feat(skills): add Codex discovery for Brain agents

AGENTS.md

@@ -126,12 +126,12 @@ humans invoke identically, so behaviour isn't re-derived from prose each time.
   leg → re-judge — until Heart goes GREEN. Delegates all dispatch to the
   release conductor. Current scope is *validate + recommend*, checkpointing
   every dispatch; edit-in fixes are an explicit follow-up. (Skeleton.)
-- **`agents/conductors/clone/`** — the *Mitosis Agent* (**design only** — see
-  its `DESIGN.md`; no CLI yet): will reproduce a mature domain assistant
+- **`agents/conductors/clone/`** — the *Mitosis Agent* (**analysis-only v0** —
+  see its `DESIGN.md`): analyzes how to reproduce a mature domain assistant
   (reference: `autolens_assistant`) into a new specialised assistant cell —
   domain analysis, template-boundary partition, a `CloneDecision` with an
   exact-clone/sibling/seed question, generation delegated to Build, newborn
-  validation by Heart.
+  validation by Heart. Its current CLI emits the decision and writes nothing.
 
 ### Faculties
 
@@ -203,9 +203,9 @@ it resolves the sibling `pyauto-heart` and `autobuild` binaries from PATH or the
 ## The command surface (Brain implicit)
 
 The `bin/pyauto-brain <agent>` CLI above is the machinery; humans drive it through
-short verb commands installed into `~/.claude/commands/`. The Brain stays
-**implicit** — you type a verb (or plain natural language) and the Brain routes it
-to the right agent; normal usage never says "PyAutoBrain".
+short verb commands in Claude Code or discoverable skills in Claude and Codex.
+The Brain stays **implicit** — you type a verb (or plain natural language) and
+the Brain routes it to the right agent; normal usage never says "PyAutoBrain".
 
 > **Users speak in short commands; PyAutoBrain performs the routing.**
 
@@ -228,10 +228,11 @@ never for symmetry — the Refactor Agent earned its promotion via the `ideas.md
 backlog bullet, the skill's own recorded follow-up, and the autonomy series.
 Every command routes **through** the Brain; none replaces it.
 
-The command bodies live in `skills/<verb>/<verb>.md` (thin; installed as flat
-commands by `bin/install.sh`); the shared architecture prose is in
-[`skills/COMMANDS.md`](skills/COMMANDS.md). The work-type taxonomy the router and
-work-type entries use is `PyAutoMind/ROUTING.md`.
+The command bodies live in `skills/<verb>/<verb>.md`; thin `SKILL.md` wrappers
+make the same canonical workflows discoverable to skill-aware harnesses.
+`bin/install.sh` installs both surfaces without duplicating their bodies. Shared
+architecture prose is in [`skills/COMMANDS.md`](skills/COMMANDS.md). The
+work-type taxonomy the router and work-type entries use is `PyAutoMind/ROUTING.md`.
 
 ## Never rewrite history
 

bin/README.md

@@ -3,7 +3,7 @@
 Executable tooling for the PyAuto organism.
 
 - **`pyauto-brain`** — the PyAutoBrain CLI (reasoning-layer entry point).
-- **`install.sh`** — the cross-organ Claude skill/command **installer**.
+- **`install.sh`** — the cross-organ Claude/Codex skill and command installer.
 - **`check_skill_line_counts.sh`** — the skill line-count **guard** (primary
   skill `.md` files must stay < 200 lines).
 
@@ -14,7 +14,8 @@ admin_jammy hosts no skills and is slated to leave `PyAutoLabs/`.
 ## install.sh
 
 Auto-discovers (no hardcoded skill list) by scanning each organ repo's
-`skills/` dir for `*/` subdirs, then symlinks each into `~/.claude/`:
+`skills/` dir for `*/` subdirs, then symlinks each supported surface into its
+harness home:
 
 - **`PyAutoMind/skills/`** — registry-coupled skills (`create_issue`).
 - **`PyAutoBrain/skills/`** — development-workflow skills (`start_dev`,
@@ -32,15 +33,17 @@ Auto-discovers (no hardcoded skill list) by scanning each organ repo's
 - **`admin_jammy/skills/`** — vestigial: hosts no skills; scanned only so an old
   checkout still resolves, auto-skipped once admin_jammy leaves.
 
-Discovery rule per root:
+Discovery rules per root are independent, so a directory may expose both:
 
-- A directory with `SKILL.md` → **skill**, symlinked to `~/.claude/skills/<name>/`.
-- A directory with `<dirname>.md` (and no `SKILL.md`) → **command**, symlinked to
-  `~/.claude/commands/<name>.md`.
+- `SKILL.md` → skill directory symlinked to Claude under its source-directory
+  name and to Codex under the hyphenated frontmatter `name`.
+- `<dirname>.md` → command symlinked to `~/.claude/commands/<name>.md`.
 
 Roots that aren't checked out are skipped. It is idempotent — existing symlinks
 are replaced, non-symlink files are left alone, and broken symlinks pointing into
-a PyAuto root are pruned. Re-run it after pulling updates.
+a PyAuto root are pruned. `CLAUDE_HOME` and `CODEX_HOME` may override the
+destinations for isolated tests or non-default installations. Re-run it after
+pulling updates.
 
 ## Bootstrap on a new machine
 

bin/check_skill_line_counts.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 # Report any *primary* skill .md file over the line limit (default 200).
 #
-# Claude skill guidance keeps primary skill files short; long background,
+# Agent skill guidance keeps primary skill files short; long background,
 # templates and examples belong in supporting docs (`reference.md`, shared
 # `WORKFLOW.md`), which are exempt here. A "primary" file is a skill's `SKILL.md`
 # or its command body `<dirname>.md`.

bin/install.sh

@@ -1,10 +1,11 @@
 #!/usr/bin/env bash
-# Bootstrap Claude Code skills/commands into ~/.claude/ from every PyAuto
-# organism repo that hosts skills.
+# Bootstrap Claude Code commands and Claude/Codex skills from every PyAuto
+# organism repo that hosts them.
 #
 # This installer lives in PyAutoBrain (the reasoning/orchestration organ). It
-# scans every organ repo's skills/ dir and symlinks their skills+commands into
-# ~/.claude/. Roots that aren't checked out are simply skipped.
+# scans every organ repo's skills/ dir and symlinks skills into both harnesses;
+# Claude-only command files remain available as slash commands. Roots that
+# aren't checked out are simply skipped.
 #
 # Discovery roots (scanned in order):
 #   - admin_jammy/skills/   — vestigial: admin_jammy hosts no skills and is
@@ -19,10 +20,10 @@
 # (PyAutoBuild's root is for its own release/packaging-execution skills only — it
 #  owns NO dev-workflow skills; the ship_* skills merely call its release step.)
 #
-# Auto-discovers skills in each root:
-#   - A directory with SKILL.md → installed as ~/.claude/skills/<name>/
-#   - A directory with <name>.md (and no SKILL.md) → installed as a flat
-#     command at ~/.claude/commands/<name>.md
+# Auto-discovers surfaces in each root independently:
+#   - SKILL.md → installed in both the Claude and Codex skill roots
+#   - <name>.md → installed as a flat Claude command
+# A directory may contain both; neither surface suppresses the other.
 #
 # Safe to re-run — existing symlinks are replaced, non-symlink files are skipped.
 #
@@ -32,13 +33,16 @@
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
-PYAUTO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+DEFAULT_PYAUTO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+PYAUTO_ROOT="${PYAUTO_ROOT:-$DEFAULT_PYAUTO_ROOT}"
 ADMIN_SKILLS_DIR="$PYAUTO_ROOT/admin_jammy/skills"
 MIND_SKILLS_DIR="$PYAUTO_ROOT/PyAutoMind/skills"
 BRAIN_SKILLS_DIR="$PYAUTO_ROOT/PyAutoBrain/skills"
 HEART_SKILLS_DIR="$PYAUTO_ROOT/PyAutoHeart/skills"
 BUILD_SKILLS_DIR="$PYAUTO_ROOT/PyAutoBuild/skills"
 PROFILING_SKILLS_DIR="$PYAUTO_ROOT/autolens_profiling/skills"
+CLAUDE_HOME="${CLAUDE_HOME:-$HOME/.claude}"
+CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
 
 # ---------- Execution-environment note ----------
 #
@@ -53,7 +57,7 @@ else
 fi
 echo ""
 
-mkdir -p "$HOME/.claude/skills" "$HOME/.claude/commands"
+mkdir -p "$CLAUDE_HOME/skills" "$CLAUDE_HOME/commands" "$CODEX_HOME/skills"
 
 # ---------- Prune stale symlinks ----------
 #
@@ -76,8 +80,9 @@ prune_stale_symlinks() {
   done
 }
 
-prune_stale_symlinks "$HOME/.claude/skills"
-prune_stale_symlinks "$HOME/.claude/commands"
+prune_stale_symlinks "$CLAUDE_HOME/skills"
+prune_stale_symlinks "$CLAUDE_HOME/commands"
+prune_stale_symlinks "$CODEX_HOME/skills"
 
 # ---------- Install one source dir's skills/commands ----------
 
@@ -101,17 +106,27 @@ install_from_dir() {
     # Skip the install.sh dir itself, or any non-skill dirs
     [ "$name" = "skills" ] && continue
 
+    local found=0
     if [ -f "$entry/SKILL.md" ]; then
-      # Skill — symlink the directory
-      local dst="$HOME/.claude/skills/$name"
-      _link_symlink "$entry" "$dst" "skill"
+      _link_symlink "$entry" "$CLAUDE_HOME/skills/$name" "Claude skill"
       installed_count=$((installed_count + 1))
-    elif [ -f "$entry/$name.md" ]; then
-      # Command — symlink the flat .md file
-      local dst="$HOME/.claude/commands/$name.md"
-      _link_symlink "$entry/$name.md" "$dst" "command"
+      found=1
+
+      local skill_name
+      skill_name="$(sed -n 's/^name:[[:space:]]*//p' "$entry/SKILL.md" | head -1)"
+      if [ -z "$skill_name" ] || [[ "$skill_name" == *[!a-z0-9-]* ]]; then
+        echo "  SKIP $name (Codex skill name invalid or missing: ${skill_name:-<empty>})"
+      else
+        _link_symlink "$entry" "$CODEX_HOME/skills/$skill_name" "Codex skill"
+        installed_count=$((installed_count + 1))
+      fi
+    fi
+    if [ -f "$entry/$name.md" ]; then
+      _link_symlink "$entry/$name.md" "$CLAUDE_HOME/commands/$name.md" "Claude command"
       installed_count=$((installed_count + 1))
-    else
+      found=1
+    fi
+    if [ "$found" -eq 0 ]; then
       echo "  SKIP $name (no SKILL.md or $name.md found)"
     fi
   done
@@ -149,4 +164,4 @@ install_from_dir "$PROFILING_SKILLS_DIR" "autolens_profiling/skills/ — science
 
 # ---------- Summary ----------
 
-echo "Done. Restart Claude Code to pick up changes."
+echo "Done. Restart Claude Code and Codex to pick up changes."

skills/COMMANDS.md

@@ -61,11 +61,18 @@ symmetry. The taxonomy they tag is `PyAutoMind/ROUTING.md`.
 - **`/brain <agent> [args]`** — explicit, un-veneered passthrough to
   `bin/pyauto-brain` for debugging. Free-text `/brain` defers to `/route`.
 
+Codex skills also expose the remaining public CLI agents directly: the
+analysis-only `clone` conductor, the `release` conductor, and the read-only
+`vitals`, `review`, `memory`, and `samplers` faculties. They do not gain new
+slash commands; `brain` remains Claude's low-level passthrough.
+
 ## How these are installed
 
-Each command is a directory `skills/<verb>/` containing only `<verb>.md` (no
-`SKILL.md`). `bin/install.sh` turns that into a flat `~/.claude/commands/<verb>.md`
-symlink — a typed slash command, not an auto-triggered skill. This file
-(`COMMANDS.md`) sits at the `skills/` root, so the installer's directory scan
-skips it: it is reference-only. Keep command bodies short and this file the single
-place architecture prose lives (guarded by `bin/check_skill_line_counts.sh`).
+Each command directory keeps `<verb>.md` as its canonical command body and may
+add a thin `SKILL.md` discovery wrapper. `bin/install.sh` treats the two files
+independently: it installs the command into `~/.claude/commands/` and the skill
+into both `~/.claude/skills/` and `~/.codex/skills/` (using the skill's
+hyphenated frontmatter name for Codex). This file (`COMMANDS.md`)
+sits at the `skills/` root, so the directory scan skips it; it is reference-only.
+Keep wrappers and command bodies short and keep shared architecture prose here
+(guarded by `bin/check_skill_line_counts.sh`).

skills/WORKFLOW.md

@@ -75,15 +75,13 @@ file and `PyAutoBrain/AGENTS.md`, and note that the agent was emulated.
 
 ## Model delegation (judgment tier plans, execution tier ships)
 
-The workflow skills split work across **model tiers**, not named models — the
-doctrine survives model access changing:
+The workflow skills split work across **capability tiers**, not named models or
+harness-specific tool names, so the doctrine survives model access changing:
 
-- **Judgment tier** — the strongest model available to the session (currently
-  **Fable 5**; previously, and as fallback, **Opus**). Planning, orchestration,
-  risk judgment, anything user-facing.
-- **Execution tier** — a fast, cheap model (currently **Sonnet**) for
-  mechanical shell/git phases, delegated as subagents (`Agent` tool,
-  `model: "sonnet"`).
+- **Judgment tier** — the strongest reasoning model available to the session.
+  Planning, orchestration, risk judgment, and anything user-facing.
+- **Execution tier** — a fast, lower-cost model for mechanical shell/git phases,
+  delegated through the harness's subagent mechanism when available.
 
 The main session stays on the judgment tier; bulk execution moves to the
 execution tier — no manual model toggling.
@@ -170,8 +168,8 @@ the vitals faculty rather than re-deriving pass/fail criteria in the skill.
 - If the user gives a development task with **no** PyAutoMind prompt path,
   first write a concise prompt under the right `<work-type>/<target>/` folder
   (original request verbatim), then continue with that path.
-- Where a body delegates mechanical execution to an execution-tier subagent, a harness
-  without subagents performs the same steps directly, preserving the
+- Where a body delegates mechanical execution to an execution-tier subagent, a
+  harness without subagents performs the same steps directly, preserving the
   judgment/mechanical split above.
 
 ## Execution environments

skills/brain/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: brain
+description: Run a named PyAutoBrain conductor or faculty directly through bin/pyauto-brain. Use for low-level agent invocation, agent help, or debugging the Brain router.
+---
+
+# Brain
+
+Follow [`brain.md`](brain.md) exactly. Use [`../COMMANDS.md`](../COMMANDS.md)
+for the shared command-surface contract. Keep agent behavior in the deterministic
+CLI and its adjacent `AGENTS.md`; do not reimplement it here.

skills/brain/brain.md

@@ -8,8 +8,8 @@ Shared routing context: `PyAutoBrain/skills/COMMANDS.md`.
 
 ## Do
 
-- `/brain <agent> [args...]` where `agent ∈ {feature, build, release, health,
-  vitals}` → run `bin/pyauto-brain <agent> [args...]` and report the output
-  verbatim.
+- `/brain <agent> [args...]` where `agent ∈ {intake, feature, bug, refactor,
+  profiling, clone, build, release, health, vitals, review, memory, samplers}` →
+  run `bin/pyauto-brain <agent> [args...]` and report the output verbatim.
 - `/brain help [agent]` → run `bin/pyauto-brain help [agent]`.
 - `/brain <free text>` with no known agent → defer to **`/route`**.

skills/bug/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: bug
+description: Classify and plan PyAutoLabs bugs, regressions, failing tests, incorrect output, or PyAutoHeart findings through the PyAutoBrain Bug Agent before starting development.
+---
+
+# Bug
+
+Follow [`bug.md`](bug.md) exactly. Use [`../COMMANDS.md`](../COMMANDS.md) for
+shared routing. Let `bin/pyauto-brain bug` produce the decision; do not bypass
+the Brain or implement the repair inside this skill.

skills/bug/bug.md

@@ -24,9 +24,9 @@ Shared routing context: `PyAutoBrain/skills/COMMANDS.md`.
    carrying forward). Then tell the user it is **safe to `/clear` before the next
    `/bug`**: the transcript now holds only disposable investigation scratch (file
    reads, repro runs, tool output), so a fresh session loses nothing and context
-   stays lean across many bugs in one sitting. Claude cannot self-clear — this is a
-   deliberate one-keystroke user step, and the externalised state above is what makes
-   it lossless.
+   stays lean across many bugs in one sitting. An agent cannot clear its own host
+   transcript — this is a deliberate one-keystroke user step, and the externalised
+   state above is what makes it lossless.
 
 The Bug Agent **reasons; it never edits source**, and its first question is always
 *where the fix belongs*: prefer a general library-source fix, and **never degrade a

skills/build/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: build
+description: Coordinate PyAuto build, deploy, or release execution through the PyAutoBrain Build Agent, including its vitals consultation and PyAutoBuild delegation. Use when the user asks to build, publish, deploy, or release.
+---
+
+# Build
+
+Follow [`build.md`](build.md) exactly. Preserve the Brain -> vitals -> Heart ->
+Build boundary and all human release gates.

skills/clone/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: clone
+description: Analyze how a mature PyAuto domain assistant should be reproduced as an exact clone, sibling, or seed through the PyAutoBrain Clone Agent. Use for assistant-cell cloning decisions; the current agent is analysis-only and must not generate files.
+---
+
+# Clone
+
+Read [`../../agents/conductors/clone/AGENTS.md`](../../agents/conductors/clone/AGENTS.md)
+completely, then run `bin/pyauto-brain clone` with the documented arguments.
+Return its `CloneDecision` and preserve the agent's no-write boundary.

skills/clone/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Clone Agent"
+  short_description: "Analyze assistant cloning boundaries"
+  default_prompt: "Use $clone to analyze how a PyAuto domain assistant should be cloned."

skills/docs/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: docs
+description: Start PyAutoLabs documentation, tutorial, notebook, example, or narrative-prose work through the Brain development workflow with the docs work type fixed.
+---
+
+# Docs
+
+Follow [`docs.md`](docs.md) exactly. Route the task through `start-dev`; this is
+a work-type entry, not a separate documentation conductor.

skills/docs/docs.md

@@ -13,4 +13,5 @@ Treat the request as PyAutoMind work-type **`docs/`** — documentation, example
 notebooks, or tutorial prose. If no prompt path exists, create one under
 `PyAutoMind/docs/<target>/<name>.md` (original request verbatim), then run
 **`/start_dev`** on it. `start_dev` routes through the Brain, and the tutorial
-Opus/Sonnet split in `../WORKFLOW.md` applies. Taxonomy: `PyAutoMind/ROUTING.md`.
+The judgment/execution tier split in `../WORKFLOW.md` applies. Taxonomy:
+`PyAutoMind/ROUTING.md`.

skills/feature/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: feature
+description: Select, size, phase, or plan a PyAutoLabs feature through the PyAutoBrain Feature Agent, then route an approved task into start-dev. Use for new capabilities and next-task selection.
+---
+
+# Feature
+
+Follow [`feature.md`](feature.md) exactly. Let `bin/pyauto-brain feature`
+produce the structured decision before development starts.

skills/health/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: health
+description: Use the single PyAuto health door for readiness assessment, validation sweeps, active-work status, release-run status, or worktree status through the Brain and Heart-owned procedures.
+---
+
+# Health
+
+Follow [`health.md`](health.md) exactly. Keep health judgment in PyAutoHeart and
+dispatch through the documented Brain conductor or reference leg.

skills/intake/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: intake
+description: Turn raw PyAutoLabs ideas, bug reports, or loose requirements into classified and sized PyAutoMind prompt files through the PyAutoBrain Intake Agent. Use before start-dev when intent is not yet formalized.
+---
+
+# Intake
+
+Follow [`intake.md`](intake.md) exactly. Run the deterministic intake agent as a
+dry run first and apply only after reviewing its decision.

skills/memory/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: memory
+description: Recall cited PyAuto scientific, architectural, assistant, or task-history context through the read-only PyAutoBrain Memory Faculty. Use before substantial planning when prior decisions or domain knowledge could change the approach.
+---
+
+# Memory
+
+Read [`../../agents/faculties/memory/AGENTS.md`](../../agents/faculties/memory/AGENTS.md)
+completely, then run `bin/pyauto-brain memory "<topic>"`. Read only the cited
+pages needed for the task and preserve the documented privacy boundary.