Adapter capability declarations

Per-adapter declaration of side effects: which shell commands the adapter invokes, which paths it writes outside its engine directory, which network hosts it contacts, and whether it relies on background processes. These declarations are the source of truth that the D4 capability verifier (aidokit verify --capabilities) compares against the emitted hook scripts.

Schema: see packages/core/src/schema.ts — CapabilitiesSchema. Per-project projection: .aidokit/capabilities.json at Strict tier (see ADR-0017).

How these declarations are used #

1. Authoritative source. Each adapter's manifest.capabilities field in its package's manifest.ts declares the surface area the adapter accepts responsibility for.
2. Projected to disk at Strict tier. aidokit init --tier strict writes .aidokit/capabilities.json derived from the declared capabilities of every adapter in the project.
3. Cross-checked at runtime. aidokit verify --capabilities scans the emitted hook scripts under .claude/scripts/** (etc.) and flags any command, path, or host that appears in the scripts but isn't declared in the manifest.
4. Cited in audit packets. aidokit audit export references the declarations as evidence for SOC2 CC3.1 / CC5.2 / CC9.1 and EU AI Act Articles 9 / 15.

@aidokit/adapter-codex — Minimum conformance #

Why these:

• Codex CLI's MCP wiring is not yet a stable shell-command interface; the adapter does not invoke a Codex CLI command from postInstall. (This will change when Codex stabilises its MCP surface — tracked in ROADMAP §v1.0.)
• AGENTS.md is the Codex agent-rules file (per the Codex spec). The other write paths (docs/, agent-artifacts/) are the shared substrate every adapter writes, so they are identical across all three first-party adapters.

@aidokit/adapter-copilot — Minimum conformance #

Why these:

• Copilot CLI exposes no hook mechanism we can wire MCPs through; the adapter declares Minimum conformance because that is the honest ceiling its target CLI supports, not a quality gap (see ADR-0016).
• .github/copilot-instructions.md is the Copilot project rules file per GitHub's documented convention. The other write paths follow the shared-substrate pattern every adapter uses.

@aidokit/adapter-claude-code — Strict conformance #

Why these:

• claude mcp add / claude mcp remove are run by postInstall and aidokit mcp add/remove to wire MCP servers into the Claude Code CLI itself. They are the only shell commands the adapter invokes in user-controlled scope.
• CLAUDE.md is emitted at the project root; docs/ carries the shared documentation skeleton; agent-artifacts/ carries change summaries, blockers, metrics. All three are outside the adapter's declared engine directory (.claude/) by design — see ADR-0005.
• No network endpoints: the adapter contacts no host directly. (npm, Claude Code itself, and any MCPs you opt into have their own network posture; see wiki/privacy/index.md.)
• No background processes: every hook script is short-lived and exits.
• Claude Code exposes deterministic hooks and output styles, which is what lets this adapter declare Strict.

How to verify #

# Local — checks the on-disk emitted hooks against the artifact.
aidokit verify --capabilities

# In CI — same machinery, exits non-zero on any drift.
aidokit manifest --verify-capabilities

Both commands surface drift such as:

• a shell command appearing in a hook script but not in capabilities.json
• a network call (fetch / https.get) appearing in a hook script when networkEndpoints is empty
• a setInterval in a hook script when backgroundProcesses: false

Adding a capability #

When you make an adapter change that legitimately needs a new shell command, write path, or network host:

1. Update manifest.capabilities in the adapter package's manifest.ts.
2. Run pnpm -r build && aidokit sync --tier strict (in a Strict- tier dogfood project) to refresh .aidokit/capabilities.json.
3. Run aidokit verify --capabilities to confirm the declaration matches the new behaviour.
4. Update this page so the documented capability table stays in sync with the manifest.

See also #

• ADR-0017 — the artifact emission design
• Strict tier — when the artifact is emitted
• Security model — broader context
• aidokit verify — the integrity command
• Audit export — SOC2 mapping — evidence usage