v1.16 — A feature fits in a day

From a field project where one feature (22 acceptance criteria, four fix rounds) took over a
working day. The root cause was size, magnified by repeated full runs and a dirty e2e database.
Text only — no script changes; every rule here is Documented.

• Size signal: the spec's self-critique recommends a split when a spec holds more than one user
  capability or more than ~8–10 criteria; keeping it as one is recorded in the spec. R-08: a new
  capability mid-work is a new spec, not a change.
• Data counts too (testing.md, Resource hygiene rule 5; ANEW-0007 revised): every run starts
  from a known state; what a run writes never reaches the next one. Bootstrap's inventory names it.
• A full green is evidence for its commit — run it once (Two speeds rule 6; ANEW-0009
  revised): review, verify and gates cite the existing CHECK GREEN (full, …) of the exact commit
  (handoff, or a CI run whose SHA equals HEAD) instead of re-running. Writing one test runs that
  test (rule 7): authoring, never evidence; repeat runs only for R-03.
• Fix rounds are counted: fix commits carry [plan NNNN/fix-k]; the review report opens with
  Review round n · fix rounds so far: k; a fourth round is not started before R-06.

Upgrading

• docs/testing.md: intent — Resource hygiene rule 5 (data); Two speeds rules 6 (cite, don't
  re-run, same SHA) and 7 (one test while writing it; repeats only in R-03).
• prompts/spec.md: intent — SIZE step in the self-critique. specs/TEMPLATE.md: the same hint
  in the Acceptance criteria comment.
• prompts/build.md: intent — rule 7 while iterating; [plan NNNN/fix-k] on fix commits, no
  fourth round before R-06. docs/git.md: the fix-commit format.
• prompts/review.md: intent — cite an existing full green for the commit; open with the round
  count; flag R-06 after round 3. prompts/verify.md: intent — cite, don't re-run.
• prompts/recovery/spec-change.md (R-08): intent — a new capability is a new spec.
• prompts/bootstrap.md: intent — the inventory names data left for the next run.
• workflows/feature-development.md, prompts/recovery/README.md, README: the matching lines.
• docs/decisions/anew/ANEW-0007, ANEW-0009: take as they are (each gains a Revised: line).
• .anew-version: v1.16.
• Full diff: v1.15...v1.16

v1.15 — Prompts close their gaps

A review of every prompt against the roles, workflows and testing rules. Text only — no script
changes; every rule added here is Documented.

• review.md: loads the approved plan too (the Tests dimension checks its criterion ↔ test map);
  new Plan conformance dimension — changed files outside the plan's list are findings (R-07);
  Correctness covers every User experience state and accessibility.
• build.md: an error from the current plan step is fixed inside BUILD; R-01 starts when the error
  comes from outside the step or survives two attempts. The ramp list adds R-03, R-10 and R-09.
• plan.md and specs/plans/TEMPLATE.md: the criterion ↔ test map has a Profile column
  (fast or full); a new test must be caught by its step's selector (testing.md, rule 4).
• spec.md, plan.md, adr.md, review.md, verify.md: write in the docs language from
  AGENTS.md; field labels stay English.
• spec.md: ask for a link to a fixed version of the design (or a dated export); a design that
  changes later is a spec change (R-08).
• clarify.md: at most ~7 questions, ordered by impact; the rest wait for the self-critique.
• verify.md: the evidence paragraph is a list; "a test that can't fail" is checked by reading
  the assert or by breaking the code in a temporary copy outside the repo, never committed.
• bootstrap.md: the testing interview asks, for UI work, the e2e tool and who owns the
  app-under-test lifecycle.
• R-09: the fresh session deletes the state file once it has taken over; nothing outlives it
  unless moved into the plan. R-03, R-08, R-12 end with ./scripts/check (full) green. R-05, R-06,
  R-09 and R-10 change no code (R-05's red reproduction is the point), so they carry no check exit.

Upgrading

• Documents (intent, apply in your own words if you translated them): each bullet above is the
  intent for its file; docs/commands.md — /review reads the plan.
• specs/plans/TEMPLATE.md: add the Profile column; existing plans need no change.
• .anew-version: v1.15.
• Full diff: v1.14...v1.15

v1.14 — Fast output is caught as evidence

From the same field report; it is also ANEW-0009's first revisit trigger.

• scripts/doctor warns (every mode, never fails) when a shipped spec, or in CI the pull request
  description, shows CHECK GREEN (fast, …) and no CHECK GREEN (full, …). The description is
  read from the event payload ($GITHUB_EVENT_PATH), no network; it is checked when the PR is
  opened, reopened or pushed to, not on a description-only edit. An untagged conf's
  (fast = full: …) run counts as full. The rule moves from Documented to Validated.
• R-02 (red test): many unrelated tests red at once → check the environment first (daemons,
  containers, ports); a crashed environment is not a code fault.
• scripts/selftest covers the warning for specs and PR descriptions, and that full output, or an
  untagged --fast run, is not reported.

Upgrading

• scripts/doctor (behavior): for each specs/done/*.md, and for $GITHUB_EVENT_PATH when it is
  a pull-request payload, warn if the text contains CHECK GREEN (fast, and not
  CHECK GREEN (full,.
• scripts/selftest (behavior): the doctor wrapper also unsets GITHUB_EVENT_PATH; new
  doctor_pr <payload> helper; the scenarios above.
• docs/testing.md: intent — Two speeds rule 2 names the doctor warning.
• prompts/recovery/red-test.md: intent — environment first when many unrelated tests are red.
• README: the "Evidence cites the full profile" row is Validated.
• .anew-version: v1.14.
• Full diff: v1.13...v1.14

v1.13 — Upgrades without collisions

From the same field report: upgrading a derived project meant guessing its base, renumbering the
template's ADRs around the project's own, and losing local script extensions to "copy".

• .anew-version records the ANEW version a workspace is based on; scripts/doctor prints it
  (ok ANEW base version: v1.13) and warns when it is missing. In the template, scripts/selftest
  checks it matches the newest CHANGELOG entry.
• The template's decisions move to docs/decisions/anew/ as ANEW-0001 … ANEW-0009 (each
  with a Revised: line; decisions unchanged), plus ANEW-0010 recording why. Project ADRs number
  from 0001 in docs/decisions/ with no collisions; prompts/adr.md counts only those.
• CHANGELOG opens with "How to upgrade": base from .anew-version, behavior lists for scripts,
  one-line intents for documents, a diff link per release.
• scripts/selftest checks usage output by its usage: prefix and exit code, not its full text,
  so a legitimate extension of a script's arguments is not reported as a difference.

Upgrading

• .anew-version (new): one line, v1.13.
• scripts/doctor (behavior): print ANEW base version: <first line of .anew-version> as ok;
  warn when the file is missing. Never fail on it.
• scripts/selftest (behavior): usage assertions match usage: only; fixture copies
  .anew-version; the version scenarios above.
• docs/decisions/: move the template decisions you have (ADR 0001–0009 as shipped by ANEW, if
  you kept them in your own range) to docs/decisions/anew/ under their ANEW-NNNN names, or take
  the folder from v1.13 and delete your copies. Your own ADRs stay where they are; one that had
  to be renumbered around a template ADR may keep its number. Add ANEW-0010.
• docs/decisions/README.md: intent — two owners, two places; the next project number counts
  only docs/decisions/.
• prompts/adr.md: intent — NNNN is the highest in docs/decisions/ + 1, anew/ excluded.
• README: the docs/decisions/ row names the two places.
• Full diff: v1.12...v1.13

v1.12 — Check tells you how long it took

From a field report (a two-stack project that adopted v1.7): the split between fast and full is
decided on step durations, and nothing in ANEW measured them.

• scripts/check prints <== <step> (n s) after every step (red ones too) and the total in the
  green line: CHECK GREEN (full, 14 steps, 1080 s). date +%s, no bc; where the clock gives
  no number, times are left out and the run still counts. The ==> <step> lines and the
  CHECK GREEN (<profile>, n steps / CHECK FAILED (<profile>) at step: prefixes are unchanged.
• docs/testing.md, Two speeds: "~a minute" is a per-step threshold, not a fast total; when fast
  grows too long, drop lint, then formatting, then unit tests. New rule 4: split a test step by a
  selector (category, trait, tag, namespace) and its inverse, never by a list — a test added
  later then still runs in one profile. New rule 5: decide on the median of at least 3 runs.
• prompts/bootstrap.md (3d) applies rules 4 and 5 at the gate.
• scripts/selftest checks that durations are present (never their value) and that a missing
  clock leaves them out.

Upgrading

scripts/check (behavior, for copies you extended): time each step with date +%s and print
<== <step> (n s) after it, red or green; append , <total> s inside the green line; print no
time when date +%s is not a number. Anything that parses the green line must accept the extra
field. docs/testing.md: Two speeds rule 3 (per-step threshold, removal order), new rules 4 and 5.
prompts/bootstrap.md 3d: selector split, median of 3. Full diff: v1.11...v1.12

v1.11 — Codex becomes a first-class adapter

• adapters/codex/: every command as a project-level Codex skill (skills/<name>/SKILL.md, the
  same pointers as the Cursor commands), invoked as $analyze, $plan 0003, … No pointer file:
  Codex reads AGENTS.md natively. Each skill is explicit only (agents/openai.yaml:
  allow_implicit_invocation: false), so Codex never starts a gated stage by itself. Review runs
  in a fresh Codex session; specs/done/ is backed by CI. The deprecated user-level custom
  prompts (~/.codex/prompts/) are not used.
• ./scripts/init codex installs the skills to .agents/skills/ (kept unless --force).
• scripts/doctor: adapter parity covers Codex (adapter parity: <x> missing in codex) and fails
  on a Codex skill that is not explicit-only; .agents/skills/ counts as an installed adapter.
  scripts/selftest covers both checks.
• README (quickstart, command sentences, a Codex column in "How strongly is each rule held?"),
  adapters/README.md and docs/commands.md list Codex next to the other tools.

Upgrading

Where Codex is used, run ./scripts/init codex and commit .agents/skills/. Template users copy
adapters/codex/ and the scripts/init, scripts/doctor and scripts/selftest changes.

v1.10 — The commands document themselves

• docs/commands.md: one page for all 17 commands (5 scripts, 12 slash commands) — command map
  with the four human gates, then for each: what it does, when to use it, reads / writes, stops
  at, lite / strict, refusal (verbatim), common mistake; a quick-reference table and the two
  golden rules. It points at the files that define the rules instead of restating them; the
  recovery catalog stays in prompts/recovery/README.md.
• README links it at the top ("New here?").
• scripts/doctor requires docs/commands.md like any core file and, where adapters/ exists,
  fails when a command or script has no section in it
  (command reference out of date: /<x> missing from docs/commands.md). scripts/selftest proves
  the guard fires and clears.
• No behavior changes.

Upgrading

Copy docs/commands.md, the README "New here?" line, scripts/doctor and scripts/selftest.
If you added your own commands, give each a ### `/<name>` section in docs/commands.md.

v1.9.1 — README stands on its own

• README drops the "Origin" section that pointed to the course: ANEW is now its own environment,
  separate from it. No change to the core.

Upgrading

Nothing to copy.

v1.9 — The residue test is one command

• scripts/residue runs the residue test (docs/testing.md, Resource hygiene): it measures processes
  and available memory, runs ./scripts/check twice (--fast for the fast profile), and names
  every survivor with its command line; memory over the tolerance (ANEW_RESIDUE_MB, 512 MB) is
  residue too. A process only counts if it is still alive at a second look. Unix-likes via ps,
  Windows (Git Bash) via PowerShell/CIM. A red check aborts it: the test needs a green check.
• docs/testing.md, prompts/bootstrap.md and workflows/bootstrap.md name the command, for both
  profiles; docs/testing.md adds: rerun it whenever scripts/check.conf changes.
• scripts/selftest covers it (usage, red check, --fast, a leaked process named, a waited-for
  child not reported). Its assertions look for the scenario's own marker, so a busy machine cannot
  make them flaky.

Upgrading

Copy scripts/residue (keep it executable) and scripts/selftest, then run ./scripts/residue and
./scripts/residue --fast against your scripts/check.conf.

v1.8 — ANEW proves its own scripts

• scripts/selftest: 50 scenario checks that scripts/check (profiles, tags, failures, arguments,
  stdin) and scripts/doctor (every gate, --strict, immutability, twin specs, mini-specs,
  evidence hygiene, adapter parity, missing core files) behave as documented. Each scenario runs
  in a throwaway copy of the ANEW core outside the repo; a project's own specs and code are never
  copied, so it behaves the same in the template and in bootstrapped projects. CI runs it after
  doctor.
• README catches up with v1.4–v1.7: the enforcement table gains evidence hygiene (Validated),
  resource hygiene and the UX section (Documented); --fast, evidence pointers and the selftest
  are described; the ADR range and the script count no longer go stale.
• docs/decisions/README.md: accepted ADRs may take editorial corrections that leave the decision
  unchanged, recorded as a Revised: line. ADR 0007 records its v1.6.1 wording change that way.

Upgrading

Copy scripts/selftest (keep it executable) and add its step to your CI's doctor job
(.github/workflows/check.yml). If you changed scripts/check or scripts/doctor locally, the
selftest tells you where your copy now differs from the documented behavior.