x402 market recon: measure before you change anything

Trigger, when the question is about an x402 endpoint's discoverability or its price:

• "why is nobody buying my x402 endpoint", "should I rewrite my endpoint descriptions"
• "where do I list this", "is my price wrong", "does anyone actually see this endpoint"
• a plan to change a payment path, a discovery document or a listing that currently rests on a hunch

Do not invoke this skill for: debugging an unrelated payment failure, rewriting copy on a marketplace whose rules you already know, or building a new paid endpoint from scratch - that last one is an implementation task, not a measurement one. This skill measures; it does not build or spend.

The whole skill is one habit: turn the question into a number you can re-measure, before editing anything you would have to test with real money. In practice this reorders the usual plan. Retrofit projects get cancelled, and the real blocker usually turns out to be one of three things: your discovery document does not publish route/price/description, your price sits above the buyer's hard cap, or the directory you care about never sees a settlement from you.

Declared permissions, and how each one is enforced

What this skill will NOT do

• It will not spend, sign or settle anything. There is no wallet code here: a tool that can move money deserves its own description, its own declared permissions and its own review, so it belongs in a separate artifact rather than inside a measurement tool.
• It will not read a private key, a seed phrase, a token file or a password.
• It will not send your data anywhere. The only bytes that leave the machine are the URLs and the empty bodies of the reads listed above: no telemetry, no analytics, no callback.
• It will not write outside --out, and it will not touch a service you own.

The five measurements, in this order

Do them in this order; each one either kills the plan or sharpens it.

1. Census the registry and find where demand actually is (census) GET https://api.cdp.coinbase.com/platform/v2/x402/discovery/resources?limit=100&offset=N is offset-paginated with no server-side filter, and each item carries quality: {l30DaysTotalCalls, l30DaysUniquePayers, lastCalledAt}. That field is the difference between guessing what sells and counting it. Answers: median demand per listing, how concentrated the demand is, which price bands and tags get picked, and what it takes to reach a visible rank.

2. Probe live 402s and see how terms are actually delivered (probe) Ask each endpoint unpaid, read the answer, and record whether the payment terms arrived in the PAYMENT-REQUIRED header or in the body, and what fields accepts[0] carries. This is what ends arguments about "maybe the directory reads the body" - with a count, not an opinion.

3. Read the directory's own check list (see references/decision-rules.md) Directories publish exactly which checks they run and what each badge requires. Read the checks, not a prose summary of them, and never quote a browsing agent's paraphrase of a method page as fact.

4. Read the router's published ranking rule and its caps A router that sends buyers your way ranks candidates by a published sequence and often caps what it will spend on a seller with no settlement history. Find that sentence before promising anyone that a listing equals sales: those caps are the usual reason a healthy endpoint is never bought.

5. Compare your discovery document against live truth (manifest) Crawlers read /.well-known/x402 for route, price and description. If it publishes bare URLs, the crawler has nothing to match a task against, even though every paid endpoint is perfect. Build the missing part from measured live challenge values, keep every other key byte-identical, back up outside the docroot, and re-fetch the public URL to compare hashes.

Decision rules

The full table with the measured anchors is in references/decision-rules.md. The short version:

• Metadata completeness is table stakes, not a differentiator. Most of the catalogue already has examples and schemas; if nearly everyone has it, it cannot be what separates you.
• Description length is not quality. Long descriptions correlate with LOWER demand. Lead with what the call does and what it returns; add one honest "not for" clause if it prevents misuse.
• Price decides who can buy you. Buyers and routers have hard caps (a default client cap, a router's bootstrap cap). A route priced above the cap is invisible in practice, however good.
• Settlements, not requests, build every published rank. Two callers rank a listing: distinct payers over a window, and settled volume. A free endpoint creates neither.
• A manifest is not evidence of demand, and demand is not evidence of a good manifest. Measure both; never infer one from the other.

Pitfalls that cost real time

1. Reading only accepts[0].amount invents free resources. v1-style sellers put the price in maxAmountRequired. Read both keys, or you will "discover" a market of free endpoints that is actually paid.
2. HTTP headers are case-insensitive. PAYMENT-REQUIRED and payment-required are the same header. A probe that looks for one exact spelling reports "no terms" about a healthy endpoint.
3. A single probe can report a 404/405 that only means the wrong verb. Try the verb the route declares, and try twice: a trial or consumed preview can answer 200 once and 402 on the retry.
4. Registry numbers have an unpublished definition and are not clean. One sampled amount was larger than the entire ecosystem's volume; use medians, and label the count as registry-reported.
5. A browsing agent's "quote" is data, not evidence. Cross-check anything that would change a decision against a machine-readable source: a check list, a raw fetch, an API.
6. Never edit an audited payment path for a cosmetic metadata claim. Mirroring accepts[] into the body buys v1-reader compatibility only, and the graders already pass header-only endpoints.
7. Anything left under the docroot is published. Backup files, .bak, ~ copies: move backups outside and re-check that the old URLs answer 404.
8. A payment from your own wallet is a test, not revenue. Label it, keep the tx hash, and read the payer from the USDC Transfer log (from), never from tx.from - a facilitator relays the transfer and pays the gas.
9. The cheapest listed price is not the cheapest path to a rank. Manufacturing transactions is penny-scale and detectable; if you build loop traffic, disclose it next to the number.

Hand back a report, not an opinion

Deliver: the measured baseline (with the date, the sample size and each figure's source), the one or two blockers the numbers exposed, and a short list of next actions with their exact cost, cheapest first. Mark anything you did not verify as unverified rather than leaving it out.

Files

• scripts/recon_x402_market.py - census, probe and manifest comparison. Read-only, Python standard library only. Off-domain reads need --allow-any-host, and internal addresses stay refused.
• references/decision-rules.md - thresholds, badge semantics, router order, what each finding means.
• references/findings-baseline.md - a dated snapshot of the ecosystem, so drift is distinguishable from a measurement error.