Spike

Spike

Resolve one technical kill-risk with executable evidence. A spike is not a rough first implementation: its evidence survives, its code does not.

Run this workflow only after an explicit user request for $safeword:spike. A BDD checkpoint offer is not invocation; wait for the user to choose it.

Eligibility gate

Use a spike only after behavior is validated, when documentation and repository code cannot settle a kill-risk, failure would change the plan, and a bounded executable proof can answer it.

Otherwise route the uncertainty before writing experimental code:

• answerable from documentation or code → research it;
• dependent on user-only knowledge → $safeword:elicit;
• a choice among researchable alternatives → $safeword:figure-it-out;
• known implementation work → continue to plan-implementation.

Charter

Define the experiment before writing code. Record all five fields:

1. Question — one precise technical uncertainty.
2. Hypothesis — the result expected and why.
3. Kill criterion — the observable result that rejects the direction.
4. Proof — the exact command or walkthrough and expected signal.
5. Budget — one vertical slice with a time or effort ceiling.

If any field is missing, name the missing field and stop. Do not create a worktree, run a proof command, or spend the spike budget until the charter is complete.

Keep it question-sized

• Default to one experiment, one worker, and the smallest kill-risk vertical slice.
• Permit parallel worktrees only for independent comparison variants using the same charter and proof.
• Reject feature-wide component work as production implementation, not a spike.

Isolation

Inspect the repository root, current worktrees, branch, and status. Validated scenario and ticket-state changes must be included in one commit. If there are uncommitted validated scenarios or ticket state, name the affected files and stop: do not record PRE_SPIKE_BASE, and do not create a spike branch or worktree.

Once that state is committed, record the current production commit as PRE_SPIKE_BASE. Create a uniquely named spike/<question-slug> branch and sibling worktree from that exact commit; never reuse an existing path or branch. Before running the proof, verify the spike worktree is at PRE_SPIKE_BASE and contains the exact validated scenario and ticket-state files from that commit.

Keep every experimental commit in the spike worktree. After classifying the result, create the fresh production worktree from PRE_SPIKE_BASE. Run plan-implementation there: create impl-plan.md, map the structured handoff, commit the plan and ticket state, and review the plan. Only then begin production implementation in that same worktree.

Do not merge, rebase, cherry-pick, or copy spike commits or files into the production worktree. The spike branch remains unmerged while production work begins.

Keep the spike worktree until its evidence is distilled and the implementation plan is reviewed, then remove it through the repository's normal recoverable cleanup process.

Run the proof

Build only enough to execute the charter's proof. If the first attempt exposes a setup defect rather than answering the question, spend at most one bounded fix round. Stop when the proof answers the question or the budget expires.

Evidence distillation

Classify the result exactly once:

• VALIDATED — the hypothesis survived the kill criterion;
• PARTIAL — the direction works only under named constraints;
• INVALIDATED — the proof hit the wall and the direction is rejected.

Return a concise report:

## Spike result: <VALIDATED | PARTIAL | INVALIDATED>

- Question:
- Hypothesis:
- Pre-spike base:
- Proof command or walkthrough:
- Evidence:
- Constraints or wall:
- Useful shortcuts:
- Decision:
- Production consequences:

Return this structured handoff to the BDD orchestrator. Do not assume impl-plan.md exists yet: plan-implementation creates it, maps every handoff value to its documented destination, and never reuses experimental code.

Stop and report instead of widening scope when the budget expires, user-only knowledge or external authority is required, the experiment grows beyond one kill-risk slice, or the result changes validated behavior. Route those cases to $safeword:elicit or back to BDD scenario validation as appropriate.