v1.7.0 — Build with tests, respond with evidence

Use the workflow while building and responding to feedback. v1.7.0 adds two everyday development skills and connects them to the existing task flow.

• Implementing with tests: choose a check of the requested behavior, observe a meaningful baseline, implement the correction and verify relevant regressions. Missing fixtures are not proof of the bug; already-correct behavior and prose edits need no manufactured failure.
• Responding to review: verify each comment against current code and requirements, fix supported defects, explain stale or disputed suggestions, and keep external replies within existing authorization.
• Keep the task: existing skills now shape rough requests from project evidence, preserve unfinished work when newer messages steer it, and retain failed attempts with a reason to retry.
• Try a real task: the review-feedback exercise has runnable code, a contract, narrow passing tests and mixed review comments. Daily-use requests show how to start.

The package now includes 15 skills, with four added synthetic decision cases (22 total). Existing project copies can use the conservative update path in the installation guide; native users can follow plugin updates. Core18 and the approved artwork are unchanged.

Two fresh Codex sessions exercised implementation and review-only paths. The implementation probe used an initial skill candidate; the revised final skill bytes received the review-only probe. Observed behavior and exact limits. These are bounded author observations, not comparative effectiveness results.

Validation includes 211 tool tests (9 local Windows privilege skips), package/catalog/document checks, 3 narrow public-fixture tests, an independent original-code regression control, native observations and independent review. Windows/Linux Python3.10 CI passed before merge.

v1.6.0 — Bring a receipt to your next PR

Bring a receipt to your next PR. v1.6.0 turns a local command receipt into a Markdown evidence summary.

• Shareable reports: tools/report_receipt.py separates historical command outcome from current selected-file, log and Git HEAD consistency. Failures stay nonzero. Default output omits arguments, raw logs, paths, filenames and freeform diagnostics; optional argument disclosure is escaped and bounded.
• Your existing project: the quickstart shows how to capture your real check and produce the summary from a separate toolkit clone.
• Try it and explain it: updated English/Turkish README entry points, a launch kit, and use reports that distinguish synthetic demonstrations, author-run sessions and independent reports.
• Catalog navigation: skills.sh metadata groups all 13 skills by task. Display refresh depends on the directory; no listing/ranking outcome is claimed.

The existing 18 rules, 13 skill bodies, recorded false-green demo and stamp artwork are preserved. Give your agent the repository link and installation request to install the workflow and get a plain explanation of what changed. Optional Python tools still require a toolkit clone.

Validation: 211 local tests, 202 executed successfully with 9 Windows symlink-privilege skips; structure/package checks; separate-project successful, stale and failing-test controls; GitHub Markdown rendering; independent review. Windows/Linux Python 3.10 CI passed before merge.

Receipts remain editable, non-atomic observations with selected scope. A match is not proof that the task is complete, that a command's tests were adequate, or that evidence is authentic.

v1.5.0 — Give your agent the link

Give your coding agent the repository link and ask it to install, then get a simple explanation of what changed and how to begin.

• Both READMEs now lead with that request. Root INSTALL.md guides supported Codex/Claude hosts through scope selection, existing-install checks, installation, readback and a short welcome in the user's language.
• The copier prints what was added, the selected capabilities and a next step. Optional --json reports distinguish planned work, completed copies/updates, failures and retained paths; they do not claim runtime activation.
• Existing project instructions, custom skills and unrelated settings are preserved. Claude defaults to local scope. Host permission boundaries are reported and use normal approval; no permission bypass or paid fallback is added.
• Ordinary inherited-permission staging replaces restrictive temporary staging. Cleanup removes only verified staged material and retains unexpected content or cleanup failures.

Validation includes 191 repository tests, independent review, a verified 13-skill authorized installation, and a real Codex read-only reuse/welcome case producing a concise Turkish explanation. The native install failure is kept separate; the welcome fixture was already installed, and those prepared-checkout tests did not exercise HTTP cloning or establish universal sandbox write access.

Agent installation guide · Copier and JSON reports

v1.4.0 — Real native sessions, independent acceptance

Test the installed workflow on a real coding task, then check what it actually delivered.

• New tools/native_smoke.py create/verify builds a small broken CSV project and independently checks eight behavior cases plus protected input identity. Two green starter tests deliberately do not establish acceptance.
• Malformed results, crashes, timeouts and changed inputs stay nonpassing. An exclusive project result directory works with the tested Windows Codex sandbox; cleanup removes only known output and retains unexpected contents.
• Real Codex sessions loaded the installed plugin, fixed median and CSV fixtures, and passed independent acceptance. A fresh native session ran the exact corrected CSV verifier successfully without relaxing sandbox permissions.
• Real Claude installation, discovery and native slash expansion were observed. Weekly subscription quota blocked model execution; Claude task repair remains unverified.
• Native session instructions, observed limits and bilingual onboarding now link to the repeatable exercise. Corrected the stale skills count in CLAIMS.md.

The 13 skill files, original 18 rules and approved stamp artwork are unchanged. Initial native observations used v1.3.0's identical skill payload. After the main merge, Codex was updated to v1.4.0 and a fresh native session read its actual 1.4.0 cache entry and passed the exact eight-case verifier with scratch removed. Claude's local installation was also updated to v1.4.0; its 13 cached skill bodies matched, but no quota-blocked model run was retried. Existing global skills also participated, so these observations do not isolate this package's causal effect. The tool does not launch models, install plugins, access credentials or enable paid services.

Run your own native test · Native installation

v1.3.0 — One workflow, thirteen skills

Done is a claim now connects acceptance, planning, implementation, diagnosis, review and delivery in one development workflow.

• One entry point: using-done-is-a-claim chooses a proportionate path. Tiny changes stay small; substantial work gets a plan and a review/fix/recheck loop.
• Five new skills: entry, planning-changes, executing-plans, debugging-with-evidence and reviewing-changes. Thirteen standalone skills in total; no mandatory subagents or Grit services.
• Native packages: Claude Code and Codex plugin manifests and repository marketplaces. Installation and scope.
• Project installation: --list, composable workflow/evidence/review profiles, and explicit --update that refuses local edits and retains replaced copies as backups. Updates are non-atomic; provenance hashes are unsigned local records.
• Practical guidance: a workflow diagram, change/diagnosis/review templates, bilingual onboarding and six new synthetic decision scenarios.

The original 18 rules and stamped cover remain. Source material inspired the organization; all new skills and implementation are original.

Validation includes Windows/Linux Python tool tests, package/repository checks, canonical skill metadata checks, independent code review, native Claude package validation and a bounded fresh-context scenario exercise. Claude reports the documented warning that plugin-root CLAUDE.md is not adopted as project context. Structural validation does not establish native installation, live skill activation or improved agent outcomes.

Start the workflow · Project copier · Source decisions

v1.2.0 — Recheck receipts, resume work, verify delivery

Recheck saved command evidence before reusing it, resume unfinished work from the actual saved boundary, and inspect what was delivered at the destination.

• Receipt rechecks: new check_receipt.py compares the closed log, selected current files and applicable recorded Git HEAD against an explicitly chosen project. It never executes recorded commands or follows historical absolute paths.
• Bound log bytes: new receipts include additive log_sha256 in schema v1. Older receipts remain readable; an unbound log stays incomplete. Historical command failure and input mutation are not upgraded by matching bytes.
• Eight focused skills: add resuming-work and checking-delivery, adapted from Grit runbooks and static implementation/test lessons in original portable wording.
• Delivery exercise: real temporary copies demonstrate a stale same-named file, a correct positive control, and identical bytes that still fail JSON parsing.
• Twelve scenario packets: E11/E12 cover acknowledgement correlation and saved resume boundaries, with separate evaluator rubrics. Handoff and completion templates now include the corresponding observations.

The existing receipt producer, safe installer and 18-rule core remain available. Tools use Python 3.10+ and its standard library. The approved stamped cover is unchanged.

Windows and Linux checks cover the implementation; platform-specific cases run on their applicable host. Independent code and documentation reviews found no remaining actionable defect. These are editable, scoped, non-atomic evidence comparisons—not authentication, a signed attestation, in-toto/SLSA conformance, an agent-effectiveness benchmark or proof that a person read a delivery.

Recheck guide · Delivery and receipt recipes · Sources and limits

v1.1.0 — Local receipts and safe skill installation

Install selected skills without overwriting project customizations, and capture the evidence behind a command result with two optional Python tools.

• Safe project installation: preview destinations, refuse collisions and links, preserve existing instructions/settings, and copy license/provenance with each selected skill.
• Local command receipts: save the full output, actual child exit, Git observations and selected input hashes. Changed inputs and lost observations remain distinct from command success.
• Six focused skills: add acceptance design, worker-result collection and evidence freshness; refine failure attribution and public-claim handling.
• Practical evaluation: ten participant packets with separate evaluator rubrics, plus a runnable false-green example and reusable acceptance, completion and handoff templates.
• Bilingual onboarding: English and Turkish READMEs with the approved animated receipt stamp, installation routes, recipes and a use-report form.

The core rules and skills remain plain Markdown. The installer and receipt tools need Python 3.10+ and its standard library; optional artwork renderers have separate dependencies.

Validation runs on Windows and Ubuntu in GitHub Actions. Platform-specific cases are skipped on the other operating system. These checks establish tool behavior and repository structure, not an agent-effectiveness benchmark. Receipts are editable local records and selected-file snapshots, not signed attestations or atomic isolation.

Start with installation or capture a receipt.

v1.0.0 — 18 rules and 3 skills

First release.

• AGENTS.md: 18 rules for coding agents, grouped as before you start, before you say done, when you report, when you write tests and fixes, and never.
• INCIDENTS.md: the failure behind every rule, from running Claude Code and Codex agents on our own repository in September 2026.
• CLAUDE.md: one line, @AGENTS.md, so Claude Code imports the rules.
• skills/: reading-measurements, whose-red, public-claims.

Propose a rule with the incident behind it: Issues → Propose a rule.