can repeat them.
7. When the answer turns into work, hand back to the flow skill (Skill(kf:flow)) to reclassify the request as bounded or architectural before any change is made.

Report path

The researcher's report goes where the brief says: with no work item, $CLAUDE_JOB_DIR/tmp/research-<slug>.md when $CLAUDE_JOB_DIR is set, otherwise research-<slug>.md under the OS temp dir. A spike never writes a report under .works/; if the findings are worth keeping, say so and let the user decide where they go.