@turmind/halo 2.2.3 → 2.2.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "2.2.3",
+  "version": "2.2.5",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",
@@ -60,6 +60,7 @@
     "semver": "^7.8.5",
     "sherpa-onnx-node": "1.13.8",
     "simple-git": "^3.27.0",
+    "undici": "6.29.1",
     "ws": "^8.21.3",
     "yaml": "^2.8.3",
     "yazl": "3.3.1"

package/templates/agents/default/agent.yaml

@@ -28,7 +28,7 @@ tools:
   - grep
   - glob
   - web_fetch
-  # Linked machines (`halo node`): grants all 8 node_* tools, full-access sessions only.
+  # Linked machines (`halo-node`): grants all 8 node_* tools, full-access sessions only.
   - node_list
 # Delegation: a non-empty `team` is what grants the whole session-tool bundle
 # (start_session / query_session / … / archive_session / query_agent) + the

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

@@ -1,6 +1,6 @@
 id: aws-bedrock-claude-invoke
 runtime: bedrock-invoke
-revision: 2026110401
+revision: 2026111001
 displayName: AWS Bedrock Claude (Invoke API)
 displayName_zh: AWS Bedrock Claude（Invoke API）
 description: Invokes Bedrock via InvokeModelWithResponseStream (streaming), sends raw Anthropic Messages API format
@@ -31,12 +31,73 @@ secrets:
     default: <<AWS_SECRET_ACCESS_KEY>>
     secret: true
 
+# Claude Mythos 5.2 (anthropic.claude-mythos-5-2) is deliberately not listed:
+# the same model as Fable 5.2 but a Preview limited to organisations approved
+# by Anthropic's Cyber / Life Sciences Verification Programs. It is absent
+# from list-foundation-models, and its anthropic. / global. / us. ids all
+# answer "provided model identifier is invalid", on this account (2026-11-10).
+
 # capability defaults (per model):
 #   promptCaching.default      → auto-on value when toggled / on provider switch
 #   thinking.defaultEnabled    → start with Thinking checked for new agents
 #   thinking.default           → effort selected when Thinking is on (adaptive)
 #   thinking.defaultBudgetTokens → budget selected when Thinking is on (manual)
 models:
+  # Claude Fable 5.2 (launch 2026-11-09; global. / us. profiles only, no
+  # in-Region id; the card lists bedrock-runtime only, and Mantle /v1/models in
+  # us-east-1 / us-east-2 / us-west-2 (2026-11-10) has no 5.2 — us-east-1
+  # carries only anthropic.claude-fable-5). Live-probed 2026-11-10 via
+  # InvokeModel, us-east-1 (global.*; us.* plain call only).
+  # DATA RETENTION: the card says to opt in to `aws_review` (a Covered Model).
+  # What was observed: us-east-1 here is set to provider_data_share and answers
+  # 200; eu-west-1 / ap-northeast-1 / us-west-2 are `inherit` and answer 400
+  # "data retention mode 'default' is not available for this model" (Fable 5.1
+  # does the same in eu-west-1, Opus 5.5 does not). Other modes weren't tried —
+  # account settings were left alone. Set the mode per Region
+  # (put-account-data-retention) before pointing an endpoint preset at one.
+  # Behaviour matches 5.1, re-probed: an invalid output_config.effort lists
+  # low/medium/high/xhigh/max and all five return 200; thinking is adaptive
+  # only — thinking.type enabled (budget_tokens) and disabled both 400, an
+  # omitted thinking field still thinks (the card's default effort is high;
+  # xhigh stays our pick, as for 5.1). Image OK (64×64 red PNG → "Red", blue →
+  # "Blue"). tool_use OK ({"city":"Paris"}, no-arg ping{}), but tool_choice
+  # any / tool 400 — only auto, which Halo sends. temperature other than 1 and
+  # top_p below 0.99 / any top_k 400 as deprecated (bedrock-agent.ts forwards
+  # them only if an agent sets them). Streaming events: message_start,
+  # content_block_start (thinking / text), thinking_delta + signature_delta,
+  # text_delta, content_block_stop, message_delta, message_stop. Prompt caching
+  # 5m and 1h both honoured (a ~19K-token system: 1h write 18913 in the 1h
+  # bucket, then 18913 read; 5m the same in the 5m bucket). max_tokens 128000
+  # → 200, 128001 → 400 "maximum allowed number of output tokens". 1M context:
+  # a 404K-token prompt → 200 (past our 272K default), 1,042,065 → 400
+  # "prompt is too long: … > 1000000 maximum", so maxContextWindow 1000000; a
+  # prompt near the limit wasn't sent (cost).
+  - id: global.anthropic.claude-fable-5-2
+    displayName: Claude Fable 5.2
+    maxOutputTokens: 128000
+    contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: xhigh
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+
   - id: global.anthropic.claude-fable-5-1
     displayName: Claude Fable 5.1
     maxOutputTokens: 128000

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

@@ -1,10 +1,10 @@
 id: aws-bedrock-openai
 runtime: bedrock-mantle
-revision: 2026110501
-displayName: AWS Bedrock OpenAI-compatible (GPT-6.1 / 6 / 5.6 · Grok 4.7 / 4.6 · Kimi K3 · GLM 5.3)
-displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6.1 / 6 / 5.6 · Grok 4.7 / 4.6 · Kimi K3 · GLM 5.3）
-description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6.1 (Sol / Terra), GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.7 / 4.6, Moonshot Kimi K3, Z.AI GLM 5.3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).
-description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/v1）调用 Bedrock 上的前沿模型——OpenAI GPT-6.1（Sol / Terra）、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）、xAI Grok 4.7 / 4.6、月之暗面 Kimi K3、智谱 GLM 5.3。仅支持跨区域推理 profile（global.*）——每个区域都能调所有模型，无区域分裂。不支持服务端 web_search（那是 Mantle 独有）。
+revision: 2026110801
+displayName: AWS Bedrock OpenAI-compatible (GPT-6.1 / 6 / 5.6 · Grok 4.7 / 4.6 · Kimi K3.1 / K3 · GLM 5.3)
+displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6.1 / 6 / 5.6 · Grok 4.7 / 4.6 · Kimi K3.1 / K3 · GLM 5.3）
+description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6.1 (Sol / Terra), GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.7 / 4.6, Moonshot Kimi K3.1 / K3, Z.AI GLM 5.3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).
+description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/v1）调用 Bedrock 上的前沿模型——OpenAI GPT-6.1（Sol / Terra）、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）、xAI Grok 4.7 / 4.6、月之暗面 Kimi K3.1 / K3、智谱 GLM 5.3。仅支持跨区域推理 profile（global.*）——每个区域都能调所有模型，无区域分裂。不支持服务端 web_search（那是 Mantle 独有）。
 # Why a second Bedrock/OpenAI provider next to aws-bedrock-mantle: this is the
 # bedrock-runtime host (https://bedrock-runtime.<region>.amazonaws.com/openai/v1),
 # which AWS now recommends for new applications; bedrock-mantle is the legacy
@@ -12,8 +12,11 @@ description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/
 # class (MantleAgent), different rules:
 #   - model ids MUST be cross-Region inference profiles (`global.*` / `us.*`);
 #     the bare foundation-model id 400s with "on-demand throughput isn't
-#     supported". Kimi K3 isn't on Mantle at all; Grok 4.6 is Mantle us-west-2
-#     only — here everything is reachable from any region.
+#     supported". Kimi K3.1 / K3 aren't on Mantle at all (the K3.1 card marks
+#     bedrock-mantle unsupported; live /v1/models in us-east-1 / us-east-2 /
+#     us-west-2, 2026-11-08: only kimi-k2.5 and kimi-k2-thinking, and the
+#     K3.1 / K3 ids 404 there); Grok 4.6 is Mantle us-west-2 only — here
+#     everything is reachable from any region.
 #   - live-probed 2026-09-24 with SigV4: all nine global.* ids below return
 #     200 (gpt-6.1-sol probed 2026-09-30) in us-east-1 / us-west-2 /
 #     eu-west-1 / ap-northeast-1 alike, so the endpoint is purely a latency /
@@ -426,6 +429,77 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
+  # Kimi K3.1 (Bedrock card launch 2026-11-04, startOfLifeTime 2026-11-06;
+  # INFERENCE_PROFILE only — global. / us. profiles, no in-Region id). The
+  # Bedrock model card for moonshotai.kimi-k3.1 is titled "Kimi K3.1 Pro -
+  # Highspeed" (the control plane just says "Kimi K3.1"; there is no plain
+  # "Kimi K3.1 Pro" card), so by the card this is Moonshot's
+  # kimi-k3.1-pro-highspeed (same intelligence as K3.1 Pro in kimi.yaml, ~2-3x
+  # output speed per Moonshot's K3.1 Pro quickstart); nothing beyond the card
+  # title confirms the mapping. 1M context: input + max_output_tokens is
+  # accepted up to exactly 1048576 (input 21 + 1048555 → 200, + 1048556 →
+  # 400; K3 identical), matching Moonshot's native K3.1 Pro context_length, so
+  # maxContextWindow 1048576. That bound is the context, not an output cap —
+  # 131072 / 131073 / 262144 / 1000000 all return 200 — so 131072 stays
+  # Moonshot's official default output (as K3.1 Pro / K3), not endpoint-proven.
+  # Global CRIS $4 / $20 per 1M (cache read $0.20, cache write 30 min $5; US
+  # CRIS $4.40 / $22); output tokens burn 10x of the TPM quota.
+  # Live-probed 2026-11-08 via /openai/v1 Responses, SigV4: reasoning.effort
+  # accepts none/low/medium/high/xhigh/max (the validation error lists them;
+  # "Is 9991 prime?" reasoning_tokens 0 / 54 / 73 / 64 / 88 / 79 — advisory,
+  # within noise like K3). Thinking off omits the field and the model still
+  # reasons (88 tokens; the card's default is max), so `none` gets no preset.
+  # text.verbosity low/medium/high accepted, no visible effect on a short
+  # prompt. Image input OK (64×64 red PNG → "Red", blue → "Blue"; the card
+  # wants images placed before text; a 1×1 red PNG came back "Black"). Tools:
+  # no-arg function → function_call ping{} (auto and required), get_weather
+  # {"city":"Paris"}. Streaming emits response.reasoning.delta (raw
+  # reasoning_text, no summary) like K3. Implicit caching OK (an ~8.6K-token
+  # system sent 3x: run 1 cached_tokens 0, runs 2-3 cached_tokens 8600 of 8607);
+  # the card's explicit checkpoints (>=256 tokens, >=30 min) aren't sent by
+  # MantleAgent, so the single 5m preset is a form hint only.
+  # ACCOUNT QUOTA CAVEAT: the card default for "Global cross-region model
+  # inference tokens per minute for Moonshot AI Kimi K3.1" is 1M, but this
+  # probing account's applied value is 0 in us-east-1 / us-east-2 / us-west-2
+  # / eu-west-1 / ap-northeast-1 (K3 and GLM 5.3 too), where every call —
+  # even a bogus-value request — answers 429 "Too many tokens, please wait".
+  # That is the quota, not the model. So the probes ran in ap-southeast-1
+  # (everything above), eu-north-1 (plain, image, no-arg tool) and eu-west-3
+  # (plain), where the quota is 1M; the four presets above weren't verified
+  # for K3.1 from this account, though list-inference-profiles shows
+  # global.moonshotai.kimi-k3.1 ACTIVE in all of them (us.* in us-east-1 / 2
+  # and us-west-2). A 429 on every call = request a quota increase.
+  - id: global.moonshotai.kimi-k3.1
+    displayName: Kimi K3.1
+    maxOutputTokens: 131072
+    contextWindow: 272000
+    maxContextWindow: 1048576
+    compressAt: 0.9
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: xhigh
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+      verbosity:
+        default: medium
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
   # Kimi K3 (Bedrock launch 2026-09-18, INFERENCE_PROFILE only — no in-Region
   # id). 1M context per the Bedrock model card; 131072 max output carried over
   # from the native kimi.yaml (Moonshot's official default cap). Live-probed