laya-cli 0.2.0__tar.gz → 0.2.1__tar.gz

{laya_cli-0.2.0 → laya_cli-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: laya-cli
-Version: 0.2.0
+Version: 0.2.1
 Summary: Ergonomic CLI for Laya — typed decisions for humans and AI agents (predict, batch, presets, shortlist, router)
 Project-URL: Homepage, https://github.com/MIt9/laya-cli
 Project-URL: Repository, https://github.com/MIt9/laya-cli.git
@@ -51,6 +51,7 @@ A modern, high-performance CLI for [Laya](https://github.com/NandhaKishorM/laya)
 * 🔍 **Preset Library**: `triage` / `email` / `guard` / `moderation` / `router` — direct `laya.*_questions()` passthrough, mergeable with `--questions file.json` and `--questions-inline`
 * 📦 **Streaming Batch Mode**: `classify`/`predict --input` loads model **once** + warmup, streams JSONL (`--state-field` verbatim, no silent `(photographer: Name)` injection)
 * 🔧 **Post-Filter & Eval**: `filter --where "on_topic>=0.4" --sort -on_topic` and `evaluate` (accuracy, passing @ threshold, precision, escalation rate)
+* ⚡ **Resident Daemon (optional)**: `laya-cli serve` keeps model in RAM — `predict`/`classify`/`evaluate` auto-hit `127.0.0.1` and skip 10-35s `laya.load()` on repeated calls (hash → `~/.cache/laya-cli/daemons/<hash>.json`, idle-timeout 1800s)
 
 ---
 
@@ -160,32 +161,68 @@ laya-cli evaluate --questions questions.json --labeled labeled.jsonl --label-fie
 
 ### 4. Resident Daemon Mode (optional, speeds up repeated calls)
 
-Measured: `laya.load()` 10–35s (MPS) + ~120ms × 82 candidates. For one pipeline run it's fine; for iterative tuning of `questions.json` each `predict` pays 10–35s again. `serve` keeps the model in RAM — subsequent calls skip the load and hit `127.0.0.1` (<1ms overhead).
+**Виміряний факт (сесія 2026-09-22, 82 кандидати, MPS):**
+
+| етап | час |
+|------|-----|
+| `px` (мережа, 12 запитів) | 9.3s |
+| `laya.load()` (модель у пам'ять) | 10–35s |
+| inference, 82 кандидати | ~10s (~120ms/шт) |
+| `filter` | 0.08s |
+
+Для одного пайплайну `px → classify → filter` 35–50s норм. Проблема — коли за сесію кілька разів викликаєш `predict`/`classify`/`evaluate` (підбір `questions.json`: прогнав → подивився → поправив → знову), кожен раз платиш 10–35s за ту саму модель. `laya-cli serve` тримає модель в RAM, наступні виклики летять на `127.0.0.1` (<1ms overhead).
+
+#### Транспорт (як у `~/.claude/skills/laya-integration/SKILL.md` "Anything else... HTTP sidecar")
+
+HTTP на loopback `127.0.0.1` (не `0.0.0.0`), один потік з `threading.Lock` (SKILL.md: "one GPU serves one forward pass at a time"):
+
+- `POST /predict` → `{"state": ..., "questions": {...}, "lang": "...", "shortlist_k": 20}` → те саме що `agent.predict()` (той самий JSON що `laya-cli predict --format json`)
+- `GET /status` → `{"model": "...", "subfolder": "...", "device": "...", "loaded_at": 123..., "idle_seconds": 42, "requests_served": 17, "pid": 12345, "port": 8765}`
+- `POST /shutdown` → graceful stop (лише з localhost, також `serve stop` шле сигнал за pid-файлом)
+
+#### Lifecycle
 
 ```bash
-# Start daemon in background (one daemon per model/device/router config)
+# Старт (без --foreground — форк у фон, пише pid+port в ~/.cache/laya-cli/daemons/<hash>.json)
 laya-cli serve --model convaiinnovations/laya --device mps
-laya-cli serve --model convaiinnovations/laya --subfolder multilingual --device cpu --idle-timeout 60  # short idle for test
-laya-cli serve --router --foreground --idle-timeout 0  # foreground, no auto-exit, for logs
+laya-cli serve --model convaiinnovations/laya --subfolder multilingual --device cpu --idle-timeout 60
+laya-cli serve --router --foreground --idle-timeout 0 --port 8765  # форграунд для логів, 0=disabled idle
+
+# Один daemon = один конфіг (hash(model|subfolder|device|router|lang)), інший конфіг — окремий файл/порт
+laya-cli serve status                                          # дефолтний конфіг (як у predict без прапорців)
+laya-cli serve status --model convaiinnovations/laya --subfolder multilingual --device cpu
+laya-cli serve stop                                            # graceful (POST /shutdown → pid файл видаляється)
+laya-cli serve stop --all                                      # всі daemon-и
+# Також: curl http://127.0.0.1:<port>/status
+```
 
-# Check status (also shows port, pid, idle_seconds, requests_served)
-laya-cli serve status
-laya-cli serve status --model convaiinnovations/laya --subfolder multilingual
+Перед прийомом запитів — прогрів throwaway `predict` (як у `TASK.md`). `idle-timeout 1800s` дефолт, `--idle-timeout 5` для тесту → daemon сам виходить через 5s без запитів і `serve status` каже `not running`.
 
-# Use it — predict/classify/evaluate automatically hit the daemon if live for that config
-laya-cli predict "hello" --preset guard --format json          # -> via daemon, no 10s load
-laya-cli predict "hello" --preset guard --no-daemon --format json  # force in-process, ignore daemon
+#### Клієнт (`predict`/`classify`/`evaluate`)
 
-# Batch also benefits (each line -> POST /predict on loopback, serialized with lock)
-cat candidates.jsonl | laya-cli classify --questions q.json | laya-cli filter --where "on_topic>=0.4" --sort -on_topic
+Перед `laya.load()` перевіряє `~/.cache/laya-cli/daemons/<hash>.json` для поточного конфігу; якщо файл є і daemon відповідає на `/status` — шле туди, інакше мовчки падає назад на in-process (поведінка `v1` без змін). `--no-daemon` форсує in-process (для відтворюваності/дебагу). Для батчу (`classify`/`predict --input`) — кожен рядок окремий `POST /predict` в циклі (loopback мілісекунди, batch-ендпоінт не потрібен).
 
-# Stop
-laya-cli serve stop                 # stop daemon for default config
-laya-cli serve stop --all           # stop all daemons
-curl http://127.0.0.1:<port>/status  # GET /status, POST /predict, POST /shutdown also work directly
+```bash
+# AI workflow (повністю автоматичний):
+laya-cli serve --device mps &                          # один раз на сесію
+laya-cli predict "hello" --preset guard --format json  # -> via daemon, без 10s
+laya-cli predict "hello2" --preset guard --format json # -> знову via daemon
+laya-cli serve status                                  # {"loaded_at":..., "requests_served": 2}
+laya-cli predict "hello" --preset guard --no-daemon    # форс in-process (знову 10s, для ізоляції)
+laya-cli serve stop
+
+# Human tuning loop (типовий):
+laya-cli serve &                                       # фон
+cat candidates.jsonl | laya-cli classify --questions q.json | laya-cli filter --where "on_topic>=0.4" --sort -on_topic  # via daemon
+# ...поправив q.json...
+cat candidates.jsonl | laya-cli classify --questions q.json | laya-cli filter ...  # знову via daemon, без перезавантаження
+laya-cli serve stop
+
+# 5 паралельних predict — не падають, результати не плутаються (серіалізація Lock)
+seq 1 5 | xargs -P5 -I{} laya-cli predict "text {}" --preset guard --format json
 ```
 
-**Details:** HTTP on `127.0.0.1` only (SKILL.md: "Bind to 127.0.0.1"), `POST /predict {"state":..., "questions":{...}}` → same as `agent.predict()`, `GET /status` → `{model, device, loaded_at, idle_seconds, requests_served}`, `POST /shutdown` (localhost only). One daemon = one checkpoint hash(`model|subfolder|device|router|lang`) → pid+port in `~/.cache/laya-cli/daemons/<hash>.json`, pid file removed on exit. Idle timeout 1800s default, `0` disables. If daemon not running, `predict`/`classify`/`evaluate` silently fall back to in-process load — no new step for scripts. `--no-daemon` forces fallback. Requests are serialized with a `threading.Lock` (one GPU = one forward pass).
+**Безпека/межі:** лише `127.0.0.1`, без auth (однокористувацька машина, як у SKILL.md), один потік, stateless крім моделі.
 
 ---
 

{laya_cli-0.2.0 → laya_cli-0.2.1}/README.md

@@ -16,6 +16,7 @@ A modern, high-performance CLI for [Laya](https://github.com/NandhaKishorM/laya)
 * 🔍 **Preset Library**: `triage` / `email` / `guard` / `moderation` / `router` — direct `laya.*_questions()` passthrough, mergeable with `--questions file.json` and `--questions-inline`
 * 📦 **Streaming Batch Mode**: `classify`/`predict --input` loads model **once** + warmup, streams JSONL (`--state-field` verbatim, no silent `(photographer: Name)` injection)
 * 🔧 **Post-Filter & Eval**: `filter --where "on_topic>=0.4" --sort -on_topic` and `evaluate` (accuracy, passing @ threshold, precision, escalation rate)
+* ⚡ **Resident Daemon (optional)**: `laya-cli serve` keeps model in RAM — `predict`/`classify`/`evaluate` auto-hit `127.0.0.1` and skip 10-35s `laya.load()` on repeated calls (hash → `~/.cache/laya-cli/daemons/<hash>.json`, idle-timeout 1800s)
 
 ---
 
@@ -125,32 +126,68 @@ laya-cli evaluate --questions questions.json --labeled labeled.jsonl --label-fie
 
 ### 4. Resident Daemon Mode (optional, speeds up repeated calls)
 
-Measured: `laya.load()` 10–35s (MPS) + ~120ms × 82 candidates. For one pipeline run it's fine; for iterative tuning of `questions.json` each `predict` pays 10–35s again. `serve` keeps the model in RAM — subsequent calls skip the load and hit `127.0.0.1` (<1ms overhead).
+**Виміряний факт (сесія 2026-09-22, 82 кандидати, MPS):**
+
+| етап | час |
+|------|-----|
+| `px` (мережа, 12 запитів) | 9.3s |
+| `laya.load()` (модель у пам'ять) | 10–35s |
+| inference, 82 кандидати | ~10s (~120ms/шт) |
+| `filter` | 0.08s |
+
+Для одного пайплайну `px → classify → filter` 35–50s норм. Проблема — коли за сесію кілька разів викликаєш `predict`/`classify`/`evaluate` (підбір `questions.json`: прогнав → подивився → поправив → знову), кожен раз платиш 10–35s за ту саму модель. `laya-cli serve` тримає модель в RAM, наступні виклики летять на `127.0.0.1` (<1ms overhead).
+
+#### Транспорт (як у `~/.claude/skills/laya-integration/SKILL.md` "Anything else... HTTP sidecar")
+
+HTTP на loopback `127.0.0.1` (не `0.0.0.0`), один потік з `threading.Lock` (SKILL.md: "one GPU serves one forward pass at a time"):
+
+- `POST /predict` → `{"state": ..., "questions": {...}, "lang": "...", "shortlist_k": 20}` → те саме що `agent.predict()` (той самий JSON що `laya-cli predict --format json`)
+- `GET /status` → `{"model": "...", "subfolder": "...", "device": "...", "loaded_at": 123..., "idle_seconds": 42, "requests_served": 17, "pid": 12345, "port": 8765}`
+- `POST /shutdown` → graceful stop (лише з localhost, також `serve stop` шле сигнал за pid-файлом)
+
+#### Lifecycle
 
 ```bash
-# Start daemon in background (one daemon per model/device/router config)
+# Старт (без --foreground — форк у фон, пише pid+port в ~/.cache/laya-cli/daemons/<hash>.json)
 laya-cli serve --model convaiinnovations/laya --device mps
-laya-cli serve --model convaiinnovations/laya --subfolder multilingual --device cpu --idle-timeout 60  # short idle for test
-laya-cli serve --router --foreground --idle-timeout 0  # foreground, no auto-exit, for logs
+laya-cli serve --model convaiinnovations/laya --subfolder multilingual --device cpu --idle-timeout 60
+laya-cli serve --router --foreground --idle-timeout 0 --port 8765  # форграунд для логів, 0=disabled idle
+
+# Один daemon = один конфіг (hash(model|subfolder|device|router|lang)), інший конфіг — окремий файл/порт
+laya-cli serve status                                          # дефолтний конфіг (як у predict без прапорців)
+laya-cli serve status --model convaiinnovations/laya --subfolder multilingual --device cpu
+laya-cli serve stop                                            # graceful (POST /shutdown → pid файл видаляється)
+laya-cli serve stop --all                                      # всі daemon-и
+# Також: curl http://127.0.0.1:<port>/status
+```
 
-# Check status (also shows port, pid, idle_seconds, requests_served)
-laya-cli serve status
-laya-cli serve status --model convaiinnovations/laya --subfolder multilingual
+Перед прийомом запитів — прогрів throwaway `predict` (як у `TASK.md`). `idle-timeout 1800s` дефолт, `--idle-timeout 5` для тесту → daemon сам виходить через 5s без запитів і `serve status` каже `not running`.
 
-# Use it — predict/classify/evaluate automatically hit the daemon if live for that config
-laya-cli predict "hello" --preset guard --format json          # -> via daemon, no 10s load
-laya-cli predict "hello" --preset guard --no-daemon --format json  # force in-process, ignore daemon
+#### Клієнт (`predict`/`classify`/`evaluate`)
 
-# Batch also benefits (each line -> POST /predict on loopback, serialized with lock)
-cat candidates.jsonl | laya-cli classify --questions q.json | laya-cli filter --where "on_topic>=0.4" --sort -on_topic
+Перед `laya.load()` перевіряє `~/.cache/laya-cli/daemons/<hash>.json` для поточного конфігу; якщо файл є і daemon відповідає на `/status` — шле туди, інакше мовчки падає назад на in-process (поведінка `v1` без змін). `--no-daemon` форсує in-process (для відтворюваності/дебагу). Для батчу (`classify`/`predict --input`) — кожен рядок окремий `POST /predict` в циклі (loopback мілісекунди, batch-ендпоінт не потрібен).
 
-# Stop
-laya-cli serve stop                 # stop daemon for default config
-laya-cli serve stop --all           # stop all daemons
-curl http://127.0.0.1:<port>/status  # GET /status, POST /predict, POST /shutdown also work directly
+```bash
+# AI workflow (повністю автоматичний):
+laya-cli serve --device mps &                          # один раз на сесію
+laya-cli predict "hello" --preset guard --format json  # -> via daemon, без 10s
+laya-cli predict "hello2" --preset guard --format json # -> знову via daemon
+laya-cli serve status                                  # {"loaded_at":..., "requests_served": 2}
+laya-cli predict "hello" --preset guard --no-daemon    # форс in-process (знову 10s, для ізоляції)
+laya-cli serve stop
+
+# Human tuning loop (типовий):
+laya-cli serve &                                       # фон
+cat candidates.jsonl | laya-cli classify --questions q.json | laya-cli filter --where "on_topic>=0.4" --sort -on_topic  # via daemon
+# ...поправив q.json...
+cat candidates.jsonl | laya-cli classify --questions q.json | laya-cli filter ...  # знову via daemon, без перезавантаження
+laya-cli serve stop
+
+# 5 паралельних predict — не падають, результати не плутаються (серіалізація Lock)
+seq 1 5 | xargs -P5 -I{} laya-cli predict "text {}" --preset guard --format json
 ```
 
-**Details:** HTTP on `127.0.0.1` only (SKILL.md: "Bind to 127.0.0.1"), `POST /predict {"state":..., "questions":{...}}` → same as `agent.predict()`, `GET /status` → `{model, device, loaded_at, idle_seconds, requests_served}`, `POST /shutdown` (localhost only). One daemon = one checkpoint hash(`model|subfolder|device|router|lang`) → pid+port in `~/.cache/laya-cli/daemons/<hash>.json`, pid file removed on exit. Idle timeout 1800s default, `0` disables. If daemon not running, `predict`/`classify`/`evaluate` silently fall back to in-process load — no new step for scripts. `--no-daemon` forces fallback. Requests are serialized with a `threading.Lock` (one GPU = one forward pass).
+**Безпека/межі:** лише `127.0.0.1`, без auth (однокористувацька машина, як у SKILL.md), один потік, stateless крім моделі.
 
 ---
 

{laya_cli-0.2.0 → laya_cli-0.2.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "laya-cli"
-version = "0.2.0"
+version = "0.2.1"
 description = "Ergonomic CLI for Laya — typed decisions for humans and AI agents (predict, batch, presets, shortlist, router)"
 readme = "README.md"
 requires-python = ">=3.10"

{laya_cli-0.2.0 → laya_cli-0.2.1}/src/laya_cli/__init__.py

@@ -1,3 +1,3 @@
 """laya-cli — streaming JSONL classifier over Laya."""
 
-__version__ = "0.2.0"
+__version__ = "0.2.1"

{laya_cli-0.2.0 → laya_cli-0.2.1}/src/laya_cli/cli.py

@@ -53,7 +53,11 @@ except ImportError:
 def _build_parser() -> argparse.ArgumentParser:
     p = argparse.ArgumentParser(
         prog="laya-cli",
-        description="Ergonomic CLI for Laya — typed decisions (choice/score/noul) in one forward pass. Works for humans (table output) and agents (JSON/JSONL).",
+        description=(
+            "Ergonomic CLI for Laya — typed decisions (choice/score/noul) in one forward pass. "
+            "Works for humans (table) and agents (JSON/JSONL). "
+            "Optional resident daemon (laya-cli serve) keeps model in RAM to avoid 10-35s laya.load() on repeated predict/classify/evaluate (same model|subfolder|device|router|lang → ~/.cache/laya-cli/daemons/<hash>.json, 127.0.0.1 only, auto fallback if not running)."
+        ),
     )
     p.add_argument("--version", action="version", version=f"%(prog)s {__version__}")
     sub = p.add_subparsers(dest="cmd", required=True)
@@ -61,8 +65,11 @@ def _build_parser() -> argparse.ArgumentParser:
     # ---- predict (primary, human+AI friendly) -----------------------------
     pr = sub.add_parser(
         "predict",
-        help="Run Laya on one text/JSON state or a batch (JSONL). Human-friendly & agent-friendly.",
-        description="Predict with Laya on a single state or a batch. Supports plain text, JSON state, presets, shortlist for high-cardinality choices, and table/JSON output.",
+        help="Run Laya on one text/JSON state or a batch (JSONL). Human-friendly & agent-friendly. Auto-uses daemon if live.",
+        description=(
+            "Predict with Laya on a single state or a batch. Supports plain text, JSON state, presets, shortlist for high-cardinality choices, and table/JSON output. "
+            "If a daemon is live for the same --model/--subfolder/--device/--router/--lang (hash → ~/.cache/laya-cli/daemons/<hash>.json), predict hits 127.0.0.1:<port>/predict and skips 10-35s laya.load(); otherwise it loads in-process. Use --no-daemon to force in-process."
+        ),
         epilog=(
             "Examples (human):\n"
             '  laya-cli predict "I was charged twice, refund please" --preset triage\n'
@@ -74,6 +81,13 @@ def _build_parser() -> argparse.ArgumentParser:
             "  cat candidates.jsonl | laya-cli predict --questions q.json --state-field state --format jsonl > scored.jsonl\n"
             "  laya-cli predict --input candidates.jsonl --questions q.json --shortlist-k 20 --format jsonl\n"
             '  laya-cli predict --preset triage --text "my payment failed" --model convaiinnovations/laya --device cpu --full-probs\n'
+            "\n"
+            "Daemon (optional, speeds up repeated calls):\n"
+            "  laya-cli serve --model convaiinnovations/laya --device mps  # start daemon (background, writes ~/.cache/laya-cli/daemons/<hash>.json)\n"
+            "  laya-cli serve status                                        # check live daemon (GET /status)\n"
+            '  laya-cli predict "hello" --preset guard --format json      # auto-uses daemon (<1ms overhead), no 10s load\n'
+            '  laya-cli predict "hello" --preset guard --no-daemon       # force in-process, ignore daemon\n'
+            "  laya-cli serve stop                                          # graceful shutdown (POST /shutdown)\n"
         ),
         formatter_class=argparse.RawDescriptionHelpFormatter,
     )
@@ -190,14 +204,19 @@ def _build_parser() -> argparse.ArgumentParser:
     # ---- classify (legacy batch, kept for TASK.md compat) -----------------
     c = sub.add_parser(
         "classify",
-        help="Batch classify JSONL from stdin via Laya (legacy, use predict for new code).",
-        description="Classify JSONL from stdin via Laya; append flattened answers. Kept for pipeline compatibility (px | laya-cli classify). For new code prefer `predict`.",
+        help="Batch classify JSONL from stdin via Laya (legacy, use predict for new code). Auto-uses daemon if live.",
+        description=(
+            "Classify JSONL from stdin via Laya; append flattened answers. Kept for pipeline compatibility (px | laya-cli classify). "
+            "For new code prefer `predict`. If a daemon is live for the same config (hash → ~/.cache/laya-cli/daemons/<hash>.json), classify hits 127.0.0.1 and skips load; else in-process. Use --no-daemon to force."
+        ),
         epilog=(
             "Examples:\n"
             "  cat candidates.jsonl | laya-cli classify --questions questions.json > scored.jsonl\n"
             '  px videos --queries "..." --state --dedupe keep-first \\\n'
             "    | laya-cli classify --questions questions.json \\\n"
             '    | laya-cli filter --where "on_topic>=0.4" --sort -on_topic > shortlist.jsonl\n'
+            "  # with daemon (start once, then all classify hit daemon):\n"
+            "  laya-cli serve --model convaiinnovations/laya & laya-cli classify --questions q.json < candidates.jsonl\n"
         ),
         formatter_class=argparse.RawDescriptionHelpFormatter,
     )
@@ -279,12 +298,16 @@ def _build_parser() -> argparse.ArgumentParser:
     # ---- evaluate ---------------------------------------------------------
     e = sub.add_parser(
         "evaluate",
-        help="Evaluate on a labelled JSONL set and print accuracy / threshold report.",
-        description="Evaluate before shipping: run classify on a labelled set and report accuracy per question, pass rate at threshold, and precision @ threshold.",
+        help="Evaluate on a labelled JSONL set and print accuracy / threshold report. Auto-uses daemon if live.",
+        description=(
+            "Evaluate before shipping: run classify on a labelled set and report accuracy per question, pass rate at threshold, and precision @ threshold. "
+            "If a daemon is live for the same --model/--device/--router, evaluate hits it and skips load; else in-process. Use --no-daemon to force."
+        ),
         epilog=(
             "Examples:\n"
             "  laya-cli evaluate --questions questions.json --labeled labeled.jsonl --label-field label --threshold 0.5\n"
             "  laya-cli evaluate --questions q.json --labeled dev.jsonl --field on_topic --state-field state\n"
+            "  laya-cli serve & laya-cli evaluate --questions q.json --labeled dev.jsonl  # via daemon\n"
         ),
         formatter_class=argparse.RawDescriptionHelpFormatter,
     )
@@ -314,16 +337,28 @@ def _build_parser() -> argparse.ArgumentParser:
     s = sub.add_parser(
         "serve",
         help="Resident daemon that keeps model in memory (optional, speeds up repeated predict).",
-        description="Start a resident daemon that holds the Laya model in RAM. Subsequent predict/classify/evaluate automatically use it if live (same model/device/router). No change to pipelines if daemon not running.",
+        description=(
+            "Start a resident daemon that holds the Laya model in RAM (10-35s laya.load() once). "
+            "Subsequent `predict`/`classify`/`evaluate` with the SAME config (--model/--subfolder/--device/--router/--lang → hash → ~/.cache/laya-cli/daemons/<hash>.json) "
+            "automatically hit 127.0.0.1:<port>/predict and skip the load (<1ms overhead). If no daemon, they fall back to in-process load — no pipeline change. "
+            "Daemon binds only 127.0.0.1 (SKILL.md: Bind to 127.0.0.1), single-threaded lock (one GPU = one forward pass), idle-timeout auto-exit (default 1800s, 0=disabled), warmup throwaway predict on start."
+        ),
         epilog=(
-            "Examples:\n"
-            "  laya-cli serve --model convaiinnovations/laya --device mps  # start daemon (background)\n"
-            "  laya-cli serve --foreground --idle-timeout 60                # foreground, for debugging\n"
-            "  laya-cli serve status                                        # show live daemon\n"
-            "  laya-cli serve status --model convaiinnovations/laya --subfolder multilingual\n"
-            "  laya-cli serve stop                                          # graceful shutdown\n"
-            "  laya-cli serve stop --all                                    # stop all daemons\n"
-            '  laya-cli predict "hello" --preset guard --no-daemon       # force in-process, ignore daemon\n'
+            "Workflow for AI/human (typical tuning loop):\n"
+            "  laya-cli serve --model convaiinnovations/laya --device mps         # start once (background, writes ~/.cache/laya-cli/daemons/<hash>.json)\n"
+            "  laya-cli serve status                                              # GET /status → {model,device,loaded_at,idle_seconds,requests_served,pid,port}\n"
+            '  laya-cli predict "test" --preset triage --format json             # auto via daemon (no 10s load)\n'
+            '  laya-cli predict "test2" --preset triage --format json            # still via daemon\n'
+            '  laya-cli predict "test" --preset triage --no-daemon --format json # force in-process (ignore daemon)\n'
+            "  cat candidates.jsonl | laya-cli classify --questions q.json        # batch: each line POST /predict (serialized)\n"
+            "  laya-cli serve stop                                                # POST /shutdown or SIGTERM, removes pid file\n"
+            "  laya-cli serve stop --all                                          # stop all configs\n"
+            "\n"
+            "Other examples:\n"
+            "  laya-cli serve --foreground --idle-timeout 5 --port 0              # foreground for logs, 5s idle test\n"
+            "  laya-cli serve --model convaiinnovations/laya --subfolder multilingual --device cpu\n"
+            "  laya-cli serve status --model convaiinnovations/laya --subfolder multilingual  # check that config\n"
+            '  curl http://127.0.0.1:<port>/status ; curl -X POST http://127.0.0.1:<port>/predict -d \'{"state":"hi","questions":{...}}\'\n'
         ),
         formatter_class=argparse.RawDescriptionHelpFormatter,
     )

{laya_cli-0.2.0 → laya_cli-0.2.1}/uv.lock

@@ -499,7 +499,7 @@ wheels = [
 
 [[package]]
 name = "laya-cli"
-version = "0.2.0"
+version = "0.2.1"
 source = { editable = "." }
 dependencies = [
     { name = "huggingface-hub" },