Repository navigation
Releases
v3.1 — hardened by the first live adoption
The first release since v3: gate records and cross-checks, optional Jev checkpoints and a dashboard (Sep 27–28), then the fixes from the framework's first run on a real product it was not written for — an iOS app and its backend (Sep 29). Each of those live-run fixes came from an overnight run and ships with a test that fails on the old code.
Highlights
scripts/loop-kit.sh install | update | status— install the framework into a product repo and update it later without overwriting your changes.- The driver trusts gate records over the round's own report, reads gates run in the background, never counts a usage limit as a failure, and passes a trajectory REDIRECT's direction to the next round.
- Optional Jev checkpoints that can only block or trigger an LLM gate, never approve — with a shadow report that measures whether they are worth it.
- Evidence a check can read: per-test results from the iOS adapter, and a validator that must show the test covering its claim.
- 17 hard lessons, 11–17 from the first live adoption: docs/06-lessons.md.
Upgrading an adopted repo: from a checkout of this release, scripts/loop-kit.sh status <repo> then scripts/loop-kit.sh update <repo>. New config keys are listed, not written. Then bash scripts/test-driver.sh.
Full changelog: v3...v3.1
2026-09-29 — hardened by the first live adoption
The framework ran on its first external product — an iOS app and its backend, one framework in two repos. Each item below was found by a live run and fixed the same day, with a test that fails on the old code (lessons 11–17).
Packaging
scripts/loop-kit.sh install | update | status— install the framework into a product repo and update it later. A lock file records what was installed: unchanged files are updated, files you changed are never overwritten (the new version waits as.kit-new), executable bits are restored, your own files are never touched, new config keys are listed. Tested byscripts/test-kit.shin CI.
Driver
- A Claude usage/session limit is never counted as a NOOP. It stops with its own reason, or waits and re-runs the same call (
LIMIT_WAIT,LIMIT_MAX_WAIT).80bb83d - Gate agents started with
run_in_backgroundare recorded from their task notification; the spec and round prompt require foreground gates. Before: validated ships counted as NOOPs and stopped a run.ec26966 - A trajectory
REDIRECTpasses its whole line to the next round ("go here") instead of the generic RESET note ("not here").74db75a - A timed-out round is labelled by how far it got —
timeout mid-fix,timeout before validation— inloop.logand the dashboard.1fc315f
Evidence and checks
- iOS adapter:
IOS_TEST=1runs the tests and writes one line per test to/tmp/loop-tests.txt;--list-tests <xcresult>for custom adapters. The validator must show the test that covers the CLAIM, not a total.3f12806 - Jev
claim-evidenceacceptsBUILD_CMDby its script, not its output path, or any command inCHECK_CMDS.3f12806
Jev data integrity
- Every
jev.mjsoutcome is logged,UNAVAILABLEand timeouts included;shadow-reportshows per-task availability and latency.30a6b28 - The Jev test suite writes to a temp log, never the real
.loop/jev.jsonl;shadow-reportcounts only rows from real loop runs.c8a8a18
Tests
test-driver.shbrings its own config, so an adopter'sloop.config.envcannot leak into the scenarios.2054581- The dashboard test skips the web-v2 replay when
examples/is absent (as in every adopting repo).80bb83d
2026-09-28 — records, cross-checks, Jev checkpoints
- Gate records: every gate's full report, verdict and evidence to
.loop/gates.jsonl, across runs.df7eb61 - The driver trusts the records over the round's self-report: under-reported rejects are corrected, a SHIPPED round without a validator PASS is a NOOP; protected paths park the run; timeouts kill the whole process group; the pre-push hook is required.
02a3140390381e(gate agent names configurable, only the loop's own scripts protected) - Jev checkpoints:
claim-evidence(a fresh validator re-validates a PASS its own evidence does not back)ef936c4;same-failure(spec Step 7b) andnoop-cause(early audit)7559578; shadow calls paired with value-critic verdicts7010a5a. - CI: driver, dashboard and Jev tests plus shellcheck on every push and PR.
f213159 - Security:
@claudelimited to repo members; dashboard Host allow-list.8cdaec3 - Docs in English, README rewritten as the entry point, dashboard in English.
ccebdcf1a7d485