fix: CLI namer 不再往会话列表里堆「起名会话」

ARCHITECTURE.md

@@ -152,14 +152,18 @@ For each discovered session, in order:
 3. **No-activity short-circuit** — if `last_active` is unchanged since we last
    fully evaluated it (`seen_active` in state), skip without re-reading the
    transcript. Cheap; keeps the poll loop light.
-4. **Substance** — skip if there are fewer than `min_user_messages` non-trivial
+4. **Namer-artifact** — skip if the current title or a user message is the
+   naming prompt itself (`util.is_namer_artifact`). CLI namer calls and Claude
+   Code's built-in tab-titler can leak into the session list; feeding them back
+   into the namer creates more junk sessions.
+5. **Substance** — skip if there are fewer than `min_user_messages` non-trivial
    user messages (acknowledgements, slash-commands and harness/tool noise are
    filtered out in `util.is_trivial` / `is_noise`).
-5. **Unchanged content** — hash the recent transcript (`util.signature`); if it
+6. **Unchanged content** — hash the recent transcript (`util.signature`); if it
    matches the hash tied to the title we last wrote, skip. This makes runs
    idempotent and **respects titles you edit by hand** — until the conversation
    moves on.
-6. Otherwise generate a title, shape it (`util.shape_title`), and write it if it
+7. Otherwise generate a title, shape it (`util.shape_title`), and write it if it
    differs from the current one.
 
 ### Baseline timestamp (v0.6.0+)
@@ -206,7 +210,8 @@ baseline they didn't want to commit to.
   next pass to clobber a hand-edited title.
 - **Titling via your own logged-in CLI by default.** The default `auto` namer
   reuses the `claude`/`codex` CLI you're signed into (no API key); a short excerpt
-  goes to that provider. Set `namer = "heuristic"` for a fully offline run. See
+  goes to that provider. Those CLI calls are ephemeral (no extra session in your
+  Claude Code / Codex list). Set `namer = "heuristic"` for a fully offline run. See
   [SECURITY.md](SECURITY.md).
 
 ## Adding a tool

CHANGELOG.md

@@ -5,6 +5,30 @@ All notable changes to this project are documented here. Format loosely follows
 
 ## [Unreleased]
 
+## [1.0.1] - 2026-08-19
+
+### Fixed
+- **CLI namers no longer flood Claude Code / Codex with junk sessions.** The
+  default `auto` / `claude` / `codex` namer shells out to `claude -p` or
+  `codex exec` to write a title. Those CLIs were persisting the call as a
+  *real* session whose title is the naming prompt itself — `You name
+  coding-assistant sessions…`, or Claude Code's own `Generate a concise tab
+  title for this coding chat…`. After the idle window, rename treated those as
+  new work and called the namer again, so the session list grew with every
+  pass (and with every project you ran from).
+  - `claude`: `--bare --no-session-persistence` plus
+    `CLAUDE_CODE_SKIP_PROMPT_HISTORY=1` (the flag is a no-op on some Claude
+    Code versions; the env var still suppresses transcripts)
+  - `codex`: `--ephemeral --skip-git-repo-check`
+  - both run in an isolated scratch directory, not the user's project
+  - sessions whose title or user text looks like a namer / auto-title prompt
+    are skipped, so leftover junk cannot retrigger the loop
+  - older CLIs that reject the new flags are retried without them
+
+The Codex namer still defaults to **`gpt-5.3-codex-spark`**. Default
+`namer = "auto"` still prefers the `claude` CLI (Haiku) when installed, then
+Codex Spark.
+
 ## [1.0.0] - 2026-06-02
 
 ### Changed — breaking

README.en.md

@@ -145,7 +145,8 @@ This makes the whole thing **idempotent** and **safe to run continuously**.
 
 **Where the title comes from.** By default rename shells out to the `claude`
 (or `codex`) CLI you're already logged into — `claude --model haiku -p "…"` — so
-titles are real LLM summaries of the conversation, with no API key. No CLI
+titles are real LLM summaries of the conversation, with no API key. The call is
+ephemeral and does not leave an extra session in your chat list. No CLI
 installed? It falls back to the offline heuristic.
 
 **Safe by default on existing machines.** When you install rename, the daemon
@@ -234,7 +235,10 @@ back to a fully-offline heuristic if neither is installed. You never paste a key
 | `openai` | OpenAI API directly, with **your own key** | `api_key` or `OPENAI_API_KEY` |
 
 Out of the box — nothing to configure, no key to paste — you get LLM-quality titles
-using credits you already have. Prefer zero cost / fully offline? Set `namer = "heuristic"`.
+using credits you already have. `auto` **prefers the logged-in `claude` CLI (Haiku)**
+and only uses `codex` (default model id **`gpt-5.3-codex-spark`**) if `claude` is not
+installed. CLI naming calls are ephemeral and must not show up as extra sessions in
+Claude Code / Codex. Prefer zero cost / fully offline? Set `namer = "heuristic"`.
 Want to force Codex? Set `namer = "codex"` and change `[codex] model = "..."`.
 
 **Bring your own key.** Want to use your own Anthropic/OpenAI account instead of a

README.md

@@ -139,7 +139,7 @@ $ rename search postgres --content   # 连消息正文一起搜,带匹配片段
 
 **标题是怎么来的。** 默认 rename 会调用你已经登录的 `claude`（或 `codex`）命令行
 ——`claude --model haiku -p "…"`——所以标题是对话的真实 LLM 总结，且无需 API key。
-没装 CLI？就退回离线启发式。
+这次调用是一次性的，不会在会话列表里再留下一条「起名会话」。没装 CLI？就退回离线启发式。
 
 **装上就用，不会偷偷改你的历史。** 第一次跑的时候 rename 会记一个"基线时间戳"，
 之后后台只会改"基线之后才活跃"的会话——你装 rename 之前的旧聊天不会被自动碰，
@@ -221,7 +221,9 @@ Antigravity 有两个形态——**IDE 版**(基于 VS Code 的客户端,带 Gem
 | `openai` | 直连 OpenAI API,用**你自己的 key** | `api_key` 或 `OPENAI_API_KEY` |
 
 开箱即用、零配置、不用粘贴任何 key，你就能得到 LLM 质量的标题（花的是你已有的
-额度）。想要零成本/完全离线？设 `namer = "heuristic"`。
+额度）。`auto` **优先用已登录的 `claude`（Haiku）**，没有 `claude` 才用 `codex`
+（默认模型 id：**`gpt-5.3-codex-spark`**）。CLI 起名是一次性调用，不会在 Claude Code /
+Codex 会话列表里再留一条「起名会话」。想要零成本/完全离线？设 `namer = "heuristic"`。
 想固定使用 Codex？设 `namer = "codex"`；模型可在 `[codex] model = "..."` 里改。
 
 **用自己的 API key。** 想用自己的 Anthropic / OpenAI 账号而不是已登录的 CLI?把

SECURITY.md

@@ -11,7 +11,9 @@ privacy are a first-class concern.
 - **Titling uses your own logged-in CLI by default.** The default `auto` namer
   asks the `claude` / `codex` tool you're already signed into to write the title,
   so a short transcript excerpt is sent through that provider — there is no API
-  key to paste. The `anthropic` / `openai` namers do the same via a key you set.
+  key to paste. Those calls are ephemeral and must not create extra sessions in
+  your Claude Code / Codex history. The `anthropic` / `openai` namers do the same
+  via a key you set.
 - **Fully offline option.** Set `namer = "heuristic"` and nothing ever leaves
   your machine.
 - **Conservative writes.** Reads use read-only SQLite connections; writes use a

pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "hatchling.build"
 # still `rename` (see [project.scripts]); only the published name differs,
 # because the bare `rename` name on PyPI is owned by an unrelated project.
 name = "rename-cli"
-version = "1.0.0"
+version = "1.0.1"
 description = "Keep your AI coding sessions named after what they actually became — Claude Code, Codex, Cursor and more."
 readme = "README.md"
 license = { text = "MIT" }

src/rename/config.py

@@ -51,7 +51,7 @@
 #                 neither is installed. (default)
 #   "heuristic" - instant, fully offline, no LLM, no token cost
 #   "claude"    - always use the `claude` CLI (defaults to the fast Haiku model)
-#   "codex"     - always use the `codex` CLI
+#   "codex"     - always use the `codex` CLI (defaults to gpt-5.3-codex-spark)
 #   "anthropic" - Anthropic API directly, with your OWN key (set api_key in the
 #                 [anthropic] table below, or export ANTHROPIC_API_KEY)
 #   "openai"    - OpenAI API directly, with your OWN key (set api_key in the
@@ -69,7 +69,7 @@
 
 # Model overrides for the CLI namers (optional). These reuse your existing
 # login — no API key. Defaults are the fast/cheap models, which are plenty for
-# a short title.
+# a short title. CLI namers run ephemerally (no extra Claude/Codex session).
 [claude]
 model = "haiku"
 

src/rename/engine.py

@@ -5,9 +5,11 @@
   1. Idle gate — skip if used within `idle_seconds` (still in use).
   2. No-activity short-circuit — skip if `last_active` is unchanged since we last
      evaluated it (cheap: no transcript read needed).
-  3. Substance — skip if too few real user messages.
-  4. Unchanged — skip if the content hash matches the title we last wrote.
-  5. Otherwise it's a *candidate*: ask the namer for a title and write it back
+  3. Namer-artifact — skip sessions whose title or user text is the naming
+     prompt itself (a CLI namer / auto-titler side-effect, not real work).
+  4. Substance — skip if too few real user messages.
+  5. Unchanged — skip if the content hash matches the title we last wrote.
+  6. Otherwise it's a *candidate*: ask the namer for a title and write it back
      if it differs.
 
 Assessment (fast, local) is deliberately separated from naming (slow — it shells
@@ -52,7 +54,7 @@ def _assess(
         include_historical: bool = False,
     ):
         """Return (status, sig, msgs) without naming. status is one of
-        historical | active | no-activity | thin | unchanged | candidate.
+        historical | active | no-activity | namer-artifact | thin | unchanged | candidate.
 
         ``historical`` means the session existed before the daemon's first
         run on this machine — we leave those alone so rename never
@@ -69,7 +71,11 @@ def _assess(
         prev = self.state.get(adapter.name, s.id)
         if prev and prev.get("seen_active") == s.last_active:
             return ("no-activity", None, None)
+        if util.is_namer_artifact(s.title):
+            return ("namer-artifact", None, None)
         msgs = adapter.read_transcript(s)
+        if any(m.role == "user" and util.is_namer_artifact(m.text) for m in msgs):
+            return ("namer-artifact", None, msgs)
         substantive = [
             m for m in msgs if m.role == "user" and not util.is_trivial(m.text)
         ]
@@ -88,7 +94,7 @@ def _name(self, adapter: Adapter, s: Session, msgs: list) -> str:
 
     def _record_skip(self, adapter, s, status, sig, now_ts) -> None:
         fields: dict = {"last_seen": now_ts}
-        if status in ("thin", "unchanged"):
+        if status in ("thin", "unchanged", "namer-artifact"):
             fields["seen_active"] = s.last_active
         if status == "unchanged" and sig:
             fields["content_sig"] = sig
@@ -117,6 +123,10 @@ def _plan_one(
             )
         if status == "no-activity":
             return RenamePlan(s, "skip", reason="no activity since last check")
+        if status == "namer-artifact":
+            return RenamePlan(
+                s, "skip", mark_seen=True, reason="namer/CLI side-effect session"
+            )
         if status == "thin":
             return RenamePlan(
                 s, "skip", mark_seen=True, reason="no substantive user messages"

src/rename/namers/cli_namer.py

@@ -3,6 +3,11 @@
 Reuses whatever login the user already has for that tool — no API key wiring,
 no extra cost beyond the tool's own usage. This is the default (via ``auto``),
 which prefers ``claude`` then ``codex``.
+
+Calls are ephemeral: they must not land in the user's Claude Code / Codex
+session list. A persisted namer call shows up as a real session titled with
+the naming prompt itself, and rename would then try to rename *that* session
+by calling the CLI again — a loop.
 """
 
 from __future__ import annotations
@@ -17,11 +22,76 @@
 
 _TIMEOUT = 90  # codex with reasoning can take a while; keep generous
 
-# The fast Codex model used for titling unless the user overrides it.
+# Official Codex CLI model id (OpenAI, Feb 2026). Fast enough for a 6-word title.
 _CODEX_DEFAULT_MODEL = "gpt-5.3-codex-spark"
 # Claude's small/fast model — plenty for a 6-word title, and cheap.
 _CLAUDE_DEFAULT_MODEL = "haiku"
 
+# Flags that stop the namer call from showing up as a real coding session.
+# Some CLI versions don't know them; we retry without them if rejected.
+_CLAUDE_EPHEMERAL_FLAGS = ("--bare", "--no-session-persistence")
+_CODEX_EPHEMERAL_FLAGS = ("--ephemeral", "--skip-git-repo-check")
+
+
+def _scratch_cwd() -> str:
+    """Isolated cwd so a leaking CLI cannot dump sessions into a user project."""
+    path = util.state_dir() / "namer-scratch"
+    path.mkdir(parents=True, exist_ok=True)
+    return str(path)
+
+
+def _unknown_flag(stderr: str) -> bool:
+    low = (stderr or "").lower()
+    return any(
+        needle in low
+        for needle in (
+            "unknown option",
+            "unknown argument",
+            "unexpected argument",
+            "unrecognized",
+            "invalid option",
+            "unexpected option",
+        )
+    )
+
+
+def _run(
+    argv: list[str],
+    *,
+    env: dict[str, str] | None = None,
+    cwd: str | None = None,
+):
+    merged = os.environ.copy()
+    if env:
+        merged.update(env)
+    try:
+        return subprocess.run(
+            argv,
+            capture_output=True,
+            text=True,
+            timeout=_TIMEOUT,
+            env=merged,
+            cwd=cwd,
+        )
+    except (subprocess.TimeoutExpired, OSError) as exc:
+        util.log(f"{argv[0]} namer call failed: {exc}", level="debug")
+        return None
+
+
+def _run_ephemeral(
+    required: list[str],
+    optional: tuple[str, ...],
+    trailing: list[str],
+    *,
+    env: dict[str, str] | None = None,
+    cwd: str | None = None,
+):
+    """Run with persistence-killing flags; retry without them on older CLIs."""
+    proc = _run(required + list(optional) + trailing, env=env, cwd=cwd)
+    if proc is not None and proc.returncode != 0 and _unknown_flag(proc.stderr):
+        proc = _run(required + trailing, env=env, cwd=cwd)
+    return proc
+
 
 class CliNamer(Namer):
     def __init__(self, name: str, options: dict | None = None):
@@ -43,13 +113,15 @@ def _generate_claude(self, prompt: str) -> str | None:
         model = self.options.get("model", _CLAUDE_DEFAULT_MODEL)
         if model:
             argv += ["--model", str(model)]
-        argv += ["-p", prompt]
-        try:
-            proc = subprocess.run(
-                argv, capture_output=True, text=True, timeout=_TIMEOUT
-            )
-        except (subprocess.TimeoutExpired, OSError) as exc:
-            util.log(f"claude namer call failed: {exc}", level="debug")
+        env = {"CLAUDE_CODE_SKIP_PROMPT_HISTORY": "1"}
+        proc = _run_ephemeral(
+            argv,
+            _CLAUDE_EPHEMERAL_FLAGS,
+            ["-p", prompt],
+            env=env,
+            cwd=_scratch_cwd(),
+        )
+        if proc is None:
             return None
         if proc.returncode != 0:
             util.log(
@@ -72,13 +144,13 @@ def _generate_codex(self, prompt: str) -> str | None:
             model = self.options.get("model", _CODEX_DEFAULT_MODEL)
             if model:
                 argv += ["-m", str(model)]
-            argv += ["--output-last-message", out_path, prompt]
-            try:
-                proc = subprocess.run(
-                    argv, capture_output=True, text=True, timeout=_TIMEOUT
-                )
-            except (subprocess.TimeoutExpired, OSError) as exc:
-                util.log(f"codex namer call failed: {exc}", level="debug")
+            proc = _run_ephemeral(
+                argv,
+                _CODEX_EPHEMERAL_FLAGS,
+                ["--output-last-message", out_path, prompt],
+                cwd=_scratch_cwd(),
+            )
+            if proc is None:
                 return None
             if proc.returncode != 0:
                 util.log(

src/rename/util.py

@@ -146,6 +146,22 @@ def log(msg: str, *, level: str = "info") -> None:
 )
 _JSONISH = re.compile(r'^[\[{]\s*["{\[]')
 
+# Side-effect sessions created by a CLI namer (or Claude Code's own tab-titler).
+# Matching is case-insensitive and works on truncated titles in the session list.
+_NAMER_ARTIFACT_MARKERS = (
+    "you name coding-assistant sessions",
+    "generate a concise tab title for this coding chat",
+    "concise title of 3 to 6 words capturing what the user",
+)
+
+
+def is_namer_artifact(text: str | None) -> bool:
+    """True for a title or user message that is a session-naming prompt, not work."""
+    if not text:
+        return False
+    low = text.casefold()
+    return any(marker in low for marker in _NAMER_ARTIFACT_MARKERS)
+
 
 def is_noise(text: str) -> bool:
     """True for harness/tool artifacts that are not genuine conversation."""