Conversation
Add jev/, a TypeScript package run directly by Node 24 that asks the typed evaluation model a fixed set of questions about an issue or a pull request through the Vercel AI Gateway and returns calibrated decisions. Every question and threshold lives in src/policy.ts; fields are written only at confidence 0.90 or above, area labels at 0.85, and everything below is left for the agent. Tests replay recorded fixtures and never touch the network. Add actions/jev-decide, which runs the classifier from a private prefix, no-ops with a notice when AI_GATEWAY_API_KEY is empty, and never fails the job on a gateway error. Add actions/set-issue-fields, the one implementation of repo-scoped issue-field, type and label writes, with a regression test against a fake gh. No workflow calls either action yet, so this release changes nothing for callers. Self test covers the package tests, the no-key path, the pinned thresholds, synthetic fixtures and the field-writer. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A later step in the same job reads a step output; the outputs block is evaluated with the action's own context and is the wrong place to name a path the agent will call. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This was referenced Sep 18, 2026
…|| C as if-else Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
The first of three releases that put a typed, probabilistic classifier in front of the agent workflows.
jev/: a TypeScript package Node 24 runs directly (no build). It asks the evaluation model a fixed set ofchoice,scoreandbooleanquestions about an issue or a pull request through the Vercel AI Gateway and turns the calibrated answers into decisions. Every question, option description and threshold lives injev/src/policy.ts. Fields are written only at confidence 0.90 or above, area labels at 0.85, prompt-injection and spam routes at 0.90, a trivial-diff skip at 0.95. Below threshold, the field stays empty and the agent decides, as today.actions/jev-decide: runs the classifier from a private prefix underRUNNER_TEMPand exposes its decisions as outputs. An emptyAI_GATEWAY_API_KEYormode: offis a::noticeandran=false. A gateway error is a::warningandran=false. It never fails the job.actions/set-issue-fields: the one implementation of repo-scoped issue-field, type and label writes, with a regression test against a fakegh. The org endpoint is refused, never offered as a fallback.No workflow calls either action yet, so merging this release changes nothing for callers. Triage wiring and review wiring follow as separate PRs.
Why
Nothing in the pipeline parses agent output today. Fields and labels are side effects a prompt asks for and bash asserts afterwards, with no confidence anywhere. This gives the job a typed answer with a number attached, so "leave unset rather than guess" becomes a threshold in one file instead of a sentence in a prompt, and every decision is recorded for later calibration.
Verification
Self testgains: package tests on recorded fixtures (no key, no network), the no-key path ofjev-decide, pinned policy floors, a synthetic-fixtures check, and the field-writer test.tsc, shellcheck, actionlint and YAML parse all pass.jev/scripts/record.tsre-records the fixtures once a key exists, and the pilot rollout in the next PR is where thresholds get checked against real issues.🤖 Generated with Claude Code