agent-pipeline

A reusable, autonomous issue-to-PR pipeline for GitHub Actions: an issue becomes a draft PR, CI failures are fixed in a loop, a multi-lens review panel gates the result, and a clean PR is promoted to ready-for-review. Humans still approve and merge — nothing here can merge a pull request.

Status: not yet consumable. This repo is mid-extraction. The scripts are here and their tests pass, but the workflows are still wafflebase-shaped copies and are not yet callable via workflow_call. Do not point a repo at this yet. The adoption kit lands with the first v1.0.0 tag.

Layout

Path	What
packages/pipeline/	The pipeline: review-panel orchestrator, rebuttal adjudication, round guards, state machine, metrics ledger. A standalone npm package, not part of any workspace.
packages/pipeline/lenses/	The review lens manifest (lenses.json) and one rubric per lens.
.github/workflows/agent-*.yml	Verbatim copies from wafflebase, kept so the script↔workflow mirror tests keep working. Inert here — every entry job gates on vars.AGENT_PIPELINE_ENABLED, which is unset. Converted to workflow_call later in the extraction.
.github/workflows/tests.yml	This repo's own CI. Deliberately not named "CI".

Running the tests

npm install          # root only. Installs the linter and points git at .githooks
npm run verify:self  # lint + structural invariants + the suite

Or the suite on its own, with no install at all:

cd packages/pipeline
node --test-timeout=60000 --test '**/*.test.mjs'

Do not npm install first. The suite is designed to pass with node_modules absent — third-party dependencies (the Agent SDK, zod) are reachable only through await import() inside the function that needs them, and running without them installed is what enforces that. A module that statically imported the SDK would take the whole suite down, which is the intended signal.

Note the recursive glob. A flat *.test.mjs silently matches nothing in subdirectories, which is how wafflebase lost a whole tier of suites for a while.

The two dependencies are exact-pinned. The SDK's option names are verified against that exact version, so a bump needs a smoke test and an end-to-end run.

Design notes worth knowing before changing anything

• The reviewer never runs branch code. Review and fix jobs execute pipeline scripts from a trusted checkout, never from the pull request under review, so a PR cannot alter its own review. Preserving that across the repo boundary is the central problem of this extraction.
• Check runs are the gate; labels are decoration. agent:<state> labels are forgeable by the author agent, so nothing gates on them — see the header of packages/pipeline/set-state.mjs.
• Fail closed. Lens checks that cannot complete are failed, not skipped; round and attempt caps page a human via latch comments rather than retrying.
• Markers are load-bearing. Model output that will be embedded in a comment is neutralised first, because a planted <!-- agent-review-paged --> inside a finding summary would forge a latch.
• Five of the moved scripts have no test suite of their own: disclosure, gh-checks, mark-ready, review-round-guard, summarize-ci. (auth-smoke was the sixth and now has nine.) mark-ready.mjs is the notable gap — its exit codes 1/2/3 are a contract the panel workflow branches on, and nothing pins them.