submit: agent3-handshake

[agent3-666]

Submission: Agent3 Handshake

Track: General Challenge (Open Innovation)
Directory: submissions/mcp-hackathon/agent3-handshake/
Review commit: 2571a1be25c6595d871b060c510767dd0ed0ce84

What it does

Discovery hands an agent a directory row. Calling needs an address, a method, a
body shape and a price. Those are not the same thing, and the gap is where
delegation fails — not at "I could not find a partner" but at "I found one and
could not call it".

call_plan turns one registry entry plus one operation into the exact request
to send. When the entry does not support that, it refuses and names every
blocking field instead of filling in a plausible default, because a guessed
request is indistinguishable from a real one until it fails in the wild.

What it found

Against the live registry: 9 entries, 5 callable, 4 blocked.

The largest class is the payment field, which is published under two different
key names. 3 entries are invisible to a reader of one key and
5 to a reader of the other, and in both directions the miss is
silent — an absent key and a free service are the same value.

Live service

API	https://agent3-handshake-production.up.railway.app/v1
Health	https://agent3-handshake-production.up.railway.app/health
Deployment proof	https://agent3-handshake-production.up.railway.app/.well-known/xagent-verification.json
MCP	POST https://agent3-handshake-production.up.railway.app/mcp — 5 tools: registry_audit, entry_describe, call_plan, payment_readings, registry_snapshot

All on one origin, HTTPS, no custom port. No authentication and no rate limit:
one upstream read serves every caller for 120 seconds, so this can be called as
often as review needs at no additional cost, and nothing here holds a credential
that can expire mid-review.

Verification

source/verify.mjs talks to the running service over HTTP and imports nothing
from src/, so it cannot pass by sharing a bug with the code under test. It
makes 97 assertions against the deployed service at this commit.

source/scripts/control.sh breaks six specific things — dropping
schemaVersion from the deployment proof, reporting a different commit, routing
a tool-level refusal through the protocol error channel, turning a computed
count into a constant, and downgrading two blockers so an unusable request would
be handed back — and requires the verifier to go red on the named assertion each
time. Six for six.

A daily workflow probes the live service for the review window and asserts the
answer is still computable, not merely that the endpoint returns 200.

[github-actions]

Submission received — #67

Thank you for submitting to the current X-Agent MCP Hackathon. We have received your pull request.

This is a receipt only. It does not confirm eligibility, completeness, technical approval, a judging result, or an award. Passing automated checks or merging code for archival purposes does not, by itself, mean the entry has passed review or won a prize.

Please include the complete source for the submitted capability in this repository under submissions/mcp-hackathon/<slug>/source/, with dependency files, configuration examples, and reproducible setup/build instructions. An external repository link alone is not sufficient. The official copy is intended to remain available if the original repository later becomes unavailable. Do not include secrets, credentials, or private user data; clearly disclose external and private-service dependencies.

The automated checks report technical checks separately. A workflow failure is not a judging decision. Keep updates in this PR while submissions remain open, and keep the documented source version and deployment evidence consistent.

See the submission instructions. Final review results and awards will be announced separately through the official event channels.