@turmind/halo 0.1.9

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

@@ -0,0 +1,61 @@
+id: mimo-token-plan-china
+displayName: Mimo Token Plan (China)
+displayName_zh: Mimo Token Plan（中国）
+description: Mimo Token Plan — Xiaomi's Anthropic-compatible gateway for MiMo models in China. Supports 1M context window.
+description_zh: Mimo Token Plan — 小米中国区 Anthropic 兼容网关，支持 MiMo 模型，1M 上下文窗口。
+defaultEndpoint: https://token-plan-cn.xiaomimimo.com/anthropic
+defaultModelId: mimo-v2.5
+endpointPresets:
+  - https://token-plan-cn.xiaomimimo.com/anthropic
+
+secrets:
+  - key: api_key
+    description: Mimo Token Plan API Key. Format starts with tp-.
+    description_zh: Mimo Token Plan API Key，tp- 开头
+    default: <<MIMO_API_KEY>>
+    secret: true
+
+models:
+  - id: mimo-v2.5-pro
+    displayName: MiMo v2.5 Pro
+    maxOutputTokens: 131072
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: high
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }
+
+  - id: mimo-v2.5
+    displayName: MiMo v2.5
+    maxOutputTokens: 131072
+    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
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }

package/templates/models/minimax.yaml

@@ -0,0 +1,80 @@
+id: minimax
+displayName: MiniMax
+displayName_zh: MiniMax
+description: MiniMax M2.7 family via the Anthropic-compatible Messages API
+description_zh: MiniMax M2.7 系列，通过 Anthropic 兼容的 Messages API 调用
+defaultEndpoint: https://api.minimaxi.com/anthropic
+defaultModelId: MiniMax-M2.7
+
+secrets:
+  - key: api_key
+    description: MiniMax API Key. Get from https://platform.minimaxi.com/user-center/basic-information/interface-key
+    description_zh: MiniMax API Key。从 https://platform.minimaxi.com/user-center/basic-information/interface-key 获取
+    default: <<MINIMAX_API_KEY>>
+    secret: true
+
+# Capability notes for MiniMax M2.7 family — verified against
+# platform.minimaxi.com/docs/api-reference/text-anthropic-api on 2026-05-26
+# plus live probes against MiniMax-M2.7:
+#   - image / video / audio: documented as not supported
+#   - thinking: forced on — `thinking:{type:'disabled'}` and omitting the
+#     field both still produce thinking blocks; only budget_tokens influences
+#     output. So we don't expose an off toggle.
+#   - cache_control: both models support caching. Doc says TTL is 5min,
+#     refreshing automatically on each hit. cache_control:{type:'ephemeral'}
+#     is the only shape — no longer-TTL knob exposed by MiniMax.
+#       - MiniMax-M2.7-highspeed: turn 1 reports cache_creation_input_tokens=N,
+#         turn 2 reports cache_read_input_tokens=N. Standard Anthropic shape.
+#       - MiniMax-M2.7: turn 1 reports cache_creation=0 (write happens
+#         implicitly without surfacing the count); turn 2 still reports
+#         cache_read_input_tokens=N. Read badge will appear after the first
+#         repeat, but the "write" badge never shows on M2.7 — quirk of
+#         MiniMax's reporting, not a Halo bug. Verified 2026-05-26.
+#     Cache is per-model (no cross-model sharing observed).
+#   - effort presets: pure Halo-side abstraction. `low` (1024 budget) was
+#     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:
+  - id: MiniMax-M2.7
+    displayName: MiniMax M2.7
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        # MiniMax thinking is always-on — there's no off switch. We still
+        # surface a Thinking control because budget_tokens shapes output
+        # length / latency. `effortPresets` are pure Halo-side labels;
+        # minimax-agent.ts maps them via effortToBudget() before sending.
+        mode: manual
+        defaultEnabled: true
+        default: medium
+        defaultBudgetTokens: 8192
+        effortPresets:
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  - id: MiniMax-M2.7-highspeed
+    displayName: MiniMax M2.7 Highspeed
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: manual
+        defaultEnabled: true
+        default: medium
+        defaultBudgetTokens: 8192
+        effortPresets:
+          - { value: medium, label: Medium }
+          - { value: high, label: High }

package/templates/models/openai.yaml

@@ -0,0 +1,51 @@
+id: openai
+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.
+description_zh: 通用 OpenAI 兼容协议。可用于原生 OpenAI 或任何第三方 OpenAI 兼容网关。能力字段开放给用户自选——由用户对端点能力自行负责。
+defaultEndpoint: https://api.openai.com/v1
+defaultModelId: gpt-4o
+endpointPresets:
+  - https://api.openai.com/v1
+
+secrets:
+  - key: api_key
+    description: API key for the chosen OpenAI-compatible endpoint. Format depends on provider (sk-... for OpenAI, ark-... for Volcengine, etc.)
+    description_zh: 所选 OpenAI 兼容端点的 API Key。前缀格式因厂商而异（OpenAI 是 sk-、火山方舟是 ark- 等）
+    default: <<OPENAI_API_KEY>>
+    secret: true
+
+# This provider is a "user-fills-everything" placeholder. The single
+# placeholder model below exists so the registry has something to render
+# in the form's model dropdown — replace its id/endpoint with whatever
+# you actually want to call. capabilities here are deliberately
+# permissive (image/video/audio all true, full thinking + caching
+# presets) so the form doesn't lock anything; if your real endpoint
+# doesn't support, e.g., audio, just don't send audio.
+#
+# Custom model ids (anything not listed under `models:`) are treated as
+# fully-permissive in the admin form — the user can flip image/video/
+# audio toggles, pick caching TTL, and set thinking effort or budget.
+models:
+  - id: gpt-4o
+    displayName: GPT-4o (placeholder)
+    maxOutputTokens: 16384
+    capabilities:
+      image: true
+      video: true
+      audio: true
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        mode: manual
+        defaultEnabled: false
+        default: medium
+        defaultBudgetTokens: 8192
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: max, label: Max }

package/templates/models/qwen.yaml

@@ -0,0 +1,78 @@
+id: qwen
+displayName: Qwen (Aliyun)
+displayName_zh: 通义千问（阿里云）
+description: Qwen3 family on Aliyun Bailian via the Anthropic-compatible Messages API
+description_zh: 阿里云百炼通义千问，通过 Anthropic 兼容的 Messages API 调用
+defaultEndpoint: https://dashscope.aliyuncs.com/apps/anthropic
+defaultModelId: qwen3.7-max
+endpointPresets:
+  - https://dashscope.aliyuncs.com/apps/anthropic         # Beijing
+  - https://dashscope-intl.aliyuncs.com/apps/anthropic    # Singapore
+  - https://dashscope-us.aliyuncs.com/apps/anthropic      # Virginia
+
+secrets:
+  - key: api_key
+    description: Aliyun Bailian (DashScope) API Key. Get from https://bailian.console.aliyun.com
+    description_zh: 阿里云百炼 API Key。从 https://bailian.console.aliyun.com 获取
+    default: <<DASHSCOPE_API_KEY>>
+    secret: true
+
+# Capability notes for Qwen — verified end-to-end on 2026-05-26 against
+# help.aliyun.com/zh/model-studio/anthropic-api-messages plus live probes:
+#   - cache_control: ephemeral (5min default per doc) — verified by sending
+#     same long system twice; turn1 reports cache_creation_input_tokens=N,
+#     turn2 reports cache_read_input_tokens=N. Standard Anthropic shape.
+#     Doc only mentions 5m; no 1h variant — qwen-agent.ts deliberately
+#     omits the `ttl` attribute.
+#   - thinking: honored both ways. type=enabled emits a thinking block,
+#     type=disabled really suppresses it (unlike MiniMax which forces it on).
+#   - image: probed live — qwen3.6-plus correctly identified colors (works
+#     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().
+#   - 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.
+models:
+  - id: qwen3.7-max
+    displayName: Qwen3.7 Max
+    maxOutputTokens: 16384
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        # Anthropic legacy/manual shape: thinking:{type:'enabled', budget_tokens:N}.
+        # qwen-agent.ts maps effort labels via effortToBudget().
+        mode: manual
+        defaultEnabled: true
+        default: medium
+        defaultBudgetTokens: 8192
+        effortPresets:
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  - id: qwen3.6-plus
+    displayName: Qwen3.6 Plus
+    maxOutputTokens: 16384
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: manual
+        defaultEnabled: true
+        default: medium
+        defaultBudgetTokens: 8192
+        effortPresets:
+          - { value: medium, label: Medium }
+          - { value: high, label: High }

package/templates/optional-skills/tavily-web-search/SKILL.md

@@ -0,0 +1,91 @@
+---
+name: Tavily Web Search
+description: Real-time web search using Tavily Search API. Use this skill when the user asks about current events, news, latest updates, prices, or any real-time information that requires up-to-date web results.
+---
+
+# Tavily Web Search
+
+Perform real-time web search via [Tavily Search API](https://docs.tavily.com/documentation/api-reference/endpoint/search). Returns structured results with content snippets and source URLs.
+
+## How to Use
+
+Call `shell_exec` with the command below. Only change the query string in `-d`; keep `{{params.api_key}}` exactly as-is (the system substitutes it at runtime).
+
+```bash
+curl -s -X POST https://api.tavily.com/search \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer {{params.api_key}}" \
+  -d '{"query":"your natural language query here","max_results":5,"include_answer":true}' \
+  | python3 -c "
+import sys,json
+r=json.load(sys.stdin)
+if r.get('answer'): print('Answer:',r['answer'],'\n')
+for x in r.get('results',[]):
+    print(f\"- [{x['title']}]({x['url']})\")
+    print(f\"  {x['content'][:300]}\n\")
+"
+```
+
+For news queries, add `"topic":"news"` to `-d`. For deeper search, use `"search_depth":"advanced"` (costs 2 credits instead of 1).
+
+## Query Formulation
+
+Use **natural language questions**.
+
+### Language choice — important
+
+Tavily's index is dominated by English-language sources for international topics. Pick the query language by topic, not by the user's input language:
+
+| Topic | Use English query | Use Chinese query |
+|---|---|---|
+| International sports / celebrities (e.g. Lionel Messi, Taylor Swift) | ✅ | ❌ Chinese hits niche aggregators only |
+| International tech / products / pricing (e.g. iPhone, RTX 5080) | ✅ | ❌ |
+| Global news / politics / finance | ✅ (or both) | Often weak |
+| Domestic Chinese topics (A 股, 国内政策, 中文媒体) | ❌ | ✅ |
+| Chinese tech ecosystem (e.g. 字节跳动, 小米) | Either | ✅ |
+
+If the user asks in Chinese about an international topic, **translate the entity to its English form first**: 「梅西最新消息」→ `"Lionel Messi latest news"`. Don't pass `"梅西最新消息"` directly to the API for non-Chinese topics — you'll mostly miss ESPN / Bleacher Report / Reuters etc.
+
+### Time-sensitive queries: prefer `time_range`, not year keywords
+
+Don't bake the year into the query string. Two reasons:
+1. You may be wrong about the current year. (The skill itself doesn't see today's date.)
+2. The index is fresher than your knowledge cutoff — pinning a year filters out the most relevant recent results.
+
+**Use `time_range` instead**: add `"time_range":"day" | "week" | "month" | "year"` to `-d`. For breaking news, `"week"` is usually right; for a monthly review, `"month"`.
+
+If the user explicitly anchors to a past year ("how was Apple's 2023?"), then a year keyword is correct.
+
+### Templates
+
+| User Intent | Query template |
+|---|---|
+| Latest news on a person/product | `"<English entity name> latest news"` + `time_range: "week"` |
+| Current product/pricing/specs | `"Tell me about <X>, including specs, price, and latest updates."` |
+| Domestic Chinese news | `"<中文实体> 最新动态"` + `time_range: "week"` |
+| General reference question | `"Search for information about X."` |
+
+**Examples**:
+- ✅ `"Lionel Messi latest news"` + `time_range: "week"`
+- ✅ `"梅西最新动态"` (only if specifically targeting Chinese-language sources)
+- ❌ `"梅西最新新闻 2025"` — wrong language for international topic *and* hard-coded year
+- ❌ `"Messi news"` — too terse, no recency hint
+
+## Optional Parameters
+
+Add these fields to the JSON body as needed:
+
+| Parameter | Default | Description |
+|---|---|---|
+| `max_results` | 5 | Number of results (1–20) |
+| `include_answer` | false | AI-generated summary answer (`true`, `"basic"`, or `"advanced"`) |
+| `search_depth` | `"basic"` | `"basic"` / `"advanced"` (2 credits, higher relevance) |
+| `topic` | `"general"` | `"news"` for time-sensitive queries |
+| `time_range` | null | `"day"`, `"week"`, `"month"`, `"year"` |
+| `include_raw_content` | false | Full page content (`true`, `"markdown"`, `"text"`) |
+| `include_domains` | [] | Restrict to specific domains |
+| `exclude_domains` | [] | Exclude specific domains |
+
+## Requirements
+
+- API key configured at `tavily-web-search.params.api_key` in `~/.halo/secrets/settings.yaml`, or set env var `TAVILY_API_KEY`. Set it from the Settings page → Skills → Tavily Web Search.

package/templates/optional-skills/tavily-web-search/config.yaml

@@ -0,0 +1,10 @@
+displayName_zh: Tavily 联网搜索
+description: Real-time web search via Tavily Search API.
+description_zh: 通过 Tavily 搜索 API 进行实时联网搜索。
+
+params:
+  - key: api_key
+    description: Tavily Search API Key. Get one at https://app.tavily.com.
+    description_zh: Tavily 搜索 API Key。从 https://app.tavily.com 获取。
+    default: <<TAVILY_API_KEY>>
+    secret: true

package/templates/prompts/all/TOOL_GUIDELINES.md

@@ -0,0 +1,35 @@
+## Tool Usage Guidelines
+
+### Tool Choice
+
+Default priority:
+- **Search content** → `grep` over shelling out (`grep` / `rg` / `find` in `shell_exec`)
+- **Find files by name/pattern** → `glob` over `file_list` walking
+- **Edit a file** → `file_edit` (targeted) over `file_write` (only for new files or full rewrites)
+- **Read a large file** → `grep` first to locate the section, then `file_read` with `offset`/`limit`
+
+Worked examples (query → tool):
+- "Rename function `foo` everywhere" → `grep` for callers across the repo, then `file_edit` each hit.
+- "Where does config X get loaded?" → `grep` the symbol, not a directory walk.
+- "List all yaml under `.halo/agents`" → `glob({ pattern: "**/*.yaml", path: ".halo/agents" })`.
+- "Find TODOs in the API layer" → `grep({ pattern: "TODO", path: "src/api", include: "*.ts" })`.
+
+### glob / grep — narrow the scope first
+Both walk every file under their `path`. Pick the narrowest `path` that contains what you want, and use `include` (grep) or a specific pattern (glob) to filter further. Unsure where a file is? `file_list` a likely parent first to figure out the right subdirectory, then `glob` inside it.
+
+### Output
+- Don't dump whole files — grep or summarize
+- For >100 lines, focus on errors, stack traces, status codes
+
+### When a Tool Fails
+- Read the error, don't blind-retry
+- After 2 same-approach failures, change approach
+- Common causes: wrong path (use `glob` to verify), stale content (re-read), `file_edit` mismatch (re-read first to copy `old_string` exactly)
+
+### file_edit staleness — re-read between edits to the same file
+After any successful `file_edit`, your earlier `file_read` of that file is **stale**. Don't reach for `old_string` from memory or from an old read; the bytes have shifted. Re-read the relevant range, then build the next `old_string` from that fresh read. The same applies after any external change to a file (another agent, a build step, a shell command).
+
+### Execution
+- If asked to do something, do it — don't just describe
+- Text-only reply when tools fit the task = incomplete
+- Report concrete results (files changed, commands run, output seen), not "done"

package/templates/prompts/all/TOOL_SHELL.md

@@ -0,0 +1,4 @@
+### Shell
+- `shell_exec` runs via the system shell (bash/sh) and has unrestricted access — install missing tools yourself (`apt`, `pip`/`pip3`, `npm`, `brew`, etc.). Missing tools / libraries are install problems, not refusal triggers — provision your own runtime (ffmpeg / pandoc / imagemagick for media, the right packages for ML/data work).
+- Chain with `&&`; check exit status before continuing
+- Unix conventions apply: `~/` for home, `/tmp` for scratch, forward-slash paths, `python3` for Python

package/templates/prompts/all/TOOL_SHELL.windows.md

@@ -0,0 +1,7 @@
+### Shell (Windows)
+- Running on **Windows** — `shell_exec` runs commands through `cmd.exe`, not bash. Missing tools / libraries are install problems, not refusal triggers — provision your own runtime.
+- Use Windows commands, not unix ones: `dir` (not `ls`), `type` (not `cat`), `findstr` (not `grep`), `del` (not `rm`), `copy`/`xcopy` (not `cp`), `move` (not `mv`), `mkdir` works but `mkdir -p` does not.
+- Chain commands with `&&`; check `%ERRORLEVEL%` before continuing. (Prefer the `grep`/`glob`/`file_*` tools over shelling out — they're cross-platform.)
+- **Paths**: home is `%USERPROFILE%` (not `~/`); scratch dir is `%TEMP%` (not `/tmp`). Backslash `\` is the separator, though forward slashes usually work inside quoted strings — prefer forward slashes for cross-platform tools.
+- **Python**: the interpreter is usually `python` (sometimes the `py` launcher), not `python3`. When a skill's docs show `python3 …`, run `python …` instead.
+- **Skill scripts**: some built-in skills ship `.py` helpers (e.g. cron, workspace, acp) and assume unix. If `python` isn't found, install Python first; the workspace skill's `stage.py` additionally needs `pip install pyyaml`. Skip any `chmod +x` step — it's a no-op on Windows.

package/templates/prompts/bootstrap/BOOTSTRAP.md

@@ -0,0 +1,26 @@
+## First Conversation — Getting to Know You
+
+No user profile exists yet. Before doing anything else, have a brief, warm getting-to-know-you exchange:
+
+1. Greet the user naturally
+2. Ask what they'd like to call you (your name as the AI)
+3. Ask how you should address them (their name or nickname)
+4. Ask about their preferred communication style — casual/formal, humor level, language preference, or anything else that matters to them
+
+Keep it natural and conversational — 2-3 exchanges, not a form. The 4 items above are topics, not a questionnaire; weave them in as the conversation flows, don't ask them one by one. Adapt to whatever language the user responds in.
+
+After collecting enough info, use `file_write` to create `~/.halo/global/USER.md` (global user profile, shared across all projects):
+
+```markdown
+---
+user_name: [their name]
+ai_name: [the name they chose for you]
+lang: [preferred language, e.g. zh-CN, en]
+---
+
+## Communication Style
+
+[Their preferences — humor, formality, etc.]
+```
+
+Once USER.md is saved, naturally transition to helping with whatever the user needs. Do not mention this bootstrap process explicitly — just be friendly and curious.

package/templates/prompts/root/WORKSPACE_MEMORY.md

@@ -0,0 +1,11 @@
+## Workspace Long-Term Memory
+
+Each workspace has a `.halo/` directory whose persistent files outlive any single session — this is how the workspace remembers across conversations:
+
+| File | Purpose | When to write |
+|------|---------|---------------|
+| `.halo/INSTRUCTIONS.md` | Project conventions | When writing normative rules |
+| `.halo/INDEX.md` | Project documentation index | After project changes — remind the user to sync |
+| `.halo/memory/YYYY-MM-DD.md` | Past work worth keeping | When something should outlive this session |
+
+**New workspace has no INDEX.md**: If the user starts discussing this project's goals/structure and it looks like real work, proactively offer to draft one. Don't ask for casual browsing.

package/templates/schema.sql

@@ -0,0 +1,30 @@
+CREATE TABLE IF NOT EXISTS sessions (
+  id TEXT PRIMARY KEY,
+  title TEXT NOT NULL DEFAULT '',
+  messages TEXT NOT NULL,
+  message_count INTEGER NOT NULL DEFAULT 0,
+  created_at INTEGER NOT NULL,
+  updated_at INTEGER NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS agent_sessions (
+  id TEXT PRIMARY KEY,
+  parent_id TEXT,
+  agent_id TEXT NOT NULL,
+  agent_name TEXT NOT NULL DEFAULT '',
+  description TEXT NOT NULL DEFAULT '',
+  working_dir TEXT,
+  access_level TEXT,
+  created_at INTEGER NOT NULL,
+  updated_at INTEGER NOT NULL,
+  stopped_at INTEGER,
+  archived_at INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS disabled_items (
+  item_type TEXT NOT NULL CHECK(item_type IN ('agent', 'skill')),
+  item_id TEXT NOT NULL,
+  scope TEXT NOT NULL CHECK(scope IN ('global', 'workspace')),
+  disabled_at INTEGER NOT NULL,
+  PRIMARY KEY (item_type, item_id, scope)
+);