diff --git a/.archgate/adrs/GEN-003-tool-invocation-via-scripts.md b/.archgate/adrs/GEN-003-tool-invocation-via-scripts.md
new file mode 100644
index 00000000..b9698fd3
--- /dev/null
+++ b/.archgate/adrs/GEN-003-tool-invocation-via-scripts.md
@@ -0,0 +1,103 @@
+---
+id: GEN-003
+title: Tool Invocation via Package Scripts
+domain: general
+rules: false
+---
+
+# Tool Invocation via Package Scripts
+
+## Context
+
+### Problem Statement
+
+AI agents and contributors working across multiple Archgate repositories encounter different toolchains: the plugins repo uses Prettier for formatting, while the CLI repo uses oxfmt. Both use OxLint for linting, but with different flags (`oxlint src/` vs `oxlint --deny-warnings .`). When agents invoke tools directly (e.g., `bunx prettier --write`, `bunx oxfmt --write`, `npx eslint`), they must know which tool each repository uses — and they routinely get it wrong, causing CI failures that require multiple push-fix-push cycles.
+
+### Pain Points
+
+Without a standardized invocation pattern:
+
+- Agents run `bunx prettier --write` in a repository that uses oxfmt, reformatting files with the wrong tool and producing CI failures on every push
+- Different repositories use different linter flags (e.g., `--deny-warnings` in the CLI repo but not in the plugins repo); invoking linters directly bypasses these project-specific configurations
+- Contributors copy-paste tool commands from one project's documentation and use them in another, introducing subtle formatting differences that pass locally but fail in CI
+- CI failures from wrong-tool invocations waste time: each push-amend-force-push cycle adds 2-5 minutes of CI latency before the real issue is identified
+- The `validate` script aggregates lint, format, typecheck, test, and build checks in a project-specific order — running individual tools misses steps and gives false confidence
+
+### Alternatives Analysis
+
+**Direct tool invocation with per-repo knowledge**: Agents and contributors memorize which tool each repository uses and invoke it directly (`bunx prettier` here, `bunx oxfmt` there). This is fragile — it requires maintaining a mental mapping of repositories to tools, breaks when a project switches formatters, and fails silently when the wrong tool is used (both produce valid output, just different output).
+
+**Wrapper scripts in each repo**: Each repository provides a `scripts/format.sh` or similar. This adds unnecessary indirection when `package.json` scripts already serve this purpose. Shell scripts are also less portable across Windows and Unix.
+
+**Package.json scripts as the sole invocation layer**: Every repository already defines `lint`, `format`, `format:check`, and `validate` scripts in `package.json`. These scripts encapsulate the correct tool, flags, and targets. Agents and contributors run `bun run format` regardless of which formatter the project uses. This is the simplest, most portable, and most robust approach.
+
+### Project-Specific Motivation
+
+For the Archgate ecosystem, the plugins repository and the CLI repository are developed in parallel, often in the same coding session. An agent working on a VS Code extension feature may commit to the plugins repo (Prettier) and the CLI repo (oxfmt) within minutes of each other. The cost of invoking the wrong formatter is not a one-time mistake — it recurs every time an agent switches context between repositories. Standardizing on `bun run format` eliminates this entire class of errors.
+
+## Decision
+
+All linting, formatting, and validation MUST be invoked through `package.json` scripts, never by running tools directly. The canonical commands are `bun run lint`, `bun run format`, `bun run format:check`, and `bun run validate`. Direct invocation of formatting or linting binaries (`bunx prettier`, `bunx oxfmt`, `npx eslint`, `oxlint .`) is prohibited in agent workflows and discouraged for manual use.
+
+**Scope**: This ADR governs how linting and formatting tools are invoked — not which tools are used. It applies to all repositories in the Archgate ecosystem, not just this one.
+
+**Required scripts**: Every `package.json` that contains lintable or formattable source code MUST define at minimum:
+
+| Script | Purpose |
+|--------|---------|
+| `lint` | Run the project's linter with project-specific flags |
+| `format` | Run the project's formatter in write mode |
+| `format:check` | Run the project's formatter in check mode (CI-safe) |
+| `validate` | Run the full validation suite (lint + format:check + typecheck + test + any other checks) |
+
+## Do's and Don'ts
+
+### Do
+
+- **DO** run `bun run lint` to lint — never invoke the linter binary directly
+- **DO** run `bun run format` to format — never invoke the formatter binary directly
+- **DO** run `bun run validate` before pushing or opening a PR — it runs the full project-specific validation suite
+- **DO** run `bun run format:check` in CI and when verifying formatting without mutating files
+- **DO** check `package.json` scripts when working in an unfamiliar repository to understand what tools and flags are configured
+- **DO** define `lint`, `format`, `format:check`, and `validate` scripts in every `package.json` that contains source code
+
+### Don't
+
+- **DON'T** run `bunx prettier`, `bunx oxfmt`, `npx eslint`, or any linter/formatter binary directly — always use `bun run format` or `bun run lint`
+- **DON'T** assume which formatter a repository uses — the `package.json` scripts abstract this; rely on the abstraction
+- **DON'T** run individual validation steps when `bun run validate` is available — it ensures the correct order and complete coverage
+- **DON'T** pass custom flags to linters or formatters outside of `package.json` — project-specific flags (e.g., `--deny-warnings`, target directories) are encoded in the scripts and MUST NOT be overridden ad hoc
+- **DON'T** add new linting or formatting tools without defining corresponding `package.json` scripts for them
+
+## Consequences
+
+### Positive
+
+- **Repository-agnostic workflow**: Agents and contributors use the same commands (`bun run format`, `bun run lint`) in every repository, regardless of which specific tools are configured
+- **Zero CI regressions from tool confusion**: The wrong formatter can never be invoked because the script abstraction selects the correct one automatically
+- **Single point of change**: When a project switches formatters (e.g., from Prettier to oxfmt), only the `package.json` scripts change — all invocation sites continue working unchanged
+- **Flag consistency**: Project-specific flags (`--deny-warnings`, target directories, ignore patterns) are defined once in `package.json` and applied uniformly
+- **Complete validation**: `bun run validate` ensures all checks run in the correct order, preventing the false confidence of running only one check
+
+### Negative
+
+- **Indirection**: Contributors cannot see which tool is running without inspecting `package.json` — the command `bun run format` does not reveal whether it invokes Prettier, oxfmt, or another formatter
+- **Script maintenance**: Every repository must maintain four scripts (`lint`, `format`, `format:check`, `validate`), adding a small amount of boilerplate to each `package.json`
+
+### Risks
+
+- **Missing scripts in new repositories**: A new Archgate repository might omit the required scripts, causing agents to fall back to direct invocation. **Mitigation:** The `archgate init` scaffolding MUST include these scripts in the generated `package.json`. Code review MUST verify their presence in any new `package.json` file.
+- **Script divergence across repos**: The `validate` script may have different steps in different repos (e.g., some include `bun run build:check`, others don't), creating inconsistent validation depth. **Mitigation:** This is acceptable — each repository's `validate` script reflects its specific needs. The invariant is that `bun run validate` always runs the *complete* set of checks for that repository.
+
+## Compliance and Enforcement
+
+**Automated enforcement**: The Archgate developer agent definition (`agents/developer.md`) MUST instruct agents to use `bun run format`, `bun run lint`, and `bun run validate` — never direct tool invocation. Agent memory records reinforce this rule across sessions.
+
+**Manual enforcement**: Code reviewers MUST reject PRs that introduce direct tool invocations (e.g., `bunx prettier --write`, `npx oxlint`) in scripts, CI workflows, or documentation. The only acceptable place for direct tool invocation is inside the `package.json` scripts themselves.
+
+**Exceptions**: No exceptions. If a tool must be invoked with non-standard flags for a one-off debugging session, that invocation MUST NOT be committed, pushed, or documented as a recommended workflow.
+
+## References
+
+- [ARCH-006: Dependency Policy](./ARCH-006-dependency-policy.md) — governs which dependencies are permitted; tools referenced by scripts must comply
+- [GEN-001: Documentation Site](./GEN-001-documentation-site.md) — related general governance ADR