Fix/codex profile responses

assets/README_ZH.md

@@ -211,7 +211,7 @@ skillclaw setup
 第一次最小化验证时，推荐这样选：
 
 - `CLI agent` 选 `none`，先不要自动改外部 agent 配置
-- `skills` 目录保持默认值 `~/.skillclaw/skills`；如果你选了 Hermes，默认技能库会变成 `~/.hermes/skills`
+- `skills` 目录保持默认值 `~/.skillclaw/skills`；如果你选了 Hermes、Codex 或 Claude Code，默认技能库会变成 `~/.hermes/skills`、`~/.codex/skills` 或 `~/.claude/skills`
 - 如果你只是想先验证代理能不能正常用，可以先关闭 shared storage
 - 如果你后面想在同一台机器上继续跑本地 evolver 闭环，就把 shared storage 打开并选 `local` backend，例如 `~/.skillclaw/local-share`
 - 如果你想先把成本压到最低，可以先关闭 PRM

pyproject.toml

@@ -17,10 +17,6 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-# Tokenizer for prompt truncation and PRM/OPD token-level data
-tokenizer = [
-    "transformers>=4.51.1",
-]
 # Embedding-based skill retrieval
 embedding = [
     "numpy",
@@ -44,7 +40,7 @@ server = [
 ]
 # Everything
 all = [
-    "skillclaw[tokenizer,embedding,evolve,sharing,server]",
+    "skillclaw[embedding,evolve,sharing,server]",
 ]
 
 [project.scripts]

skillclaw/claw_adapter.py

@@ -6,7 +6,7 @@
   openclaw  — runs `openclaw config set …` + `openclaw gateway restart`
   opencode  — patches ~/.config/opencode/opencode.json to register SkillClaw provider
   hermes    — patches ~/.hermes/config.yaml to point model traffic at SkillClaw
-  codex     — patches ~/.codex/config.toml to register SkillClaw as a provider
+  codex     — patches ~/.codex/config.toml to register an opt-in SkillClaw profile
   claude    — patches ~/.claude/settings.json to route Anthropic traffic via SkillClaw
   qwenpaw   — patches QwenPaw model config, selects SkillClaw as active model
   ironclaw  — patches ~/.ironclaw/.env, runs `ironclaw service restart`
@@ -328,6 +328,30 @@ def _upsert_top_level_toml_keys(text: str, updates: dict[str, object]) -> str:
     return "\n".join(merged).rstrip() + "\n"
 
 
+def _remove_top_level_toml_keys(text: str, keys: set[str]) -> str:
+    """Remove selected top-level assignments before the first TOML table."""
+    lines = text.splitlines()
+    first_table_index = len(lines)
+    for idx, line in enumerate(lines):
+        stripped = line.strip()
+        if stripped.startswith("[") and stripped.endswith("]"):
+            first_table_index = idx
+            break
+
+    preamble = lines[:first_table_index]
+    remainder = lines[first_table_index:]
+    kept: list[str] = []
+    for line in preamble:
+        stripped = line.strip()
+        if stripped.startswith("#") or "=" not in stripped:
+            kept.append(line)
+            continue
+        key = stripped.split("=", 1)[0].strip()
+        if key not in keys:
+            kept.append(line)
+    return "\n".join(kept + remainder).rstrip() + "\n"
+
+
 def _remove_toml_table(text: str, table_name: str) -> str:
     """Remove a TOML table and its body, if present."""
     lines = text.splitlines()
@@ -630,8 +654,21 @@ def _build_codex_provider_block(base_url: str, api_key: str) -> str:
     return "\n".join(lines) + "\n"
 
 
+def _build_codex_profile_block(model_id: str) -> str:
+    lines = [
+        "[profiles.skillclaw]",
+        f"model = {_format_toml_value(model_id)}",
+        'model_provider = "skillclaw"',
+    ]
+    return "\n".join(lines) + "\n"
+
+
 def _configure_codex(cfg: "SkillClawConfig") -> None:
-    """Auto-configure Codex CLI to use the SkillClaw proxy."""
+    """Register SkillClaw as an opt-in Codex profile.
+
+    Do not change Codex's global ``model`` / ``model_provider`` defaults.
+    Users opt in explicitly with ``codex --profile skillclaw``.
+    """
     model_id = cfg.served_model_name or cfg.llm_model_id or "skillclaw-model"
     api_key = cfg.proxy_api_key or "skillclaw"
     base_url = f"http://127.0.0.1:{cfg.proxy_port}/v1"
@@ -645,15 +682,13 @@ def _configure_codex(cfg: "SkillClawConfig") -> None:
         except Exception as e:
             logger.warning("[ClawAdapter] Failed to read Codex config %s: %s", config_path, e)
 
-    updated = _upsert_top_level_toml_keys(
-        existing_text,
-        {
-            "model": model_id,
-            "model_provider": "skillclaw",
-        },
-    )
+    updated = existing_text
+    if str(_extract_top_level_toml_value(updated, "model_provider") or "") == "skillclaw":
+        updated = _remove_top_level_toml_keys(updated, {"model", "model_provider"})
     updated = _remove_toml_table(updated, "model_providers.skillclaw").rstrip() + "\n\n"
+    updated = _remove_toml_table(updated, "profiles.skillclaw").rstrip() + "\n\n"
     updated += _build_codex_provider_block(base_url, api_key)
+    updated += "\n" + _build_codex_profile_block(model_id)
 
     _backup_codex_config_if_changed(config_path, updated)
     _write_text_atomic(config_path, updated, "Codex config")
@@ -683,10 +718,13 @@ def inspect_codex_config(cfg: "SkillClawConfig") -> dict[str, object]:
     configured_base_url = str(provider_cfg.get("base_url") or "")
     configured_wire_api = str(provider_cfg.get("wire_api") or "")
     configured_token = str(provider_cfg.get("experimental_bearer_token") or "")
+    profile_cfg = _extract_toml_table(text, "profiles.skillclaw")
+    configured_profile_model = str(profile_cfg.get("model") or "")
+    configured_profile_provider = str(profile_cfg.get("model_provider") or "")
 
     proxy_match = (
-        configured_model == expected_model
-        and configured_provider == "skillclaw"
+        configured_profile_model == expected_model
+        and configured_profile_provider == "skillclaw"
         and configured_base_url == expected_base_url
         and configured_wire_api == "responses"
         and configured_token == expected_api_key
@@ -696,7 +734,8 @@ def inspect_codex_config(cfg: "SkillClawConfig") -> dict[str, object]:
     skills_dir_match = configured_skillclaw_skills_dir == expected_skills_dir
     issues: list[str] = []
     notes: list[str] = [
-        "Codex uses the OpenAI Responses-compatible SkillClaw endpoint via `model_providers.skillclaw`.",
+        "Codex can opt into SkillClaw with `codex --profile skillclaw`.",
+        "SkillClaw registers a Codex profile and does not change Codex's global model defaults.",
         "Codex session boundaries fall back to proxy-side heuristics because"
         " Codex does not send SkillClaw session headers.",
     ]
@@ -705,8 +744,11 @@ def inspect_codex_config(cfg: "SkillClawConfig") -> dict[str, object]:
     if not config_path.exists():
         issues.append("Codex config is missing: ~/.codex/config.toml")
     if not proxy_match:
-        issues.append("Codex model routing is not pointing at the local SkillClaw proxy.")
-        next_steps.append("Start SkillClaw once with `claw_type=codex` so it can rewrite ~/.codex/config.toml.")
+        issues.append("Codex SkillClaw profile is missing or not pointing at the local SkillClaw proxy.")
+        next_steps.append("Start SkillClaw once with `claw_type=codex` so it can register ~/.codex/config.toml.")
+    if configured_provider == "skillclaw":
+        issues.append("Codex global model_provider still points at SkillClaw; normal Codex runs may be intercepted.")
+        next_steps.append("Remove top-level `model_provider = \"skillclaw\"` or run `skillclaw restore codex`.")
     if not expected_skills_dir.is_dir():
         issues.append(f"Codex skills directory is missing: {expected_skills_dir}")
         next_steps.append(f"Create or prepare the Codex skills directory: {expected_skills_dir}")
@@ -723,9 +765,12 @@ def inspect_codex_config(cfg: "SkillClawConfig") -> dict[str, object]:
         "status": "ok" if not issues else "warning",
         "config_path": str(config_path),
         "config_exists": config_path.exists(),
-        "integration_scope": "codex-only",
+        "integration_scope": "codex-profile-only",
         "expected_model": expected_model,
         "configured_model": configured_model or "(unset)",
+        "expected_profile": "skillclaw",
+        "configured_profile_model": configured_profile_model or "(unset)",
+        "configured_profile_provider": configured_profile_provider or "(unset)",
         "expected_base_url": expected_base_url,
         "configured_base_url": configured_base_url or "(unset)",
         "configured_provider": configured_provider or "(unset)",

skillclaw/config_store.py

@@ -19,6 +19,10 @@
 _DEFAULT_CODEX_SKILLS_DIR = Path.home() / ".codex" / "skills"
 _DEFAULT_CLAUDE_SKILLS_DIR = Path.home() / ".claude" / "skills"
 _DEFAULT_OPENCODE_SKILLS_DIR = Path.home() / ".config" / "opencode" / "skills"
+_DEFAULT_LLM_API_MODE_BY_CLAW = {
+    "codex": "responses",
+}
+_FALLBACK_LLM_API_MODE = "chat"
 
 _DEFAULTS: dict = {
     "llm": {
@@ -161,6 +165,12 @@ def default_skills_dir_for_claw(claw_type: str) -> Path:
     return _DEFAULT_SKILLS_DIR
 
 
+def default_llm_api_mode_for_claw(claw_type: str) -> str:
+    """Return the default upstream API mode for the selected agent."""
+    normalized = str(claw_type or "").strip().lower()
+    return _DEFAULT_LLM_API_MODE_BY_CLAW.get(normalized, _FALLBACK_LLM_API_MODE)
+
+
 def resolve_skills_dir(skills_dir: Any, *, claw_type: str) -> str:
     """Normalize a configured skills dir, applying agent-native defaults.
 
@@ -254,13 +264,14 @@ def to_skillclaw_config(self) -> SkillClawConfig:
         llm_api_base = llm.get("api_base", "")
         llm_api_key = llm.get("api_key", "")
         llm_model_id = llm.get("model_id", "")
-        llm_api_mode = str(llm.get("api_mode", "chat") or "chat")
+        raw_claw_type = str(data.get("claw_type", "openclaw") or "openclaw")
+        default_api_mode = default_llm_api_mode_for_claw(raw_claw_type)
+        llm_api_mode = str(llm.get("api_mode", default_api_mode) or default_api_mode)
         proxy = data.get("proxy", {})
         skills = data.get("skills", {})
         orouter = data.get("openrouter", {})
         prm = data.get("prm", {})
         configure_openclaw = bool(data.get("configure_openclaw", True))
-        raw_claw_type = str(data.get("claw_type", "openclaw") or "openclaw")
         if not configure_openclaw:
             raw_claw_type = "none"
 

skillclaw/setup_wizard.py

@@ -8,7 +8,7 @@
 from pathlib import Path
 
 from .claw_adapter import CLAW_TYPES
-from .config_store import CONFIG_DIR, ConfigStore, resolve_skills_dir
+from .config_store import CONFIG_DIR, ConfigStore, default_llm_api_mode_for_claw, resolve_skills_dir
 
 _PROVIDER_PRESETS = {
     "kimi": {
@@ -202,7 +202,13 @@ def run(self):
                 f"Recommended directory: {default_skills_dir}"
             )
         elif claw_type == "codex":
-            print(f"Codex reads native skills from ~/.codex/skills.\nRecommended directory: {default_skills_dir}")
+            print(
+                "Codex will get a SkillClaw profile without changing its global defaults.\n"
+                "After starting SkillClaw, run: codex --profile skillclaw\n"
+                "Normal `codex` runs remain unchanged.\n"
+                "Codex reads native skills from ~/.codex/skills.\n"
+                f"Recommended directory: {default_skills_dir}"
+            )
         elif claw_type == "claude":
             print(
                 f"Claude Code reads native skills from ~/.claude/skills.\nRecommended directory: {default_skills_dir}"
@@ -343,7 +349,8 @@ def run(self):
         proxy_config["port"] = proxy_port
         proxy_config.setdefault("host", "0.0.0.0")
         proxy_config["served_model_name"] = served_model_name or "skillclaw-model"
-        llm_api_mode = str(current_llm.get("api_mode", "chat") or "chat")
+        default_api_mode = default_llm_api_mode_for_claw(claw_type)
+        llm_api_mode = str(current_llm.get("api_mode", default_api_mode) or default_api_mode)
         data = {
             "claw_type": claw_type,
             "llm": {
@@ -375,4 +382,7 @@ def run(self):
 
         print(f"\nConfig saved to: {cs.config_file}")
         print("\nRun 'skillclaw start' to launch SkillClaw.")
+        if claw_type == "codex":
+            print("Then run 'codex --profile skillclaw' to use Codex through SkillClaw.")
+            print("Use 'skillclaw doctor codex' if the profile does not work as expected.")
         print("=" * 60 + "\n")

tests/test_anthropic_messages_api.py

@@ -9,8 +9,7 @@
 
 
 @pytest.fixture
-def anthropic_server(monkeypatch, tmp_path):
-    monkeypatch.setattr(SkillClawAPIServer, "_load_tokenizer", lambda self: None)
+def anthropic_server(tmp_path):
     return SkillClawAPIServer(
         SkillClawConfig(
             proxy_api_key="skillclaw",
@@ -43,14 +42,6 @@ async def test_anthropic_count_tokens_endpoint_returns_local_estimate(anthropic_
 
 @pytest.mark.asyncio
 async def test_anthropic_count_tokens_accounts_for_image_content(anthropic_server):
-    class FakeTokenizer:
-        def apply_chat_template(self, messages, tools=None, tokenize=False, add_generation_prompt=False):
-            return "user: screenshot"
-
-        def __call__(self, text, add_special_tokens=False):
-            return {"input_ids": [1, 2, 3]}
-
-    anthropic_server._tokenizer = FakeTokenizer()
     png_header = (
         b"\x89PNG\r\n\x1a\n"
         + struct.pack(">I", 13)