@turmind/halo 1.5.7 → 1.5.9

package/templates/models/anthropic.yaml

@@ -1,5 +1,6 @@
 id: anthropic
 runtime: anthropic-messages
+revision: 2026100501
 displayName: Anthropic (Generic)
 displayName_zh: Anthropic（通用）
 description: Generic Anthropic Messages API. Use for api.anthropic.com or any Anthropic-compatible third-party gateway.

package/templates/models/aws-bedrock-claude-invoke.yaml

@@ -1,5 +1,6 @@
 id: aws-bedrock-claude-invoke
 runtime: bedrock-invoke
+revision: 2026100501
 displayName: AWS Bedrock Claude (Invoke API)
 displayName_zh: AWS Bedrock Claude（Invoke API）
 description: Invokes Bedrock via InvokeModel (non-streaming), sends raw Anthropic Messages API format

package/templates/models/aws-bedrock-mantle.yaml

@@ -1,5 +1,6 @@
 id: aws-bedrock-mantle
 runtime: bedrock-mantle
+revision: 2026100501
 displayName: AWS Bedrock Mantle (OpenAI GPT-6.1 / 6 / 5.6)
 displayName_zh: AWS Bedrock Mantle（OpenAI GPT-6.1 / 6 / 5.6）
 description: OpenAI GPT-6.1 Sol, GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna) on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). Region-split — GPT-6.1 Sol and GPT-6 Sol / Luna are on us-east-1 only, Astra on us-west-2 only.

package/templates/models/aws-bedrock-openai.yaml

@@ -1,5 +1,6 @@
 id: aws-bedrock-openai
 runtime: bedrock-mantle
+revision: 2026100501
 displayName: AWS Bedrock OpenAI-compatible (GPT-6.1 / 6 / 5.6 · Grok 4.6 · Kimi K3)
 displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6.1 / 6 / 5.6 · Grok 4.6 · Kimi K3）
 description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6.1 Sol, GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.6, Moonshot Kimi K3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).

package/templates/models/deepseek.yaml

@@ -1,5 +1,6 @@
 id: deepseek
 runtime: deepseek
+revision: 2026100501
 displayName: DeepSeek
 displayName_zh: DeepSeek
 description: DeepSeek V4 via OpenAI-compatible chat completions API

package/templates/models/doubao.yaml

@@ -1,5 +1,6 @@
 id: doubao
 runtime: doubao
+revision: 2026100501
 displayName: Doubao (Volcengine)
 displayName_zh: 豆包（火山方舟）
 description: ByteDance Doubao Seed-2.0 family on Volcengine Ark — pro / lite / mini / code, OpenAI-compatible

package/templates/models/hunyuan.yaml

@@ -1,5 +1,6 @@
 id: hunyuan
 runtime: hunyuan
+revision: 2026100501
 displayName: Hunyuan (Tencent)
 displayName_zh: 混元（腾讯）
 description: Tencent Hunyuan Hy3 preview — 295B/21B MoE, agent-oriented with no_think / think_low / think_high modes

package/templates/models/kimi.yaml

@@ -1,5 +1,6 @@
 id: kimi
 runtime: kimi
+revision: 2026100501
 displayName: Kimi (Moonshot AI)
 displayName_zh: Kimi（月之暗面）
 description: Kimi K3 / K2.6 via OpenAI-compatible chat completions API

package/templates/models/mimo-token-plan-china.yaml

@@ -1,5 +1,6 @@
 id: mimo-token-plan-china
 runtime: anthropic-messages
+revision: 2026100501
 displayName: Xiaomi MiMo
 displayName_zh: 小米 MiMo
 description: Xiaomi MiMo via its Anthropic-compatible gateway — Token Plan (China / Singapore / Europe clusters) or pay-as-you-go. Supports 1M context window.

package/templates/models/minimax.yaml

@@ -1,5 +1,6 @@
 id: minimax
 runtime: minimax
+revision: 2026100501
 displayName: MiniMax
 displayName_zh: MiniMax
 description: MiniMax M3 / M2.7 family via the Anthropic-compatible Messages API

package/templates/models/openai.yaml

@@ -1,5 +1,6 @@
 id: openai
 runtime: openai-chat
+revision: 2026100501
 displayName: OpenAI (Generic)
 displayName_zh: OpenAI（通用）
 description: Generic OpenAI-compatible chat completions. Use for native OpenAI or any third-party OpenAI-compatible gateway. Capabilities are exposed openly — the user is expected to know what their endpoint supports.

package/templates/models/qwen.yaml

@@ -1,5 +1,6 @@
 id: qwen
 runtime: qwen
+revision: 2026100501
 displayName: Qwen (Aliyun)
 displayName_zh: 通义千问（阿里云）
 description: Qwen3 family on Aliyun Bailian via the Anthropic-compatible Messages API

package/templates/models/zhipu.yaml

@@ -1,5 +1,6 @@
 id: zhipu
 runtime: zhipu
+revision: 2026100501
 displayName: Zhipu AI (GLM)
 displayName_zh: 智谱 AI（GLM）
 description: Zhipu GLM-5.x family on open.bigmodel.cn, OpenAI-compatible chat completions

package/templates/skills/cron/SKILL.md

@@ -38,6 +38,12 @@ loop). No restart needed after edits.
   the result back to the current chat with `{{channel.chat_id}}` — the user
   almost always means "send it to me here"
 - **edit / pause / resume** → `update`, `enable`, `disable`
+- **skip / pause until a date / resume** ("no A-share report over the holiday",
+  "skip this week", "send until month end", "start again") → `update <id>
+  --active-from …` / `--active-until …` (see *Active window* in step 2).
+  A pause with a known end date goes through `--active-from`, **not**
+  `disable` — a disabled job only comes back if someone remembers to
+  re-enable it
 - **delete / remove** → `delete`. From inside a channel the user won't
   know the cron `id`; first run `list --chat-id {{channel.chat_id}}` to find
   the matching job, confirm which one, then delete by id
@@ -54,7 +60,7 @@ Sensible defaults if user didn't say:
 | targets    | none (log only — the run shows in admin Cron tab)             |
 | label      | summarize from prompt + schedule                              |
 | timeout    | leave unset (3600s) — see `--timeout-sec` below               |
-| session    | leave unset (the job's own `cron-<jobId>`) — see `--session` below |
+| session    | leave unset (the job's own session) — see `--session` below   |
 | prompt     | pass the user's task as-is — the runner prepends an "unattended run, don't ask questions" notice to every fire, so don't add one yourself |
 
 ### 2. Translate the schedule
@@ -89,7 +95,9 @@ heavy analysis run that needs more than an hour (`--timeout-sec 7200`). On
 `update`, pass `--timeout-sec ""` to clear back to the default.
 
 **Run inside an existing session (`--session`, optional)** — by default every
-job accumulates its own `cron-<jobId>` session. Pass `--session <rootSessionId>`
+job accumulates its own session (same id as the job; older jobs keep the
+`cron-cron-…` session they already have, shown as their `session_id`).
+Pass `--session <rootSessionId>`
 only when the user wants the runs to continue a specific conversation (the
 agent then sees that session's history). Rules the user should hear once:
 - the session keeps **its own agent and access level** — `--agent` only
@@ -100,7 +108,31 @@ agent then sees that session's history). Rules the user should hear once:
 
 Root session ids only (no `>` sub-session paths). You don't know session ids
 yourself — ask the user (admin → Sessions) rather than guessing. On `update`,
-pass `--session ""` to go back to the job's own session.
+pass `--session ""` to go back to the job's own session (for an older job
+pinned to `cron-cron-…` that starts a fresh session — leave it alone unless
+the user asks).
+
+**Active window (`--active-from` / `--active-until`, optional, recurring
+jobs only)** — the schedule fires only inside `[from, until)`; fires outside
+it are skipped silently (no run recorded). "Run now" ignores the window.
+- Values: `YYYY-MM-DD`, ISO-8601 (`2026-10-08T09:00`) or unix ms. A time
+  without an offset is read in the **job's timezone** (on `update`: the new
+  `--timezone` if you pass one, else the job's stored one; host time if
+  the job has none).
+- A bare date means **that day 00:00** for `--active-from`, and **the next
+  day 00:00** for `--active-until` — the date itself is included
+  (`--active-until 2026-10-31` still fires on the 31st).
+- `--active-from` may be in the past (= active now); `--active-until` must
+  be in the future and after `--active-from`. `""` clears either bound.
+- Past `--active-until` the job stays enabled but stops firing; the admin
+  shows it as "ended". Extending `--active-until` brings it back.
+- Rejected together with `--run-at` (one-shot jobs have no window).
+- There is **no "skip a middle stretch, keep firing before it"**: one
+  window per job. If the user asks for that (e.g. "keep sending this week,
+  skip next week, then resume"), say so plainly and offer a workaround —
+  usually: leave it running now and push `--active-from` past the gap just
+  before the gap starts (a one-shot cron can remind you), or create a
+  second job for the earlier stretch with an `--active-until`.
 
 ### 3. Channels (only when delivery is wanted)
 
@@ -188,12 +220,27 @@ shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py update <id> \
   --schedule "0 10 * * 1"
 ```
 
-**Pause / unpause:**
+**Pause / unpause (no end date in mind):**
 ```bash
 shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py disable <id>
 shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py enable <id>
 ```
 
+**Skip / pause until a date / resume (active window):** find the job first
+(`list`, or `list --chat-id {{channel.chat_id}}` inside a chat), then:
+```bash
+# "Don't send the English weekly this week" → active from next Monday
+shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py update <id> --active-from 2026-10-12
+# "No A-share report over the National Day holiday" → resume on Oct 8
+shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py update <id> --active-from 2026-10-08
+# "Send it until the end of the month" → Oct 31 still fires
+shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py update <id> --active-until 2026-10-31
+# "Start sending again" → clear the pause
+shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py update <id> --active-from ""
+```
+Prefer this over `disable` whenever the pause has a known end — it resumes
+on its own.
+
 **Delete:**
 ```bash
 shell_exec: python3 ~/.halo/global/skills/cron/manage-cron.py delete <id>
@@ -229,6 +276,14 @@ picks it up within ~10s. Example:
 > Created cron job `cron-mphsy-abc123`: runs the `default` agent every day at 9am,
 > output will be pushed to this chat. The runner picks it up within ~10s.
 
+For an active-window change, say when it resumes / stops — read
+`active_from_iso` / `active_until_iso` back with `get <id>` and name the
+first fire on or after it from the schedule (`get` has no computed next-run
+field; the admin Cron tab shows it):
+
+> Paused `A-share daily brief` over the holiday: it resumes on Oct 8 —
+> next run Oct 8 09:30 (Asia/Shanghai). Nothing is sent before that.
+
 ## Patterns that go sideways
 
 - **telegram / slack / feishu / wecom target without a chatId** — every fire errors with

package/templates/skills/cron/manage-cron.py

@@ -32,6 +32,12 @@ Subcommands:
 
 `--targets` arg: comma-separated `channelType:accountId` pairs.
   example: `--targets telegram:halo_agent_bot,wechat:alice`
+
+`--active-from` / `--active-until` (recurring jobs only): the schedule fires
+only inside [from, until); fires outside it are skipped silently. Accepts
+YYYY-MM-DD, ISO-8601 or unix ms; a time without offset is read in the job's
+timezone. A bare date means that day 00:00 for --active-from and the NEXT
+day 00:00 for --active-until (so the date itself is included). `""` clears.
 """
 import argparse
 import json
@@ -126,8 +132,8 @@ SESSION_ID_RE = re.compile(r'^[A-Za-z0-9_:-]{1,200}$')
 
 def parse_session(raw: str | None) -> str | None:
     """Validate --session: a root session id (same rule as the admin REST
-    routes — no `>` sub-session path). Empty = None = the job's own
-    `cron-<jobId>` session."""
+    routes — no `>` sub-session path). Empty = None = the job's own session
+    (same id as the job)."""
     if raw is None or not raw.strip():
         return None
     v = raw.strip()
@@ -157,6 +163,90 @@ def parse_run_at(raw: str | None) -> int | None:
     return int(dt.timestamp() * 1000)
 
 
+DATE_RE = re.compile(r'^\d{4}-\d{2}-\d{2}$')
+
+
+def _tzinfo(tz_name: str | None):
+    """ZoneInfo for an IANA name; None for unset or unknown names."""
+    if not tz_name:
+        return None
+    from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
+    try:
+        return ZoneInfo(tz_name)
+    except (ZoneInfoNotFoundError, ValueError):
+        return None
+
+
+def job_tz(tz_name: str | None):
+    """tzinfo for the job's timezone; None = host-local."""
+    tz = _tzinfo(tz_name)
+    if tz_name and tz is None:
+        die(f'unknown timezone "{tz_name}"; expected an IANA name like Asia/Shanghai')
+    return tz
+
+
+def parse_active_bound(flag: str, raw: str, tz_name: str | None) -> int | None:
+    """Parse --active-from / --active-until into epoch ms; "" = None (clear).
+
+    YYYY-MM-DD → that day 00:00 (from) or the NEXT day 00:00 (until, so the
+    date is included). ISO without offset / bare dates are read in the job's
+    timezone (host-local when the job has none)."""
+    raw = raw.strip()
+    if not raw:
+        return None
+    if raw.isdigit():
+        return int(raw)
+    from datetime import datetime, timedelta
+    tz = job_tz(tz_name)
+    if DATE_RE.match(raw):
+        try:
+            dt = datetime.strptime(raw, '%Y-%m-%d')
+        except ValueError:
+            die(f'invalid {flag} "{raw}": not a calendar date')
+        if flag == '--active-until':
+            dt += timedelta(days=1)
+    else:
+        try:
+            dt = datetime.fromisoformat(raw)
+        except ValueError:
+            die(f'invalid {flag} "{raw}": expected YYYY-MM-DD, ISO-8601 (e.g. 2026-10-08T09:00) or unix ms')
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=tz) if tz else dt.astimezone()
+    return int(dt.timestamp() * 1000)
+
+
+def validate_active_window(run_at, active_from, active_until, new_until: bool):
+    """Same rules as the admin REST routes, on the merged row: recurring jobs
+    only; a newly set --active-until must be in the future and after from."""
+    if active_from is None and active_until is None:
+        return
+    if run_at is not None:
+        die('--active-from/--active-until only apply to recurring jobs, not --run-at one-shots '
+            '(on update, clear them with --active-from "" --active-until "")')
+    if new_until and active_until is not None and active_until <= now_ms():
+        die('--active-until must be in the future')
+    if active_from is not None and active_until is not None and active_until <= active_from:
+        die('--active-until must be after --active-from')
+
+
+def require_active_columns(conn: sqlite3.Connection):
+    """cron.db gains active_from/active_until via the server's boot migration;
+    a server that hasn't restarted onto this version yet won't have them."""
+    cols = {r[1] for r in conn.execute('PRAGMA table_info(cron_jobs)').fetchall()}
+    if 'active_from' not in cols or 'active_until' not in cols:
+        die('cron.db has no active_from/active_until columns yet — restart the halo server '
+            'once so it migrates cron.db, then retry')
+
+
+def ms_to_iso(ms, tz_name: str | None) -> str | None:
+    if ms is None:
+        return None
+    from datetime import datetime
+    tz = _tzinfo(tz_name)  # display only — never die on a stored bad tz
+    dt = datetime.fromtimestamp(ms / 1000, tz=tz) if tz else datetime.fromtimestamp(ms / 1000).astimezone()
+    return dt.isoformat(timespec='minutes')
+
+
 def row_to_job(row: sqlite3.Row) -> dict:
     j = dict(row)
     try:
@@ -164,6 +254,11 @@ def row_to_job(row: sqlite3.Row) -> dict:
     except Exception:
         j['targets'] = []
     j['enabled'] = bool(j['enabled'])
+    # Active window: ms (as stored) + ISO in the job's timezone. Pre-migration
+    # dbs have no columns — report them as unset.
+    for k in ('active_from', 'active_until'):
+        j.setdefault(k, None)
+        j[f'{k}_iso'] = ms_to_iso(j[k], j.get('timezone'))
     return j
 
 
@@ -222,6 +317,9 @@ def cmd_create(args):
         die('--run-at must be in the future')
     timeout_sec = parse_timeout_sec(args.timeout_sec)
     session_id = parse_session(args.session)
+    active_from = parse_active_bound('--active-from', args.active_from or '', args.timezone)
+    active_until = parse_active_bound('--active-until', args.active_until or '', args.timezone)
+    validate_active_window(run_at_ms, active_from, active_until, new_until=True)
 
     job_id = args.id or gen_job_id()
     now = now_ms()
@@ -232,25 +330,31 @@ def cmd_create(args):
     if existing:
         die(f'job id {job_id} already exists; pick another with --id', 2)
 
+    values: dict[str, object] = {
+        'id': job_id,
+        'label': args.label,
+        'workspace_path': args.workspace,
+        'agent_id': args.agent or 'default',
+        'user_prompt': args.prompt,
+        'schedule': args.schedule or '',
+        'run_at': run_at_ms,
+        'timezone': args.timezone,
+        'timeout_sec': timeout_sec,
+        'session_id': session_id,
+        'targets': json.dumps(targets, ensure_ascii=False),
+        'enabled': 0 if args.disabled else 1,
+        'created_at': now,
+        'updated_at': now,
+    }
+    # Only name the window columns when used, so jobs without a window can
+    # still be created on a cron.db the server hasn't migrated yet.
+    if active_from is not None or active_until is not None:
+        require_active_columns(conn)
+        values['active_from'] = active_from
+        values['active_until'] = active_until
     conn.execute(
-        'INSERT INTO cron_jobs(id, label, workspace_path, agent_id, user_prompt, schedule, '
-        'run_at, timezone, timeout_sec, session_id, targets, enabled, created_at, updated_at) '
-        'VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)',
-        (
-            job_id,
-            args.label,
-            args.workspace,
-            args.agent or 'default',
-            args.prompt,
-            args.schedule or '',
-            run_at_ms,
-            args.timezone,
-            timeout_sec,
-            session_id,
-            json.dumps(targets, ensure_ascii=False),
-            0 if args.disabled else 1,
-            now, now,
-        ),
+        f'INSERT INTO cron_jobs({", ".join(values)}) VALUES ({", ".join("?" for _ in values)})',
+        list(values.values()),
     )
     conn.commit()
     print(json.dumps({'id': job_id, 'created': True}, ensure_ascii=False))
@@ -279,12 +383,29 @@ def cmd_update(args):
         # same convention as `--run-at ""`.
         sets.append(('timeout_sec', parse_timeout_sec(args.timeout_sec)))
     if args.session is not None:
-        # `--session ""` clears back to the job's own cron-<jobId> session.
+        # `--session ""` clears back to the job's own session (id = job id).
         sets.append(('session_id', parse_session(args.session)))
     if args.targets is not None:
         sets.append(('targets', json.dumps(parse_targets(args.targets), ensure_ascii=False)))
+    cur = dict(row)
+    if args.active_from is not None or args.active_until is not None:
+        require_active_columns(conn)
+        # Bare dates / offset-less times read in the job's timezone after
+        # this update: the new --timezone if given, else the stored one.
+        tz = args.timezone if args.timezone is not None else cur.get('timezone')
+        if args.active_from is not None:
+            sets.append(('active_from', parse_active_bound('--active-from', args.active_from, tz)))
+        if args.active_until is not None:
+            sets.append(('active_until', parse_active_bound('--active-until', args.active_until, tz)))
+    # Window rules on the merged row (also catches --run-at on a job that
+    # still has a window).
+    merged = {**cur, **dict(sets)}
+    validate_active_window(
+        merged.get('run_at'), merged.get('active_from'), merged.get('active_until'),
+        new_until=bool(args.active_until and args.active_until.strip()),
+    )
     if not sets:
-        die('nothing to update — pass at least one --label/--workspace/--agent/--prompt/--schedule/--run-at/--timezone/--timeout-sec/--session/--targets')
+        die('nothing to update — pass at least one --label/--workspace/--agent/--prompt/--schedule/--run-at/--timezone/--timeout-sec/--session/--targets/--active-from/--active-until')
     sets.append(('updated_at', now_ms()))
 
     placeholders = ', '.join(f'{k} = ?' for k, _ in sets)
@@ -432,10 +553,16 @@ def main():
     c.add_argument('--timeout-sec', dest='timeout_sec',
                    help='max run time in seconds, 60–21600 (default: unset = 3600)')
     c.add_argument('--session',
-                   help="run in this existing root session id (default: the job's own cron-<jobId>)")
+                   help="run in this existing root session id (default: the job's own session, same id as the job)")
     c.add_argument('--targets',
                    help='comma-separated channelType:accountId[:chatId] list, or JSON array. '
                         'chatId pins delivery to a specific chat (e.g. when scheduling from inside a chat).')
+    c.add_argument('--active-from', dest='active_from',
+                   help='recurring only: start firing from this time — YYYY-MM-DD (= that day 00:00), '
+                        'ISO-8601 or unix ms; no offset = the job timezone. Use it to "pause until" a date.')
+    c.add_argument('--active-until', dest='active_until',
+                   help='recurring only: stop firing at this time (exclusive) — YYYY-MM-DD (= the NEXT day 00:00, '
+                        'so the date itself still fires), ISO-8601 or unix ms; must be in the future')
     c.add_argument('--disabled', action='store_true', help='create paused (default: enabled)')
     c.set_defaults(func=cmd_create)
 
@@ -451,8 +578,15 @@ def main():
     u.add_argument('--timeout-sec', dest='timeout_sec',
                    help='max run time in seconds, 60–21600; pass "" to clear back to default 3600')
     u.add_argument('--session',
-                   help='root session id to run in; pass "" to clear back to cron-<jobId>')
+                   help='root session id to run in; pass "" to clear back to the job\'s own session '
+                        '(same id as the job — an older job pinned to cron-cron-… then starts a fresh one)')
     u.add_argument('--targets')
+    u.add_argument('--active-from', dest='active_from',
+                   help='recurring only: YYYY-MM-DD (= that day 00:00), ISO-8601 or unix ms, in the job timezone; '
+                        'pass "" to clear (= active now)')
+    u.add_argument('--active-until', dest='active_until',
+                   help='recurring only: YYYY-MM-DD (= the NEXT day 00:00, date included), ISO-8601 or unix ms, '
+                        'in the job timezone; pass "" to clear (= no end)')
     u.set_defaults(func=cmd_update)
 
     e = sub.add_parser('enable')

package/templates/skills/extension/SKILL.md

@@ -1,19 +1,20 @@
 ---
 name: extension
-description: Install / list / remove admin preview extensions (`~/.halo/global/extensions/<id>/`) — static viewers the admin editor loads for file types it has no built-in preview for (e.g. `.glb`). `/extension install <id|zip|url>` pulls the latest `<id>-v*` release from the configured hub (default turmind/halo-hub). Activate when the user wants to preview a file type the editor can't open, or asks to install / update / remove an extension.
+description: Install / list / remove admin preview extensions (`~/.halo/global/extensions/<id>/`) — static viewers the admin editor loads for file types it has no built-in preview for (e.g. `.glb`) — and update the model provider list (new models, changed provider endpoints) from the hub. `/extension install <id|zip|url>` pulls the latest `<id>-v*` release from the configured hub (default turmind/halo-hub); `/extension models` pulls the latest `models-v*` provider configs. Activate when the user wants to preview a file type the editor can't open, asks to install / update / remove an extension, or asks to update the model list / get a newly released model ("更新模型列表").
 command: /extension
 requiresAccess: full
 verbs:
   - { name: install, desc: "Install or upgrade an extension from the hub (by id), a local zip path, or a zip URL" }
   - { name: list,    desc: List installed extensions with version and status }
   - { name: remove,  desc: Remove an installed extension by id }
+  - { name: models,  desc: "Update the model provider list from the hub (newest models-v* release)" }
 ---
 
 # extension
 
 The requested action is **`$1`**; everything after it is the payload
 (`$ARGUMENTS` minus the first token). With natural language ("install the glb
-viewer"), infer both.
+viewer", "更新模型列表"), infer both.
 
 Preview extensions are static bundles the admin editor loads in a sandboxed
 iframe to preview file types it has no built-in viewer for (e.g. `.glb`).
@@ -70,6 +71,33 @@ short `-` list; `(none)` when empty.
 `bash EXT remove <id>` → deletes `~/.halo/global/extensions/<id>/`. Confirm with
 the user first if they didn't name the id explicitly.
 
+## models
+
+Updates the model provider list (the providers / models / endpoints the agent
+editor offers) from the hub's newest `models-v*` release. Also run it for
+natural language like "更新模型列表" / "update the model list" / "is there a
+newer model". Run it right away:
+
+`HALO_HUB_REPO='{{params.hub_repo}}' bash EXT models update`
+
+The script prints one line per provider (`install` / `up-to-date` / `skip` /
+`refuse` + reason), then a summary as its last line. By exit code:
+- `0` — done. Relay the summary line; name any `refuse`d provider and its reason
+  (e.g. "needs a newer halo" → run `halo upgrade`). Takes effect immediately
+  — no restart, no page reload. Existing agents keep their own model / endpoint
+  (they live in each agent.yaml); only the choices offered change.
+- `3` — nothing was written: a provider is new or its endpoints change. Show
+  the user the indented endpoint lines verbatim and ask whether to apply them.
+  Only after they agree, run the same command with `--yes` appended. Never add
+  `--yes` on your own.
+- anything else — relay the error. Model configs come only from an https
+  GitHub / Gitea / Forgejo / GitLab hub with releases; a local / plain-git /
+  http hub is refused. "release API failed" / "no release for models" → same
+  hints as for `install`.
+
+`bash EXT models list` shows each provider, which copy is in effect (`bundled`
+= shipped with halo, `hub` = installed by this command) and its revision.
+
 ## Notes
 
 - Extensions are global (all workspaces on this server). Say so if the user

package/templates/skills/extension/config.yaml

@@ -4,6 +4,6 @@ description_zh: 安装 / 列出 / 删除管理后台的预览扩展。
 
 params:
   - key: hub_repo
-    description: "Hub that `install <id>` pulls from: a GitHub / Gitea / Forgejo / GitLab repo URL (newest `<id>-v*` release zip), `owner/repo` for GitHub, or any git URL / local repo path (newest `<id>-v<x.y.z>` tag). Empty = https://github.com/turmind/halo-hub."
-    description_zh: "`install <id>` 的来源仓库：GitHub / Gitea / Forgejo / GitLab 仓库地址（取最新的 `<id>-v*` release zip），GitHub 可简写为 `owner/repo`；也可以是任意 git 地址或本地仓库路径（取最新的 `<id>-v<x.y.z>` 标签）。留空 = https://github.com/turmind/halo-hub。"
+    description: "Hub that `install <id>` and `models` pull from: a GitHub / Gitea / Forgejo / GitLab repo URL (newest `<id>-v*` release zip), `owner/repo` for GitHub, or any git URL / local repo path (newest `<id>-v<x.y.z>` tag; not for `models`, which needs an https release hub). Empty = https://github.com/turmind/halo-hub."
+    description_zh: "`install <id>` 和 `models` 的来源仓库：GitHub / Gitea / Forgejo / GitLab 仓库地址（取最新的 `<id>-v*` release zip），GitHub 可简写为 `owner/repo`；也可以是任意 git 地址或本地仓库路径（取最新的 `<id>-v<x.y.z>` 标签；`models` 不支持，它只认 https 的 release 仓库）。留空 = https://github.com/turmind/halo-hub。"
     default: https://github.com/turmind/halo-hub