Native Tools

codexclaw’s discipline is only half the story — the Codex runtime already ships powerful native surfaces, and the skills now route to them by exact tool id. The verified inventory lives in structure/60_native_capabilities.md (re-verified per Codex release).

The two collab surfaces

V1 is codexclaw’s default, but the model catalog selects V2 for sol/terra and V1 for luna; features.multi_agent_v2 is the fallback selector for other models. The chosen surface pins on the session’s first turn. V1 collab tools (multi_agent_v1.spawn_agent / wait_agent / send_input / resume_agent / close_agent) are deferred behind tool_search; V2 tools are direct. If spawn_agent is not visible, run tool_search for it first.

V1 parses skill mentions natively. On plaintext V2 provider/proxy paths, the spawn hook normalizes mentions and inlines recognized SKILL.md bodies. Native ChatGPT-backend V2 gives the hook encrypted message ciphertext, so both operations are no-ops there. When no body can be inlined, the hook appends a plaintext [CXC-SKILL-AFFORDANCE] block telling the child to self-load any $cxc-<folder> / $codexclaw:cxc-<folder> mention from <skillsDir>/<folder>/SKILL.md; fork inheritance remains a secondary channel. The other reliable native V2 hook channels are the leaf guard and omitted configured model/effort injection on non-full-history spawns.

Hook trust

Codex pins each hook identity as hooks.state.<key>.trusted_hash in ~/.codex/config.toml and silently skips a hook whose current hash drifts. After any edit, commit, or merge touching plugins/codexclaw/hooks/*.json, run cxc doctor. If it reports a drifted or untrusted hook, run cxc hooks retrust to back up the config and atomically record safely recomputed hashes, then rerun cxc doctor.

Browse-use ladder (owned by cxc-search)

Proof-of-source escalates through named tools, stopping at the first rung that yields primary evidence:

agbrowse is the primary surface while it resolves; the native tools are its fallback tier:

1. agbrowse fetch --json --browser never — scripted HTTP proof; the mandatory first attempt (its JSON envelope is the evidence artifact)
2. agbrowse CDP — one-shot fetch --browser auto for JS/blocked pages, or an interactive session (start --headed → navigate → snapshot --interactive → click eN → stop) when steps must act on the page
3. Native fallback — browser:control-in-app-browser (JS/PDF/visual) and chrome:control-chrome (conversational real-profile CDP), used when agbrowse is unresolvable or cannot complete the flow (state why)
4. computer-use:computer-use — GUI-only last resort (per-app approval)

Every interactive rung follows the verification loop: inspect → act → re-inspect, with screenshot + view_image fallback when DOM inspection fails.

Computer-use QA (owned by cxc-dev-testing §4.6)

Playwright owns deterministic suites; the native tools own exploratory QA — “does this change work in the real UI right now”. Drive the flow, capture screenshots, read them back with view_image, and attach them as PABCD C-phase evidence. Promote flows that must stay guarded into Playwright.

Other natives now wired into skills

Tool	Where it’s used
update_plan	cxc-pabcd PLAN-TRACK-01 — mirror plan items, keep statuses live through B
imagegen	cxc-dev-frontend / cxc-dev-uiux-design — real bitmap assets instead of placeholders
view_image	design reads, screenshot evidence, blocked-source captures
multi_tool_use.parallel	cxc-sparksearch / cxc-ultraresearch parallel lanes
list_available_plugins_to_install / request_plugin_install	cxc-skill-hub capability discovery

CSV batch fan-out via spawn_agents_on_csv and memories remain flag-gated and are documented as future surfaces. V2 is live through catalog selection or the fallback feature flag; it is not part of this “not shipped” set.