160-tool MCP server for real-time global intelligence from free public APIs: markets, liquidity, prediction markets, SEC filings, conflict, military, aircraft-cohort early warning, migration and border flows, cyber, climate, news and more. Dated observation history, 29 research prompts, cited briefs, geofences, SSE dashboard. MIT, no paid API keys.
Python ai-toolsai-watchanthropicclaudecybersecuritydashboardgeofencinggeopoliticalgeospatialintelligencemarketsmcpmigrationmilitarymodel-context-protocolnuclearosintreal-timespace-weatherthreat-intelligence
Guide tests: a step's purpose must match its tool's own summary
Phoenix Agent (mac-studio)bd3f75262bc4
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 80857d8..e697414 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -420,6 +420,14 @@ source here, and no proxy is substituted for any of these measures.years in history-recall; and the central bank, hotspot, airport andYahoo maturity counts. The test claim above is narrowed to the countsthe tests pin.+- The check that a prompt step's purpose agrees with its tool let wrong+ tools through on one shared parameter word, or on a tool's own+ disclaimer (internet outages repointed to the cyclone tool passed). A+ second check now requires a shared word with the tool's name or first+ sentence, without generic schema words, and three such swaps are pinned+ as failures. Seven step purposes were reworded to name what their tool+ returns. Measured on 118 steps and 160 tools, 1,141 of 18,762 one-step+ swaps still pass both checks.### Added: tropical cyclone forecast tracks- `intel_cyclones` and `intel_jtwc_cyclones` take `include_forecast`diff --git a/src/world_intel_mcp/guide_prompts.py b/src/world_intel_mcp/guide_prompts.pyindex 1be4e69..fa3d057 100644--- a/src/world_intel_mcp/guide_prompts.py+++ b/src/world_intel_mcp/guide_prompts.py@@ -537,7 +537,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = ("the daily effective federal funds rate history (needs FRED_API_KEY)",args=(("series_id", "DFF"),),),- ("intel_major_crosses", "how the major currencies sit against each other today"),+ ("intel_major_crosses", "the major currency cross rates: how they sit against each other today"),Step("intel_history_backfill","only if FRED_API_KEY is not set: loads the effective federal "@@ -709,7 +709,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = (steps=(Step("intel_earthquakes",- "the events, newest first and at most 50 by default, so a "+ "the earthquakes, newest first and at most 50 by default, so a ""count of 50 means older ones may have been cut off; call again ""with a higher limit to check",args=(("min_magnitude", "{min_magnitude}"), ("hours", "{hours}")),@@ -1056,7 +1056,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = (question="Which cloud providers report active incidents, and which of their regions are involved?",steps=(("intel_service_status", "active incidents at AWS, Azure, GCP, Cloudflare and GitHub"),- ("intel_cloud_regions", "where the named regions are"),+ ("intel_cloud_regions", "where the named cloud regions are"),),limits=("Status pages are published by the providers themselves and can "@@ -1142,7 +1142,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = (Step("intel_aoi_brief", "the cited picture inside {name} today", args=(("name", "{name}"),)),Step("intel_aoi_changes",- "the first call records a baseline for {name}",+ "the first change detection call records a baseline for {name}",args=(("name", "{name}"),),),),@@ -1316,7 +1316,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = (question="Does {query} appear on the US Treasury OFAC SDN list, and under which programs?",arguments=(PromptArg("query", "Name to search for", required=True),),steps=(- Step("intel_sanctions_search", "name matches on the SDN list", args=(("query", "{query}"),)),+ Step("intel_sanctions_search", "name matches on the SDN sanctions list", args=(("query", "{query}"),)),Step("intel_company_profile","only if a match is a listed company: who it is",@@ -1382,7 +1382,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = (Step("intel_domain_summary","the same over ten years: its earliest timestamps show when "- "storage began for each category",+ "storage began for each domain",args=(("hours", 87600),),),Step(@@ -1393,7 +1393,7 @@ PROMPTS: tuple[ResearchPrompt, ...] = (("intel_timeline", "the stored entries in time order for the matching domain"),Step("intel_similar_events",- "earlier stored entries that resemble the strongest match",+ "earlier stored events similar to the strongest match",args=(("text", Fill("the text of the strongest match from step 3")),),),(diff --git a/src/world_intel_mcp/tests/test_guide.py b/src/world_intel_mcp/tests/test_guide.pyindex fdfeda7..752752a 100644--- a/src/world_intel_mcp/tests/test_guide.py+++ b/src/world_intel_mcp/tests/test_guide.py@@ -738,9 +738,10 @@ def test_no_prompt_repeats_a_call() -> None:# the tool says about itself: it shares a word with the tool's name,# description or parameter descriptions, and any upper-case name it uses (a# provider such as AAA or NWS) appears there too, as written or spelled out.-# A step pointed at a wrong but registered tool fails one or both (review F14-# M7). Measured 2026-10-17: the check catches 12,006 of the 13,524 ways of-# pointing one step at another registered tool.+# It catches many steps pointed at a wrong but registered tool (review F14+# M7), not all: measured 2026-10-18 on 118 steps and 160 tools, 2,731 of the+# 18,762 ways of pointing one step at another tool pass it, and 1,141 also+# pass the summary check further down (review R8)._COMMON_WORDS = frozenset("about after also and any each every for from into its latest most only ""over recent that the their them then this today were what when where "@@ -2105,3 +2106,54 @@ def test_prompt_counts_match_their_tables() -> None:words = {3: "three", 4: "four", 5: "five"}flat = " ".join(_text("yield-curve-check").split())assert f"falls back to {words[len(maturities)]} Yahoo maturities ({', '.join(maturities)})" in flat+++# A stronger purpose check (review R8). One shared word anywhere in a tool's+# text let wrong tools through: parameter words such as country or days, and+# a tool's own disclaimer (intel_cyclones says "not an outage"). Here the+# purpose must share a word with the tool's own summary, its name and the+# first sentence of its description, leaving out words most tools share.+_GENERIC_SCHEMA_WORDS = frozenset(+ "active call country current data days default filter first flag intel "+ "limit list major name recent region tool".split()+)+++def _summary_words(tool: Tool) -> set[str]:+ summary = guide._first_sentence(" ".join((tool.description or "").split()))+ return _content_words(f"{tool.name.replace('_', ' ')} {summary}") - _GENERIC_SCHEMA_WORDS+++def _summary_problem(tool: Tool, purpose: str) -> str | None:+ if _content_words(purpose) & _summary_words(tool):+ return None+ return f"{tool.name}: the purpose shares no word with the tool's name or first sentence"+++def test_step_purposes_share_a_word_with_their_tools_summary(tools_pkg) -> None:+ by_name = {t.name: t for t in tools_pkg.ALL_TOOLS}+ problems = {+ f"{prompt.name} step {i + 1}": found+ for prompt in guide_prompts.PROMPTS+ for i, step in enumerate(prompt.steps)+ if (found := _summary_problem(by_name[step.tool], step.purpose))+ }+ assert problems == {}+++ ("prompt_name", "index", "wrong_tool"),+ [+ ("internet-disruption", 0, "intel_cyclones"),+ ("internet-disruption", 0, "intel_acled_events"),+ ("yield-curve-check", 0, "intel_disease_outbreaks"),+ ],+)+def test_summary_check_catches_swaps_the_word_check_let_through(+ tools_pkg, prompt_name: str, index: int, wrong_tool: str+) -> None:+ """Each of these passed the shared-word check; the summary check fails it."""+ by_name = {t.name: t for t in tools_pkg.ALL_TOOLS}+ purpose = guide_prompts.get(prompt_name).steps[index].purpose+ assert _purpose_problems(by_name[wrong_tool], purpose) == []+ assert _summary_problem(by_name[wrong_tool], purpose) is not None