@turmind/halo 2.0.7 → 2.1.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "2.0.7",
+  "version": "2.1.0",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",

package/templates/agents/default/agent.yaml

@@ -28,6 +28,8 @@ tools:
   - grep
   - glob
   - web_fetch
+  # 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
 # roster — they're no longer hand-listed under `tools:`. These ids also scope

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

@@ -1,10 +1,10 @@
 id: aws-bedrock-mantle
 runtime: bedrock-mantle
-revision: 2026101005
+revision: 2026102701
 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.
-description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6.1 Sol、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。按区域分布：GPT-6.1 Sol 与 GPT-6 Sol / Luna 仅 us-east-1，Astra 仅 us-west-2。
+description: OpenAI GPT-6.1 (Sol / Terra), 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.
+description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6.1（Sol / Terra）、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。按区域分布：GPT-6.1 Sol 与 GPT-6 Sol / Luna 仅 us-east-1，Astra 仅 us-west-2。
 # Mantle is a distinct endpoint from bedrock-runtime; the OpenAI surface lives
 # under /openai/v1. Region split (live-probed 2026-09-24, gpt-6.1-sol
 # 2026-09-30; 404 on the wrong endpoint):
@@ -14,6 +14,7 @@ description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6.1 Sol、GPT-6（
 #   gpt-6-sol        ✓           -           -
 #   gpt-6-luna       ✓           -           -
 #   gpt-6.1-sol      ✓           -           -
+#   gpt-6.1-terra    ✓           -           -   (probed 2026-10-27)
 #   gpt-5.6-sol      ✓           ✓           -
 #   gpt-5.6-terra    ✓           ✓           ✓
 #   gpt-5.6-luna     ✓           ✓           ✓
@@ -28,7 +29,7 @@ description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6.1 Sol、GPT-6（
 # us-east-2 configs (5.6-era) keep working; picking Astra moves the agent to
 # us-west-2. The endpoint is a per-agent pick.
 #
-# The same seven models are also in aws-bedrock-openai.yaml (bedrock-runtime
+# The same eight models are also in aws-bedrock-openai.yaml (bedrock-runtime
 # host, `global.openai.*` ids, no region split) — AWS's recommended endpoint
 # for new agents. Stay here only for Mantle-only features (server-side
 # web_search, background inference), which the web-search skill relies on.
@@ -193,6 +194,45 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
+  # GPT-6.1 Terra (Bedrock GA 2026-10-26) — us-east-1 endpoint ONLY (404 "model
+  # does not exist" on us-east-2 / us-west-2, live-probed 2026-10-27). On
+  # us-east-1: plain, reasoning.effort none/max, image ("Red"), tool call
+  # (function_call) and implicit cache (turn 2 read 6014 tokens) all 200. Specs
+  # follow the Bedrock model card (1,050,000 context / 128,000 max output);
+  # effort / verbosity presets as GPT-6.1 Sol (no "none" preset).
+  - id: openai.gpt-6.1-terra
+    displayName: GPT-6.1 Terra
+    endpoints:
+      - https://bedrock-mantle.us-east-1.api.aws/openai/v1
+    maxOutputTokens: 128000
+    contextWindow: 272000
+    maxContextWindow: 1050000
+    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 }
+
   - id: openai.gpt-6-luna
     displayName: GPT-6 Luna
     endpoints:

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

@@ -1,10 +1,10 @@
 id: aws-bedrock-openai
 runtime: bedrock-mantle
-revision: 2026101005
+revision: 2026102701
 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, 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、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）。
+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）。
 # 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
@@ -59,7 +59,7 @@ secrets:
 #   - Reasoning summaries are never surfaced (output[].reasoning.summary
 #     empty) — same as GPT on Mantle; reasoning_tokens > 0 proves it ran.
 models:
-  # OpenAI GPT-6.1 / GPT-6 / GPT-5.6 — the same seven models as
+  # OpenAI GPT-6.1 / GPT-6 / GPT-5.6 — the same eight models as
   # aws-bedrock-mantle.yaml (see that file for the per-model background:
   # 1,050,000 context for GPT-6 and 5.6 per the Bedrock model cards, 128,000 /
   # 64,000 max output, 2x input price past 272K — hence contextWindow 272000
@@ -136,6 +136,46 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
+  # GPT-6.1 Terra (Bedrock GA 2026-10-26; model card: 1,050,000 context /
+  # 128,000 max output, EOL no sooner than 2027-10-26). Live-probed 2026-10-27
+  # with SigV4 on global.openai.gpt-6.1-terra: 200 from us-east-1 / us-west-2 /
+  # eu-west-1 / ap-northeast-1; the bare openai.gpt-6.1-terra id 400s (needs the
+  # profile). reasoning.effort accepts none/low/medium/high/xhigh/max (`minimal`
+  # 400s; presets follow GPT-6 Sol, no "none"); text.verbosity low/medium/high;
+  # image → "Red"; tool call → function_call; implicit cache: turn 1 wrote 6014
+  # tokens, turn 2 read 6014. max_output_tokens 128001 is not rejected (proves
+  # nothing), so 128000 is the card's figure.
+  - id: global.openai.gpt-6.1-terra
+    displayName: GPT-6.1 Terra
+    maxOutputTokens: 128000
+    contextWindow: 272000
+    maxContextWindow: 1050000
+    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 }
+
   - id: global.openai.gpt-6-astra
     displayName: GPT-6 Astra
     maxOutputTokens: 128000

package/templates/models/doubao.yaml

@@ -1,12 +1,12 @@
 id: doubao
 runtime: doubao
-revision: 2026101005
+revision: 2026101901
 displayName: Doubao (Volcengine)
 displayName_zh: 豆包（火山方舟）
-description: ByteDance Doubao Seed-2.1 / 2.0 family on Volcengine Ark — pro / lite / mini / code, OpenAI-compatible
+description: ByteDance Doubao Seed-2.1 / 2.0 family on Volcengine Ark — pro / lite / turbo / mini / code, OpenAI-compatible
 description_zh: 字节豆包 Seed-2.1 / 2.0 系列（火山方舟），OpenAI 兼容协议
 defaultEndpoint: https://ark.cn-beijing.volces.com/api/v3
-defaultModelId: doubao-seed-2-1-pro-260628
+defaultModelId: doubao-seed-2-1-pro-260915
 
 secrets:
   - key: api_key
@@ -45,12 +45,74 @@ secrets:
 # 229376 max input, so contextWindow = 229376; every 2.x model takes up to
 # 131072 output (max_tokens 131073 → 400) and Ark trims the reply to the
 # remaining quota instead of erroring. 2.1-pro-260628 live-probed: thinking
-# enabled/disabled, tool calls, image. Not carried: 2.1-pro / lite-260915 and
-# doubao-seed-evolving (1M ctx) return ModelNotOpen on our account — add them
-# after a live probe on an account that has them activated; 2.1-turbo-260628
-# likewise. Ark's newer `reasoning_effort` knob (deep-thinking doc) is not
+# enabled/disabled, tool calls, image.
+# 2026-10-19 census: 2.1-pro / lite-260915 and 2.1-turbo-260628 are now
+# callable on our account (official list docs/82379/1330310: 260915 models →
+# 1024k ctx / 1024k input / 256k output; turbo → 256k / 256k / 256k) and were
+# live-probed: plain, thinking enabled/disabled (type=auto → 400), image
+# (red PNG → red), tool call, automatic cache hits, max_tokens 262145 → 400
+# on lite / turbo (pro-260915 never rejects, so its 262144 is the official
+# figure). contextWindow = min(272000, max input). doubao-seed-evolving still
+# returns ModelNotOpen — add after a probe on an account that has it.
+# Ark's newer `reasoning_effort` knob (deep-thinking doc) is not
 # wired — doubao-agent still sends only thinking:{type}.
 models:
+  - id: doubao-seed-2-1-pro-260915
+    displayName: Doubao Seed 2.1 Pro (260915)
+    maxOutputTokens: 262144
+    contextWindow: 272000
+    maxContextWindow: 1048576
+    compressAt: 0.9
+    capabilities:
+      # Live probe 2026-10-19: named the red 64×64 PNG correctly; thinking enabled/disabled, tool call OK; cache hit on the 6th identical turn.
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        # Doubao caches automatically (no cache_control parameter needed
+        # in the request). The badge is just a visibility hint — UI shows
+        # the read/write counts so users see when cache is helping.
+        # Doc says default TTL is the platform default; observed empirically
+        # to behave like ~5min ephemeral. Cache only kicks in for prompts
+        # above ~4096 tokens; first 1-3 turns of a fresh prompt commonly
+        # miss while the cache is warming up.
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: medium, label: On }
+
+  - id: doubao-seed-2-1-lite-260915
+    displayName: Doubao Seed 2.1 Lite
+    maxOutputTokens: 262144
+    contextWindow: 272000
+    maxContextWindow: 1048576
+    compressAt: 0.9
+    capabilities:
+      # Live probe 2026-10-19: named the red 64×64 PNG correctly; thinking enabled/disabled, tool call OK; cache hit; max_tokens 262145 → 400.
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        # Doubao caches automatically (no cache_control parameter needed
+        # in the request). The badge is just a visibility hint — UI shows
+        # the read/write counts so users see when cache is helping.
+        # Doc says default TTL is the platform default; observed empirically
+        # to behave like ~5min ephemeral. Cache only kicks in for prompts
+        # above ~4096 tokens; first 1-3 turns of a fresh prompt commonly
+        # miss while the cache is warming up.
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: medium, label: On }
+
   - id: doubao-seed-2-1-pro-260628
     displayName: Doubao Seed 2.1 Pro
     maxOutputTokens: 131072
@@ -82,6 +144,34 @@ models:
         effortPresets:
           - { value: medium, label: On }
 
+  - id: doubao-seed-2-1-turbo-260628
+    displayName: Doubao Seed 2.1 Turbo
+    maxOutputTokens: 262144
+    contextWindow: 262144
+    maxContextWindow: 262144
+    compressAt: 0.9
+    capabilities:
+      # Live probe 2026-10-19: named the red 64×64 PNG correctly; thinking enabled/disabled, tool call OK; cache hit from turn 2; max_tokens 262145 → 400.
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        # Doubao caches automatically (no cache_control parameter needed
+        # in the request). The badge is just a visibility hint — UI shows
+        # the read/write counts so users see when cache is helping.
+        # Doc says default TTL is the platform default; observed empirically
+        # to behave like ~5min ephemeral. Cache only kicks in for prompts
+        # above ~4096 tokens; first 1-3 turns of a fresh prompt commonly
+        # miss while the cache is warming up.
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: medium, label: On }
+
   - id: doubao-seed-2-0-lite-260428
     displayName: Doubao Seed 2.0 Lite
     maxOutputTokens: 131072

package/templates/models/kimi.yaml

@@ -1,10 +1,10 @@
 id: kimi
 runtime: kimi
-revision: 2026101004
+revision: 2026102701
 displayName: Kimi (Moonshot AI)
 displayName_zh: Kimi（月之暗面）
-description: Kimi K3 / K2.7 Code / K2.6 via OpenAI-compatible chat completions API
-description_zh: Kimi K3 / K2.7 Code / K2.6，通过 OpenAI 兼容的 chat completions API 调用
+description: Kimi K3.1 Pro / K3 / K2.7 Code / K2.6 via OpenAI-compatible chat completions API
+description_zh: Kimi K3.1 Pro / K3 / K2.7 Code / K2.6，通过 OpenAI 兼容的 chat completions API 调用
 defaultEndpoint: https://api.moonshot.cn/v1
 
 secrets:
@@ -25,6 +25,44 @@ secrets:
 #     k2.6). An earlier comment here claiming K2.6 accepts
 #     reasoning_effort { minimal, low, medium, high } was wrong.
 models:
+  # Kimi K3.1 Pro (platform.kimi.com/docs/models + docs/guide/kimi-k3-1-pro-quickstart:
+  # 1M context, top-level reasoning_effort low / high / max, default max,
+  # max_completion_tokens default 131072). /v1/models: context_length 1048576,
+  # image in. Live probe 2026-10-27: plain, effort low/high/max/medium/xhigh all
+  # 200 (reasoning returned; medium/xhigh are silently accepted, only the three
+  # documented values are offered); red 64×64 PNG → "Red"; tool call →
+  # tool_calls; automatic cache: turn 2 read 4864 of 5110 tokens. Docs say
+  # thinking can't be turned off, though thinking:disabled / effort none were
+  # accepted and returned no reasoning — not exposed. max_tokens 1048577 isn't
+  # rejected (proves nothing); output stays at the official 131072 default.
+  # prompt_cache_options.ttl '1h' is documented and accepted, but usage never
+  # reports a TTL bucket, so 5m stays the only preset. The sibling
+  # kimi-k3.1-pro-highspeed is on the docs page but 404s for this account —
+  # add after a probe by someone with access.
+  - id: kimi-k3.1-pro
+    displayName: Kimi K3.1 Pro
+    maxOutputTokens: 131072
+    contextWindow: 272000
+    maxContextWindow: 1048576
+    compressAt: 0.9
+    capabilities:
+      # Video input exists upstream; Halo has no video path (see K3).
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        # Always on server-side; same top-level reasoning_effort as K3.
+        defaultEnabled: true
+        default: high
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: high, label: High }
+          - { value: max, label: Max }
+
   - id: kimi-k3
     displayName: Kimi K3
     # Official default output cap; the model can go up to 1M but we don't max out.

package/templates/models/minimax.yaml

@@ -1,12 +1,12 @@
 id: minimax
 runtime: minimax
-revision: 2026101002
+revision: 2026101901
 displayName: MiniMax
 displayName_zh: MiniMax
-description: MiniMax M3 / M2.7 family via the Anthropic-compatible Messages API
-description_zh: MiniMax M3 / M2.7 系列，通过 Anthropic 兼容的 Messages API 调用
+description: MiniMax M3.1 / M3 / M2.7 family via the Anthropic-compatible Messages API
+description_zh: MiniMax M3.1 / M3 / M2.7 系列，通过 Anthropic 兼容的 Messages API 调用
 defaultEndpoint: https://api.minimaxi.com/anthropic
-defaultModelId: MiniMax-M3
+defaultModelId: MiniMax-M3.1-Flash
 
 secrets:
   - key: api_key
@@ -38,6 +38,37 @@ secrets:
 #     observed to truncate the run with stop=max_tokens before any thinking
 #     completes — leaving it out of the preset list. medium/high are safe.
 models:
+  # MiniMax-M3.1-Flash (listed 2026-09-16, platform.minimaxi.com/docs/guides/models-intro
+  # + /api-reference/text-anthropic-api): 1,000,000 context, native multimodal.
+  # Live probe 2026-10-19 on /anthropic: plain, effort low / medium / high /
+  # xhigh / max all 200 (bogus value 400s with "allowed: low, medium, high,
+  # xhigh, max"); thinking:{type:'adaptive'} 200; thinking:{type:'disabled'}
+  # 400 ("requires adaptive thinking") — thinking can't be switched off, and
+  # omitting it leaves it on, so Halo's "off" (field omitted) is safe. Red PNG
+  # → "Red"; tool_use OK; max_tokens 524288 OK, 524289 → 400, so the hard cap
+  # is 524288 — kept at 131072 like M3. Halo's minimax-agent only sends
+  # thinking:{type:'adaptive'} (no output_config.effort), so the API default
+  # (max) applies; the doc's effort levels aren't exposed until the runtime
+  # sends them. Caching is passive: cache_read_input_tokens rose on turn 2
+  # with or without cache_control, cache_creation stayed 0 and `ttl:'1h'`
+  # made no difference — no promptCaching key, as for M3.
+  - id: MiniMax-M3.1-Flash
+    displayName: MiniMax M3.1 Flash
+    maxOutputTokens: 131072
+    contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: adaptive
+        effortPresets:
+          - { value: adaptive, label: Adaptive }
+
   - id: MiniMax-M3
     displayName: MiniMax M3
     # Official recommended value; hard cap is 524288 but we don't max out.