feat(codex): add native runtime support

.gitignore

@@ -186,3 +186,4 @@ nf-benchmark/baseline.json
 # they accumulate one file per session forever. Local state, never committed.
 .claude/nf-session-seen-*.flag
 .copilot/nf-session-seen-*.flag
+.codex/nf-session-seen-*.flag

CHANGELOG.md

@@ -7,6 +7,7 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 ## [Unreleased]
 
 ### Added
+- feat(runtime): **native OpenAI Codex support.** The installer now accepts `--codex` for global or project-local installs, converts every `/nf:*` command into a discoverable `$nf:*` Codex skill, emits native TOML custom agents, writes Codex `hooks.json`, and manages nForma-owned MCP server blocks in `config.toml` without replacing user settings. Reinstall and uninstall paths migrate legacy Codex files, preserve unrelated configuration, and keep generated skills and MCP registrations idempotent. New conversion/unit coverage and a Codex virgin-install suite exercise install, hook adaptation, quorum context, MCP configuration, reinstall, and uninstall behavior.
 - feat(quorum): **opt-in thread persistence per slot** (`quorum.persistent_threads`, default false). The CE-5 full-convergence loop was stateless by design: each round spawns a fresh CLI invocation with prior-round outputs spliced into the prompt, and the convergence check diffs that prompt-injected state. With `persistent_threads: true`, slot CLIs keep a native conversation across rounds — `codex exec resume <thread_id>`, `-p --resume <session_id>`, or `-c` for CWD-scoped continue (agy, kimi). Round 1 captures the session id from stdout (codex JSONL `thread.started`, claude `--output-format json`) and persists it to `.planning/quorum/sessions/<slot>-<invocation>.json`; round 2+ reads the file and replaces the fresh argv template with the resume argv before `{prompt}` substitution. GC: every invocation sweeps session files >24h old, fail-open. Drop-guard mirrors the existing pattern (garbage `persistent_threads` warns + falls back to false; absent restores silently). New: `bin/quorum-resume.cjs` (helper), `bin/quorum-sessions-store.cjs` (scratch store), `bin/quorum-resume.test.cjs` (28 PURE tests, red-proven). Also fixes a latent path bug: `bin/call-quorum-slot.cjs` required `./config-loader` which resolves to a nonexistent path; corrected to `../hooks/config-loader`.
 
   **Empirical findings (3 task classes, codex/gpt-5.5, judged on the FINAL design produced, not on word-reuse proxies):** the prevailing assumption "thread memory always helps" is **not what the data shows**. Across a 2-round essay expansion, a 3-round plan review, and a 5-round complex synthesis, the **default `false` (stateless) is at least as good as `true`** (persistent) in 2 of 3 task classes and ties the third. `FRESH` (no context at all) always refuses with a safety message — better than hallucinating. The right rule of thumb: enable persistent when the reasoning chain spans 10+ rounds, the orchestrator's prompt-injection window is too narrow for prior rounds, or the model is *evolving* its own prior reasoning rather than answering fresh; otherwise stateless preserves the documented CE-5 "team has nothing left to add" semantics and beats the model's noisier thread memory in mid-complexity tasks. **No change for existing callers** — default `false` preserves the current stateless CE-5 semantics.

README.md

@@ -156,9 +156,15 @@ npm install @huggingface/transformers
 > Runs a local [all-MiniLM-L6-v2](https://huggingface.co/Xenova/all-MiniLM-L6-v2) model on CPU (~23MB). No API keys or cloud calls. The `/nf:proximity` pipeline automatically uses embeddings when this dependency is present.
 
 The installer prompts you to choose:
-1. **Runtime** — Claude Code, OpenCode, Gemini, or all
+1. **Runtime** — Claude Code, OpenCode, Gemini, Codex, or another supported runtime
 2. **Location** — Global (all projects) or local (current project only)
 
+> [!NOTE]
+> **Using Codex?** Choose Codex in the installer, or run
+> `npx @nforma.ai/nforma --codex --global`. nForma installs native Codex skills,
+> custom agents, hooks, and MCP servers. Invoke workflows with `$nf:help`,
+> `$nf:new-project`, and the other `$nf:*` skill names.
+
 ### 2. Use Skip-Permissions Mode
 
 nForma is designed for frictionless automation. Run Claude Code with:
@@ -201,6 +207,7 @@ If you prefer not to use that flag, add this to your project's `.claude/settings
 ```
 
 You should see the full command list. If not, restart your runtime to reload commands.
+In Codex, use `$nf:help` (or open `/skills` and select `nf:help`) instead.
 
 ### 4. Set Up Your Quorum (Optional)
 
@@ -272,12 +279,16 @@ npx @nforma.ai/nforma --opencode --global # Install to ~/.config/opencode/
 # Gemini CLI
 npx @nforma.ai/nforma --gemini --global   # Install to ~/.gemini/
 
+# OpenAI Codex
+npx @nforma.ai/nforma --codex --global    # Skills: ~/.agents/skills; config: ~/.codex/
+npx @nforma.ai/nforma --codex --local     # Skills: ./.agents/skills; config: ./.codex/
+
 # All runtimes
 npx @nforma.ai/nforma --all --global      # Install to all directories
 ```
 
 Use `--global` (`-g`) or `--local` (`-l`) to skip the location prompt.
-Use `--claude`, `--opencode`, `--gemini`, or `--all` to skip the runtime prompt.
+Use `--claude`, `--opencode`, `--gemini`, `--codex`, or `--all` to skip the runtime prompt.
 
 </details>