@turmind/halo 1.6.1 → 1.6.2

package/templates/models/deepseek.yaml

@@ -1,6 +1,6 @@
 id: deepseek
 runtime: deepseek
-revision: 2026101002
+revision: 2026101101
 displayName: DeepSeek
 displayName_zh: DeepSeek
 description: DeepSeek V4 via OpenAI-compatible chat completions API
@@ -15,9 +15,18 @@ secrets:
     secret: true
 
 # Capabilities verified via live probes (2026-05-27, re-probed 2026-09-24).
-# reasoning_effort accepts { low, medium, high, max, xhigh } — bogus values
-# 400 with the full valid list in the error. Caching is automatic;
-# cached_tokens shows up from turn 2.
+# Caching is automatic; cached_tokens shows up from turn 2.
+#
+# Thinking (re-probed 2026-10-11, api-docs.deepseek.com/guides/thinking_mode +
+# GET /models): `thinking:{type:enabled|disabled}` toggles, on by default;
+# `reasoning_effort` grades it. The API's real levels are low / high / max
+# (/models `effort.supported_levels`, default high); it also accepts none /
+# minimal / medium / xhigh / ultra and folds them itself (minimal→low,
+# medium/xhigh→high, ultra→max) — all 200 live, a bogus value 422s with that
+# list. So the presets are the three real levels and deepseek-agent.ts passes
+# the label through unmapped: an older agent.yaml with medium / xhigh keeps
+# working at high. Before 2026-10-11 the runtime sent no reasoning_effort at
+# all (every agent ran at the API default, high).
 #
 # Vision is per model: `image_url` blocks are no longer rejected at the
 # deserialize layer on either id, but only `deepseek-flash` (V4.1-Flash,
@@ -42,13 +51,11 @@ models:
           - { value: 5m, label: 5min }
       thinking:
         defaultEnabled: true
-        default: xhigh
+        default: high
         effortPresets:
           - { value: low, label: Low }
-          - { value: medium, label: Medium }
           - { value: high, label: High }
           - { value: max, label: Max }
-          - { value: xhigh, label: XHigh }
 
   # DeepSeek retired the versioned flash ids (deepseek-v4-flash,
   # deepseek-v4-flash-vision-exp) in favour of a rolling `deepseek-flash`
@@ -70,10 +77,8 @@ models:
           - { value: 5m, label: 5min }
       thinking:
         defaultEnabled: true
-        default: xhigh
+        default: high
         effortPresets:
           - { value: low, label: Low }
-          - { value: medium, label: Medium }
           - { value: high, label: High }
           - { value: max, label: Max }
-          - { value: xhigh, label: XHigh }

package/templates/models/hunyuan.yaml

@@ -1,6 +1,6 @@
 id: hunyuan
 runtime: hunyuan
-revision: 2026101003
+revision: 2026101101
 displayName: Hunyuan (Tencent)
 displayName_zh: 混元（腾讯）
 description: Tencent Hunyuan Hy4 preview / Hy3 on TokenHub — agent-oriented MoE with no_think / think_low / think_high modes
@@ -42,14 +42,18 @@ secrets:
 #   - Caching re-verified on hy3 / hy4-preview (turn 2 cached_tokens ≈ prompt).
 #   - Tool calls verified on hy3 / hy4-preview. Image: hy3 / hy4-preview both
 #     answered wrongly about a solid red PNG — keep image:false.
-#   - Max output: no official per-model number found (search-only claims of
-#     64K, unconfirmed: https://www.tencentcloud.com/techpedia/147980); the
-#     gateway accepted max_tokens 2,000,000 without error, so it isn't a hard
-#     check either — maxOutputTokens stays at the conservative 16384.
+#   - Max output: the gateway accepted max_tokens 2,000,000 without error, so
+#     a rejection probe proves nothing.
+# 2026-10-11: the TokenHub model list (cloud.tencent.com/document/product/1823/130051,
+# updated 2026-09-28, 语言模型 table: 上下文窗口 / 最大输入 / 最大输出) gives
+# Hy4 preview 1M / 960k / 64k and Hy3 256k / 192k / 128k → maxOutputTokens
+# 65536 / 131072. hy3-preview is not on that page → stays at 16384. (One
+# long-output run on hy3 with max_tokens 200000 stopped at 147,389 completion
+# tokens, finish_reason length — above the doc figure; the doc value is used.)
 models:
   - id: hy3
     displayName: Hunyuan Hy3
-    maxOutputTokens: 16384
+    maxOutputTokens: 131072
     contextWindow: 262144
     maxContextWindow: 262144
     compressAt: 0.9
@@ -77,7 +81,7 @@ models:
 
   - id: hy4-preview
     displayName: Hunyuan Hy4 Preview
-    maxOutputTokens: 16384
+    maxOutputTokens: 65536
     contextWindow: 272000
     maxContextWindow: 1048576
     compressAt: 0.9

package/templates/models/qwen.yaml

@@ -1,6 +1,6 @@
 id: qwen
 runtime: qwen
-revision: 2026101003
+revision: 2026101101
 displayName: Qwen (Aliyun)
 displayName_zh: 通义千问（阿里云）
 description: Qwen3 family on Aliyun Bailian via the Anthropic-compatible Messages API
@@ -32,20 +32,31 @@ secrets:
 #     despite docs saying a vision-specific model is required);
 #     qwen3.7-max returned `Unexpected item type in content` (not
 #     supported). So image:false on max, image:true on plus.
-#   - effort presets: pure Halo-side abstraction. Same effort→budget
-#     table as bedrock-agent.ts via qwen-agent.ts's effortToBudget().
+#   - effort presets on the manual (3.7 / 3.6) models: pure Halo-side
+#     abstraction. Same effort→budget table as bedrock-agent.ts via
+#     qwen-agent.ts's effortToBudget().
 #   - temperature range: doc says [0, 2) instead of Anthropic's [0, 1] —
 #     doesn't affect Halo since we don't expose temperature in the form.
 # 2026-10-10 census (help.aliyun.com/zh/model-studio/anthropic-api-messages +
 # the text-generation model guide, both list 3.8-max / 3.8-flash / 3.7-plus as
 # current, 1M context): live-probed on the Anthropic endpoint — manual
 # thinking.budget_tokens still accepted (the doc marks it "即将废弃" in favour
-# of output_config.effort; qwen-agent still sends budget_tokens), thinking
+# of output_config.effort — see 2026-10-11 below), thinking
 # disabled honored, tool_use OK, image OK, cache_control ephemeral hit on
 # turn 2. `ttl: '1h'` is accepted but the entry is still written as
 # ephemeral_5m, so 5m stays the only TTL preset. Max output isn't published
 # per model on these pages — maxOutputTokens stays 16384. 3.7-max / 3.6-plus
 # are not on the deprecation page, kept below the new generation.
+# 2026-10-11 (help.aliyun.com/zh/model-studio/anthropic-api-messages,
+# `output_config.effort`): the doc lists effort values for qwen3.8-max /
+# qwen3.8-flash only — xhigh (default) / medium / low, with max and high
+# folded onto xhigh by the API. So 3.8 is `mode: adaptive`: qwen-agent sends
+# thinking:{type:'enabled'} + output_config.effort. An agent.yaml that already
+# carries an explicit budget_tokens keeps sending it (still accepted). For 3.8
+# the same doc says max_tokens caps reply + thinking together. The 3.7 / 3.6
+# models have no documented effort values → they stay `mode: manual`
+# (budget_tokens). Thinking-field omitted = thinking on for every model here,
+# so off is always sent as type:'disabled'.
 models:
   - id: qwen3.8-max
     displayName: Qwen3.8 Max
@@ -63,13 +74,13 @@ models:
         ttlPresets:
           - { value: 5m, label: 5min }
       thinking:
-        mode: manual
+        mode: adaptive
         defaultEnabled: true
-        default: high
-        defaultBudgetTokens: 8192
+        default: xhigh
         effortPresets:
+          - { value: low, label: Low }
           - { value: medium, label: Medium }
-          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
 
   - id: qwen3.7-plus
     displayName: Qwen3.7 Plus
@@ -111,13 +122,13 @@ models:
         ttlPresets:
           - { value: 5m, label: 5min }
       thinking:
-        mode: manual
+        mode: adaptive
         defaultEnabled: true
-        default: high
-        defaultBudgetTokens: 8192
+        default: xhigh
         effortPresets:
+          - { value: low, label: Low }
           - { value: medium, label: Medium }
-          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
 
   - id: qwen3.7-max
     displayName: Qwen3.7 Max

package/templates/skills/cron/SKILL.md

@@ -79,7 +79,8 @@ Sensible defaults if user didn't say:
 
 **One-shot (at-mode)** — pass `--run-at` instead of `--schedule`. The job
 fires once at that instant and auto-disables. Useful for "remind me at 3pm
-today" or "send the report next Monday morning".
+today" or "send the report next Monday morning". On `update`, passing one
+of the two switches the job's mode and clears the other; never pass both.
 
 | Spoken                            | Flag                                      |
 |-----------------------------------|-------------------------------------------|

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

@@ -96,7 +96,10 @@ def parse_targets(raw: str) -> list[dict]:
     if not raw:
         return []
     if raw.startswith('['):
-        return json.loads(raw)
+        try:
+            return json.loads(raw)
+        except ValueError as err:
+            die(f'invalid --targets JSON: {err}')
     out = []
     for token in raw.split(','):
         token = token.strip()
@@ -127,7 +130,8 @@ def parse_timeout_sec(raw: str | None) -> int | None:
     return v
 
 
-SESSION_ID_RE = re.compile(r'^[A-Za-z0-9_:-]{1,200}$')
+# Always fullmatch: with `$`, re.match also accepts one trailing "\n".
+SESSION_ID_RE = re.compile(r'[A-Za-z0-9_:-]{1,200}')
 
 
 def parse_session(raw: str | None) -> str | None:
@@ -137,7 +141,7 @@ def parse_session(raw: str | None) -> str | None:
     if raw is None or not raw.strip():
         return None
     v = raw.strip()
-    if not SESSION_ID_RE.match(v):
+    if not SESSION_ID_RE.fullmatch(v):
         die(f'invalid --session "{raw}": expected a root session id (letters, digits, _ : - only; no ">" sub-session path)')
     return v
 
@@ -393,6 +397,10 @@ def cmd_create(args):
     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)
 
+    # A custom id becomes the job's own session id (`cron-<id>`), so it takes
+    # the session-id charset: a `>` would make it a sub-session path.
+    if args.id and not SESSION_ID_RE.fullmatch(args.id):
+        die(f'invalid --id "{args.id}": letters, digits, _ : - only (it becomes the job\'s session id)')
     job_id = args.id or gen_job_id()
     check_session_rules({
         'id': job_id, 'workspace_path': args.workspace, 'agent_id': args.agent or 'default',
@@ -451,12 +459,19 @@ def cmd_update(args):
     if args.workspace is not None: sets.append(('workspace_path', args.workspace))
     if args.agent is not None: sets.append(('agent_id', args.agent))
     if args.prompt is not None: sets.append(('user_prompt', args.prompt))
+    # Recurring vs one-shot, same as the REST PUT: both in one call is an
+    # error; one side alone is a mode switch that clears the other (else the
+    # runner prefers run_at and its post-fire auto-disable kills the schedule).
+    has_schedule = bool(args.schedule and args.schedule.strip())
+    run_at_ms = parse_run_at(args.run_at) if args.run_at is not None else None
+    if run_at_ms is not None and run_at_ms <= now_ms():
+        die('--run-at must be in the future')
+    if has_schedule and run_at_ms is not None:
+        die('--schedule and --run-at are mutually exclusive')
     if args.schedule is not None: sets.append(('schedule', args.schedule))
-    if args.run_at is not None:
-        run_at_ms = parse_run_at(args.run_at)
-        if run_at_ms is not None and run_at_ms <= now_ms():
-            die('--run-at must be in the future')
-        sets.append(('run_at', run_at_ms))
+    elif run_at_ms is not None: sets.append(('schedule', ''))
+    if args.run_at is not None: sets.append(('run_at', run_at_ms))
+    elif has_schedule: sets.append(('run_at', None))
     if args.timezone is not None: sets.append(('timezone', args.timezone))
     if args.timeout_sec is not None:
         # `--timeout-sec ""` clears back to the runner default (NULL),
@@ -498,6 +513,8 @@ def cmd_update(args):
     # Window rules on the merged row (also catches --run-at on a job that
     # still has a window).
     merged = {**cur, **dict(sets)}
+    if not (merged.get('schedule') or '').strip() and merged.get('run_at') is None:
+        die('--schedule or --run-at required (this update would leave the job with neither)')
     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()),