@turmind/halo 0.1.9

package/templates/config.yaml

@@ -0,0 +1,69 @@
+# Halo System Configuration
+# Self-documenting: each leaf is { value, default, description, description_zh }.
+#
+# Env var overrides (highest priority):
+#   server.port               → HALO_PORT
+#   server.password           → HALO_PASSWORD
+#   server.cors_origins       → HALO_CORS_ORIGINS  (comma-separated)
+#   timeout.shell_exec        → HALO_SHELL_TIMEOUT
+#   timeout.web_fetch         → HALO_WEB_FETCH_TIMEOUT
+#   timeout.session_grace     → HALO_SESSION_GRACE
+#   timeout.terminal_grace    → HALO_TERMINAL_GRACE
+#   logging.max_file_size     → HALO_LOG_MAX_SIZE
+#   logging.max_files         → HALO_LOG_MAX_FILES
+
+server:
+  port:
+    value: 9527
+    default: 9527
+    description: "HTTP/WS port the server listens on."
+    description_zh: "服务监听的 HTTP/WS 端口。"
+  password:
+    value: ""
+    default: ""
+    description: "Admin password — stored as a scrypt hash, set via `halo setup`. To reset a forgotten password, blank this field and re-run `halo setup`."
+    description_zh: "管理员密码（scrypt hash），通过 `halo setup` 设置。忘记密码时清空此字段并重新运行 `halo setup`。"
+  jwt_secret:
+    value: ""
+    default: ""
+    description: "Signing key for session JWTs. Auto-generated by `halo setup`; rotated on password change so old cookies are invalidated."
+    description_zh: "Session JWT 签名密钥。由 `halo setup` 自动生成，密码更改时一同轮换以使旧 cookie 失效。"
+  cors_origins:
+    value: []
+    default: []
+    description: "CORS allowlist. Empty = reflect any origin (recommended for self-hosted single-user setups). Set a non-empty list of explicit origins to enforce strict CORS."
+    description_zh: "CORS 白名单。空 = 反射任意来源（自部署单用户场景推荐）。填入显式来源列表则启用严格 CORS。"
+
+timeout:
+  shell_exec:
+    value: 120000
+    default: 120000
+    description: "shell_exec tool timeout (ms)."
+    description_zh: "shell_exec 工具的超时（毫秒）。"
+  web_fetch:
+    value: 10000
+    default: 10000
+    description: "web_fetch tool timeout (ms)."
+    description_zh: "web_fetch 工具的超时（毫秒)。"
+  session_grace:
+    value: 300000
+    default: 300000
+    description: "Idle session grace period before unloading from memory (ms)."
+    description_zh: "空闲会话保留在内存中的时长（毫秒）。"
+  terminal_grace:
+    value: 300000
+    default: 300000
+    description: "Terminal grace period after disconnect before killing the pty (ms)."
+    description_zh: "终端断连后保留 pty 的时长（毫秒）。"
+
+logging:
+  max_file_size:
+    value: 10485760
+    default: 10485760
+    description: "Per-log-file size cap before rotation (bytes)."
+    description_zh: "单个日志文件的大小上限（字节）。"
+  max_files:
+    value: 5
+    default: 5
+    description: "Number of rotated log files to keep."
+    description_zh: "保留的日志文件数量。"

package/templates/models/anthropic.yaml

@@ -0,0 +1,47 @@
+id: anthropic
+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.
+description_zh: 通用 Anthropic Messages API。可用于 api.anthropic.com 原生端点或任何 Anthropic 兼容的第三方网关。
+defaultEndpoint: https://api.anthropic.com
+defaultModelId: claude-sonnet-4-6
+endpointPresets:
+  - https://api.anthropic.com
+
+secrets:
+  - key: api_key
+    description: Anthropic API Key. Format starts with sk-ant-.
+    description_zh: Anthropic API Key，sk-ant- 开头
+    default: <<ANTHROPIC_API_KEY>>
+    secret: true
+
+# Placeholder model — the user is expected to set the actual model id
+# (claude-sonnet-4-6, claude-opus-4-7, etc.) and the matching endpoint.
+# capabilities are deliberately permissive; flip what your real endpoint
+# doesn't support.
+#
+# Custom model ids (anything not listed under `models:`) are treated as
+# fully-permissive in the admin form — image/video/audio toggles,
+# caching TTL, thinking effort/budget all become editable.
+models:
+  - id: claude-sonnet-4-6
+    displayName: Claude Sonnet 4.6 (placeholder)
+    maxOutputTokens: 64000
+    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: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }

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

@@ -0,0 +1,156 @@
+id: aws-bedrock-claude-invoke
+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
+description_zh: 通过 Bedrock InvokeModel（非流式）调用，发送原生 Anthropic Messages API 格式
+defaultEndpoint: https://bedrock-runtime.us-east-1.amazonaws.com
+endpointPresets:
+  - https://bedrock-runtime.us-east-1.amazonaws.com
+  - https://bedrock-runtime.us-west-2.amazonaws.com
+  - https://bedrock-runtime.eu-west-1.amazonaws.com
+  - https://bedrock-runtime.ap-northeast-1.amazonaws.com
+
+# When the user switches to this provider, the form lands on this model id
+# (falls back to models[0].id if missing). Sonnet 4.6 is the daily-driver pick:
+# Opus 4.6/4.7 cost 5–10x more, Haiku 4.5 is best for narrow agents.
+defaultModelId: global.anthropic.claude-sonnet-4-6
+
+# Server-side credentials. Stored at <provider-id>.secrets.<key> in
+# settings.yaml; never injected into agent prompts or shell_exec.
+secrets:
+  - key: access_key_id
+    description: AWS Access Key ID. Leave empty to use the AWS credential chain (env / ~/.aws / EC2 role).
+    description_zh: AWS Access Key ID。留空则使用凭证链（环境变量 / ~/.aws/credentials / EC2 角色）
+    default: <<AWS_ACCESS_KEY_ID>>
+    secret: true
+  - key: secret_access_key
+    description: AWS Secret Access Key. Leave empty to use the credential chain.
+    description_zh: AWS Secret Access Key。留空则使用凭证链
+    default: <<AWS_SECRET_ACCESS_KEY>>
+    secret: true
+
+# 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:
+  - id: global.anthropic.claude-sonnet-4-6
+    displayName: Claude Sonnet 4.6
+    maxOutputTokens: 64000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # adaptive = `thinking: { type: 'adaptive' }` + `output_config.effort`
+        # manual   = `thinking: { type: 'enabled', budget_tokens: N }` (effort
+        #            translated to a budget number at runtime)
+        # See bedrock-agent.ts for the effort→budget table used in manual mode.
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }
+
+  - id: global.anthropic.claude-opus-4-6-v1
+    displayName: Claude Opus 4.6
+    maxOutputTokens: 128000
+    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: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }
+
+  - id: global.anthropic.claude-opus-4-7
+    displayName: Claude Opus 4.7
+    maxOutputTokens: 128000
+    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: high       # Opus 4.7 is the heavy-thinking model — high pays off
+        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-opus-4-8
+    displayName: Claude Opus 4.8
+    maxOutputTokens: 128000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # Bedrock-confirmed adaptive effort set for opus-4-8 (verified by an
+        # invalid-value probe): low / medium / high / xhigh / max.
+        mode: adaptive
+        defaultEnabled: true
+        default: high       # Opus 4.8 is the latest flagship — high by default; xhigh/max on demand
+        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-haiku-4-5-20251001-v1:0
+    displayName: Claude Haiku 4.5
+    maxOutputTokens: 64000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # Haiku 4.5 only supports legacy budget_tokens; UI shows a number input.
+        # bedrock-agent.ts translates effort labels via the effort→budget table.
+        mode: manual
+        defaultEnabled: false   # Haiku is the cheap-fast model — thinking off by default
+        defaultBudgetTokens: 4000
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }

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

@@ -0,0 +1,89 @@
+id: aws-bedrock-mantle
+displayName: AWS Bedrock Mantle (OpenAI GPT-5.x)
+displayName_zh: AWS Bedrock Mantle（OpenAI GPT-5.x）
+description: OpenAI GPT-5.5 / 5.4 on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). us-east-2 only.
+description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-5.5 / 5.4（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。仅 us-east-2。
+# Mantle is a distinct endpoint from bedrock-runtime; the OpenAI surface lives
+# under /openai/v1. GPT-5.5 is us-east-2 only (GPT-5.4 also allows us-west-2,
+# but we ship the us-east-2 endpoint that serves both).
+defaultEndpoint: https://bedrock-mantle.us-east-2.api.aws/openai/v1
+endpointPresets:
+  - https://bedrock-mantle.us-east-2.api.aws/openai/v1
+
+defaultModelId: openai.gpt-5.5
+
+# Auth: a Bedrock long-term API key passed as a Bearer token. Generate at
+# console → Bedrock → API keys → long-term. The default placeholder reads the
+# AWS_BEARER_TOKEN_BEDROCK env var; leave the env unset and fill the key in
+# Settings to store it in secrets/settings.yaml instead. resolveApiKey() reads
+# `api_key`, so the bearer token flows straight to the agent with no extra wiring.
+secrets:
+  - key: api_key
+    description: Bedrock long-term API key (sent as a Bearer token). Defaults to the AWS_BEARER_TOKEN_BEDROCK env var.
+    description_zh: Bedrock 长期 API Key（作为 Bearer 令牌发送）。默认读取 AWS_BEARER_TOKEN_BEDROCK 环境变量。
+    default: <<AWS_BEARER_TOKEN_BEDROCK>>
+    secret: true
+
+# capability notes for Mantle/OpenAI:
+#   - image: input only (text/image in, text-only out — no audio/video/image gen)
+#   - promptCaching: AUTOMATIC (prefix ≥1024 tok, ~5min idle TTL). No cache_control
+#     is sent and there's no 1h option — single 5m preset reflects reality, the
+#     runtime never writes cache control, it only reads cached_tokens back.
+#   - thinking: adaptive via reasoning.effort = low|medium|high (no budget_tokens
+#     mode; xhigh/max would be mapped down to high by the runtime).
+models:
+  - id: openai.gpt-5.5
+    displayName: GPT-5.5
+    maxOutputTokens: 64000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+      # Output length of the final answer (Responses API `text.verbosity`),
+      # separate from reasoning effort. Default low = terse + cheaper, best for
+      # tool-driven agent turns. Override per-agent via agent.yaml model.verbosity.
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  - id: openai.gpt-5.4
+    displayName: GPT-5.4
+    maxOutputTokens: 64000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+      # See GPT-5.5 above — final-answer length, default low, per-agent overridable.
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }

package/templates/models/deepseek.yaml

@@ -0,0 +1,62 @@
+id: deepseek
+displayName: DeepSeek
+displayName_zh: DeepSeek
+description: DeepSeek V4 via OpenAI-compatible chat completions API
+description_zh: DeepSeek V4，通过 OpenAI 兼容的 chat completions API 调用
+defaultEndpoint: https://api.deepseek.com
+
+secrets:
+  - key: api_key
+    description: DeepSeek API Key. Get from https://platform.deepseek.com/api_keys
+    description_zh: DeepSeek API Key。从 https://platform.deepseek.com/api_keys 获取
+    default: <<DEEPSEEK_API_KEY>>
+    secret: true
+
+# Capabilities verified 2026-05-27 via live probes. DeepSeek's image_url
+# blocks are rejected at the JSON-deserialize layer ("unknown variant
+# image_url"), so image:false is correct. 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.
+models:
+  - id: deepseek-v4-pro
+    displayName: DeepSeek V4 Pro
+    maxOutputTokens: 65536
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }
+          - { value: xhigh, label: XHigh }
+
+  - id: deepseek-v4-flash
+    displayName: DeepSeek V4 Flash
+    maxOutputTokens: 65536
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }
+          - { value: xhigh, label: XHigh }

package/templates/models/doubao.yaml

@@ -0,0 +1,137 @@
+id: doubao
+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
+description_zh: 字节豆包 Seed-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-2-0-pro-260215
+
+secrets:
+  - key: api_key
+    description: Volcengine Ark API Key. Format starts with ark-. Get from https://console.volcengine.com/ark
+    description_zh: 火山方舟 API Key，ark- 开头。从 https://console.volcengine.com/ark 获取
+    default: <<ARK_API_KEY>>
+    secret: true
+
+# Capability notes for Doubao Seed-2.0 — verified end-to-end on 2026-05-26:
+#   - All four models share the same wire shape (OpenAI chat completions
+#     + Volcengine `thinking:{type:'enabled'/'disabled'}`).
+#   - thinking type=auto is rejected as "Unsupported" — only enabled/disabled.
+#   - reasoning_content is returned in message.reasoning_content when
+#     thinking is enabled; suppressed when disabled.
+#   - Caching is automatic — no cache_control parameter required. Verified
+#     2026-05-26 with a >4096-token system prompt and 5 same-prompt turns:
+#     all four models (pro/lite/mini/code) hit cache at least once.
+#     Caveats observed empirically:
+#       - prompts shorter than ~4096 tokens generally don't cache at all;
+#       - first 1-3 turns of a fresh prompt commonly miss while the cache
+#         warms up;
+#       - cache can drop unexpectedly mid-conversation (turn N hits, turn
+#         N+1 misses, turn N+2 hits again) — not stable steady-state.
+#     Treat the cache as opportunistic; UI badges are accurate when shown.
+#   - Tool calls return standard OpenAI tool_calls shape.
+#   - Image: pro/lite/mini all accepted image_url blocks; in our smoke
+#     test all four answered "Black" for what was actually a red 64×64 PNG,
+#     so reliability is uncertain. We default image:false; user can flip
+#     after their own verification.
+#   - Doubao 2.0 docs aren't fully published; model id strings sourced
+#     from the Volcengine Ark console (the activated-model list).
+models:
+  - id: doubao-seed-2-0-pro-260215
+    displayName: Doubao Seed 2.0 Pro
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      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
+        # effort label is Halo-side only — Doubao only sees enabled/disabled.
+        # We still expose Off/On as toggles by mapping both options to the
+        # same flag (keeps the form consistent across providers).
+        effortPresets:
+          - { value: medium, label: On }
+
+  - id: doubao-seed-2-0-lite-260428
+    displayName: Doubao Seed 2.0 Lite
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      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-mini-260428
+    displayName: Doubao Seed 2.0 Mini
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      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-code-preview-260215
+    displayName: Doubao Seed 2.0 Code Preview
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      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 }

package/templates/models/hunyuan.yaml

@@ -0,0 +1,56 @@
+id: hunyuan
+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
+description_zh: 腾讯混元 Hy3 preview，295B/21B 激活 MoE，面向 Agent 工作负载，支持 no_think / think_low / think_high 三档思考
+defaultEndpoint: https://tokenhub.tencentmaas.com/v1
+defaultModelId: hy3-preview
+
+secrets:
+  - key: api_key
+    description: Tencent Hunyuan Hy3 API Key. Format starts with sk-.
+    description_zh: 腾讯混元 Hy3 API Key，sk- 开头
+    default: <<HUNYUAN_API_KEY>>
+    secret: true
+
+# Capability notes for Hunyuan Hy3 — verified end-to-end on 2026-05-26:
+#   - Endpoint hosted on tokenhub.tencentmaas.com (NOT api.hunyuan.cloud.tencent.com,
+#     which is the older OpenAI-compat surface that doesn't have Hy3).
+#   - Caching is automatic (no cache_control needed). Verified: same long
+#     system prompt across turns shows usage.prompt_tokens_details
+#     .cached_tokens > 0 from turn 2 onward.
+#   - Thinking modes per Hy3 product naming → `reasoning_effort`:
+#       no_think    → omit reasoning_effort (zero reasoning_tokens)
+#       think_low   → reasoning_effort: 'low'
+#       think_high  → reasoning_effort: 'high'
+#     hunyuan-agent.ts maps thinking.enabled=false → no_think (no field),
+#     thinking.enabled=true + effort=low|medium|high → reasoning_effort.
+#   - Image input: gateway accepts image_url blocks but model says "I don't
+#     see an image" — image:false.
+#   - 256K context per product description (not enforced here; model just
+#     accepts long inputs).
+models:
+  - id: hy3-preview
+    displayName: Hunyuan Hy3 Preview
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        # Automatic caching, no cache_control parameter. cached_tokens reliably
+        # appears from turn 2 once prompt is long enough. The TTL knob isn't
+        # exposed by Hy3 — value is hint-only for the form.
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        # Maps Halo effort labels to Hy3's reasoning_effort. The Hy3
+        # product surface markets these as no_think / think_low / think_high;
+        # we expose Off / Low / High since "Off" is the natural opposite of
+        # an enable toggle.
+        defaultEnabled: true
+        default: low
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: high, label: High }

package/templates/models/kimi.yaml

@@ -0,0 +1,41 @@
+id: kimi
+displayName: Kimi (Moonshot AI)
+displayName_zh: Kimi（月之暗面）
+description: Kimi K2.6 via OpenAI-compatible chat completions API
+description_zh: Kimi K2.6，通过 OpenAI 兼容的 chat completions API 调用
+defaultEndpoint: https://api.moonshot.cn/v1
+
+secrets:
+  - key: api_key
+    description: Kimi (Moonshot AI) API Key. Get from https://platform.kimi.com/console/api-keys
+    description_zh: Kimi (月之暗面) API Key。从 https://platform.kimi.com/console/api-keys 获取
+    default: <<KIMI_API_KEY>>
+    secret: true
+
+# Capabilities verified 2026-05-27 via live probes. Kimi-k2.6 returns
+# "prepare image failed" on image_url blocks despite past docs claiming
+# vision — set image:false until/unless a vision-specific model id is
+# tested here. reasoning_effort accepts { minimal, low, medium, high };
+# `max` errors with explicit message listing the four valid values.
+# Caching is automatic; cached_tokens shows up reliably from turn 2 once
+# the prompt is long enough (~2K+).
+models:
+  - id: kimi-k2.6
+    displayName: Kimi K2.6
+    maxOutputTokens: 32768
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: minimal, label: Minimal }
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }