grounding

name

grounding

description

Establish what is actually true and current for the surface you are about to change — not just search. Grounding = locate the authoritative source and reconcile sources that disagree (code vs doc, migration vs schema, memory vs current code, live vs archived), not take the first hit. Use before any grep/find/explore of the codebase or docs, when deciding which of several definitions is the real one, or when a doc or memory conflicts with the code. Covers the source-of-truth hierarchy, reconciliation discipline, scoped ripgrep, and the docs/tags.yml + docsearch.py index.

Grounding

Grounding is establishing what is actually true and current for the surface you are about to change, then acting on that — not on a guess, a memory, or the first grep hit. The hard part is not search; it is picking the authoritative source and reconciling sources that disagree.

Intentional seed. This holds only Grida-specific facts that are not obvious from CLAUDE.md/AGENTS.md. Grow it from real, discovered specifics (concept→file anchors that bite repeatedly, conflicts that actually happened and how they resolved). Do not pad it with generic search advice a competent agent already knows.

Source of truth (scoped to the surface)

Authority is per-surface — the engine model and its TS mirror can each be right for their own surface and still disagree. Name the surface, then trust:

• Canvas render/node model → Rust engine crates/grida/src/ (node/schema.rs). The TS mirror editor/grida-canvas/ is authoritative for editor behavior and can lag the engine.
• DB schema → supabase/migrations/ (applied, immutable); supabase/schemas/*.sql is a readable projection that can lag — use the database skill.
• Directory contract → the nearest AGENTS.md/README.md.
• "I remember API X…" → re-read current code; a memory is a claim about a past state, verify before acting.

Disagreement → decide which wins and why (git log -1 recency, what the running entrypoint imports, what tests assert); don't average; surface a material conflict to the user. Never authoritative even when they match: docs/_history/, docs/@designto-code/ (synced — truth is upstream), docs/cli/ (deprecated), .ref/, vendored third_party/.

Dead-tree traps

Some directories are git-tracked but dead — a bare rg from repo root returns obsolete hits that look like confirmation: docs/_history/, .ref/, vendored third_party/. (The legacy editor trees .legacy/ and packages/.legacy/ were retired May 2026 in #759; recover from the snapshot/legacy-with-2023-grida-code-editor-at-202505 branch.) rg already skips gitignored build dirs (node_modules/, target/, .next/, …); don't waste flags there. Scope positively, or exclude the dead trees:

rg PATTERN editor/grida-canvas crates/grida/src packages
rg PATTERN -g '!**/third_party/**' -g '!**/docs/_history/**'

Grounding the docs

~360 markdown files; only docs/wg/** and docs/reference/** (plus the user-facing trees in docs/AGENTS.md) are maintained. docs/tags.yml is the controlled vocabulary; tags: frontmatter is the navigation signal. Use the index, don't grep bodies — scripts/docsearch.py reads frontmatter only, self-installs via uv, runs from any cwd:

S=.agents/skills/grounding/scripts/docsearch.py
uv run $S tags                          # vocabulary + usage counts + drift
uv run $S find --tag canvas --tag svg   # AND (--any=OR); + --has K --field K=V
uv run $S show wg/feat-svg/pattern.md   # one file's frontmatter only

Before trusting a doc over code: draft: true = proposal not built yet (intent ahead); doc_tasks: or stale git log -1 = likely behind.

Related skills

database (DB source-of-truth), research (upstream/peer projects), naming (where new things belong).