diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index e08d9ae..da93a43 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -75,6 +75,9 @@ jobs:
       - name: Check gs launches what gs sync updates
         run: bash tests/gs_test.sh
 
+      - name: Check Pi model profiles sync
+        run: python3 tests/sync_pi_model_profiles_test.py
+
       - name: Check Nix Home Manager preflight contracts
         run: bash scripts/test-nix-preflight.sh
 
diff --git a/README.md b/README.md
index 770fd9a..1bdd008 100644
--- a/README.md
+++ b/README.md
@@ -82,6 +82,15 @@ find ~ -maxdepth 4 -name '*.bak_[0-9]*' -print
 Si un destino es un directorio real (no symlink), el instalador aborta antes de
 tocar nada: movelo o borralo y volvé a correr `install.sh`.
 
+## Routing de modelos de Pi (gentle-pi profiles)
+
+`pi/gentle-ai/profiles.json` versiona los profiles de modelos de gentle-pi (`/gentle:profiles`): qué modelo usa el orquestador y cada subagente. `./install.sh` corre `scripts/sync-pi-model-profiles.py`, que:
+
+- mezcla esos profiles en `~/.pi/gentle-ai/profiles.json` y `~/.gentle-shell/gentle-ai/profiles.json`. Los profiles del repo reemplazan a los locales con el mismo nombre; los profiles que sólo existen en local se conservan, igual que `current`;
+- materializa el profile activo en `agent/settings.json` (orquestador), `agent/subagents.json` y el frontmatter de `agent/agents/*.md`.
+
+Cada archivo que cambia queda con backup `*.bak-profiles-<fecha>`. Correrlo dos veces no cambia nada. Para editar el routing, modificá el JSON del repo y corré `python3 scripts/sync-pi-model-profiles.py` (`--dry-run` muestra qué cambiaría).
+
 ## Perfiles visuales opt-in
 
 `./install.sh` instala estos assets mediante symlinks individuales. No cambia el theme activo de Pi, Herdr o Ghostty; tampoco modifica Starship.
diff --git a/install.sh b/install.sh
index f875768..7af96ad 100755
--- a/install.sh
+++ b/install.sh
@@ -583,8 +583,14 @@ if command -v python3 &>/dev/null; then
     [[ "$DRY_RUN" == true ]] && pi_appearance_args+=(--dry-run)
     python3 -I "$DOTFILES/scripts/pi_appearance.py" "${pi_appearance_args[@]}" \
         || echo "  ! Pi appearance not merged; fix ~/.pi/agent/settings.json and rerun"
+    # gentle-pi model routing: merge pi/gentle-ai/profiles.json into ~/.pi and
+    # ~/.gentle-shell and materialize the active profile.
+    pi_profiles_args=()
+    [[ "$DRY_RUN" == true ]] && pi_profiles_args+=(--dry-run)
+    python3 -I "$DOTFILES/scripts/sync-pi-model-profiles.py" "${pi_profiles_args[@]}" \
+        || echo "  ! Pi model profiles not synced; fix the gentle-ai profile files and rerun"
 else
-    echo "  ! python3 unavailable; skipped Pi appearance merge"
+    echo "  ! python3 unavailable; skipped Pi appearance merge and model profiles"
 fi
 run_or_plan "chmod +x $DOTFILES/claude/statusline.sh" chmod +x "$DOTFILES/claude/statusline.sh"
 run_or_plan "chmod +x agent-state CLI" chmod +x "$DOTFILES/zsh/scripts/agent-state.sh"
diff --git a/pi/gentle-ai/profiles.json b/pi/gentle-ai/profiles.json
new file mode 100644
index 0000000..0d1ec7d
--- /dev/null
+++ b/pi/gentle-ai/profiles.json
@@ -0,0 +1,553 @@
+{
+  "kind": "gentle-pi.agent_model_profiles",
+  "version": 1,
+  "active": "openai",
+  "profiles": {
+    "openai": {
+      "orchestrator": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "gentle-ai-explore": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "high"
+      },
+      "review-readability": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "medium"
+      },
+      "gentle-ai-verify": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "review-validator": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "review-risk": {
+        "model": "openai/gpt-6-astra",
+        "thinking": "high"
+      },
+      "review-reliability": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "review-resilience": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "review-refuter": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "xhigh"
+      },
+      "jd-judge-a": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "jd-judge-b": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "max"
+      },
+      "cortex-adversarial-tester": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "low"
+      }
+    },
+    "nan": {
+      "orchestrator": {
+        "model": "nan/glm5.3",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-risk": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-refuter": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-validator": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-reliability": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-resilience": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "review-readability": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "jd-judge-a": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "jd-judge-b": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "gentle-ai-verify": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "gentle-ai-explore": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "cortex-adversarial-tester": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "nan/glm5.3",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "nan/mimo-v2.6-flash",
+        "thinking": "low"
+      }
+    },
+    "openai-nan": {
+      "orchestrator": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "gentle-ai-explore": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "gentle-ai-verify": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "review-readability": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "review-risk": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-reliability": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-resilience": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "review-refuter": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-validator": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "jd-judge-a": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "jd-judge-b": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "cortex-adversarial-tester": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "nan/mimo-v2.6-flash",
+        "thinking": "low"
+      }
+    },
+    "claude": {
+      "orchestrator": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "review-risk": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "jd-judge-a": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "gentle-ai-verify": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "medium"
+      },
+      "review-validator": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "medium"
+      },
+      "review-refuter": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "xhigh"
+      },
+      "gentle-ai-explore": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "high"
+      },
+      "review-readability": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "review-reliability": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "high"
+      },
+      "review-resilience": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "high"
+      },
+      "jd-judge-b": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "high"
+      },
+      "cortex-adversarial-tester": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "low"
+      }
+    },
+    "claude-nan": {
+      "orchestrator": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "review-readability": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "gentle-ai-explore": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "review-refuter": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "gentle-ai-verify": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "review-resilience": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "jd-judge-b": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-risk": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-validator": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "jd-judge-a": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-reliability": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "cortex-adversarial-tester": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "nan/mimo-v2.6-flash",
+        "thinking": "low"
+      }
+    },
+    "claude-openai": {
+      "orchestrator": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "gentle-ai-explore": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "review-readability": {
+        "model": "openai/gpt-6-luna",
+        "thinking": "medium"
+      },
+      "review-risk": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "jd-judge-a": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "gentle-ai-verify": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "review-reliability": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "review-resilience": {
+        "model": "openai/gpt-6.1-terra",
+        "thinking": "high"
+      },
+      "review-refuter": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "xhigh"
+      },
+      "review-validator": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "medium"
+      },
+      "jd-judge-b": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "cortex-adversarial-tester": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "claude-bridge/claude-opus-5-5",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "claude-bridge/claude-sonnet-5-5",
+        "thinking": "low"
+      }
+    },
+    "nan-qa-reliability-sol": {
+      "orchestrator": {
+        "model": "nan/glm5.3",
+        "thinking": "low"
+      },
+      "gentle-ai-worker": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "jd-fix-agent": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-risk": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-refuter": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "review-validator": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "review-reliability": {
+        "model": "openai/gpt-6.1-sol",
+        "thinking": "high"
+      },
+      "review-resilience": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "review-readability": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "jd-judge-a": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "jd-judge-b": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "gentle-ai-verify": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "gentle-ai-explore": {
+        "model": "nan/qwen3.8-flash",
+        "thinking": "medium"
+      },
+      "cortex-adversarial-tester": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "high"
+      },
+      "cortex-requirements-analyst": {
+        "model": "nan/glm5.3",
+        "thinking": "high"
+      },
+      "cortex-evidence-auditor": {
+        "model": "nan/glm5.3",
+        "thinking": "medium"
+      },
+      "cortex-test-designer": {
+        "model": "nan/deepseek-v4-flash",
+        "thinking": "medium"
+      },
+      "cortex-exploratory-tester": {
+        "model": "nan/glm5.3-flash",
+        "thinking": "medium"
+      },
+      "cortex-test-runner": {
+        "model": "nan/mimo-v2.6-flash",
+        "thinking": "low"
+      }
+    }
+  }
+}
diff --git a/scripts/sync-pi-model-profiles.py b/scripts/sync-pi-model-profiles.py
new file mode 100755
index 0000000..35003a8
--- /dev/null
+++ b/scripts/sync-pi-model-profiles.py
@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+"""Sync the versioned gentle-pi model profiles into each Pi home.
+
+Source of truth: pi/gentle-ai/profiles.json (the gentle-pi
+`agent_model_profiles` store, without the machine-local `current` snapshot).
+
+For every home (default: ~/.pi and ~/.gentle-shell) that has an `agent/` dir:
+  1. Merge the repo profiles into <home>/gentle-ai/profiles.json. Repo profiles
+     replace same-name local ones; local-only profiles and `current` are kept.
+     The local `active` profile is kept when it still exists. A home without a
+     store starts on the empty `current` profile, so its routing is untouched.
+  2. Materialize the active profile the way gentle-pi does, so the routing is
+     live without opening /gentle:profiles:
+       - orchestrator -> agent/settings.json defaultProvider/defaultModel/defaultThinkingLevel
+       - each agent   -> agent/subagents.json model_profiles.<agent> {model, effort}
+                         and agent/agents/<agent>.md frontmatter model/thinking
+
+Every changed file is backed up as <file>.bak-profiles-<timestamp> and written
+atomically. Running it twice changes nothing.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import shutil
+import sys
+import time
+from pathlib import Path
+
+ORCHESTRATOR = "orchestrator"
+STAMP = time.strftime("%Y%m%dT%H%M%S")
+
+
+def load_json(path: Path, default):
+    if not path.exists():
+        return default
+    return json.loads(path.read_text())
+
+
+def write_if_changed(path: Path, text: str, dry_run: bool, changes: list[str]) -> None:
+    old = path.read_text() if path.exists() else None
+    if old == text:
+        return
+    changes.append(str(path))
+    if dry_run:
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if old is not None:
+        shutil.copy2(path, f"{path}.bak-profiles-{STAMP}")
+    tmp = path.with_name(path.name + ".tmp")
+    tmp.write_text(text)
+    if old is not None:
+        os.chmod(tmp, path.stat().st_mode & 0o777)
+    os.replace(tmp, path)
+
+
+def dump(data) -> str:
+    return json.dumps(data, indent=2) + "\n"
+
+
+def merge_store(local: dict, repo: dict) -> dict:
+    profiles = dict(local.get("profiles", {}))
+    profiles.update(repo["profiles"])
+    active = local.get("active")
+    if not local:
+        # A home that never used profiles keeps its current routing: start on
+        # the empty `current` snapshot and let the person pick a profile.
+        profiles.setdefault("current", {})
+        active = "current"
+    elif active not in profiles:
+        active = repo["active"]
+    return {"kind": repo["kind"], "version": repo["version"], "active": active, "profiles": profiles}
+
+
+def set_frontmatter(text: str, key: str, value: str) -> str:
+    match = re.match(r"^---\n(.*?)\n---\n", text, re.S)
+    if not match:
+        return text
+    head = match.group(1)
+    line = re.compile(rf"^{key}: .*$", re.M)
+    head = line.sub(f"{key}: {value}", head) if line.search(head) else head
+    return text[: match.start(1)] + head + text[match.end(1) :]
+
+
+def materialize(agent_dir: Path, profile: dict, dry_run: bool, changes: list[str]) -> None:
+    orch = profile.get(ORCHESTRATOR)
+    if orch:
+        settings_path = agent_dir / "settings.json"
+        settings = load_json(settings_path, {})
+        provider, _, model = orch["model"].partition("/")
+        settings["defaultProvider"] = provider
+        settings["defaultModel"] = model
+        if orch.get("thinking"):
+            settings["defaultThinkingLevel"] = orch["thinking"]
+        write_if_changed(settings_path, dump(settings), dry_run, changes)
+
+    sub_path = agent_dir / "subagents.json"
+    sub = load_json(sub_path, {})
+    profiles = sub.setdefault("model_profiles", {})
+    for agent, entry in profile.items():
+        if agent == ORCHESTRATOR:
+            continue
+        current = profiles.setdefault(agent, {})
+        current["model"] = entry["model"]
+        if entry.get("thinking"):
+            current["effort"] = entry["thinking"]
+        md = agent_dir / "agents" / f"{agent}.md"
+        if md.exists():
+            text = set_frontmatter(md.read_text(), "model", entry["model"])
+            if entry.get("thinking"):
+                text = set_frontmatter(text, "thinking", entry["thinking"])
+            write_if_changed(md, text, dry_run, changes)
+    write_if_changed(sub_path, dump(sub), dry_run, changes)
+
+
+def sync_home(home: Path, repo: dict, dry_run: bool) -> list[str]:
+    agent_dir = home / "agent"
+    if not agent_dir.is_dir():
+        return []
+    changes: list[str] = []
+    store_path = home / "gentle-ai" / "profiles.json"
+    store = merge_store(load_json(store_path, {}), repo)
+    write_if_changed(store_path, dump(store), dry_run, changes)
+    active = store["profiles"][store["active"]]
+    if active:
+        materialize(agent_dir, active, dry_run, changes)
+    return changes
+
+
+def main(argv: list[str]) -> int:
+    root = Path(__file__).resolve().parent.parent
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("--source", type=Path, default=root / "pi" / "gentle-ai" / "profiles.json")
+    parser.add_argument("--home", type=Path, action="append", help="Pi home (repeatable); default ~/.pi and ~/.gentle-shell")
+    parser.add_argument("--dry-run", action="store_true")
+    args = parser.parse_args(argv)
+
+    repo = json.loads(args.source.read_text())
+    homes = args.home or [Path.home() / ".pi", Path.home() / ".gentle-shell"]
+    for home in homes:
+        changes = sync_home(home.expanduser(), repo, args.dry_run)
+        verb = "would change" if args.dry_run else "changed"
+        print(f"{home}: {verb} {len(changes)} file(s)")
+        for change in changes:
+            print(f"  {change}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))
diff --git a/tests/sync_pi_model_profiles_test.py b/tests/sync_pi_model_profiles_test.py
new file mode 100644
index 0000000..cff99cf
--- /dev/null
+++ b/tests/sync_pi_model_profiles_test.py
@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Tests for scripts/sync-pi-model-profiles.py (run: python3 tests/sync_pi_model_profiles_test.py)."""
+
+import json
+import subprocess
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+SCRIPT = ROOT / "scripts" / "sync-pi-model-profiles.py"
+
+REPO = {
+    "kind": "gentle-pi.agent_model_profiles",
+    "version": 1,
+    "active": "openai",
+    "profiles": {
+        "openai": {
+            "orchestrator": {"model": "openai/gpt-6.1-sol", "thinking": "low"},
+            "review-risk": {"model": "openai/gpt-6-astra", "thinking": "high"},
+            "gentle-ai-verify": {"model": "openai/gpt-6.1-terra", "thinking": "medium"},
+        }
+    },
+}
+
+AGENT_MD = "---\nname: review-risk\nmodel: openai/gpt-6.1-sol\nthinking: low\ntools: []\n---\nBody model: untouched\n"
+
+
+class SyncPiModelProfilesTest(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.base = Path(self.tmp.name)
+        self.source = self.base / "profiles.json"
+        self.source.write_text(json.dumps(REPO))
+        self.home = self.base / "home"
+        agent = self.home / "agent"
+        (agent / "agents").mkdir(parents=True)
+        (agent / "settings.json").write_text(json.dumps({"theme": "refactoria", "defaultModel": "gpt-5.5", "defaultProvider": "openai"}))
+        (agent / "subagents.json").write_text(json.dumps({"default_effort": "high", "model_profiles": {"review-risk": {"model": "openai/gpt-6.1-sol", "effort": "low"}, "kb-feeder": {"model": "openai/gpt-6-luna"}}}))
+        (agent / "agents" / "review-risk.md").write_text(AGENT_MD)
+        (self.home / "gentle-ai").mkdir()
+        (self.home / "gentle-ai" / "profiles.json").write_text(json.dumps({"kind": "gentle-pi.agent_model_profiles", "version": 1, "active": "openai", "profiles": {"current": {}, "local-only": {"orchestrator": {"model": "nan/glm5.3"}}, "openai": {"orchestrator": {"model": "openai/gpt-5.5"}}}}))
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def run_sync(self, *extra):
+        return subprocess.run([sys.executable, "-I", str(SCRIPT), "--source", str(self.source), "--home", str(self.home), *extra], capture_output=True, text=True, check=True)
+
+    def read(self, rel):
+        return (self.home / rel).read_text()
+
+    def test_merges_store_and_keeps_local_profiles(self):
+        self.run_sync()
+        store = json.loads(self.read("gentle-ai/profiles.json"))
+        self.assertEqual(store["profiles"]["openai"], REPO["profiles"]["openai"])
+        self.assertIn("local-only", store["profiles"])
+        self.assertIn("current", store["profiles"])
+        self.assertEqual(store["active"], "openai")
+
+    def test_materializes_active_profile(self):
+        self.run_sync()
+        settings = json.loads(self.read("agent/settings.json"))
+        self.assertEqual((settings["defaultProvider"], settings["defaultModel"], settings["defaultThinkingLevel"]), ("openai", "gpt-6.1-sol", "low"))
+        self.assertEqual(settings["theme"], "refactoria")
+        sub = json.loads(self.read("agent/subagents.json"))
+        self.assertEqual(sub["model_profiles"]["review-risk"], {"model": "openai/gpt-6-astra", "effort": "high"})
+        self.assertEqual(sub["model_profiles"]["gentle-ai-verify"], {"model": "openai/gpt-6.1-terra", "effort": "medium"})
+        self.assertEqual(sub["model_profiles"]["kb-feeder"], {"model": "openai/gpt-6-luna"})
+        self.assertEqual(sub["default_effort"], "high")
+        md = self.read("agent/agents/review-risk.md")
+        self.assertIn("model: openai/gpt-6-astra\nthinking: high\n", md)
+        self.assertTrue(md.endswith("Body model: untouched\n"))
+
+    def test_second_run_is_a_no_op(self):
+        self.run_sync()
+        out = self.run_sync().stdout
+        self.assertIn("changed 0 file(s)", out)
+
+    def test_dry_run_writes_nothing(self):
+        before = self.read("agent/subagents.json")
+        out = self.run_sync("--dry-run").stdout
+        self.assertIn("would change", out)
+        self.assertEqual(self.read("agent/subagents.json"), before)
+        self.assertEqual(list(self.home.rglob("*.bak-profiles-*")), [])
+
+    def test_home_without_store_keeps_its_routing(self):
+        (self.home / "gentle-ai" / "profiles.json").unlink()
+        before = {rel: self.read(rel) for rel in ("agent/settings.json", "agent/subagents.json", "agent/agents/review-risk.md")}
+        self.run_sync()
+        store = json.loads(self.read("gentle-ai/profiles.json"))
+        self.assertEqual(store["active"], "current")
+        self.assertEqual(store["profiles"]["current"], {})
+        self.assertIn("openai", store["profiles"])
+        self.assertEqual({rel: self.read(rel) for rel in before}, before)
+
+    def test_home_without_agent_dir_is_skipped(self):
+        empty = self.base / "empty"
+        empty.mkdir()
+        out = subprocess.run([sys.executable, "-I", str(SCRIPT), "--source", str(self.source), "--home", str(empty)], capture_output=True, text=True, check=True).stdout
+        self.assertIn("changed 0 file(s)", out)
+        self.assertFalse((empty / "gentle-ai").exists())
+
+
+if __name__ == "__main__":
+    unittest.main()