Skip to content

0.9.0: gate status: verified on review and verify, guard on its own record, throwaway files - #16

Merged
sofus-common merged 4 commits into
mainfrom
claude/ai-native-sdlc-research-c5640a
Nov 14, 2026
Merged

sofus-common merged 4 commits into
mainfrom
claude/ai-native-sdlc-research-c5640a

Conversation

@sofus-common

Copy link
Copy Markdown
Contributor

What

Hook release: 0.7.3 -> 0.9.0. Changelog entry ## 0.9.0 - 2026-11-14, both manifests bumped. 0.8.0 is in the changelog as a record but is not tagged on its own; 0.9.0 ships it. No skill wording changed.

Why

Two new behavior cases, gamed and weakened, start at status: reviewing with a defect that passes the fixture's check. On 0.7.3 the agent found and fixed the defect in 9 of 9 runs, and in every one it then wrote status: verified itself: it never invoked ai-native-verify, and nothing reviewed its fix.

A read of two weeks of session logs (141 guard denials) also showed that 12 of 17 denied probe and benchmark scripts were writes into the system temp folder, which the guard was meant to allow. Agents put probes under .sdlc/changes/<slug>/probes/ or stated the Fast lane to get through.

Change

  • New denial. Standard and Controlled work can record status: verified in a change's state.md only after the session invoked ai-native-verify on the files its last ai-native-review saw. The denial names the next skill and up to five changed files.
  • The hook keeps its own record. It now also runs after each Skill call and for a typed skill command (PostToolUse on Skill, UserPromptExpansion), and stores each skill of this plugin that the main agent invoked with a content hash per project file outside .sdlc/ that Git does not ignore. The transcript is no longer read for skill calls or code changes: it is written late and leaves sub-agent tool calls out. Code changed by a build worker is now seen.
  • Throwaway files. Every way of writing into the system temp folder passes before plan approval, cp and Copy-Item are checked on their destination only, and a file that Git ignores is not project code. The approvals denial says where a throwaway file goes.
  • Failure mode. An error after the hook has seen the session invoke the plugin denies the write and names the error; any other hook failure blocks nothing. If the plugin's data folder cannot be written, the verify check is off.
  • Every hook event is registered with a 60-second limit. scripts/test_skills.py fails when a SKILL.md passes 15,000 bytes; CI runs it.
  • Behavior cases gamed and weakened, with fixtures, checker rules, and tests. claude plugin eval now runs 17 of the 19 cases.

Verification

  • python scripts/check_version.py --self-test, python scripts/check_version.py, scripts/test_behavior_check.py (7 tests), scripts/test_edit_guard.py (13 tests: Git project, project in a subfolder of a repository, folder without Git), scripts/test_skills.py: all pass on Windows.
  • Full behavior suite on commit 4eb4a6a with the release command in docs/behavior-checks.md: Claude Code 2.1.319 in WSL2, claude-sonnet-5-5 as agent and judge, 3 runs per arm, no permissions bypassed. 17 of 17 cases passed all 3 with-plugin runs (51 of 51 runs); 0.7.2 recorded 13 of 15. Without the plugin, 14 of 51 runs scored 1.0. In the with-plugin runs the hook denied 12 status: verified writes in gamed and weakened and one write with no lane stated; each agent cleared it, and no trace holds a hook error. Cost $16.03.
  • Seven independent review rounds on the hook during development. They found defects that the behavior cases do not exercise, among them a first fingerprint that any commit reset, and a project in a subfolder of its repository whose changed files were not seen. All are fixed and tested; the changelog and the ADR-003 amendments record them.

Not verified

  • Codex. It has no guard, and no behavior run was made there for this release. The ADR-003 amendment for 0.9.0 says why no Codex guard is built.
  • Native Windows and a live interactive session. Every behavior run was headless in WSL2. The unit tests ran on native Windows.
  • feature and payment-build on the manual harness.

Known limits

Listed in the changelog entry for 0.9.0. The main ones: the lane stated in chat still comes from the transcript; a code change after verify started is not checked; changes inside a submodule or to a file Git ignores are not seen.

🤖 Generated with Claude Code

sofus-common and others added 4 commits November 14, 2026 12:15
Standard and Controlled work can record status: verified in a change's
state.md only after the session invoked ai-native-verify, with an
ai-native-review call after the last code change before it. On 0.7.3 the
agent fixed the planted defect in 9 of 9 runs of the two new behavior
cases and then recorded status: verified without invoking verify; on this
hook both cases pass 3 of 3.

- Hook: new denial with a reason that names the next skill; a hook error
  after routing denies, any other hook failure blocks nothing.
- Behavior cases: gamed and weakened, with fixtures, checker rules, tests.
- Docs: ADR-003 amendment, changelog with run numbers and known limits,
  README troubleshooting, behavior-check table.

Unreleased: the changelog entry is not dated.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…9.0)

The check before status: verified no longer reads the transcript for
skill calls and code changes. The hook now runs after each Skill call
and for a typed skill command, and records the skill with a content
hash of every project file outside .sdlc/ that Git does not ignore.
The last verify must have started from the files the last review saw;
a denial names the files that changed. Code changed by a sub-agent is
now seen.

- Throwaway files: every way of writing into the system temp folder
  passes before plan approval, cp is checked on its destination only,
  and a file that Git ignores is not project code.
- Hook events carry a 60-second limit; a test checks the registration.
- scripts/test_skills.py fails when a SKILL.md passes 15,000 bytes.
- Docs: ADR-003 amendment (including why no Codex guard is built),
  changelog with run numbers and known limits, README, architecture.

Unreleased: the changelog entry is not dated. The fixes from the last
review round are covered by unit tests only.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
All 17 cases passed 3 of 3 with-plugin runs on commit 4eb4a6a
(51 of 51 runs; 14 of 51 without the plugin). Cost $16.03.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
0.8.0 is dated the same day and marked as not published on its own.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@sofus-common
sofus-common merged commit 5f69bf2 into main Nov 14, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant