worldmonitor/tests/fixtures/jmespath-samples/README.md

JMESPath measurement fixtures

These three captured tool responses feed scripts/measure-jmespath-savings.mjs (the U6 reproducible token-savings A/B). They are real responses from the production MCP endpoint, captured ONCE via scripts/capture-mcp-fixture.mjs and committed verbatim so the script is fully deterministic — re-running on the committed fixtures produces byte-identical numbers.

Files

File	Tool	Class
fat-get-market-data.response.json	get_market_data	fat (~5–10 KB raw)
medium-get-conflict-events.response.json	get_conflict_events	medium (~2–5 KB raw)
thin-get-chokepoint-status.response.json	get_chokepoint_status	thin (~0.5–1 KB raw)

How to (re)capture

# Use either a user API key or an OAuth access token from /api/oauth/token.
export WM_MCP_KEY="wm_0123456789abcdef0123456789abcdef01234567"
# Or:
# export WM_MCP_OAUTH_TOKEN="eyJhbGciOi..."

# Fat — no filter args, full bootstrap.
node scripts/capture-mcp-fixture.mjs --tool get_market_data \
  --name fat-get-market-data

# Medium — no filter args.
node scripts/capture-mcp-fixture.mjs --tool get_conflict_events \
  --name medium-get-conflict-events

# Thin — no filter args. Deterministic cache-backed thin tool (NOT
# get_world_brief, which is a different response class and would
# mix in LLM/network variance if it were RPC).
node scripts/capture-mcp-fixture.mjs --tool get_chokepoint_status \
  --name thin-get-chokepoint-status

The capture script hits the prod MCP HTTP endpoint, unwraps the content[0].text JSON envelope, and writes pretty-printed JSON to disk. Override WM_MCP_ENDPOINT for staging.

Refresh policy

Recapture when:

• A tool's payload schema changes (new fields, renamed fields)
• A cache key is added or removed from one of the three tools
• The default limit or summary semantics change in a way that alters the unfiltered response shape

When recapturing, re-run npx tsx scripts/measure-jmespath-savings.mjs and paste the markdown table into the next PR description that touches this feature.

Why direct fixture invocation (not mcpHandler)?

executeTool() reads cache via readJsonFromUpstash which calls fetch() directly, not deps.redisPipeline. Mocking deps doesn't intercept the cache reads, so we'd have to mock globalThis.fetch keyed by each tool's _cacheKeys + _seedMetaKey to get the cache path to feed our fixtures. Calling applyJmespath directly on the already-assembled tool response is simpler AND uniquely correct for what we measure — the projection delta, not the cache-path overhead.