a digital entity named phi that roams bsky
phi.zzstoatzz.io
add web_search tool (tavily) so phi can ground recency claims
phi was confidently posting "this week" / "same week" framings about
old news because it had no way to check the open web. tavily fits the
shape: 1k free searches/month, fast, and time_range bounds the window
so phi can't accidentally cite stale articles.
tool: web_search(query, time_range, topic, max_results) — Annotated
params with descriptions the LLM sees. operational instruction nudges
phi to pass time_range BEFORE asserting recency, not after.
requires TAVILY_API_KEY env var (set as fly secret in prod).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+106
-1
+2
src/bot/agent.py
+2
src/bot/agent.py
···
42
42
check_services checks nate's infrastructure, not yours. only use during reflection or when explicitly asked about services.
43
43
44
44
check_relays reads relay-eval. modes: no args = fleet snapshot; name="<host>" = coverage history for that relay (narrow with since/until for precision); transitions=True = fleet-wide status-change log (best for "when did X happen"). use headlines verbatim — no theories about cause.
45
+
46
+
web_search hits the open web (Tavily). use BEFORE asserting anything is recent / current / "this week" — pass time_range="week" so results are bounded. for old-news pattern matching ("same week," "just announced"), search first, post second. headlines without timestamps aren't evidence of timing.
45
47
""".strip()
46
48
47
49
+6
src/bot/config.py
+6
src/bot/config.py
···
56
56
default=None, description="The search engine ID for the Google API"
57
57
)
58
58
59
+
# Tavily web search — grounds phi against the open web for currency
60
+
# checks and source-finding. Free tier covers 1k searches/month.
61
+
tavily_api_key: str | None = Field(
62
+
default=None, description="API key for Tavily web search"
63
+
)
64
+
59
65
# TurboPuffer configuration
60
66
turbopuffer_api_key: str | None = Field(
61
67
default=None, description="The API key for the TurboPuffer API"
+98
-1
src/bot/tools/search.py
+98
-1
src/bot/tools/search.py
···
1
-
"""Search tools — bluesky posts, cosmik network, trending."""
1
+
"""Search tools — bluesky posts, cosmik network, trending, open web."""
2
2
3
3
from datetime import date
4
+
from typing import Annotated, Literal
4
5
5
6
import httpx
7
+
from pydantic import Field
6
8
from pydantic_ai import RunContext
7
9
10
+
from bot.config import settings
8
11
from bot.core.atproto_client import bot_client
9
12
from bot.tools._helpers import PhiDeps, _relative_age
10
13
···
76
79
return "\n\n".join(lines)
77
80
except Exception as e:
78
81
return f"network search failed: {e}"
82
+
83
+
@agent.tool
84
+
async def web_search(
85
+
ctx: RunContext[PhiDeps],
86
+
query: Annotated[
87
+
str,
88
+
Field(description="Search query — natural language."),
89
+
],
90
+
time_range: Annotated[
91
+
Literal["day", "week", "month", "year"] | None,
92
+
Field(
93
+
description=(
94
+
"Bound results to a time window relative to today. "
95
+
"Use this BEFORE asserting recency in a post — "
96
+
"e.g. set 'week' before claiming something happened "
97
+
"this week. Without it, results may include stale items."
98
+
)
99
+
),
100
+
] = None,
101
+
topic: Annotated[
102
+
Literal["general", "news"] | None,
103
+
Field(
104
+
description=(
105
+
"'news' optimizes for recent journalism, 'general' for "
106
+
"evergreen content. Default: general."
107
+
)
108
+
),
109
+
] = None,
110
+
max_results: Annotated[
111
+
int,
112
+
Field(description="How many results to return. Default 5."),
113
+
] = 5,
114
+
) -> str:
115
+
"""Search the open web via Tavily.
116
+
117
+
Use to ground claims about the world outside atproto — current
118
+
events, primary sources, official statements, technical docs.
119
+
For atproto posts use search_posts; for the cosmik network use
120
+
search_network.
121
+
122
+
IMPORTANT: if you're about to assert something is recent, current,
123
+
or 'this week,' pass time_range first. headlines without dates
124
+
aren't evidence of when something happened."""
125
+
if not settings.tavily_api_key:
126
+
return "web_search unavailable: TAVILY_API_KEY not set"
127
+
128
+
body: dict = {
129
+
"query": query,
130
+
"max_results": max_results,
131
+
"search_depth": "basic",
132
+
}
133
+
if time_range:
134
+
body["time_range"] = time_range
135
+
if topic:
136
+
body["topic"] = topic
137
+
138
+
try:
139
+
async with httpx.AsyncClient(timeout=20) as http:
140
+
r = await http.post(
141
+
"https://api.tavily.com/search",
142
+
headers={
143
+
"Authorization": f"Bearer {settings.tavily_api_key}",
144
+
"Content-Type": "application/json",
145
+
},
146
+
json=body,
147
+
)
148
+
r.raise_for_status()
149
+
data = r.json()
150
+
except Exception as e:
151
+
return f"web search failed: {e}"
152
+
153
+
results = data.get("results", [])
154
+
if not results:
155
+
return f"no web results for '{query}'"
156
+
157
+
scope_parts = []
158
+
if time_range:
159
+
scope_parts.append(f"time_range={time_range}")
160
+
if topic:
161
+
scope_parts.append(f"topic={topic}")
162
+
scope = f" ({', '.join(scope_parts)})" if scope_parts else ""
163
+
164
+
lines = [f"web results for '{query}'{scope}:"]
165
+
for i, r_item in enumerate(results, 1):
166
+
title = r_item.get("title", "untitled")
167
+
url = r_item.get("url", "")
168
+
content = (r_item.get("content") or "").strip()
169
+
lines.append("")
170
+
lines.append(f"{i}. {title}")
171
+
if url:
172
+
lines.append(f" {url}")
173
+
if content:
174
+
lines.append(f" {content[:400]}")
175
+
return "\n".join(lines)
79
176
80
177
@agent.tool
81
178
async def get_trending(ctx: RunContext[PhiDeps]) -> str: