[ndisisnd]

Summary

kermit v4: the skill now runs natively in OpenAI Codex as well as Claude Code, from a single skill body — no fork. Codex subagents (codex exec / spawn_agent workers) can run every mode end-to-end with zero human gates and zero stalls, and report back through a structured kermit-result block.

Why this change is being made

kermit's protocols were welded to Claude Code specifics: ~25 literal AskUserQuestion calls, .claude/kermit/ paths, and hook-fed inputs. A Codex worker errors out the moment anything asks for approval, so kermit could not participate in Codex workflows at all. This PR isolates every harness-specific behaviour behind one resolution step so the same protocols serve both harnesses, interactive or not.

Specific changes

skills/kermit/SKILL.md

• New harness resolution (runs first, once): HARNESS (claude|codex), KERMIT_DIR (env → .claude/kermit/ → .codex/kermit/ → create per harness), INTERACTIVE (false under codex exec, spawned subagents, headless claude -p).
• New GATE(question, options, default) contract: interactive Claude → real AskUserQuestion tool call; interactive Codex → plain-chat question; non-interactive → stated default, silently.
• New kermit-result closing block (mode / head / changelog_entry / pushed / published / gates) emitted by every mode.

skills/kermit/refs/ (all protocols)

• Every AskUserQuestion site rewritten as GATE(...) with explicit non-interactive defaults; non-interactive runs force effective-auto gates and record gates: auto (non-interactive).
• .claude/kermit/ → $KERMIT_DIR throughout, including gitignore snippets.
• Commit step probes and seeds the changelog itself; the PreToolUse hook cache is a non-authoritative pre-warm.
• Push/PR/publish steps tolerate no-network sandboxes: pushed: failed / published: no are handled, successful terminations with remediation lines.
• protocol-init.md: zero-question bootstrap for non-interactive runs on uninitialized repos; merge-to-main guard branches per harness (Claude keeps hook wiring; Codex prints an AGENTS.md convention snippet).
• New protocol-subagent.md: the worker contract for orchestrators — preconditions per mode, result shape, handled outcomes.

skills/kermit/agents/ (new)

• openai.yaml: optional Codex skill declaration (SKILL.md alone is sufficient; verified against OpenAI docs 2026-08-07).
• kermit-agent.toml: ready-to-copy Codex subagent role (.codex/agents/kermit.toml).

Installers & CI

• install.sh / bin/kermit.js: --codex and --all flags; Codex target is ~/.agents/skills/kermit; same byte-identical skill body.
• .github/workflows/ci.yml: existence + install-sync checks for the Codex adapter files.

Docs

• README Codex section, llms.txt entry points, SECURITY.md paragraph on the non-interactive auto-gate policy (containment comes from the harness sandbox, not kermit's gates).

Other additional information

• Codex marks the subagent TOML format experimental; both adapter files carry the doc URL and verification date to make drift easy to spot.
• Statically validated against the v4 plan's acceptance criteria. Live tests still to run before merge: interactive Claude regression, Codex interactive, subagent end-to-end in a scratch repo, and the shared-repo numbering check.
• package.json description/keywords still position kermit as Claude-Code-only — left for a follow-up.