@dzhechkov/skills-bto 1.3.3 → 1.4.1

package/.dz-manifest.json

@@ -0,0 +1,157 @@
+{
+  "manifest": {
+    "version": 1,
+    "pack": "skills-bto",
+    "files": [
+      {
+        "path": "CHANGELOG.md",
+        "sha256": "3cdc74ff663f1a60b0905b26c74cb641f84e8ee89a077f479bed31b19545f359"
+      },
+      {
+        "path": "LICENSE",
+        "sha256": "3b93f7070f6843f37e0cae255f7caff22b466a4017fd2020d7d3216237e5a701"
+      },
+      {
+        "path": "README.md",
+        "sha256": "a622d32ed16bdd3b0db021ad84c5eba87da18e63fe2dd96b10ab1e3f86e24a36"
+      },
+      {
+        "path": "bin/cli.js",
+        "sha256": "a210174d6b63d053f7744757972d31acb0f4d9c59f8dacff7b62872a19ebd31a"
+      },
+      {
+        "path": "package.json",
+        "sha256": "6318b6221e222530bb00f50e978ebeed985f563b8e45f6968c0c2390354eec93"
+      },
+      {
+        "path": "src/cli.js",
+        "sha256": "a712d4840731f4974e22155dc54f7ff3478e3bd15af8e4cefeb10ee3f11b3fe7"
+      },
+      {
+        "path": "src/commands/doctor.js",
+        "sha256": "1c7d323316e502b501ce2b15f177f7cc0766c70e995092b973d7d56522bbd034"
+      },
+      {
+        "path": "src/commands/init.js",
+        "sha256": "1317cb61dc3feafaa7224fb155caaf77a7a5e1097f81070477ffdc53e5c43b22"
+      },
+      {
+        "path": "src/commands/list.js",
+        "sha256": "1ec92e70601bac6da48cddca0764663d999b0aa3cca12b785d825b88de974cfa"
+      },
+      {
+        "path": "src/commands/remove.js",
+        "sha256": "f62beed8b76b134f50d959cd70bb96674a26bcce8441baf49fe08a0780d845ea"
+      },
+      {
+        "path": "src/commands/update.js",
+        "sha256": "f3217f1601eb3fdcfb9099fa07c551e4c922544d09224c9664b4ee5ce230634c"
+      },
+      {
+        "path": "src/utils.js",
+        "sha256": "0ab07df1d28b9c2565f564e43d76c4d7f28b776d43bcb946b22ac70db821f030"
+      },
+      {
+        "path": "templates/.claude/agents/bto-judge-panel.md",
+        "sha256": "1e31c6d4caf15f0ce1202d5cbfaf17d28611e2e98f125184da9ecaa30e01d7b8"
+      },
+      {
+        "path": "templates/.claude/agents/bto-optimizer-worker.md",
+        "sha256": "b9b0522469f376135eb80d9c689c1eb0f5f07e96a3e4f4749002666a8236b64b"
+      },
+      {
+        "path": "templates/.claude/commands/bto-benchmark.md",
+        "sha256": "45a85faaaff840a44bc77e61892b3b0188ff31d4c80c2ef5438618d93a547ef7"
+      },
+      {
+        "path": "templates/.claude/commands/bto-build.md",
+        "sha256": "a1ac21e6ca097b9af1853a0cc815ff108c23eba690a76d4696e21fb4877c676c"
+      },
+      {
+        "path": "templates/.claude/commands/bto-optimize.md",
+        "sha256": "dd07ec79b88ccd3d29311325020c7fe389394e4f5d73004eadad1c605e104362"
+      },
+      {
+        "path": "templates/.claude/commands/bto-test.md",
+        "sha256": "823c509c3f0581cbd24daa1d2ce092ab4cd6f1bc60b23536ab1ce6521fadfcc1"
+      },
+      {
+        "path": "templates/.claude/commands/bto.md",
+        "sha256": "fa1a83a725e6fda68edfea7c15bce0ed688686127f2924de8983184e857fa991"
+      },
+      {
+        "path": "templates/.claude/commands/verify-chain.md",
+        "sha256": "b064f512724091908a953dde475abe1c92f63e02728846644a218144f6616e9e"
+      },
+      {
+        "path": "templates/.claude/rules/bto-quality-gates.md",
+        "sha256": "e4881acf5d15472629cea036fbea50bf2dd577b17d5c1c67730a03a5f47538b5"
+      },
+      {
+        "path": "templates/.claude/rules/witness-chain.md",
+        "sha256": "679aaed6c9b3a173d3fb2888239745d34ed1988315d4f531d14967c33193512b"
+      },
+      {
+        "path": "templates/.claude/shards/bto-evaluation.shard.md",
+        "sha256": "df1abe8d28f2a47fd24aa2da7621c9086f5208f9d675f99e25cf4889fbc6b456"
+      },
+      {
+        "path": "templates/.claude/skills/bto/SKILL.md",
+        "sha256": "cc15a07922126226776d4c48446b3d75fa8caaf16949389f448f42c9787207cb"
+      },
+      {
+        "path": "templates/.claude/skills/bto/examples/sample-benchmark-report.md",
+        "sha256": "cfea6681a75465c0372defee137a39d8fa86ed48b052f1685c96a1e96fb81f02"
+      },
+      {
+        "path": "templates/.claude/skills/bto/examples/sample-eval-report.md",
+        "sha256": "45ed20c23e42a9141de55c103a3d64917cf1f51e0d59a16fbdd9fe498632bbe5"
+      },
+      {
+        "path": "templates/.claude/skills/bto/modules/benchmark.md",
+        "sha256": "41f87e7b2cd8ba9ae2f913a586651d299052534ba4875797c79cbc577d3df4d3"
+      },
+      {
+        "path": "templates/.claude/skills/bto/modules/build.md",
+        "sha256": "d1e26e331562175b3295de3e58edfecbd80eaaed9a57d7e5cf094bc4ffae8247"
+      },
+      {
+        "path": "templates/.claude/skills/bto/modules/optimize.md",
+        "sha256": "d8737955bd02cdfa83af6d72fdbd59ac2f4a6d73b41788a4c801496654c0cc93"
+      },
+      {
+        "path": "templates/.claude/skills/bto/modules/test.md",
+        "sha256": "9daa90fd8566fe7372ad088cd489cd11caf9ba2bd2c3bf2f2b822e3d2ee0bac9"
+      },
+      {
+        "path": "templates/.claude/skills/bto/references/eval-patterns.md",
+        "sha256": "bb4e4aca4c67937cedb91e66debb0e17f84322a5d9e29d8e0ef60a1af13cc92b"
+      },
+      {
+        "path": "templates/.claude/skills/bto/references/golden-samples.md",
+        "sha256": "d98baa5d0c7816fb547ba36d2526bed6317ec9c9c6e8a7c50af2c5d7c5a22a46"
+      },
+      {
+        "path": "templates/.claude/skills/bto/references/judge-rubrics.md",
+        "sha256": "e7ee1f3802c79ce6a1966976d227a8cb060232109e487c6db45f0280cab47e20"
+      },
+      {
+        "path": "templates/.claude/skills/bto/references/optimization-methods.md",
+        "sha256": "d15962e1035e600b897384faa9ea20d3c1c687da797e63480d48e2da039b8783"
+      },
+      {
+        "path": "templates/.claude/skills/bto/references/quality-checklist.md",
+        "sha256": "611c0a0785a41b12b8dde09e9086c5e051ac0672711a4e33df5fefc3c87f7561"
+      },
+      {
+        "path": "templates/lib/judge-attestation.md",
+        "sha256": "7a5c05c6ed514256e2e394b703dc5dfd5eee292f87fe95f8f4153707ff179667"
+      },
+      {
+        "path": "templates/lib/witness-chain.md",
+        "sha256": "8d14187d213b2f4422af2699df6c940b54f5f8aba352d2e8070096b604fcba7e"
+      }
+    ]
+  },
+  "signature": "xTWXwW6XDgGgRBKacw50PZygmxm6t0dEA9kZ4FkFvI1LCzsBb7Cigip7BF4MKm4j79NBUbSvgtxwm859bKabCw=="
+}

package/CHANGELOG.md

@@ -0,0 +1,75 @@
+# skills-bto
+
+## 1.4.0 — 2026-08-18
+
+Hardening wave from the BTO improvement review (slice D). Four adopted items; one deferred.
+
+### Added
+
+- **Frontmatter fence on `SKILL.md` (P5).** The pack's own `SKILL.md` now begins with a
+  `--- name/description ---` fence. Before this, `dz list --skills-dir <pack>/templates/.claude/skills`
+  exited 1 with `SKILL.md must begin with a "---" frontmatter fence` — and because the loader walks
+  the whole tree, the unfenced file made *every* skill in the directory unlistable. The fix
+  propagates to what BTO **generates**: the BUILD Skill Template emits the fence ahead of `# <Name>`,
+  and a new `CHECK-S0` / `TEST-SK0` pair scores the property. Commands, rules and agent templates
+  deliberately keep no fence — the parser that requires it reads skills only, and that call is now
+  written down as a decision instead of an omission.
+- **`## Agent Authoring Rule` (P1).** A new normative section in `SKILL.md`, pointed at from all
+  eight authoring texts (four modules, both agent templates, the driving commands): write the
+  skeleton first, then incremental `Edit` appends; never one giant `Write`. Carries the watchdog
+  numbers (180 s / 600 s) and the consequence — a killed agent loses all unwritten output. The two
+  genuinely short emissions (B2 probe answer, one-line variant score) are declared EXEMPT at their
+  sites, so an exemption and an omission are distinguishable by grep. OPTIMIZE Step 4 gains an arity
+  tripwire that WARNs with the missing variant ids instead of silently selecting from a short pool.
+- **Layer B-1: Environment Preconditions + the INCONCLUSIVE verdict (P2).** A new first benchmark
+  layer probes that each downstream layer *can* execute and prints
+  `Preconditions: ALL_GREEN | DEGRADED(list) | ABORT` in **every** report header, clean runs
+  included. `INCONCLUSIVE` joins the verdict lattice, is evaluated ahead of every numeric threshold,
+  and is sticky downstream (an INCONCLUSIVE BENCHMARK does not enter TEST). Universal check
+  **U-13 `SKIPPED is not PASSED`** carries the same standard into Layer 0.
+- **Judge provenance lines (P4, template half).** Every evaluation report now always emits
+  `Authored by:` / `Judged by:` / `Cross-family: YES|NO`, and a same-family panel prints the
+  `SAME-FAMILY PANEL` degradation banner. The Layer-2 judge table gains a `Family` column and
+  `references/eval-patterns.md` gains anti-conformity measure 6, *Model Family Diversity*.
+- **Test suite + `npm test`.** `test/*.test.mjs` (5 spec files) wired to
+  `"test": "node --test test/*.test.mjs"`. The specs spawn the real `dz` binary as the acceptance
+  oracle and assert against the shipped `templates/` text.
+
+### Changed — user-visible recalibrations
+
+These move numbers that existing reports display. Both are deliberate:
+
+- **B1 skill-test denominator 5 → 6.** Adding `TEST-SK0` re-bases every skill's B1 ratio from fifths
+  to sixths.
+- **Universal checklist total 12 → 13.** Adding `U-13` re-bases the Layer 0 pass rate; the two
+  worked example reports were recomputed to match (`Universal (13) + Skill-specific (16) = 29`).
+- **B1 pass rate denominator is now `tests_executed`, not `tests_total`.** A check that could not run
+  is neither a pass nor a failure, and any `tests_inconclusive > 0` forces the layer to INCONCLUSIVE
+  regardless of the ratio.
+- **Partial-evaluation renormalization narrowed.** Only layers the operator deliberately skipped via
+  `level` are renormalized away. A layer that was requested and could not run is INCONCLUSIVE and is
+  never renormalized — redistributing its weight can only raise the score, turning a gap in evidence
+  into a better grade. The `or skip B0` escape in the BENCHMARK anti-pattern table is removed for the
+  same reason.
+- `references/golden-samples.md` documents the fence as a structural row but **excludes it from
+  `sections_expected`**; the required-section count stays **7**, so a conformant skill's B0 coverage
+  is unchanged.
+- `U-03` amended to accept *fence-then-H1*, so the pack no longer fails its own fixed file. `CHECK-S2`
+  was re-worded ("first heading AFTER the frontmatter fence"), never dropped — the H1 is still required.
+
+### Not changed (deliberately deferred)
+
+- **No hard cross-family requirement.** The `Cross-family:` line and its banner are *recorded and
+  advisory*: they change no score and gate nothing. Making a cross-family Critic mandatory would
+  block every user with a single model family, on evidence of n=1; it stays in backlog pending
+  3-artifact planted-defect replication. A spec asserts that nothing in the pack blocks, fails or
+  refuses on `Cross-family: NO`, so the deferred half cannot arrive unannounced.
+- The pre-existing **model-identity collusion BLOCK** is untouched. Identity blocks; family is
+  advisory; the two now read as one policy.
+- **Round narrowing (P3)** — deferred, instrument first.
+
+## 1.3.0
+
+Initial release — canonicalized into `@dzhechkov/skills-bto` in the dz-harness-hub monorepo.
+
+Build-Benchmark-Test-Optimize skill pack for Claude Code — deterministic benchmarking, quality gates, witness chain, judge attestation, and optimization

package/README.md

@@ -93,14 +93,29 @@ BUILD ──→ BENCHMARK ──→ TEST ──→ OPTIMIZE
 
 | Layer | Cost | Purpose |
 |-------|------|---------|
+| **B-1** | Zero (deterministic) | **Environment preconditions** — probes that each downstream layer *can* run, and records the answer |
 | **B0** | Zero (deterministic) | Golden sample comparison — section coverage, ordering, proportions |
-| **B1** | Zero (deterministic) | Deterministic test suite — 5 tests per artifact type, PASS/FAIL |
+| **B1** | Zero (deterministic) | Deterministic test suite — 6 tests for skills, 5 per other artifact type, PASS/FAIL |
 | **B2** | Minimal (3× haiku) | Consistency probe — 3 parallel agents, agreement measurement |
 | **B3** | Zero (deterministic) | Performance metrics — token efficiency, bloat detection, redundancy |
 
 **Scoring:** `BENCHMARK = B0×0.30 + B1×0.35 + B2×0.15 + B3×0.20`
+(pass rate denominator = checks that actually **executed**, never the declared total)
 
-**Gate:** < 0.50 BLOCK | 0.50–0.70 WARN | > 0.70 PASS → proceed to TEST
+**Gate**, first matching row wins:
+
+| Condition | Verdict |
+|-----------|---------|
+| B-1 returned ABORT | **ABORT** — no score is emitted at all |
+| Any layer INCONCLUSIVE | **INCONCLUSIVE** — not a pass and not a failure; TEST is not entered |
+| Score < 0.50 | BLOCK |
+| Score 0.50–0.70 | WARN |
+| Score > 0.70 | PASS → proceed to TEST |
+
+Layer B-1 prints `Preconditions: ALL_GREEN | DEGRADED(list) | ABORT` in the header of **every**
+report, clean runs included — an absent line is itself a finding. A layer the operator deliberately
+skipped via `--level` is renormalized away; a layer that was *requested and could not run* is
+INCONCLUSIVE and is never renormalized, because redistributing its weight can only raise the score.
 
 ### TEST Layer Model
 
@@ -117,10 +132,35 @@ BUILD ──→ BENCHMARK ──→ TEST ──→ OPTIMIZE
 - Judges never see each other's scores before submitting
 - Standard weights: Domain Expert (0.4) / Critic (0.3) / Completeness Auditor (0.3)
 - If `max_score - min_score > 3` → meta-judge escalation
+- Every evaluation report always emits three provenance lines — `Authored by:`, `Judged by:` and
+  `Cross-family: YES|NO` — so a report read six months later says whether it was cross-checked:
+
+  ```
+  Authored by:  claude-opus (anthropic)
+  Judged by:    Expert=claude-sonnet | Critic=gpt-5.5 | Auditor=claude-sonnet
+  Cross-family: YES (Critic=openai — different family than the author)
+  ```
+
+- Running the Critic seat on a **different model family** than the author is *recommended*; when no
+  second family is reachable the report prints a loud `SAME-FAMILY PANEL` banner saying CORRECTNESS
+  and ROBUSTNESS are self-assessment. The family axis is **advisory** — it changes no score and
+  blocks nothing. The separate *model-identity* rule still BLOCKS: the same model must not both
+  generate and judge an artifact.
+
+### Agent Authoring Rule
+
+Every BTO agent that writes a document writes it **incrementally**: a small `Write` carrying the
+title and section skeleton first, then one `Edit` per section. An agent that emits no tool event for
+the harness watchdog window (180 s inside a workflow runner, 600 s for a directly-spawned subagent)
+is killed, and a killed agent loses *all* unwritten output — there is no partial save. Genuinely
+short emissions (the B2 probe answer, a one-line variant score) are exempt, and each exemption is
+written down at its site rather than inferred from its absence.
 
 ### Quality Gates
 
 - **BENCHMARK** must pass (score ≥ 0.50) before TEST begins
+- A check or layer that could not execute is reported **INCONCLUSIVE** — never counted as a pass,
+  never silently dropped (`SKIPPED is not PASSED`, universal check U-13)
 - BENCHMARK score < 0.50 → BLOCK (artifact needs rework)
 - Layer 0 must pass before Layer 1
 - Layer 1 must pass before Layer 2
@@ -158,6 +198,31 @@ When installed alongside Keysarium, BTO can evaluate and optimize any skill or c
 
 ---
 
+## Generated skills carry a frontmatter fence
+
+Skills built with `/bto-build` (and the pack's own `SKILL.md`) begin with a `---` frontmatter fence
+carrying `name:` and `description:` before the `# Title` heading. This is what makes a skill loadable
+— an unfenced `SKILL.md` makes the harness refuse the *entire* skills directory, not just that one
+skill. Commands, rules and agent templates deliberately do **not** get a fence: the parser that
+requires it reads skills only.
+
+Verify a generated skill with the reproducer the pack itself uses:
+
+```bash
+dz list --skills-dir <project>/.claude/skills   # exit 0 and the skill named in the listing
+dz info <name> --skills-dir <project>/.claude/skills
+```
+
+## Development
+
+```bash
+npm test    # node --test test/*.test.mjs — pack conformance specs
+```
+
+The specs assert against the shipped `templates/` text and spawn the real `dz` binary as the
+acceptance oracle. A layout-only check (`dz skills-verify --static`) is necessary but **not**
+sufficient: it passes on a tree whose fence is missing, so `dz list` is the discriminating oracle.
+
 ## Requirements
 
 - **Claude Code CLI** — installed and configured ([installation guide](https://docs.anthropic.com/en/docs/claude-code))

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/skills-bto",
-  "version": "1.3.3",
+  "version": "1.4.1",
   "description": "Build-Benchmark-Test-Optimize skill pack for Claude Code — deterministic benchmarking, quality gates, witness chain, judge attestation, and optimization",
   "main": "src/cli.js",
   "bin": {
@@ -10,7 +10,10 @@
     "bin/",
     "src/",
     "templates/",
-    "LICENSE"
+    "CHANGELOG.md",
+    "LICENSE",
+    ".dz-manifest.json",
+    "sbom.json"
   ],
   "keywords": [
     "claude",
@@ -49,5 +52,8 @@
   },
   "publishConfig": {
     "access": "public"
+  },
+  "scripts": {
+    "test": "node --test test/*.test.mjs"
   }
 }