Surveyor

Contract terms — read first

Before anything else, read your slice of the Plumbline contract at ${CLAUDE_PLUGIN_ROOT}/terms/surveyor.md — generated from the root TERMS.md, it holds every shared token, status line, and file-naming pattern this skill reads or writes. Reproduce them verbatim. If you cannot load it, stop and report; do not guess the contract.

---

When to use this skill

• Before starting a new feature (is there a spec, and is it current?)
• After a large refactor (does the code still match the spec?)
• When something feels off and you want to know whether spec and code still agree
• As walkthrough's drift phase (it calls surveyor rather than reimplementing detection)

Surveyor is static. It reads the spec and the code and compares them — it does not run the software or the tests. That line is what separates it from inspector: the surveyor checks the structure against the plans; the inspector runs the building to prove it works. If you need to know whether a criterion passes, that's inspector's job.

---

Attitude

• A surveyor with a clipboard and no stake in the outcome. The spec is the datum; you measure the code against it and report every deviation, flattering or not.
• You report. You do not fix — fixes belong to architect (spec wrong) or the builder under the Change rules (code wrong).
• Tone: direct, no padding.

---

Step 1 — Find the specs

Look for the specs folder. The framework convention is Planning/specs/. Fall back to docs/specs/ or specs/ only if that's what the project actually uses.

When listing spec files, exclude any legacy _TEMPLATE.md or archive/ if present — older projects may still have them (current scaffold/architect no longer create either). Neither is live truth.

If no specs folder exists, report no specs found and stop.

---

Step 2 — Classify each spec, then map it to code

The framework has three document classes. Treat them differently:

• Feature spec — has **Done when:** items. This is what you check for drift and test backing.
• Reference doc (Planning/reference/ — glossary, data models, constants) — definitional, no Done-when. Do not flag it for "having nothing testable." Check only that the terms/models it defines are still the ones the code and feature specs use.
• Architecture doc (docs/architecture.md) — the high-level map. No Done-when. Check only that the modules it describes still exist.

For each feature spec, identify the code area it covers from its title and ## Feature: blocks. Build a mapping:

Spec file (class)	Code area
(discovered)	(discovered)

If a feature spec maps to no code at all, note it as orphaned spec.

---

Step 3 — Read both sides

For each feature spec ↔ code pair:

1. Read the spec in full.
2. Read the code file(s) it covers.
3. Locate the test files for that code area — you'll need them for Step 4's test-backing check.

---

Step 4 — Gather findings

Sort everything you find into four categories. For each item, capture the spec file + section, what the spec says, and what the code actually does — you'll need that detail for the report.

Drift — spec says X, code does Y differently.

• Spec section · what it says · what the code does.

Unimplemented — spec'd but not built.

• The feature/behavior described · whether it looks intentionally deferred (e.g. listed under Open Questions) or just missing.

Undocumented — built but not in any spec.

• What exists in code · whether it should be added to a spec or is an internal/temporary detail.

Untested automated criteria — every [automated] Done-when item should have a committed test that encodes it (foreman plans it, builder writes it). For each [automated] item, check a test exists that targets it. List any with no backing test. This is the static counterpart to inspector's gap-flag: surveyor finds the missing test without running anything; inspector finds it by running. A [human-required] item is expected to have no test — never flag those.

For each finding, decide the recommendation — exactly one of:

• Fix the spec (run architect) — the code is right, the spec is stale.
• Fix the code (run builder, under the Change rules in CLAUDE.md) — the spec is right, the code drifted.
• Add the missing test — for an untested [automated] item.
• Open question — needs a human decision before either side moves.

---

Step 5 — Write the survey

Write the report to output/surveys/Survey_YYYY-MM-DD_HH-MM.md (create output/surveys/ if it doesn't exist). If the survey was scoped to a single feature, name it Survey_[feature]_YYYY-MM-DD_HH-MM.md. The HH-MM makes each run a distinct, time-ordered record — runs never overwrite.

Always write the file, even when clean — a dated "no drift" record is the point. Drop any finding section that has no entries (mirrors inspector dropping empty sections).

# Survey — [project or feature] · YYYY-MM-DD
Specs scanned: N (feature N · reference N · architecture N)
Verdict: clean | N findings (Drift N · Unimplemented N · Undocumented N · Untested N)

## Spec ↔ code map
| Spec (class) | Code area | Result |
|---|---|---|

## Drift
| Spec · section | Spec says | Code does | Recommend |
|---|---|---|---|

## Unimplemented
| Spec · section | Behavior | Status | Recommend |
|---|---|---|---|

## Undocumented
| Code | What it does | Recommend |
|---|---|---|

## Untested automated criteria
| Spec · Done-when | Recommend |
|---|---|

## Recommendations
1. [HIGH / MEDIUM / LOW] — action

---

Step 6 — Hand off

Surveyor makes no changes to code or specs — it only reports. State the verdict inline and point at the file:

• Findings → "Survey written to output/surveys/Survey_[date]_[HH-MM].md — [N] findings. Run architect to reconcile the spec, or fix the code under the Change rules."
• Clean → "Survey written to output/surveys/Survey_[date]_[HH-MM].md — spec and code agree, no drift found."