ad-spike

name	ad-spike
description	Scaffold a staged spike with golden fixtures per WORKFLOW.md §14, for cases where the spec is clear but the technique is uncertain across multiple plausible approaches. Four stages — discovery, golden fixture, pipeline with gates, two-layer evaluation. Use when the unknown is *how*, not *what*. Triggers on "spike", "uncertain technique", "which library", "CV pipeline", "evaluate approaches", "ground truth", "golden fixture", "staged pipeline", "debug per stage". Routes to `ad-ground` if the *how* is routine and a single happy path is obvious. Read-and-write — creates `spikes/NNNN-<slug>/` with fixtures, debug per-stage artifacts, eval results.
summary	Staged spike with golden fixtures per WORKFLOW §14. Discovery + fixture + pipeline-with-gates + two-layer evaluation, when the *technique* is uncertain across multiple plausible approaches.
allowed-tools	Read, Write, Glob, Grep, Bash, WebFetch, WebSearch

/ad-spike

Implements WORKFLOW.md §14 (Staged Spikes With Golden Fixtures) end-to-end. The skill is for cases where the spec is clear but the technique is uncertain across multiple plausible approaches — library choice, CV approach, multi-stage transformation. WORKFLOW §9 (TDG) assumes the path is known and validates end-to-end; §14 assumes the path is unknown and validates per stage. Different uncertainty regimes; this skill is for the unknown one.

The skill creates a working directory under spikes/NNNN-<slug>/ and fills it stage-by-stage. The directory is throwaway by design — when the spike concludes, an ADR records the decision (/ad-adr) and the spike directory is deleted. See ADR-0017 for the promote-or-delete lifecycle rationale.

Step 0 — Confirm uncertainty

The skill is for unknown technique across multiple plausible approaches, not for non-trivial work in general. If a single happy path is obvious, do not start a spike. If the how is knowable from official docs / implementation references / in-repo patterns / git history, route to ad-ground and stop.

Concrete tests to run before starting:

• Could ad-ground's four-source research surface a single happy path with a defensible deviation gate? If yes, run that instead.
• Are there ≥2 candidate techniques with materially different trade-offs that no source resolves? If no, this is not a spike.
• Is end-to-end validation against expected outputs feasible without per-stage debug? If yes, this is ad-task + ad-philosophy Goal-Driven Execution territory, not a spike.

If the spike is warranted, confirm with the user the recortte (the specific surface where uncertainty sits — not the whole feature) and proceed.

Step 1 — Discovery

List canonical approaches grounded in official docs and real examples. Pick one (or a small set, ≤3) by an explicit criterion.

Candidate-listing process:

1. Search official documentation for the language / library / domain in question. Cite URL + version.
2. Search public implementation references (open-source repos, Stack Overflow / forum answers, blog posts, gists) for solutions to the same technical recortte. Cite <source>:<locator> — <repo>:<path>:<line-range> for repos, <URL> for Stack Overflow / blog / gist — and fetch via tools; never paraphrase from training memory.
3. Survey in-repo for analogous patterns the codebase already uses. Cite <file>:<line> or "no analog found".
4. Survey git history for prior attempts at the same problem. Cite <commit-sha> or "no prior attempt".

Output format:

## Discovery — <recortte>

### Candidate techniques
1. **<name>** — <one-line description>. Source: <URL or repo:path>. Trade-offs: <pros / cons>.
2. **<name>** — ...
3. **<name>** — ...

### Selection criterion
<one-line criterion: latency / accuracy / readability / dependencies / etc>

### Picked
<technique X>, picked by criterion <Y>. Alternatives held in reserve: <list>.

The output of this step is information, not code. No spike directory is created yet. The user reviews the candidate list and confirms the picked approach (or revises) before Step 2.

Step 2 — Golden fixture

Curate inputs with rich expected outputs. The fixture is the ground truth the staged pipeline validates against; richer fixtures catch more failure modes.

Create the spike directory:

mkdir -p spikes/NNNN-<slug>/{fixtures,debug,eval}

Where NNNN is the next available 4-digit number (mirrors ADR / task / spec numbering). List spikes/ and pick the next slot.

The fixture format is JSON keyed by input path (recommended) or whatever shape the domain demands. For computer vision: bounding boxes, sizes, lighting condition, difficulty tag, edge case markers. For multi-stage transformations: intermediate states. For library choice: representative inputs covering typical and edge cases.

Example fixture file (spikes/0001-detect-circles/fixtures/golden.json):

{
  "inputs/easy-01.jpg": {
    "expected": [
      { "bbox": [120, 80, 240, 200], "label": "circle", "size": "large", "lighting": "even" }
    ],
    "difficulty": "easy",
    "edge_cases": []
  },
  "inputs/hard-01.jpg": {
    "expected": [
      { "bbox": [50, 60, 90, 100], "label": "circle", "size": "small", "lighting": "low" },
      { "bbox": [200, 80, 260, 140], "label": "circle", "size": "medium", "lighting": "even", "occluded": true }
    ],
    "difficulty": "hard",
    "edge_cases": ["low-light", "partial-occlusion", "multiple-objects"]
  }
}

Curation principles:

• Include edge cases (low light, partial occlusion, malformed inputs, large inputs, empty inputs) — not just "happy path" examples. The fixture's job is to surface where a technique fails, not just whether it succeeds on easy cases.
• Include difficulty tags so per-stage evaluation can report performance segmented by difficulty.
• Keep the fixture as data, not code. JSON / YAML / CSV — anything that diffs cleanly and survives a refactor.

The fixture is the contract the pipeline validates against. Treat it like spec text — it should not change once the spike runs unless ground truth itself changes.

Step 3 — Pipeline with gates

One technique per stage. Each stage emits a debug artifact that makes its output inspectable.

Pipeline structure:

spikes/NNNN-<slug>/
├── README.md          # spike framing (Step 1 output)
├── fixtures/          # golden inputs + expected outputs
│   └── golden.json
├── pipeline/          # one file per stage
│   ├── 01-preprocess.<ext>
│   ├── 02-detect.<ext>
│   └── 03-postprocess.<ext>
├── debug/             # per-stage debug artifacts
│   ├── 01-preprocess/
│   ├── 02-detect/
│   └── 03-postprocess/
└── eval/              # evaluation results (Step 4)

Each stage's debug artifact format depends on the domain:

• CV pipelines: image saved to debug/NN-<stage>/<input-name>.png showing the stage's output.
• Multi-stage transformations: intermediate JSON saved to debug/NN-<stage>/<input-name>.json.
• Library evaluation: log row per (input, library) saved to debug/NN-<stage>/log.csv.

The discipline: each stage's output must be inspectable independently. End-to-end output alone tells you that it failed; per-stage debug tells you where.

Implementation pattern (any language):

• Stage takes (input, context) and returns (output, debug-record). Debug-record is written to debug/NN-<stage>/.
• Pipeline runs stages sequentially. Failure at any stage halts the pipeline and reports the stage where divergence happened.

Step 4 — Two-layer evaluation

Run the pipeline against the fixture and emit two layers of results:

• End-to-end: how many fixture inputs produced expected outputs? Reported as pass / fail per input, plus aggregate pass rate.
• Per-stage: for each fixture input, where did the pipeline diverge? Stage NN's output vs the expected intermediate. Reported as pass / fail per (input, stage).

Output to spikes/NNNN-<slug>/eval/results.json:

{
  "fixture": "fixtures/golden.json",
  "pipeline_version": "<commit-sha or timestamp>",
  "end_to_end": {
    "total": 10,
    "passed": 7,
    "failed": 3
  },
  "per_stage": {
    "01-preprocess": { "passed": 10, "failed": 0 },
    "02-detect": { "passed": 8, "failed": 2 },
    "03-postprocess": { "passed": 7, "failed": 1 }
  },
  "failures": [
    {
      "input": "inputs/hard-02.jpg",
      "diverged_at": "02-detect",
      "expected": [...],
      "actual": [...],
      "debug_artifact": "debug/02-detect/hard-02.png"
    }
  ]
}

The per-stage layer is what makes the spike actionable. End-to-end says that it failed; per-stage + debug artifact says where and why.

Step 5 — Conclude (promote or delete)

When the spike concludes — either the picked technique works or it does not — record the outcome via /ad-adr and delete the spike directory. The ADR is the persistent artifact; the spike code is throwaway.

ADR template for spike outcomes:

# ADR-NNNN: We will use technique X for <recortte>

## Context

<why the spike was needed — what was uncertain>

## Decision

We will use technique X. The spike at `spikes/NNNN-<slug>/` (now deleted) showed:
- End-to-end pass rate: <%>
- Failures concentrated at stage <NN>, root cause <Y>
- Mitigation: <Z>

Alternatives held in reserve and rejected:
- Technique A: rejected because <reason from spike eval>
- Technique B: rejected because <reason from spike eval>

## Consequences

<follow-on work this decision unblocks; rails to maintain>

Then:

rm -rf spikes/NNNN-<slug>/
git add doc/adr/NNNN-<slug>.md
git commit -m "feat: adopt technique X for <recortte> per spike NNNN"

Spikes that conclude inconclusively get an ADR too — Decision: defer; the spike at NNNN inconclusive because Y — and the directory is deleted. Inconclusive spikes are real signal; preserving the framing in an ADR prevents re-litigation.

Output contract

A spike directory at spikes/NNNN-<short-slug>/ with the four-stage layout above (discovery README, fixtures, pipeline, debug per stage, eval results). The directory is throwaway by design — promote-or-delete lifecycle' ' No Status: shipped lifecycle; spikes do not "ship" — they conclude with an ADR.

When the host exposes AskUserQuestion, use it for the Step 1 selection criterion confirmation and the Step 5 promote/delete decision.

Next

• After Step 1 (discovery output reviewed): proceed to Step 2 to create the spike directory + fixture, or abort if the discovery surfaced a single happy path (route to ad-ground).
• After Step 4 (eval results): /ad-adr to record the outcome, then delete the spike directory.
• If the spike succeeds and production work follows: /ad-task for the work units to apply the spike's findings to production code (Spec ref the original spec if applicable; cite the ADR in the task Notes).