oduffy-delphi/spike

/spike — Mechanism-Derisking Investigation

Prove — or disprove — that a specific mechanism is viable before a plan commits to it. Binary, same shape as /dogfood: converges to a viable/not-viable verdict, or switches gears — no open-ended "keep researching" path (that's coordinator:research).

One-line niche: prove a mechanism is viable before a plan commits to it.

Distinct from siblings: coordinator:research is broad, non-binary, and runs no code on this machine. coordinator:systematic-debugging is backward-looking (why did one known thing break). coordinator:dogfood is fix-through on a thing already built. coordinator:shape is problem-shaping (what to build), not mechanism-proving (can we build it this way).

Announce at start: "Running /spike <mechanism> — bounded derisking investigation. Binary outcome: viable or not-viable."

Invoke Gating — Structural, Not Self-Classified

One invocation surface, two entry paths — the gate is a structural signal, never a runtime self-classification:

• Research-heavy (default). PM-gated, never EM-initiated, never invoked from a subagent — same gate as /research//staff-session. A bare /spike <mechanism> with no structural signal present is ALWAYS this variant, full stop.
• Plan-trampoline. EM-reachable, no PM round-trip — only when coordinator:plan Branch B substrate verification carries an explicit trampoline: true-style signal on an unproven mechanism gate (coordinator/skills/plan/SKILL.md, the eighth verification dimension). Absence of that signal is dispositive — never infer the trampoline path from context.

Both Halves Required

1. External web research — how others solve this, what official docs say, known pitfalls. Same lookup shape as coordinator:research, scoped narrowly to this mechanism.
2. Local empirical study — install the dependency, run the probe, measure. The half coordinator:research structurally cannot do.

A web-only investigation is /research wearing a spike costume — dispatch research instead, or finish the empirical half before declaring a verdict.

Probes Are Throwaway

Probe scripts are derisking scaffolding, not durable test artifacts — discarded once the measurement is taken and the verdict recorded. Never commit them to the persistent test suite; that confuses throwaway scaffolding with a regression guard, which is the plan's own tests' job once drafted around a proven mechanism.

Durable Output — the Verdict Record

Write to docs/research/spike-verdicts/YYYY-MM-DD-<mechanism-slug>.md, validated against spike-result.schema.json. No state/handoffs/ lifecycle fields (no status, deployment_state, pickup_ready) — a verdict record is evidence, not a baton; evidence is read and cited, never picked up.

Carries: research findings, empirical measurements, binary verdict (viable/not-viable), gated_route (the downstream plan/chunk/decision this verdict gates, when reached via the trampoline), discharged_by (required and non-null whenever gated_route is a non-empty string).

No live routing intent may leave the session unlanded. If gated_route names a downstream target, write that intent onto the real target before the session ends — the plan the spike gates, an existing baton already carrying the work, or a PM-facing surface for a coordinator:shape-reached verdict — and name it in discharged_by. The schema rejects a non-empty gated_route paired with a null discharged_by; there is no such thing as a record that exits promising to be picked up later. Rationale and the corpus evidence behind this discipline: wiki.

Exit Routing — the Verdict Routes It

| Verdict | Route | |---|---| | viable | → coordinator:plan (fresh, authored around the now-proven mechanism), OR resume the trampolined plan — Branch B re-enters with the eighth dimension GREEN off the verdict record, not a full seven-dimension re-run | | not-viable | → coordinator:shape / PM — picking a different mechanism is a problem-shaping question, not a spike question |

Not an EM judgment call once the verdict lands — the verdict is the router. This wiring is the plan⇄spike back-edge: coordinator:plan Branch B trampolines into a spike on an unproven mechanism gate, and this skill's viable-verdict exit resumes that trampolined plan.

coordinator:spike vs. scope_mode: spike

Complementary, not colliding. scope_mode: spike is a plan-header enum describing a plan's own discovery mode ("throwaway code allowed, findings + recommendation"). coordinator:spike is a distinct upstream invocation that runs before a plan exists and produces a verdict record gating whether/how one gets drafted. A /spike verdict typically precedes a plan carrying scope_mode: feature/architecture, since the spike already de-risked the mechanism; occasionally the follow-on plan still carries scope_mode: spike if it remains exploratory on a different axis than the one this spike resolved.

Destructive-Action Prohibition

/spike writes files (probes, the verdict record) and can run multi-pass autonomously. No exceptions, not proposed, not authorization-requested mid-run: rm -rf, git reset --hard, git push --force, git rebase, deleting outside the spike's own scratch/probe directory, gh pr merge, --no-verify/--no-gpg-sign, hibernate/shutdown/power-off, killing other processes.

Out of Scope

Retrofitting historical spikes (pre-dating this skill) into verdict records — they stay as-was, folded into wiki reference material, not retroactively migrated. Editing the /research or /dogfood skill bodies — this skill defines its boundary against them by citation, not by editing theirs. A generic pipeline-graph engine — the shape→spike→plan edges are wired point-to-point in the affected skills, not abstracted into a routing framework. A central PM-gated-skills registry file — the research-heavy gate is documented here and cross-referenced from discovery surfaces. A new deployment_state enum value for the trampoline path — that's a handoff body directive, not a new enum value. A new primary manifest docType beyond the verdict-record schema itself.

Discovery-Surface Integration and Test Surface

Registering /spike on /workstream-start, the PM-gated-skills registry, and pipeline-graph cross-refs is a separate concern from this authoring surface. No runtime test exists for this skill body — a skill body is prose-doctrine, and this repo has no invocation harness that runs a skill end-to-end as a unit test, same as /dogfood//shape//plan. Frontmatter validates via yaml.safe_load the same way sibling skills' does. The exit-routing table above (both verdict rows, and the plan⇄spike back-edge phrase) must stay grep-assertable from both this file and plan/SKILL.md — that cross-file assertion is the test surface for this section.