wdi-method 0.6.15 → 0.6.19

package/CHANGELOG.md

@@ -0,0 +1,244 @@
+# Changelog
+
+What changed in `wdi-method`, newest first. Every entry says what it means for a repo that already has
+the method installed — most often nothing beyond `npx wdi-method@latest update`.
+
+A version here is a **git tag**. `npm publish` is a separate, deliberate step (see
+[`CONTRIBUTING.md`](CONTRIBUTING.md)), so some versions exist as a tag and a GitHub release without ever
+having been served from npm. The registry always holds the newest published version, and a published
+version contains every fix below it.
+
+---
+
+## 0.6.19 — 2026-09-10
+
+- **One autopilot run now costs one cloud CI run, not dozens.** A run over fifteen tickets pushed often
+  enough to start GitHub Actions at every ticket commit and every spec close, and spent most of a month's
+  Actions allowance in two days — a Windows runner bills at 2x the minutes and macOS at 10x, and on a
+  private repository every one of those comes out of the allowance. **The fix is not fewer commits**: they
+  stay granular, one per ticket, plus the memlog and the registry catch-up. What changed is what a push
+  triggers.
+  - `wdi-autopilot` gains **§ Cycle-end CI**. Every intermediate push starts nothing; the cloud runner
+    fires **once**, at § Finish, when the final head is pushed and the one PR is marked ready for review.
+    § Door 2 says an iteration MUST NOT dispatch a workflow or mark a PR ready mid-ticket, and § Finish
+    says how the run is triggered at the end — ready-for-review, `workflow_dispatch`, or a final push with
+    no `[skip ci]`, in that order of preference.
+  - What a spec close is judged on until then is the run branch's **own full suite, run locally** — the
+    check that used to be "the last pushed head is green in CI". Nothing closes over red; only the place
+    the red is seen has moved to where it costs nothing.
+  - Preflight gains a row: **no workflow may fire on an intermediate push.** It is red when one does, and
+    it is checked at minute one rather than discovered on the invoice.
+  - `wdi-build` Step 5 says the same thing for a supervised run: one unit of work earns one cloud run.
+- **New guide — `.constitution/method/ci-guide.md`.** The trigger shape (`workflow_dispatch` always;
+  `pull_request` `types: [ready_for_review]` as the one automatic trigger; never a bare `on: push`),
+  `paths-ignore` for the prose and corpus layers, `concurrency` with `cancel-in-progress`, and two
+  copy-paste templates: `ci.yml` for the product's build and test, and `korpus.yml` for the corpus
+  validators, kept separate so the cheap one can still run when the expensive one is ignored.
+
+**What a repo that already has the method installed does about it.** `npx wdi-method@latest update` brings
+the guide and the two skills. The workflows themselves are yours and the installer does not write them:
+open `.constitution/method/ci-guide.md` and compare your `.github/workflows/` against § Trigger shape. A
+workflow with a bare `on: push` is the one that spends the allowance, and it is what the new preflight row
+will refuse to start a mandate over.
+
+## 0.6.18 — 2026-09-08
+
+- **This file.** The version history was only readable as `git log`, which is the wrong place to look for
+  it when you are deciding how carefully to read an `update` diff.
+- **The README had the validator count wrong.** It said twenty-seven; `validate.py` has run twenty-nine
+  named checks since `withdrawn-recorded` arrived in 0.6.14. It also now says what the validators treat as
+  corpus, which 0.6.17 changed.
+
+## 0.6.17 — 2026-09-08
+
+- **The corpus walker honours `.gitignore`.** The walk pruned only a hardcoded folder list, so a vendored
+  dependency tree inside an ignored folder was read as this product's corpus — one repo saw 172
+  `cites-resolve` findings from source code it had deliberately ignored. The ignored set is now computed
+  once per run (`git ls-files --others --ignored --exclude-standard --directory`), and the hardcoded list
+  stays as the fallback for a checkout that is not a git repository. `.gitignore` still cannot hide real
+  corpus: `corpus-in-git` fails when a folder the method commits is ignored.
+- **`timeline.py` no longer crashes on a capability that has tickets.** `cap_tickets()` returns
+  `(spec, ticket)` pairs and the span helper read each item as a ticket row, so generating the timeline
+  raised `AttributeError` and `/wdi-report progress` could not publish at all.
+
+**For a consuming repo:** `update`, then regenerate — a `timeline.md` or `report.md` written before this
+is dated and must not be read as current.
+
+## 0.6.16 — 2026-09-07
+
+- **A superseded mandate still stands for what it accepted.** `mandate-accept` asked a question about the
+  past — was there a mandate on the day this decision was accepted — and answered it with the mandate's
+  status *today*. Superseding a mandate, which `wdi-autopilot`'s own instructions call for when a run
+  ends, therefore turned every decision that run had accepted permanently red; and an `applied` decision
+  is frozen, so there was no repair the corpus was allowed to make. The check now reads the mandate's
+  state at the delegation date, and bounds the window by the **earlier** of `mandate.expires` and the date
+  of the decision that superseded it — a revoked delegation ends when it was revoked, whatever its
+  `expires` still says. `superseded_by` is read from the decision file's frontmatter as well as the
+  registry row, because the template puts it there.
+
+## 0.6.15 — 2026-09-07
+
+- The dangling-promise finding carries its own repair. `withdrawn-recorded` named the problem without
+  saying what to write, and a finding nobody can act on is a finding that gets switched off.
+
+## 0.6.14 — 2026-09-07
+
+- **New validator `withdrawn-recorded`:** a promise that was withdrawn stays in the registry and says who
+  withdrew it. Deleting the row loses the fact that it was ever promised.
+
+## 0.6.13 — 2026-09-07
+
+- A story id already scoped to its wave is not scoped twice — the pre-0.6 rename was double-prefixing ids
+  that had been written correctly.
+
+## 0.6.12 — 2026-09-07
+
+- Flattening a pre-rename wave is `wdi-upgrade`'s work, not a re-cut. The skill was inventing a new
+  decomposition where it should have been moving one that already existed.
+
+## 0.6.11 — 2026-09-07
+
+- `update` leaves a closed spec's folder where it is. Moving a finished spec's folder rewrote history for
+  nothing.
+
+## 0.6.10 — 2026-09-07
+
+- `update` reports the two things `wdi-upgrade` now has to move, instead of leaving them for a reader to
+  discover.
+
+## 0.6.9 — 2026-09-07
+
+- The TUI refuses a missing engine too, in step 2's place. The CLI had checked since 0.6.8; the
+  interactive path had not, and that is the path a first install actually takes.
+
+## 0.6.8 — 2026-09-07
+
+- **The G5 engines are the repo's own, invocable, and BMad's replacements are locked out.** `install` and
+  `update` refuse until `to-spec`, `to-tickets`, `implement`, `tdd`, `code-review` and `domain-modeling`
+  are in the repo; `wdi-build` invokes them itself, so `wdi-autopilot` can finish a spec unattended; and
+  thirteen BMad skills retired at G5 are denied model invocation. A spec also gets one predefined home,
+  `.scratch/<spec-id>-<slug>/`.
+
+**For a consuming repo:** the largest jump in the 0.6 line. The README section *"Moving a repo from 0.6.7
+or earlier to 0.6.8"* walks all four changes, and `npx wdi-method engines --fix` repairs what can be
+repaired without touching anything you wrote.
+
+## 0.6.7 — 2026-09-06
+
+- `plan-dates` reported nothing at all, because it raised before it could report; and a legacy story file
+  was looked for in the wrong folder.
+- A closed pre-rename wave stays visible to the RTM instead of disappearing from it.
+
+## 0.6.6 — 2026-09-06
+
+- **New validator `corpus-in-git`:** a folder the method commits MUST NOT be gitignored. A corpus git
+  cannot see is a corpus the clone lacks.
+- The git rule is stated where the bootstrap reads it, not only where the detail lives.
+
+## 0.6.5 — 2026-09-05
+
+- The README said three things that 0.6.2 through 0.6.4 had made untrue.
+
+## 0.6.4 — 2026-09-05
+
+- The test suite depended on the author's machine, and 0.6.3 failed CI because of it.
+
+## 0.6.3 — 2026-09-05
+
+- The ticket engines are required, and their config ships pre-answered — so `/setup-matt-pocock-skills` is
+  not part of getting started.
+- *"What now"* is answered after install and after update, in the README and in `wdi-help`.
+- Three real migrations, none of them `wdi-upgrade`'s: a mechanical rename, a printed warning, and a skill
+  that heals its own ledger.
+
+## 0.6.2 — 2026-09-05
+
+- **New skill `wdi-autopilot`:** one mandate, then every `FR` in scope delivered unattended, every answer
+  recorded in one ledger. It stops at one of three places — done, at capacity, or blocked — and one run
+  lands as one pull request. An adversarial review closed ten findings before release, two of them in the
+  validator.
+
+**For a consuming repo:** `update` renames a pre-0.6.2 ledger to `autopilot-<mandate-id>.md`. The content
+is never rewritten.
+
+## 0.6.1 — 2026-09-03
+
+- **New skill `wdi-explain-to-me`:** a decision briefing the owner reads, written in the owner's language.
+
+## 0.6.0 — 2026-09-03
+
+- **Two trees, one home per fact.** The working set points and the rendered page answers, so a human gets
+  a complete document while the corpus keeps no copies. This is the release the validators' rule against
+  comparing two copies of one fact comes from.
+- The units are renamed: wave and story are retired, `waves.yaml` becomes `specs.yaml`, and size decides
+  the spec.
+- The BMad engine layer below G5 is retired in favour of `to-spec`, `to-tickets` and `implement`;
+  `wdi-blueprint` wraps `domain-modeling` as its engine.
+- UX runs and **lands** at G2, breaking a deadlock where the container was never in its path.
+- A decision's first home is the document it governs, and no `DEC-` is mandatory.
+- The corpus is present tense, and stale is not a finding.
+- `wdi-review` stopped being a treadmill; gate *shape* and checklist *length* became separate knobs.
+
+**For a consuming repo:** run `wdi-upgrade` after updating, before any other skill. The validators read
+the new shape, and a corpus half in the old one answers them wrongly. The untagged 0.5.13 and 0.5.14
+bumps — the review panel dropping its two-CLI-family requirement, and the platform picker aligning with
+BMad — are included here.
+
+## 0.5.12 — 2026-08-22
+
+- Blueprint actor headings lose the orphaned separator a nameless actor left behind.
+
+## 0.5.11 — 2026-08-22
+
+- Version bump only; no method change.
+
+## 0.5.10 — 2026-08-19
+
+- The inventory engine littered the room it was asked to read.
+
+## 0.5.9 — 2026-08-19
+
+- The package ships no stack at all: a skeleton, and a skill that writes it. A stack baked into a generic
+  package is an assumption every consumer inherits.
+
+## 0.5.8 — 2026-08-19
+
+- `inventory.py`: the engine stays in the package, the stack moves to the room that owns it.
+
+## 0.5.7 — 2026-08-19
+
+- The method assumed a stack, and a path was only the visible half of that assumption.
+
+## 0.5.6 — 2026-08-19
+
+- The guard a real component actually needed, which the retired V24 could never have been.
+
+## 0.5.5 — 2026-08-19
+
+- V24 was unsatisfiable for every consumer, and the language sweep had missed nineteen strings.
+
+## 0.5.4 — 2026-08-19
+
+- `portability.md` was right about the path and wrong about everything else.
+
+## 0.5.3 — 2026-08-19
+
+- The 0.5.0 split was unreachable for the repos that needed it.
+- Python bytecode is kept out of the tarball — 0.5.2 shipped 123 kB of it, and a `.pyc` embeds the
+  absolute path it was compiled from.
+
+## 0.5.2 — 2026-08-19
+
+- **The direction was reversed:** a method change is authored *here* and proven against a fixture corpus
+  the registry scripts actually run against, instead of being promoted out of a live product. `promote`
+  became a rescue tool that refuses to run without `--rescue`.
+- An external audit of 0.5.0 found the `AGENTS.md` block calling the whole kit non-binding.
+- Bahasa Indonesia is swept out of the generic package; the installer's interface is English.
+- Included here: `.constitution/` becoming exactly two folders, `method/` and `project/` (0.5.0), and the
+  0.5.1 republish of a kit that had shipped pre-fix overlay content.
+
+## 0.3.0 — 2026-08-18
+
+- First public release: an interactive installer, a replaceable `AGENTS.md` method block, two language
+  settings asked separately, and `.constitution/project/` as the custom room `update` never overwrites and
+  `promote` never publishes.