Topic Radar

Topic Radar

Contract

Prereqs:

• Network access for live lookups.
• uv available from the agent-kit repository environment.
• Optional upstreams may rate-limit anonymous requests; the skill must degrade per-source instead of failing the whole digest.
• Polymarket source support prefers read-only MCP tool output when available; otherwise it falls back to the

$AGENT_HOME/skills/tools/market-research/polymarket-readonly/scripts/polymarket-readonly.sh helper.

Inputs:

• Natural-language AI or technology topics.
• Optional profile name: terry-ai-tech by default, ai-tech for the generic baseline, or aliases terry, personal, and default.
• Optional preset: radar by default, or ai-news for a faster daily AI news scan focused on official/news/HN sources.
• Optional source list: polymarket, hn, github, arxiv, hf, official, news, or all.
• Daily, weekly, or monthly report request, custom rolling day window, fixed --from/--to or --month window, result limit, parallel fetch

count, cache TTL, news provider strategy, brief mode, and output format.

• Optional Polymarket MCP JSON export path passed with --polymarket-mcp-json.

Outputs:

• Source-grounded AI/technology trend digest in Markdown or JSON.
• Optional clustered brief for fast reading across product, agent/tooling, enterprise, security/governance, and research/open-ecosystem signals.
• Ranked cross-source signal list with source metadata, URLs, timestamps, fixed or rolling window metadata, score rationale, signalTier,

signalMetrics, and per-source sections.

• Optional Early Watchlist entries for relevant but low-signal items. These entries stay in raw items output but should not be treated as

Topline or adoption evidence.

• Per-source errors when an upstream is unavailable, rate-limited, or malformed, including short response snippets when available.
• Public-response cache metadata so repeated follow-up scans are auditable.
• Sample-mode output for offline smoke checks and report-format review.

Exit codes:

• 0: success
• 1: failure
• 2: usage error
• 3: unsafe delegated source environment detected

Failure modes:

• Unknown source, report type, format, profile, invalid fixed-window shape, or invalid numeric option.
• Live source returns invalid data, times out, or rate-limits; the affected source is reported in errors, while other sources continue.
• The broad radar preset uses news-provider=auto: GDELT first, then Google News RSS fallback when GDELT is unavailable or empty.
• The faster ai-news preset uses news-provider=google by default to avoid GDELT rate-limit stalls during daily scans.
• Polymarket MCP output is missing, malformed, or has no usable records; continue to helper fallback unless --polymarket-fallback none is

set.

• Delegated Polymarket helper detects unsafe trading credentials; stop with exit 3.

Scripts (only entrypoints)

• $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh

References

• Read [references/source-strategy.md](references/source-strategy.md) before changing source selection, ranking, or report sections.

Role Boundary

• Use topic-radar directly when the user asks for source-level scanning, custom sources, raw Markdown/JSON digest output, or automation

input.

• Use daily-brief when the user wants the normal daily information entrypoint: concise user-language synthesis, preference steering,

freshness notes, and source-health explanation.

• topic-radar owns source fetching, ranking, fallback behavior, public-response caching, brief clusters, and per-source errors.
• daily-brief must not reimplement fetchers, ranking, cache, RSS/API parsing, or Polymarket fallback logic; it should consume

topic-radar --format json.

Workflow

1. Treat this skill as read-only content and trend research. Do not request or configure trading credentials, social posting credentials, or

paid news API keys by default.

1. For "AI news today / this week / these days" requests, start with the faster AI news preset. It uses a 5-day window, clustered brief

output, public-response caching, Google News RSS for mainstream coverage, and the source order official,news,hn:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh --preset ai-news ``

Use --refresh when the user asks for exact latest/current results and cached responses should be bypassed:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh --preset ai-news --refresh ``

1. Use the broad radar preset when the user asks for general AI/Tech trend scanning across market, developer, research, model, official, and

news surfaces. The default profile tracks AI agents, agentic coding, model releases, developer tools, AI infrastructure, inference serving, semiconductors, NVIDIA, robotics, embodied AI, OpenAI, Anthropic, Google DeepMind, and Hugging Face:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh --report daily ``

1. Use personal memory only as a steering layer for recurring interests and source preferences before running the script. Do not write

secrets, temporary task state, or project state into memory. Keep the report grounded in the script output and linked source items.

1. When Polymarket is requested and the current agent has the polymarket MCP tools, query read-only MCP first:

• gamma_list_events with active=true, closed=false, and a small limit
• gamma_list_markets with active=true, closed=false, and a small limit
• gamma_search_public for the user's focused topic when broad lists are too noisy

Save those raw MCP results as JSON shaped like this:

``json { "events": [], "markets": [], "search": [] } ``

Then pass the file to the radar:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh \ --sources polymarket,hn,github \ --polymarket-mcp-json /path/to/polymarket-mcp.json \ --report daily ``

1. If no MCP output is available, omit --polymarket-mcp-json; the script falls back to the read-only polymarket-readonly helper and

reports per-source errors if the local network blocks Polymarket REST.

1. For focused research, pass topics and source filters:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh \ --topic "AI agents" \ --topic "developer tools" \ --sources hn,github,arxiv,hf,polymarket \ --report weekly ``

1. For monthly trend archive work, use a fixed calendar window rather than a rolling --days window:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh \ --preset radar \ --report monthly \ --month 2026-01 \ --format json ``

For a partial current month, use explicit inclusive dates:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh \ --preset radar \ --from 2026-05-01 \ --to 2026-05-15 \ --format json ``

Treat fixed windows as source-bounded evidence. Some sources are current snapshots rather than reliable historical archives; the JSON errors field reports those gaps instead of silently inventing monthly history.

1. Use JSON when another tool, scheduled job, or agent synthesis step will consume the output:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh --preset ai-news --format json ``

1. Use sample mode for smoke checks, demos, or local validation without network access:

``bash $AGENT_HOME/skills/tools/market-research/topic-radar/scripts/topic-radar.sh --sample --format markdown ``

1. Keep the report source-grounded. Separate observed source signals from inference, and do not present heuristic ranking as objective

importance. Treat signalTier as a presentation hint: high/medium signal can feed Topline and brief clusters, while early-watchlist items are relevant topic samples that need visible metrics such as HN points/comments or GitHub stars/forks.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: graysurf
• Source: graysurf/agent-kit
• License: MIT

Install and usage instructions live in the source repository linked above.