Conversation
Two advisories published upstream since the last merge (browserslist unbounded memory growth, HIGH; postcss-selector-parser AST recursion, low) trip CI's npm-audit gate on every new push, blocking the whole open PR train. npm audit fix; 0 vulnerabilities after; tests/tsc/build green. Lockfile-only.
The June n8n refresh flow never got a host (Jetson n8n never happened). This gives it one that already exists: a weekly GitHub Actions run of Claude Code headless against a versioned research prompt (ops/prompts/gpu-price-refresh.md - two-pass: research each accelerator across the cited provider price lists, then adversarially re-check your own changes), allowed to edit exactly one file. The point is the gate, not the agent. server/gpu-price-validation.ts is a pure, unit-tested contract the run must pass before its output may even become a PR: fixed model set, every price inside its own observed low..high band, >=2 https sources per model, estimated flags never shrink, history anchors append-only (the past is immutable), and no move over 60% (blocked for human review, never averaged away). The workflow additionally rejects any run that touches other files, runs the full test suite, and only ever opens a PR - it has no path to main. Improvement over the old n8n design, which committed straight to main. Verified both ways: the shipped dataset validates against itself, and a simulated hallucinated refresh (3x price, rewritten anchor, single http source) is rejected with all five violations named, exit 1. Activation: one repo secret - CLAUDE_CODE_OAUTH_TOKEN (from `claude setup-token`, rides the Claude subscription) or ANTHROPIC_API_KEY. Until set, the run fails with a message saying exactly that. 492 tests (8 new), tsc, build green.
Owner
Author
|
Closing per Jack's call — not the direction he wants (and it needs a token). The validator + prompt work stays on the branch if ever wanted. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The answer to "what can an agent product do that this repo can't": nothing — because now the repo has its own. Weekly, GitHub Actions runs Claude Code headless against a versioned research prompt (
ops/prompts/gpu-price-refresh.md) that re-researches all 10 tracked accelerators across their cited provider price lists, two-pass (research, then adversarial self-check), allowed to edit exactly one file.The point is the gate, not the agent.
server/gpu-price-validation.tsis a pure, unit-tested contract the run must satisfy before its output may even become a PR:estimatedflags never silently shrinkThe workflow also rejects any run touching other files, runs the full test suite, and only ever opens a PR — it has no path to main (an improvement over the original June n8n design, which committed straight to main; that flow never got a host anyway since Jetson n8n was never set up).
Verified both ways in this build: the shipped dataset validates against itself, and a simulated hallucinated refresh (3x price, rewritten history anchor, single non-https source) was rejected with all five violations named and exit 1.
Activation (one secret)
Run
claude setup-tokenlocally and add the output as repo secretCLAUDE_CODE_OAUTH_TOKEN(rides your Claude subscription — no API bill), or setANTHROPIC_API_KEYfor metered billing. Until one exists, the Monday run fails with a message saying exactly that.workflow_dispatchenabled for a manual first run.Pattern, not one-off
The prompt + gate + PR-only pipeline is the template: deals and clusters refreshes are each "a prompt file and a validator" away, reusing this workflow's shape.
492/492 tests (8 new gate tests), tsc clean, build green. Carries the #30 lockfile fix. Independent of #25-#31 — merge in any order.
🤖 Generated with Claude Code
https://claude.ai/code/session_01FY4ziPhfu8VBPFVy5cDMNy