Documentation Guardian

Keeps docs/adr/ in sync with the code. Two jobs:

1. Detect ADR-worthy changes as they happen and propose drafting an ADR.
2. Check planned and in-progress changes against existing ADRs before they land.

The ADR conventions referenced here live in docs/adr/README.md. When the two disagree, docs/adr/README.md wins — update this skill to match.

Activation triggers (this repo's architecture)

Activate when you observe any of these in a planned or in-progress change:

• A new chart plugin under src/chart/plugins/, or a new entry registered in src/chart/plugins.ts.
• A new data source under src/data-source/, or a new tier added to the sunshine-source resolution chain (existing tiers are governed by an accepted ADR — re-read docs/adr/ to find which).
• A new render mode or layout (a fourth top-level mode beside Combination / Station / Forecast), or a new forecast.style variant beyond style1/style2.
• A new top-level card-config option that the user can set in their YAML — these become public API and are notoriously hard to remove.
• A build / quality gate change: ESLint rule promoted warn→error, vitest coverage threshold moved, dependency-cruiser rule added or relaxed, new CI workflow under .github/workflows/, change to the Required status checks set on master.
• A bundler or major-dependency bump: Rollup, vitest, lit, chart.js, playwright, eslint majors. The decision to take a major and the migration shape is ADR-worthy; patch/minor bumps are not.
• A boundary break: a new uplevel import in src/chart/, src/editor/, or src/utils/. Either it needs to be removed, or the boundary itself needs an ADR change.
• A change to the e2e baseline regeneration flow (the update-baselines.yml workflow, the playwright tolerance, the WSL fallback policy — pinning is governed by an accepted ADR; re-read docs/adr/ before changing the flow).
• A release-flow change — the steps listed in CLAUDE.md for cutting vX.Y.Z.
• A pattern deviation from an accepted ADR — code that contradicts an existing decision.

Skip — do not trigger on

• Bug fixes that don't change a contract (no new public option, no new module-boundary, no new dependency).
• Refactors within an existing module that keep the public surface identical.
• Adding test coverage to existing code (including v1.5 #10 work on teardown-registry.ts).
• Style fixes, lint-warning cleanup, type-narrowing inside an already-strict file.
• E2E baseline regenerations done by the update-baselines.yml GHA bot.
• Single-file documentation edits (typo, link, prose tweak) that do not change a convention.
• Patch/minor dependency bumps via Dependabot.
• Lovelace dashboard / Bubble Card / user-side YAML examples in docs (those describe usage, not internal architecture).

Out of scope (deliberately)

The skill stays narrow. It does not:

• Fire on conversational mentions of "bug" or "idea". Filing a GitHub issue is a deliberate user act, not a documentation event.
• Run gh issue create or manage labels.
• Police uncommitted working-tree state outside docs/ and the ADR check.
• Verify cross-links in issue bodies on GitHub.

Documentation locations

When suggesting where information should live, use these targets:

Gating test — apply before proposing

The activation triggers above are detection signals, not auto-suggestions. After a trigger fires, gate the proposal by the restrictive AND-of-three filter:

1. Hard to reverse — the cost of changing the decision later is meaningful.
2. Surprising without context — a future reader will look at the code and wonder "why on earth did they do it this way?"
3. Result of a real trade-off — there were genuine alternatives and you picked one for specific reasons.

All three must be true. If any one is missing, skip the ADR — the rationale belongs in a commit message, a code comment, the CHANGELOG.md, or simply in the diff itself.

• Easy to reverse → just reverse it later.
• Not surprising → nobody will wonder why.
• No real alternative → "we did the obvious thing" isn't worth recording.

This filter mirrors the one in the user-level grill-with-docs skill, so both stay aligned on what counts as ADR-worthy.

Proactive prompting

When a trigger fires and the AND-of-three gate passes, surface it before implementation, not after:

This change introduces / modifies / adds X. That's a deliberate architectural choice with genuine alternatives and meaningful reversal cost — should I draft an ADR for it before continuing?

Suggested: docs/adr/NNNN-descriptive-title.md.

If the user agrees, draft the ADR using docs/adr/template.md in the same PR as the code change. If the user defers ("not now, later"), respect that — do not nag in the same session.

For architecture suggestions in general (per CLAUDE.md): list pros/cons, then a recommendation, then act only after OK.

Compliance check

Before implementation

1. List docs/adr/ fresh — never assume a snapshot of the in-force set; new ADRs land regularly and a hardcoded list in this skill would silently drift.
2. Read each accepted ADR whose title or Decision section overlaps the staged paths or the planned change's surface area. Skip template.md and README.md.
3. Verify the planned change does not contradict any of them.
4. If the planned change contradicts an accepted ADR, raise it explicitly:

   This approach differs from ADR 000N (<title>) which decided X. Two options: (a) adjust the implementation to match the ADR, (b) write a superseding ADR. Which one?

After implementation

1. Note any significant undocumented decisions made during the change.
2. Suggest an ADR for each, naming the file path.

ADR mechanics

Required structure

Mirrors docs/adr/template.md:

# NNNN: Title

**Status:** Proposed | Accepted | Deprecated | Superseded by NNNN

**Date:** YYYY-MM-DD

## Context
## Decision
## Consequences
  - Pros / Cons / Tradeoffs
## Related

Numbering

• Sequential four-digit numbers: 0001, 0002, …
• Find the current highest number with ls docs/adr/.
• Never reuse a number, even for deprecated or superseded ADRs.

Superseding

When a new decision overrides an old one:

• The old ADR's status becomes Superseded by NNNN.
• The old ADR file is not deleted — history is preserved.
• The new ADR's status is Accepted and its Related section links back to the superseded one.

Behavioural guidelines

• Lightweight, not bureaucratic. Suggest ADRs only when the change is genuinely a decision, not when the answer is obvious from the code or already documented.
• "Should I draft an ADR?" pattern, not auto-generation. Wait for a yes.
• Respect deferrals. If the user says "later" or "no", drop it.
• Pro/contra first. When recommending an architecture or tool choice, list alternatives, then a recommendation, then wait for OK before implementing.
• Connect the dots. Link new ADRs to related ones; reference the ADR from the relevant code section in ARCHITECTURE.md or the corresponding user doc.
• English only in any file written to disk. Conversation can stay German.

Source: chriguschneider/weather-station-card — distributed by TomeVault.