Skip to content

Releases

v3.1 — hardened by the first live adoption

Choose a tag to compare

@InjayTseng InjayTseng released this 29 Sep 09:44

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 by scripts/test-kit.sh in 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_background are 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 REDIRECT passes 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 — in loop.log and the dashboard. 1fc315f

Evidence and checks

  • iOS adapter: IOS_TEST=1 runs 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-evidence accepts BUILD_CMD by its script, not its output path, or any command in CHECK_CMDS. 3f12806

Jev data integrity

  • Every jev.mjs outcome is logged, UNAVAILABLE and timeouts included; shadow-report shows per-task availability and latency. 30a6b28
  • The Jev test suite writes to a temp log, never the real .loop/jev.jsonl; shadow-report counts only rows from real loop runs. c8a8a18

Tests

  • test-driver.sh brings its own config, so an adopter's loop.config.env cannot 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. 02a3140 390381e (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) and noop-cause (early audit) 7559578; shadow calls paired with value-critic verdicts 7010a5a.
  • CI: driver, dashboard and Jev tests plus shellcheck on every push and PR. f213159
  • Security: @claude limited to repo members; dashboard Host allow-list. 8cdaec3
  • Docs in English, README rewritten as the entry point, dashboard in English. ccebdcf 1a7d485

2026-09-27 — Jev pre-checks and the dashboard

  • Optional Jev pre-checks in front of the LLM gates — prefilter, label-promise, pick, same-tactic — that can only reject or trigger, never approve; off → shadow → prefilter with an offline eval. 307f408, self-match fixes c59fdf9 2fb8f4e
  • A local, read-only developer dashboard. 8da738c dc208b5