@turmind/halo 1.1.3 → 1.1.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "1.1.3",
+  "version": "1.1.5",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",

package/templates/agents/default/agent.yaml

@@ -45,7 +45,7 @@ skills:
   - send-file
   - self
   - aws-knowledge
-  - nova-web-search
+  - web-search
   - cron
   - acp
   - halo

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

@@ -35,6 +35,34 @@ secrets:
 #   thinking.default           → effort selected when Thinking is on (adaptive)
 #   thinking.defaultBudgetTokens → budget selected when Thinking is on (manual)
 models:
+  - id: global.anthropic.claude-fable-5-1
+    displayName: Claude Fable 5.1
+    maxOutputTokens: 128000
+    contextWindow: 1000000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # Fable 5.1 — Mythos-class flagship refresh. Adaptive effort set
+        # confirmed by an invalid-value probe against Bedrock (2026-09-02):
+        # low / medium / high / xhigh / max, same as Fable 5. 1M context,
+        # 128K output. Smoke-tested on invoke (xhigh).
+        mode: adaptive
+        defaultEnabled: true
+        default: high       # flagship-of-flagships — 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-fable-5
     displayName: Claude Fable 5
     maxOutputTokens: 128000

package/templates/skills/skill/SKILL.md

@@ -46,7 +46,7 @@ simply absent. Consequences you must apply:
   global skill for one workspace, copy the WHOLE folder to workspace scope
   (including resource files — remember: no per-file fallback), then edit.
 - **Built-in skills** (`agent`, `skill`, `workspace`, `cron`, `acp`, `send-file`,
-  `self`, `aws-knowledge`, `nova-web-search`): their global copies are force re-seeded on every server
+  `self`, `aws-knowledge`, `web-search`): their global copies are force re-seeded on every server
   start — direct edits to those global files are lost. Customize at workspace
   scope instead.
 - **Disabled state** lives per-workspace in the `disabled_items` DB table, not

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

@@ -0,0 +1,44 @@
+---
+name: web-search
+description: Real-time web search with two gears. Fast (default, Amazon Nova grounding, ~3s, cheap) — current events, news, prices, any routine lookup. Deep (--deep, OpenAI GPT-5.6 web_search via Bedrock Mantle, ~20-40s, token-expensive) — multi-round retrieval with first-hand sources and per-claim citations, for verifying facts behind important decisions or double-checking weak/contradictory fast results. Default to fast; escalate to deep only when the question deserves depth.
+user-invocable: false
+---
+# Web Search
+
+One skill, two gears:
+
+| Gear | Engine | Speed / cost | Use for |
+|---|---|---|---|
+| **fast** (default) | Amazon Nova 2 Lite `nova_grounding` | ~3 s, ~500 tokens | News, quotes, prices, routine lookups — the everyday default |
+| **deep** (`--deep`) | OpenAI GPT-5.6 `web_search` (Bedrock Mantle) | ~20-40 s, 10-40k tokens | Fact verification before important conclusions; second opinion when fast results look weak, self-contradictory, or conflict with data you hold |
+
+Deep gear does multi-round real retrieval (bilingual query rewriting, `site:`-scoped
+searches against official sites), leans toward authoritative first-hand sources
+(official announcements, news agencies), and attaches a citation to every claim —
+**expensive and slow, but thorough and it doesn't make things up**.
+
+## Usage
+
+```bash
+# fast (default) — everyday lookups
+python3 ~/.halo/global/skills/web-search/search.py "What are the latest news about the China A-share market today?"
+
+# deep — verify facts that must be right
+python3 ~/.halo/global/skills/web-search/search.py --deep "What did the US Federal Reserve announce most recently about interest rates?"
+```
+
+- Output: answer text + a `Sources:` URL list (gear/model/token usage on stderr).
+- Deep gear default model is `luna` (reasoning effort=max — retrieval planning
+  maxed out). `--deep --model sol` is carpet-bombing research (auto-paired with
+  effort=xhigh; ~3x more search queries, roughly double the tokens) — reserve it
+  for "profile one stock / one topic exhaustively" jobs.
+- Full natural-language questions beat keyword stubs in both gears
+  (`"RTX 5080"` ✗ → `"Tell me about the NVIDIA RTX 5080, including specs and price."` ✓).
+  English works best; Chinese works too — deep gear rewrites queries bilingually.
+- Search output is still a lead, not gospel: for critical figures, `web_fetch`
+  the cited source URL and confirm before acting on it.
+- Auth is automatic in both gears (machine AWS credentials — no API key to
+  configure or rotate). Deep gear needs the `aws-bedrock-token-generator` pip
+  package (one-time: `pip3 install --user aws-bedrock-token-generator`).
+- A deep call can take 40 s — be patient in shell_exec. One clear question per
+  call; don't fan out parallel calls.

package/templates/skills/web-search/search.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+"""Web search with two gears.
+
+fast (default): Amazon Nova 2 Lite + nova_grounding — ~3 s, ~500 tokens.
+deep (--deep):  OpenAI GPT-5.6 on Bedrock Mantle + server-side web_search —
+                multi-round retrieval, first-hand sources, ~20-40 s, 10-40k tokens.
+
+Usage:
+  python3 search.py "your question"
+  python3 search.py --deep "your question" [--model sol|luna]
+
+Auth is automatic in both gears (machine AWS credentials): fast uses boto3's
+default credential chain; deep mints a 12-hour bearer token on every call via
+aws-bedrock-token-generator — no API key to configure or rotate.
+"""
+import json
+import sys
+import urllib.error
+import urllib.request
+
+MANTLE_URL = "https://bedrock-mantle.us-east-2.api.aws/openai/v1/responses"
+MODELS = {"luna": "openai.gpt-5.6-luna", "sol": "openai.gpt-5.6-sol"}
+# luna (default) runs effort=max to squeeze out its retrieval planning; sol's
+# retrieval is already exhaustive at medium (measured 48 search queries on one
+# question), so it gets xhigh to avoid pointless token doubling — max over
+# xhigh measured only ~6% more reasoning tokens.
+EFFORTS = {"openai.gpt-5.6-luna": "max", "openai.gpt-5.6-sol": "xhigh"}
+
+
+def fast_search(question: str) -> None:
+    import boto3
+    from botocore.config import Config
+
+    bedrock = boto3.client(
+        "bedrock-runtime", region_name="us-east-1", config=Config(read_timeout=60)
+    )
+    response = bedrock.converse(
+        modelId="us.amazon.nova-2-lite-v1:0",
+        messages=[{"role": "user", "content": [{"text": question}]}],
+        toolConfig={"tools": [{"systemTool": {"name": "nova_grounding"}}]},
+    )
+    answer = ""
+    sources: list[str] = []
+    for content in response["output"]["message"]["content"]:
+        if "text" in content:
+            answer += content["text"]
+        elif "citationsContent" in content:
+            for c in content["citationsContent"].get("citations", []):
+                url = c.get("location", {}).get("web", {}).get("url", "")
+                if url and url not in sources:
+                    sources.append(url)
+    print(answer)
+    if sources:
+        print("\nSources:")
+        for u in sources:
+            print(" -", u)
+    usage = response.get("usage", {})
+    print(f"\n[gear=fast model=nova-2-lite total_tokens={usage.get('totalTokens')}]", file=sys.stderr)
+
+
+def deep_search(question: str, model: str) -> None:
+    from aws_bedrock_token_generator import provide_token
+
+    token = provide_token(region="us-east-2")
+    body = json.dumps(
+        {
+            "model": model,
+            "input": question,
+            "tools": [{"type": "web_search"}],
+            "reasoning": {"effort": EFFORTS[model]},
+        }
+    ).encode()
+    req = urllib.request.Request(
+        MANTLE_URL,
+        data=body,
+        headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"},
+        method="POST",
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=180) as resp:
+            data = json.loads(resp.read())
+    except urllib.error.HTTPError as e:
+        print(f"HTTP {e.code}: {e.read().decode(errors='replace')[:2000]}", file=sys.stderr)
+        sys.exit(1)
+
+    texts: list[str] = []
+    sources: list[str] = []
+    for item in data.get("output", []):
+        if item.get("type") == "message":
+            for c in item.get("content", []):
+                if c.get("type") == "output_text":
+                    texts.append(c.get("text", ""))
+                    for a in c.get("annotations", []):
+                        if a.get("type") == "url_citation":
+                            u = a.get("url", "")
+                            if u and u not in sources:
+                                sources.append(u)
+    print("\n".join(texts))
+    if sources:
+        print("\nSources:")
+        for u in sources:
+            print(" -", u)
+    usage = data.get("usage", {})
+    print(f"\n[gear=deep model={model} total_tokens={usage.get('total_tokens')}]", file=sys.stderr)
+
+
+def main() -> None:
+    args = sys.argv[1:]
+    deep = False
+    model = MODELS["luna"]
+    if "--deep" in args:
+        deep = True
+        args.remove("--deep")
+    if "--model" in args:
+        i = args.index("--model")
+        try:
+            model = MODELS[args[i + 1]]
+        except (IndexError, KeyError):
+            print("--model must be 'luna' or 'sol'", file=sys.stderr)
+            sys.exit(2)
+        del args[i : i + 2]
+        if not deep:
+            print("--model only applies with --deep", file=sys.stderr)
+            sys.exit(2)
+    if not args:
+        print('usage: search.py [--deep] "<question>" [--model sol|luna]', file=sys.stderr)
+        sys.exit(2)
+    question = " ".join(args)
+
+    if deep:
+        deep_search(question, model)
+    else:
+        fast_search(question)
+
+
+if __name__ == "__main__":
+    main()

package/templates/skills/workspace/share.md

@@ -53,7 +53,7 @@ Agents (2):
 Skills (3):
   - repo-conventions (workspace)
   - aws-knowledge (global)
-  - nova-web-search (global)
+  - web-search (global)
 
 Instructions: .halo/INSTRUCTIONS.md (global fallback)
 INDEX.md → 2 docs followed: docs/architecture/auth.md, docs/dev/api.md

package/templates/skills/workspace/stage.py

@@ -74,7 +74,7 @@ BUILTIN_AGENT_IDS = {
 }
 BUILTIN_SKILL_IDS = {
     "agent", "skill", "workspace", "cron", "send-file", "acp",
-    "aws-knowledge", "nova-web-search", "self",
+    "aws-knowledge", "web-search", "self",
 }
 
 # --- helpers ---------------------------------------------------------------

package/templates/skills/nova-web-search/SKILL.md

@@ -1,65 +0,0 @@
----
-name: nova-web-search
-description: Real-time web search using Amazon Nova 2 Lite's built-in nova_grounding system tool. Use this skill when the user asks about current events, news, latest updates, prices, or any real-time information. This replaces the web_search tool.
-user-invocable: false
----
-# Nova Web Search
-
-Perform real-time web search using Amazon Nova 2 Lite's built-in `nova_grounding` system tool. Returns a synthesized answer with source URLs.
-
-## How to Use
-
-Run the following Python script via `shell_exec`:
-
-```python
-import boto3
-from botocore.config import Config
-
-bedrock = boto3.client("bedrock-runtime", region_name="us-east-1", config=Config(read_timeout=60))
-
-query = "YOUR_QUERY_HERE"  # Replace with the actual search query
-
-response = bedrock.converse(
-    modelId="us.amazon.nova-2-lite-v1:0",
-    messages=[{"role": "user", "content": [{"text": query}]}],
-    toolConfig={"tools": [{"systemTool": {"name": "nova_grounding"}}]}
-)
-
-answer = ""
-sources = []
-for content in response["output"]["message"]["content"]:
-    if "text" in content:
-        answer += content["text"]
-    elif "citationsContent" in content:
-        for c in content["citationsContent"].get("citations", []):
-            url = c.get("location", {}).get("web", {}).get("url", "")
-            if url and url not in sources:
-                sources.append(url)
-
-print(answer)
-if sources:
-    print("\nSources:", sources)
-```
-
-## Query Formulation
-
-Nova grounds better against full natural-language questions than against
-keyword stubs — raw keywords return more loosely-related sources, while
-a question gives Nova enough context to filter.
-
-| User Intent | Query Template |
-|---|---|
-| News / updates / latest developments | `"What are the latest news about X?"` |
-| Product / specs / concept lookup | `"Tell me about X, including specs, price, and latest updates."` |
-| General search | `"Search for information about X."` |
-
-**Examples:**
-- ✅ `"What are the latest news about the Middle East conflict?"`
-- ✅ `"Tell me about the NVIDIA RTX 5080, including specs and price."`
-- ✗ `"RTX 5080"` — terse keywords leave Nova guessing at intent
-
-## Requirements
-
-- Region **must** be `us-east-1`
-- English queries yield better results than other languages
-- Model ID: `us.amazon.nova-2-lite-v1:0`