Spike

/spike — Time-Boxed Exploration

Single-pass research workflow producing a SPIKE-REPORT with a go/no-go/needs-more-info recommendation. Time-cap is the primary discipline. Optional --poc flag produces a minimal proof-of-concept code alongside the writeup.

When to use

• "Can we use gRPC instead of REST?"
• "Best way to implement real-time sync?"
• "Is this library production-ready?"
• "Research: what's our path to multi-region deployment?"

Not for

• Decisions already made → just /plan then /develop
• Open-ended exploration with no question → /spike requires a stated question
• Convergence on a binary success signal → /ralph (spike is exploration, not iteration)

Invocation

/spike "Can we use gRPC instead of REST?" --time-cap 120
/spike "Best way to implement real-time sync" --poc
/spike "Is Postgres pg_vector production-ready for our scale?" --time-cap 90

Arguments

Output

• <repo>/.ai-skills-memory/spikes/<run-id>/SPIKE-REPORT.md — findings, pro/con analysis, recommendation
• (optional) <repo>/spike-poc-<run-id>/ — proof-of-concept code if --poc
• Memory write: L4 research record for future reference

Agent roster

The Lead picks the SME role based on the question domain (e.g., db-engineer for schema questions, ml-engineer for ML questions, solution-architect for architecture).

Pipeline

┌─ Lead chooses SME role based on question domain
│
├─ SME researcher:
│  - Research question exhaustively (within time-cap)
│  - Gather evidence (docs, benchmarks, examples)
│  - Build POC if --poc flag set
│  - Estimate effort if go/no-go decision involved
│  → SPIKE-REPORT.md with go/no-go/needs-more-info recommendation
│
├─ (Optional) technical-reviewer validates findings:
│  └─ Stress-test assumptions, identify gaps in evidence
│     → review.md
│
└─ Memory write: L4 research record
   Report: go/no-go + next steps

No RALF — spike is exploration, not convergence.

Decision lens & evidence requirements

• SME research + evidence gathering: every quantitative claim in the SPIKE-REPORT cites a source from references/evidence-hierarchy.md (level 1–5; level 6–7 requires "low confidence" caveat).
• SME recommendation (the go/no-go/needs-more-info writeup): apply Cynefin classification + One-Way/Two-Way Doors lens per references/decision-frameworks.md. Recommendation is one of go-two-way, go-one-way, or needs more spike before commit, with each option classified on (Cynefin domain, reversibility, ICE/RICE relative score, confidence level).

G7 spawn payloads

SME and reviewer spawns use structured G7 payloads per plugin/schemas/spawn-payload.schema.json. Returns conform to plugin/schemas/return-contract.schema.json.

Eval rubric

Pointer: plugin/eval/judge-rubrics/spike.md (B10).

Dimensions:

1. Question clarity — spike clearly answers the stated question
2. Evidence quality — findings backed by data (not opinion)
3. Completeness — considered trade-offs and alternatives
4. Feasibility assessment — if go/no-go, effort estimate is realistic
5. Actionability — next steps clear (prototype? deep dive? abandon?)

Pass: avg ≥ 4.0, no dimension < 3.

Memory writes

Hard rule on committed decisions: never auto-create .committed/decisions/ files. .committed/ goes to git and the user may want to reword the decision before committing. The skill ALWAYS asks before writing to .committed/decisions/. The pre-tool-use-committed-write.py hook enforces the allowlist on these paths regardless.

Failure modes

• Time cap hit mid-research: SME pauses at time-cap; produces partial report with explicit [INVESTIGATION INCOMPLETE] marker and "more investigation needed" section listing specific data gaps
• Findings inconclusive: report states inconclusive with specific data gaps; recommend deeper spike or prototype
• POC fails to compile/run: report failure as a finding; do NOT block the spike completing — POC failure is itself a data point for go/no-go

Observability events

• workflow_start — spike + question
• time_cap_warning — at 80% of time-cap
• workflow_end — COMPLETE with go/no-go OR INCOMPLETE if time-cap hit

Integration

• Orchestrator: feature-design-lead (only agent with tools: Task)
• Schemas: plugin/schemas/spawn-payload.schema.json, plugin/schemas/return-contract.schema.json
• Companions: /context-load (per-role context for SME), /subagent-spawn (G7 payload)
• Rules: subagent-isolation (G7 + bounded recursion), untrusted-content-wrapping (G1 wrap on subagent returns + reference docs read), memory-discipline (committed writes always-ask + allowlist)
• Hooks: pre-tool-use-committed-write.py (enforces .committed/decisions/ allowlist), tool-output-normalize.py (G2 on bash output)