Skip to content

STE-27: 7/7 GREEN, and the red steps were never the paid path - #65

Merged
Lexirieru merged 2 commits into
mainfrom
feat/27-publish-artifact-in-rehearsal
Sep 27, 2026
Merged

Lexirieru merged 2 commits into
mainfrom
feat/27-publish-artifact-in-rehearsal

Conversation

@Lexirieru

Copy link
Copy Markdown
Collaborator

Result

Run 4: 7 / 7 GREEN. docs/rehearsal/runs/2026-09-27T065414Z/, against https://api.sterish.xyz and the public dashboard https://app.sterish.xyz. 16 of 16 transaction links resolve and all 16 are successful.

# Step Runs 1–3 Run 4
3 Fresh agent checks via the dashboard → SAFE PARTIAL ×3 GREEN
4 402 → pay → licence → 200 RED ×3 GREEN — paid 0.1 USDC, licence minted, 200
5 Second call → 200 RED ×3 GREEN — served from the held licence, no payment

Steps 1, 2, 6 and 7 were already green and stayed green.

STE-48 ran for the first time. A bare X-AGENT-ADDRESS gets 401 OWNERSHIP_PROOF_REQUIRED; the same call carrying a SEP-53 signature gets 200, licence header held, balance unchanged. Run 2 called a borrowable licence the single most important thing outstanding. It is closed.

The reds were never the paid path

Steps 4–5 were red because the rehearsal tried to buy the brand-new skill it had just registered, and that skill can never be bought. Artifacts reach the API host only through intake publish-artifacts, run on that host and fed from the corpus committed in this repo — and every API route is a GET. There is no upload path, so nothing the runner does can publish anything.

In run 2 I said I would fix this by publishing the new skill's artifact as part of step 2. That plan cannot work. Recorded in FINDINGS.md rather than quietly replaced.

So steps 4–5 now buy a version the API will actually sell, chosen by asking the API rather than trusting a list in the runner: since STE-42 a 402 carrying a price is positive proof the bytes are on the server, and a 404 is a statement about publishing rather than about the verdict. The brand-new skill is still tried first, so if publishing ever reaches it this fallback stops being used on its own.

This green is a workaround in the test, not a fix in the product

Filed as STE-56 (Urgent, assigned to James, blocking STE-30). Run 4 registered a skill, audited it SAFE score 100, minted its VERIFIED badge and settled the escrow — and /use on that same skill is a plain 404. Same registry, same verdict, same badge; the only difference is whether an operator had previously copied bytes onto the box.

For the external developer STE-30 is about, the honest sequence today is: hand us the skill, we add it to our corpus and commit it, one of us opens a shell on the VPS. New finding F5 stays open until STE-56 closes, and step 4's own evidence row says this in as many words so nobody reads more into 7/7 than is there.

Also corrected

Run 2 recorded "production still answers 404 on /challenge". That was my measurement, not the deploy: the route is /use/{skill_id}/{version}/challenge, and a bare /challenge 404s whether or not the feature exists.

Tests

7 new tests in pipeline/tests/test_rehearsal_buy_target.py pinning what a /use status code is allowed to mean — in both directions, including that a network failure is reported as "could not tell" rather than as "not for sale". Full pipeline suite: 690 passed, 14 skipped. ruff check src tests clean.

Runner also points at api.sterish.xyz (STE-54); the old hostname still answers as an alias.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QdNuoJ9e954ojofNVGUezW

Lexirieru and others added 2 commits September 27, 2026 13:53
…t is not ours

Steps 4 and 5 have been RED for three runs because they tried to buy the
brand-new skill the run had just registered. I said in run 2 that I would fix
this by publishing that skill's artifact as part of step 2. That plan does not
work, and it is worth saying why rather than quietly doing something else:
artifacts reach the API host only through `intake publish-artifacts`, run on
that host and fed from the corpus committed in this repo, and every API route
is a GET — there is no upload path. The runner cannot publish anything.

So steps 4 and 5 now buy a version the API will actually sell, chosen by asking
the API rather than by trusting a list in this file. Since STE-42 a 402 carrying
a price is positive proof the bytes are on the server, and a 404 says something
about publishing rather than about the verdict. The brand-new skill is still
tried first, so if publishing ever reaches it this fallback stops being used on
its own.

The evidence now states the gap in as many words instead of recording a red
step: a developer who submits a skill and pays for the audit cannot sell it
until an operator publishes it by hand. That is a product gap, and it is exactly
the path STE-30 is about to invite external developers down.

Also points the runner at api.sterish.xyz (STE-54), with the old hostname still
answering as an alias.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…id path

runs/2026-09-27T065414Z against api.sterish.xyz and the public dashboard.
16/16 transaction links resolve, all 16 successful.

What closed the three long-standing reds:

  step 3  PARTIAL x3 -> GREEN   the dashboard is deployed and public (STE-26)
  step 4  RED x3     -> GREEN   402, paid 0.1 USDC, licence minted, 200
  step 5  RED x3     -> GREEN   served from the held licence, no payment

STE-48's ownership proof ran for the first time: a bare X-AGENT-ADDRESS gets
401 OWNERSHIP_PROOF_REQUIRED, and the same call with a SEP-53 signature gets
200 with the balance unchanged. Run 2 called a borrowable licence the single
most important thing outstanding; it is closed.

Two corrections rather than quiet edits. Run 2 recorded "production still
answers 404 on /challenge" — that was my measurement, not the deploy: the route
is /use/{skill_id}/{version}/challenge and a bare /challenge 404s regardless.
And run 2 said I would fix steps 4-5 by publishing the new skill's artifact in
step 2; that plan cannot work.

New finding F5, filed as STE-56: a developer who pays for an audit cannot sell
their skill. The rehearsal's own skill is SAFE with a VERIFIED badge and still
404s on /use, because artifacts reach the host only through a manual command fed
from this repo's corpus and every API route is a GET. F5 stays open, and the
evidence row for step 4 says in as many words that the green is a workaround in
the test rather than a fix in the product.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdNuoJ9e954ojofNVGUezW
@vercel

vercel Bot commented Sep 27, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
sterish Ready Ready Preview Sep 27, 2026 6:59am UTC

@Lexirieru
Lexirieru merged commit 8cbec1f into main Sep 27, 2026
5 checks passed
@Lexirieru
Lexirieru deleted the feat/27-publish-artifact-in-rehearsal branch September 27, 2026 07:00

This branch was successfully deployed

1 active deployment
Preview — 75af23e5 Deployed Sep 27, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant