Skip to content

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.

Guide: say that many tools do not mark a failure, and drop stale counts

Phoenix Agent (mac-studio)283f8aba8af0
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 64cec29..dae29d8 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -190,10 +190,10 @@ source here, and no proxy is substituted for any of these measures.
domain module, with a `[needs VARIABLE]` mark on each tool that returns
an error without its API key. Built from the live registry on every
read; a test fails if a registered tool is missing or listed twice. It
- says plainly that not every tool marks a failure: with every upstream
- request forced to fail (measured 2026-10-17), 53 of the 85 tools that
- call a source returned an empty list, a zero count or a neutral score
- with no error, `degraded` or `data_gaps` field.
+ says plainly that many tools do not mark a failure: when every upstream
+ request fails they can return an empty list, a zero count, a neutral
+ score, a LOW risk label, an all-clear status or a false flag with no
+ error, `degraded` or `data_gaps` field.
- `intel://guide/source-health` (JSON): open circuit breakers, stale
serves, stale cache prefixes, cache freshness, and which keyed sources
have their variables set (names only, never values), with a `verdict`
diff --git a/src/world_intel_mcp/guide.py b/src/world_intel_mcp/guide.py
index 11874a3..f8b2e53 100644
--- a/src/world_intel_mcp/guide.py
+++ b/src/world_intel_mcp/guide.py
@@ -1,7 +1,7 @@
"""Research guide served over MCP as resources and prompts.
-A model connecting to this server sees 139 tools and, before this module,
-nothing else: no map of which tool answers which question, and no
+A model connecting to this server sees every registered tool and, before
+this module, nothing else: no map of which tool answers which question, and no
statement of what the data cannot support. The guide fills that gap with
three read-only resources, a set of worked research prompts
(``guide_prompts.py``) and the server's ``instructions`` at initialize:
@@ -28,8 +28,14 @@ Gaps / not covered:
key is valid or within quota.
- A cache file moved or deleted while the server holds it open keeps
answering from the open handle; source health cannot tell.
-- The guide's count of tools that return no failure marker under a total
- outage is a measurement of one commit; it is not recomputed at runtime.
+- The guide states no count of tools that return no failure marker under
+ a total outage, because the count changes as tools are added. A review
+ sweep at 7782d22 (137 of its 139 tools, each in a fresh process with a
+ fresh breaker and cache, every request failing) found 68 of the
+ 105 tools that called a source returned no error, degraded or data_gaps
+ field; an earlier single-process sweep read 53 of 85 because shared
+ circuit breakers stopped 17 tools from calling at all. Neither figure
+ is recomputed here.
"""
import json
@@ -68,9 +74,9 @@ SERVER_INSTRUCTIONS = (
"research questions to tools, says what the data cannot support, and "
"marks the tools that need an API key. prompts/list has worked research "
"prompts. intel://guide/source-health shows which sources are failing, "
- "serving stale copies or missing keys. Not every tool marks a failure: "
- "an empty list or a zero can mean a source failed, not that nothing "
- "happened."
+ "serving stale copies or missing keys. Many tools do not mark a failure: "
+ "an empty list, a zero, a LOW label, an all-clear or a false flag can "
+ "mean a source failed, not that nothing happened."
)
# Display names for the domain modules under tools/. A module missing
@@ -260,13 +266,14 @@ with none of them invites a story.
and carry a fetch timestamp; the date a value describes is often earlier
than the fetch, so give both. Some tools return no observation date at
all; say so rather than using the fetch time.
-- Failures, most of the time. Many tools return `error`, `degraded` or
- `data_gaps` when a source is down or not configured. Not all do: some
- return empty lists, zero counts or a neutral score when every upstream
- request failed (measured 2026-10-17 with every request forced to fail:
- 53 of the 85 tools that call a source). Read an empty list or a zero
- count as nothing returned, not nothing happened, and confirm with a
- second tool before calling a place quiet.
+- Failures, sometimes. Some tools return `error`, `degraded` or
+ `data_gaps` when a source is down or not configured. Many do not: when
+ every upstream request fails they return an empty list, a zero count, a
+ neutral score, a LOW risk label, an all-clear status or a false flag,
+ with no marker. Read an empty list or a zero count as nothing returned,
+ not nothing happened, and confirm with a second tool before calling a
+ place quiet. `failing_sources` in the source-health resource names the
+ sources whose last call failed.
- Citations. Briefs and digests number their sources.
- Old copies, counted. When a live fetch fails and an older cached copy
exists, that copy is served. Dictionary responses carry `_stale` and
@@ -404,10 +411,10 @@ components with the score; a null component was not measured.
Many tools report a source that failed, is not configured, or does not
cover a place with an error, a note or a `data_gaps` entry. That means
-unknown, not none. Not every tool does: some return an empty list, a
-zero count or a neutral score when every upstream request failed, so an
-empty list or a zero means nothing was returned, not that nothing
-happened. Zero events from a fetch that worked, in a region the source
+unknown, not none. Many tools do not: when every upstream request fails
+they return an empty list, a zero count, a neutral score, a LOW risk
+label, an all-clear status or a false flag, so an empty list or a zero
+means nothing was returned, not that nothing happened. Zero events from a fetch that worked, in a region the source
covers and for a query that matched the place, is a measurement; zero
events in a region it does not cover is nothing at all.
"""
diff --git a/src/world_intel_mcp/guide_prompts.py b/src/world_intel_mcp/guide_prompts.py
index 392235a..32577a2 100644
--- a/src/world_intel_mcp/guide_prompts.py
+++ b/src/world_intel_mcp/guide_prompts.py
@@ -34,10 +34,11 @@ COMMON_LIMITS: tuple[str, ...] = (
"A tool that returns an error, a note or a data_gaps entry has told you "
"what is missing: report it as missing. Missing data is never zero, calm "
"or peace.",
- "Not every tool marks a failure. Some return an empty list, a zero count "
- "or a neutral score when every upstream request failed. Read an empty "
- "list or a zero as nothing returned, not nothing happened, and confirm "
- "with a second tool before calling a place quiet.",
+ "Many tools do not mark a failure. When every upstream request fails "
+ "they return an empty list, a zero count, a neutral score, a LOW risk "
+ "label, an all-clear status or a false flag. Read an empty list or a "
+ "zero as nothing returned, not nothing happened, and confirm with a "
+ "second tool before calling a place quiet.",
"Give each figure the date it describes (the observation date) as well "
"as when it was fetched; they are often different. If a tool gives no "
"observation date, say so.",