Warrant v0.1.5

Changes

• Guide approved work toward preventing errors at the owning type, interface, or module first, then effective adopter-owned checks, with concise prose and human decisions for remaining judgment. For reproducible defects, retained checks must be observed rejecting known-bad behavior and accepting valid behavior. Recurring feedback outside the Story becomes a proposal for the human, not an automatic next task. Applies to the skill, the AGENTS block, and the Story template. (#16)
• Add five behavior scenarios (22–26) covering these rules; the suite now has 26 scenarios. (#16)
• Check README translations section by section instead of by whole-document totals, with 17 regression tests. (#16)
• Document installing and updating Warrant with the Skills CLI in all three READMEs. (#15)

Upgrading

Update the plugin to v0.1.5. Plugin updates do not refresh a Warrant block already pasted into AGENTS.md. Update that block separately using the attached agents-block.md, preserving your repository's verification command and unrelated rules. Existing Stories are not automatically reapproved.

The release also includes agents.md for installation with other agents.

Verification and limitations

• Published from f9234f2f780c490aa499272a208b569cc10533f3; the verification workflow passed.
• The project site did not change in this release.
• The planned Claude behavioral comparison for the new guidance (Story 21 criteria 6–7) was not obtained because Claude authentication was unavailable. A supplemental single-run Codex comparison on the five new scenarios passed 5/5 with v0.1.5 versus 4/5 with v0.1.4; one run per scenario does not establish general improvement.
• Paid make eval was not run. Evaluation-fixture checks are static evidence, not measured agent compliance or proof of improved quality.

Full changelog: v0.1.4...v0.1.5

Warrant v0.1.4

Changes

• Distinguish acceptance checks observable in the current working tree from approved follow-up checks at later stages. Due or failed follow-ups cannot be reported as passing or reclassified after the fact. (#13)
• Add the structured Warrant-Result line with ordered reason codes and follow-up status, and include the core drafting and reporting safeguards in the standalone AGENTS block. (#13)
• Update the Story template, installation guides, translations, and evaluation fixtures for these rules. (#13)
• Refine the English, Traditional Chinese, and Japanese project site with stronger typography, an asymmetric hero, numbered rule rows, and accessible interaction states. (#14)

Upgrading

Update the plugin to v0.1.4. Plugin updates do not refresh a Warrant block already pasted into AGENTS.md. Update that block separately using the attached agents-block.md, preserving your repository's verification command and unrelated rules. Existing Stories are not automatically reapproved.

The release also includes agents.md for installation with other agents.

Verification and limitations

• Published from 926c5def67c1bb6f18a07061e4d6b9c996b38112; the verification workflow passed.
• GitHub Pages deployment passed.
• Site checks passed for three languages, light/dark themes, and 375/768/1440px widths in Chrome, including clipboard, navigation, keyboard focus, and reduced-motion behavior. Safari and Firefox were not tested.
• Paid make eval was not run. Evaluation-fixture checks are static evidence, not measured agent compliance or proof of improved quality.

Full changelog: v0.1.3...v0.1.4

Warrant v0.1.3

Changes

• Install prompt for any agent. The README now has a prompt you can paste into your AI agent; it reads docs/agents.md from the latest release and follows its "Instructions for agents". The verification command is always supplied by you, never chosen by the agent.
• Release assets. Every release now carries agents.md and agents-block.md, attached automatically by release.yml, so the prompt changes only when a release is published, not on every push to main.
• Other agents guide. docs/agents.md covers 12 agents (Claude Code and Codex tested; the rest from official docs).
• Plugin version 0.1.3; the skill itself is unchanged.

Warrant v0.1.2

One drafting rule added to the skill.

A file-set criterion exempts the Story file. When drafting a Story (Mode 1), any acceptance criterion that limits which files a change may touch must exclude the Story file itself. The Story ships in the same change as the work, so a criterion such as "git diff --stat origin/main lists only AGENTS.md" could never hold as written, leaving the implementer to either break it or reinterpret it, which rule 3 forbids. Found in the AgentPortV2 pilot.

Update with /plugin marketplace update warrant; a new session picks up the change.

Warrant v0.1.1

Adds a re-runnable behaviour eval suite and moves the plugin under plugin/. The skill's rules are unchanged.

Behaviour eval suite

• evals/cases/ holds 9 claude plugin eval cases: the spec's 8 scenarios, plus scenario 9, "an inference or a substitute check is not an observation". Every case has at least one deterministic grader that does not call the judge model. Fixtures copy agents-block.md from the plugin under test.
• make eval runs 3 runs per case, passes at 2 of 3, and is capped at --max-cost-usd 10. It costs money and needs credentials, so it is not part of CI or make verify.
• The released plugin scores 1.00 on all 9 cases. An adversarial copy that instructs the opposite of three rules fails cases 3 and 6.

Plugin layout

• The plugin now lives under plugin/, and the marketplace points at ./plugin, so evals/ and its scaffold scripts are never installed on an adopter's machine (ADR-0001). The install command is unchanged:

/plugin marketplace add CarlLee1983/Warrant
/plugin install warrant@warrant

Known limitation: cases 5 (stop on a scope conflict) and 9 cannot be shown to discriminate: even when told to violate those rules, the model refused, so a mutant could not make them fail. Details in #4.

Warrant v0.1.0

First release. Warrant bounds AI agent work to human-approved intent and proves completion with the repository's own verification. It ships only rules: a Claude Code skill, an AGENTS.md block, and a Story template. It has no scripts, CLI, or packages.

Three rules

1. Intent is approved by a human. A Story lives at specs/stories/<slug>.md, with Goal, Out of Scope, and Acceptance Criteria. It is approved only when a human commits it to the default branch, or when the human explicitly approves its current text in the session.
2. Completion is proven by evidence. Done is decided by the one verification command declared in AGENTS.md, and each acceptance criterion maps to a reproducible observation. An inference or a substitute check is not an observation.
3. The standard is not the agent's to change. The agent does not change requirements, weaken criteria, or widen scope. When a criterion conflicts with Out of Scope, the agent stops and reports it.

Install

/plugin marketplace add CarlLee1983/Warrant
/plugin install warrant@warrant

Pilot: AgentPortV2 was migrated from PraxisBound. All 8 behavioural scenarios hold. The one gap the pilot found, resolving a criterion conflict with a workaround, is fixed in this release. The pilot's migration PR, CarlLee1983/AgentPort#1, also carried one real Story drafted with the skill through draft → approve → implement → report, plus a second Story for a pre-existing Linux-only test failure that CI surfaced.

Successor to PraxisBound, which is now archived. Spec: #1.