Skip to content

Latest commit

 

History

History
36 lines (31 loc) · 3.99 KB

File metadata and controls

36 lines (31 loc) · 3.99 KB

Per-Runtime Capability Matrix

Spec: specs/install-path-confirmation.md AC-030 (milestone AC-027, GA criterion #3) Maintenance: hand-maintained, verified against runtimes/codex/adapter.yaml truth on every release; included in release notes from v0.10.0 onward. If this table and adapter.yaml disagree, adapter.yaml wins — fix the table.

Legend — Enforced: the runtime mechanically applies it (hook/loader; the agent cannot silently skip it). Agent-followed: shipped as instructions the agent is expected to obey; no mechanical backstop. Advisory: documentation/patterns only. : not available on that runtime.

Capability Claude Code Codex CLI Notes
Skills (27, namespaced) Enforced (plugin loader) Enforced (native Skills, ~/.codex/skills/add-*) Codex names use /add-<name>; Claude /add:<name>
Install path Marketplace (claude plugin install add@add-marketplace) curl installer (scripts/install-codex.sh) Both covered by CI install smoke as of v0.10.0
Auto-loaded rules, maturity-scaled Enforced (SessionStart load-rules.sh physically loads the maturity-appropriate set) Agent-followed (static slim manifest in AGENTS.md; maturity-loader rule read at runtime) Codex has no session hook doing physical selection
Stale-rule-copy detection Enforced (per-session warning, v0.9.11) Codex AGENTS.md merged at init can drift; check planned (/add-version), v1.1 candidate
Hooks (learnings filter, CHANGELOG, autofix) Enforced (PreToolUse / SessionStart) Agent-followed→Enforced only if user enables [features] codex_hooks = true; hook stderr not surfaced to the agent (F-012) hooks.json emitted in the ≥0.14x schema since v0.10.1 (#24); merge is manual when one already exists
Prompt-injection scanning Enforced detection, warn-only response (PostToolUse scanner emits ADD-SEC warnings + audit events at .add/security/injection-events.jsonl; never blocks by design) Advisory (pattern catalog ships at ~/.codex/add/security/; no scanner hook, no audit events) Per adapter.yaml truth-pass (v0.9.8); Codex parity revisits v1.1
Secrets scanning / redaction Enforced in hook pipeline (lib/scan-secrets.sh via scanner + learnings filter) Advisory (library + catalog ship; skills reference them, no hook auto-registration)
Sub-agents (add-test-writer, add-implementer, add-reviewer, add-verify, add-explorer) Enforced (Task tool dispatch; Claude role vocabulary stays unprefixed) Agent-followed (TOML defs; requires [features] collab = true; developer_instructions load the role's skill) TOMLs emitted in the ≥0.14x schema since v0.10.1 (#24 — prompt_skill era was silently ignored); verified accepted by CLI 0.144.5. Registered names prefixed add-* since v0.11.0 (#28); installer removes ADD-owned legacy names
Learnings persistence + active view Enforced (filter-learnings.sh hook auto-registered) Shipped but NOT auto-registered Codex user must wire the hook manually
Telemetry JSONL emission Spec'd; emission verification tracked as D6 (see milestone) Several PRD metrics depend on this
Migrations (migrations.json hop chain) Enforced via /add:version Agent-followed via /add-version Chain continuity checked by release-evidence script (v0.10.0)
GPG-signed releases n/a (repo-level) n/a (repo-level) All tags signed; see docs/release-signing.md

Pinned Codex CLI (CI target): see codex_cli_version in runtimes/codex/adapter.yaml (0.144.5 as of v0.10.0; minimum supported: 0.140.0 — the developer_instructions agent-role schema and nested hooks.json schema ADD emits since v0.10.1 do not exist on older CLIs; see #26).

Summary for release notes: on Claude Code, ADD's quality gates are mechanically enforced by hooks; on Codex CLI they are agent-followed unless the user enables Codex hook features, and prompt-injection defense is advisory-only. Anything marked warn-only here is stated identically in SECURITY.md.