v2.0.0: four phases (Intent, Plan, Build, Evidence) and one script

[cskwork]

Summary

v2.0.0 rebuilds sdlc-kit around four phases — Intent → Plan → Build → Evidence — driven by one script, sdlc.sh.

• Smaller. Prompts and templates go from about 23,700 words to 3,922. The scripts go from about 7,000 lines to 477.
• Explores before coding. Parallel scouts write refs/. One PLAN.md holds AS-IS → TO-BE, side effects, risk flags and the QA scenarios.
• Routes failures by cause, not by percentage. sdlc.sh score computes the route. It goes to Plan on a plan-caused or repeated failure. It delivers when every must scenario passes and at least 95% of all scenarios pass. It goes to the human when the scenarios that could not run are what blocks delivery. Otherwise it goes back to Build. Each kind of return is capped at three.

Why: on small tasks the old kit took 842–916 s against 85 s without it, at the same accuracy (PR #26). Local stores held 19 open features and 0 closed ones.

Details: log/changelog-v2.0.0.md.

Verification

• bash tests/selftest.sh: 95 checks PASS on macOS /bin/bash 3.2, with BSD awk, with LC_ALL=C, and with CI=1, which turns symlink skips into failures. The checks cover:
  • prompt budgets;
  • gates and stops;
  • the routing boundaries (95.0% delivers, 94.x builds, a plan cause goes to Plan, a must BLOCKED goes to the human) and the caps;
  • scenario history;
  • review binding and delivery authority;
  • legacy stores;
  • a four-phase lifecycle run on a CLI fixture.
• Each regression check written for a review finding was first run against the pre-fix script, and it failed there.
• Legacy store: on a copy of a real 12-feature v0 store, status takes 0.2 s (v0 took over 30 s). It reports every feature as legacy, and the original store is unchanged.
• Real-agent E2E: a fresh agent took one request through all four phases on a fixture with two hidden rules.
  • It found both rules during Intent.
  • It stopped for the human at intent and at the plan, where the public-API flag was set.
  • It scored 14/14, including four scenarios that QA added.
  • It shipped one local commit.
  • Its feedback led to six kit fixes.
• Adversarial review: three rounds, each by a fresh reviewer, and every round returned FAIL. All blocking findings were fixed.
  • The round-3 blocker: score exited silently with code 128 when the project root was not a git repo. A new check covers it and fails on the old script.
• Ultra review (/code-review ultra) found 10 issues. It ran on the v2 content only: 24 files, 3,896 lines. The 54 pure deletions of v0 files were left out to fit the 8,000-line limit.
  • All 10 are fixed in 4ee1186.
  • The new ultra regression group has 8 checks. 7 of them fail on the previous script. The 8th, for the Git Bash // root, cannot be reproduced on macOS.
• CI on Linux, macOS and Windows: see the checks on this PR.

Breaking change

• AGENTS.md, init.sh, gates/, tools/ and the v0 templates are removed.
• v0 features show as legacy. Finish them with v0.27.0, or close them with sdlc.sh close <slug> dropped "<reason>".
• lazymode and .sdlc/memory are still read.

Merging to main updates every clone that follows main.