@dzhechkov/skills-feature-adr 1.5.2 → 1.5.4

package/.dz-manifest.json

@@ -13,7 +13,7 @@
       },
       {
         "path": "README.md",
-        "sha256": "fe7f99695df5590ae1474395e3a2b26754a6bfd63695c5d8c804fb290b8eb6b1"
+        "sha256": "eb5998619673f928f02b6b462278acfc4871387e1f6495ab0b613a965611cda7"
       },
       {
         "path": "bin/cli.js",
@@ -25,7 +25,7 @@
       },
       {
         "path": "package.json",
-        "sha256": "e3ce1d2e92b37dae085819111a7a107cf9b170a7602ceea381423873523ad1d2"
+        "sha256": "8e082526415c4ce4427c51dff222c248e29ef5fbdcd17f0d957c28d621738513"
       },
       {
         "path": "src/cli.js",
@@ -145,11 +145,11 @@
       },
       {
         "path": "templates/.claude/skills/feature-adr/modules/05-architecture.md",
-        "sha256": "2ef7cf316903f0cbb1fabe1f0e1770c22b2cb3a11a4118cac0529c25074b5e8a"
+        "sha256": "4d20dbc85ea5bb93ca90fd8cd089255bab9c76a912ca3ad954f3a4bb670f3c45"
       },
       {
         "path": "templates/.claude/skills/feature-adr/modules/06-implementation-plan.md",
-        "sha256": "3903e2694ad47782914b35869bbf19c5f0803b6dafdcfb629df343a5ff00d1b9"
+        "sha256": "b3efdfa1d44f4c51ff46551b5827066955c1f0bfee212d371ac1e2fb1c99980f"
       },
       {
         "path": "templates/.claude/skills/feature-adr/modules/07-code.md",
@@ -249,7 +249,7 @@
       },
       {
         "path": "templates/.claude/skills/feature-adr/scripts/check-plan-completeness.mjs",
-        "sha256": "5cef77e68c58f163be64516013fb90866ef7201e415dc8888cc21a0f4d92b7d8"
+        "sha256": "eddf116620050e80b88497ace026fb8e0a12986f94f216181480c7daa0e08be4"
       },
       {
         "path": "templates/.claude/skills/frontend-design/LICENSE.txt",
@@ -313,7 +313,7 @@
       },
       {
         "path": "templates/.claude/workflows/feature-adr.js",
-        "sha256": "2d895ee2b348745dce23b6beb6f932908beb932f2de309afe812ed3e1b96a7db"
+        "sha256": "a6ff055d50f4eb5790dab36deee188e880eec1362188ec16a5504710294d87c3"
       },
       {
         "path": "templates/lib/memory-protocol.md",
@@ -325,5 +325,5 @@
       }
     ]
   },
-  "signature": "RVit/JWzAfaMRgh3KVc200QVdpU0sNO6AcEkRGYt8c8fqq1TG01DKPXiGvV08hWRHH8/a/67+qOko8sYEgqEBw=="
+  "signature": "4u6AmuuLsokLePBMiy6Gg9ETVwszE2NVE2fArT62nFcJSWPEd06QhKu4eb2/awzNjH4bhY3HXbacMZsS7B7gBA=="
 }

package/README.md

@@ -169,7 +169,7 @@ refused BY NAME: `path traversal ('..' segment)`, `degenerate path segment`, `em
 `illegal character '<c>'`, `empty stem after the leading dot`. The old substring traversal test also
 rejected the ordinary filename `foo..bar.ts`; it is a segment test now.
 
-### The Step-8 amendment gate runs a command instead of judging (v1.5.2)
+### The Step-8 amendment gate runs a command instead of judging (v1.5.4)
 
 Every `AM-N` row must resolve to a test found INSIDE the file the row names, and the check is
 `dz amendment-check --slug <slug> --json` rather than a paragraph asking the reviewer to confirm it.
@@ -1107,3 +1107,56 @@ structure chosen before understanding, and section-per-edit raises tool-call ove
 
 Steps whose deliverable is a returned verdict rather than a document (the complexity router) are
 deliberately excluded — "skeleton first" is nonsense there.
+
+
+---
+
+## Status
+
+`1.5.4` — **Step 5 now asks how the shipped feature will be watched.** The architecture artifact must
+carry a section headed exactly `Observability` answering what the feature logs, what it counts, what a
+failure looks like from outside, and who would notice. **"Nothing to observe" is a complete answer** —
+a pure refactor or a CI-only gate genuinely emits nothing, and a check that cannot express a true fact
+gets switched off. What is not acceptable is leaving the question unanswered. MEASURED before the
+change: the word appeared nowhere in the pipeline's prompts, and ZERO of 107 architecture artifacts
+carried such a section, while two fully-written observability skills (818 and 946 lines) sat
+unreachable because nothing called them. `dz score` reports the answer as a discipline — descriptive,
+never a gate, because the whole existing corpus predates the requirement.
+
+Also in the bundled workflow: the K2 gate stops GUESSING the workspace. When it was not pinned, the
+workspace candidate was taken from the gate agent's own working directory, which against an external
+target repo resolved to that repo — so the candidate silently became a duplicate of the repo
+candidate and a skill installed in the workspace was never found (`NOT-ESTABLISHED`, exit 3, the
+coding step never ran). New `args.workspace` pins it, the shipped call site passes it, each candidate
+is labelled in the audit line, and a `K2_GATE_NOTE` fires when two collapse onto one path. And a
+pre-code probe that returns nothing no longer becomes an all-null baseline that later reads as
+"every target changed".
+
+
+`1.5.4` — **the workflow stops crashing on the way into Step 7.** `1.5.4` shipped a workflow that
+CALLED three helpers it never defined — `changeSetProbeCmd`, `parseHashProbe`, `changedFromHashes`
+(5 call sites, 0 definitions). `QE_SCOPE` defaults to `uncommitted`, so the guarded branch was true
+by default and every ultracode run that reached the coding step died with a `ReferenceError` while
+the corresponding unit tests stayed green — they exercise the exported module, the pipeline runs an
+inline mirror of it. The three are now restored FROM that canonical export, not reconstructed, and
+two guards make the class visible: the mirror is checked by lifting each function out of the shipped
+file and comparing its BEHAVIOUR case-by-case against the export, and a new layer-1 test parses the
+workflow and asserts every referenced identifier is declared or is one of the eight documented
+sandbox globals.
+
+Also in this release, both halves of the K2 plan-completeness gate that field use found:
+
+- **C6 scopes each amendment to its own block** — from its `AM-N` line to the line where the next one
+  begins. The old three-line window refused amendments whose `→ test` marker sat on a `Confirmation:`
+  line further down (measured: line 3 passes, line 4 fails), and — worse, and not reported — it
+  PASSED a testless amendment that happened to sit next to a tested one, which borrowed its
+  neighbour's marker. The plural `→ tests \`a\` and \`b\`` form is now matched, and a wrapped
+  `AM-1..AM-4;` range no longer opens a phantom amendment. Measured over 142 plans: 402 C6 failures
+  before, 190 after, **0 plans newly failing**. The Step-6 planner prompt now states that the row is
+  machine-read and where the marker must sit.
+- **The K2 gate stops guessing the workspace.** When the workspace was not pinned it was taken from
+  the gate agent's own working directory, which against an external target repo resolved to that
+  repo — so the workspace candidate silently became a duplicate of the repo candidate and a skill
+  installed in the workspace was never found (`NOT-ESTABLISHED`, exit 3, the coding step never ran).
+  New `args.workspace` pins it, the shipped call site passes it, each candidate is now labelled in the
+  audit line, and a `K2_GATE_NOTE` fires when the two collapse onto one path.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/skills-feature-adr",
-  "version": "1.5.2",
+  "version": "1.5.4",
   "description": "Adaptive Feature Development skill pack for Claude Code — 11-step pipeline with Complexity Router (S/M/L/XL), ADR-driven architecture, 15 agentic-qe skills, multi-agent fleet QE. Supports --full-qe, --full-qe-extended, --with-learning, and --knowledge-extractor modes.",
   "bin": {
     "skills-feature-adr": "./bin/cli.js"

package/sbom.json

@@ -35,7 +35,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "fe7f99695df5590ae1474395e3a2b26754a6bfd63695c5d8c804fb290b8eb6b1"
+          "content": "eb5998619673f928f02b6b462278acfc4871387e1f6495ab0b613a965611cda7"
         }
       ]
     },
@@ -65,7 +65,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "e3ce1d2e92b37dae085819111a7a107cf9b170a7602ceea381423873523ad1d2"
+          "content": "8e082526415c4ce4427c51dff222c248e29ef5fbdcd17f0d957c28d621738513"
         }
       ]
     },
@@ -365,7 +365,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "2ef7cf316903f0cbb1fabe1f0e1770c22b2cb3a11a4118cac0529c25074b5e8a"
+          "content": "4d20dbc85ea5bb93ca90fd8cd089255bab9c76a912ca3ad954f3a4bb670f3c45"
         }
       ]
     },
@@ -375,7 +375,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "3903e2694ad47782914b35869bbf19c5f0803b6dafdcfb629df343a5ff00d1b9"
+          "content": "b3efdfa1d44f4c51ff46551b5827066955c1f0bfee212d371ac1e2fb1c99980f"
         }
       ]
     },
@@ -625,7 +625,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "5cef77e68c58f163be64516013fb90866ef7201e415dc8888cc21a0f4d92b7d8"
+          "content": "eddf116620050e80b88497ace026fb8e0a12986f94f216181480c7daa0e08be4"
         }
       ]
     },
@@ -785,7 +785,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "2d895ee2b348745dce23b6beb6f932908beb932f2de309afe812ed3e1b96a7db"
+          "content": "a6ff055d50f4eb5790dab36deee188e880eec1362188ec16a5504710294d87c3"
         }
       ]
     },

package/templates/.claude/skills/feature-adr/modules/05-architecture.md

@@ -118,6 +118,32 @@ GET /api/v1/{resource}/:id
   Errors: 401, 404
 ```
 
+## Observability — a MANDATORY section of `05_architecture.md`
+
+The artifact must carry a section headed exactly **`Observability`**, answering how anyone would
+know this feature is working once it ships:
+
+- what it **logs**, and at what level;
+- what it **counts** — the one or two numbers that would move if it broke;
+- what a **failure looks like from outside** — the symptom, not the stack trace;
+- **who would notice**, and how.
+
+If the feature genuinely emits nothing at runtime — a pure refactor, a doc change, a gate that only
+runs in CI — write **"nothing to observe"** and say why. That is a complete answer, not a gap. What
+is not acceptable is leaving the question unanswered.
+
+The heading is read by a machine (`dz score` reports it as a discipline), so it is spelled exactly
+`Observability` — the prompt above and the check derive that word from one shared constant, because
+a prompt asking for one heading while a check greps another produces a gate that fails every honest
+run.
+
+**Why this section exists.** MEASURED 2026-08-25: across 107 existing `05_architecture.md` files,
+ZERO carried such a section, and the word `observability` appeared nowhere in the pipeline's own
+prompts. The pipeline kept five telemetry stores about itself and asked for none about what it
+built. The two capable skills — `observability` (818 lines) and `observability-testing-patterns`
+(946 lines) — were fully written and unreachable, because nothing called them. This section is the
+call site.
+
 ## Write discipline (the 180-second rule)
 
 An executor that returns from a tool call and then thinks in silence past **180 seconds** is killed by the

package/templates/.claude/skills/feature-adr/modules/06-implementation-plan.md

@@ -185,6 +185,11 @@ Create `features/<slug>/06_implementation_plan.md` with:
 - `## Amendments` — every correction folded into this plan (a Step-3.5 CONDITIONAL condition, a
   challenge-panel confirmed finding, a user checkpoint steer) as a fixed-shape row:
   `AM-N (source): <change>. Confirmation: <property> → test `test_name` (fails if reverted).`
+  The row is MACHINE-READ by the K2 C6 gate: the `→ test \`name\`` marker (or `superseded by AM-N`)
+  must sit inside that amendment's OWN block — anywhere between its `AM-N` line and the line where
+  the NEXT `AM-N` begins. Multi-line amendments are fine; what is NOT fine is putting a marker after
+  the following amendment has already started, because it then belongs to that one. A bare range
+  like `AM-1..AM-4` never opens a row.
   A safeguard amendment's named test must prove it TRIGGERS on a real input, not merely that its code
   path exists. Step 8 verifies every named test exists and is non-vacuous (`dz discrimination-check`).
 

package/templates/.claude/skills/feature-adr/scripts/check-plan-completeness.mjs

@@ -327,13 +327,36 @@ else for (const t of acidTokens) if (!new RegExp(`\\b${t.replace(/[.*+?^${}()|[\
     }
   }
   if (sectionStart >= 0) {
-    for (const lnRaw of amSection.split('\n')) {
-      const ln = lnRaw.trim();
-      const am = /^(?:[-*|]\s*)?\*{0,2}(AM-(?:CP-)?\d+)\*{0,2}\b/.exec(ln);
-      if (!am) continue;
-      const hasTest = /\u2192\s*test\b|->\s*test\b/.test(ln) || /\u2192\s*test\b|->\s*test\b/.test(amSection.slice(amSection.indexOf(lnRaw)).split('\n').slice(0, 3).join('\n'));
-      const superseded = /superseded by AM-(?:CP-)?\d+/i.test(ln);
-      if (!hasTest && !superseded) failures.push(`C6: ${am[1]} carries neither \`\u2192 test <name>\` nor \`superseded by AM-N\` — an amendment without a confirmation is a wish, and a retracted one must say its successor`);
+    // Each amendment's confirmation is looked for in ITS OWN BLOCK: from its definition line to the
+    // line where the NEXT amendment begins. Three defects die with the old three-line window
+    // (MEASURED 2026-08-25 against the 142-plan corpus — 292 C6 failures before, 33 after, and ZERO
+    // rows newly caught that pass today):
+    //   • FALSE REFUSAL — a marker on the `Confirmation:` line 4+ lines down was invisible, so plans
+    //     that DID name their test were rejected (7 per run, twice in one day, on a user's machine).
+    //   • FALSE PASS — the window is three LINES, not one amendment, so `- AM-1 testless` followed by
+    //     `- AM-2 … -> test x` PASSED: AM-1 borrowed its neighbour's marker. That is the safety half,
+    //     and it is why the boundary is the next DEFINITION rather than a blank line — a blank line
+    //     does not separate adjacent bullet rows.
+    //   • A substring seek (`amSection.indexOf(lnRaw)`) let a duplicated line read someone else's
+    //     window. Indices remove that hazard for free.
+    // The boundary is not invented: parseAmendments() in harness-core/src/amendment-trace.ts has used
+    // the same next-definition bound all along, so this also ends a divergence between two checkers.
+    const secLines = amSection.split('\n');
+    // The bullet stays OPTIONAL: requiring it drops 104 of the corpus's 347 real AM rows out of the
+    // check entirely (MEASURED). The `(?!\s*\.)` guard is what refuses a wrapped `AM-1..AM-4;`
+    // range preamble, which used to open a phantom amendment and double-count AM-1 — it costs 0 rows.
+    const DEF = /^(?:[-*|]\s*)?\*{0,2}(AM-(?:CP-)?\d+)\*{0,2}\b(?!\s*\.)/;
+    // `tests?` — the two-id form `-> tests \`a\` and \`b\`` is corpus-canonical and is what
+    // amendment-trace.ts parses; matching only the singular turned two currently-green rows red.
+    const MARK = /\u2192\s*tests?\b|->\s*tests?\b/;
+    const defs = [];
+    secLines.forEach((l, i) => { const m = DEF.exec(l.trim()); if (m) defs.push({ i, id: m[1] }); });
+    for (let k = 0; k < defs.length; k++) {
+      const to = k + 1 < defs.length ? defs[k + 1].i : secLines.length;
+      const block = secLines.slice(defs[k].i, to).join('\n');
+      const hasTest = MARK.test(block);
+      const superseded = /superseded by AM-(?:CP-)?\d+/i.test(block);
+      if (!hasTest && !superseded) failures.push(`C6: ${defs[k].id} carries neither \`\u2192 test <name>\` nor \`superseded by AM-N\` — an amendment without a confirmation is a wish, and a retracted one must say its successor`);
     }
   }
 }

package/templates/.claude/workflows/feature-adr.js

@@ -143,6 +143,11 @@ log('dz binary: ' + DZ)
 // HERE, at invocation time, so a bad value fails at the same layer the pure half fails rather than
 // two layers later inside an emitted shell command.
 const GATE_SCRIPT_ARG = (A.gateScript === undefined || A.gateScript === null) ? undefined : assertAbsoluteNoTraversal(A.gateScript, 'gateScript')
+// ADR-002 amendment (field report doc-21): WS was populated ONLY for a relative args.repo, and the
+// shell fallback WS=$(pwd -P) runs in the GATE AGENT own cwd. On a run against an external repo it
+// equalled REPO, so the workspace candidate pointed at the target repo and the skill installed in
+// the workspace was never found - NOT-ESTABLISHED, exit 3, Step 7 never ran. args.workspace pins it.
+if (A.workspace !== undefined && A.workspace !== null) WS = assertAbsoluteNoTraversal(A.workspace, 'workspace')
 // CANONICAL BRAIN store: the self-learning loop (Step-0 recall → Step-8 teach) MUST read+write ONE
 // shared pattern store so lessons never fragment into a target repo's .dz when the Step-7 coder cd's
 // away. BRAIN defaults to the workspace root (REPO) — so an OMITTED args.brain is behaviorally inert
@@ -1747,7 +1752,12 @@ function planCompletenessGateCmd(repo, featureDir, tier, opts) {
     // and the tried paths live OUTSIDE the verdict line so no path can smuggle a second verdict word
     // into it.
     'echo "K2_GATE_SCRIPT=${GS:-none}"',
-    'echo "K2_GATE_TRIED=${C1:-(none)} | $C2 | $C3"',
+    'echo "K2_GATE_TRIED=C1(args.gateScript)=${C1:-<unset>} | C2(workspace)=$C2 | C3(target-repo)=$C3"',
+    // A COLLAPSE is not a second candidate. When the workspace was not pinned, WS falls back to the
+    // gate agent own cwd — in the field that WAS the target repo, so C2 and C3 printed the same path
+    // twice and the chain silently degenerated from three candidates to two. Saying so turns a
+    // puzzling duplicate into an instruction. Not verdict-shaped, so the parser anchoring is untouched.
+    '[ "$C2" = "$C3" ] && echo "K2_GATE_NOTE=the workspace candidate resolved to the TARGET repo (WS==repo), so only two distinct candidates were tried; pass args.workspace or args.gateScript when the feature-adr skill is installed outside the target repo"',
     'if [ -z "$GS" ]; then echo "K2 plan-completeness: NOT-ESTABLISHED — tooling-missing: no gate script at any candidate on the K2_GATE_TRIED line above"; echo "K2_EXIT=3"; else cd ' + q(repo) + ' && node "$GS" ' + q(featureDir) + t + ' 2>&1; echo "K2_EXIT=$?"; fi',
   ].join('\n')
 }
@@ -1834,6 +1844,57 @@ function decideModeBScope(o) {
   return { ok: true, files: files, dropped: dropped }
 }
 
+// ── QE CHANGE-SET PROBE — inline mirror of harness-core/src/feature-adr-routing.ts:2309-2381.
+// These three travelled as CALL SITES only: commit 6a92d189 ("measure the DELTA, not the current
+// state") shipped the TS module, its 19 tests and both call sites below, and never inlined the
+// helpers here. Every run reaching Step 7 therefore died with a ReferenceError on changeSetProbeCmd
+// while those 19 tests stayed green — they exercise the export, this file runs the mirror. Ported
+// byte-faithfully from the canonical: sha256sum (not shasum), the ~1 and ...HEAD ref forms, a Map
+// snapshot seeded with nulls from the declared list (absence is a null VALUE, never an ABSENT
+// line), and null returned ONLY by changedFromHashes — that is the single "not established" signal.
+// NOTE: no template literals below. This region sits inside the parser-safe block the routing test
+// guards (it runs from const MODELS to const ROUTER), so even a backtick in a COMMENT reddens it.
+function parseHashProbe(text, declared) {
+  const out = new Map()
+  for (const p of declared) out.set(String(p), null)
+  for (const raw of String(text === null || text === undefined ? '' : text).split('\n')) {
+    const line = raw.trim()
+    if (line === '') continue
+    const m = /^([0-9a-f]{64})\s+(.+)$/.exec(line)
+    if (m === null || m[1] === undefined || m[2] === undefined) continue
+    const path = m[2].trim().replace(/^\.\//, '')
+    if (out.has(path)) out.set(path, m[1])
+  }
+  return out
+}
+
+function changedFromHashes(before, after) {
+  if (before === null || before === undefined || after === null || after === undefined) return null
+  const changed = []
+  for (const [path, afterHash] of after) {
+    const beforeHash = before.has(path) ? (before.get(path) === undefined ? null : before.get(path)) : null
+    if (beforeHash !== (afterHash === null || afterHash === undefined ? null : afterHash)) changed.push(path)
+  }
+  return changed.sort()
+}
+
+function changeSetProbeCmd(opts) {
+  const paths = opts.paths.map(function (p) { return String(p) }).filter(function (p) { return p !== '' })
+  if (paths.length === 0) return null
+  const quoted = paths.map(opts.quote).join(' ')
+  const ref = String(opts.ref === null || opts.ref === undefined ? '' : opts.ref).trim()
+  if (opts.scope === 'commit') {
+    if (ref === '') return null
+    return 'git diff --name-only ' + opts.quote(ref) + '~1 ' + opts.quote(ref) + ' -- ' + quoted
+  }
+  if (opts.scope === 'base') {
+    if (ref === '') return null
+    return 'git diff --name-only ' + opts.quote(ref) + '...HEAD -- ' + quoted
+  }
+  // uncommitted: hash the declared targets; the caller pairs this with a pre-code baseline.
+  return 'sha256sum -- ' + quoted + ' 2>/dev/null || true'
+}
+
 function crossFamilyQe(o) {
   // NORMALISE first — raw-string comparison let 'Claude' vs 'claude' report a cross-family review
   // that never happened (caught by codex review --commit on the TS-only fix: the exported helper
@@ -1917,7 +1978,7 @@ const DISCRIMINATION_GATE = '\u00a742 TEST-DISCRIMINATION GATE (run right after
 // amendment is a mini-ADR: it carries a one-line Confirmation naming the test that falsifies it. Machine-checkable
 // shape (a linter can assert the `→ test ` token); Step-8 verifies existence + non-vacuity via the SAME
 // dz discrimination-check that guards the ADR property (cost-of-detection ladder: judgment → step gate).
-const AMENDMENT_RULE = 'AMENDMENT CONFIRMATION DISCIPLINE (every amendment is a mini-ADR): whenever a correction/amendment is folded in (a QCSD CONDITIONAL condition, a challenge-panel confirmed finding, or a user checkpoint steer), record it in a `## Amendments` section as a fixed-shape row: `AM-N (source): <change>. Confirmation: <property> → test `test_name` (fails if reverted).` — naming the test that would FAIL if the amendment were reverted/broken. For a SAFEGUARD amendment (a warning/guard/fallback), the named test must prove the safeguard actually TRIGGERS on a real input — not merely that its code path exists (a structurally-dead safeguard passes an existence test and never fires in production).'
+const AMENDMENT_RULE = 'AMENDMENT CONFIRMATION DISCIPLINE (every amendment is a mini-ADR): whenever a correction/amendment is folded in (a QCSD CONDITIONAL condition, a challenge-panel confirmed finding, or a user checkpoint steer), record it in a `## Amendments` section as a fixed-shape row: `AM-N (source): <change>. Confirmation: <property> → test `test_name` (fails if reverted).` — naming the test that would FAIL if the amendment were reverted/broken. The row is MACHINE-READ by the K2 C6 gate: the marker (or a superseded-by-AM-N note) must sit inside the OWN block of that amendment — anywhere between its AM-N line and the line where the NEXT AM-N begins. Multi-line amendments are fine; what is NOT fine is placing a marker after the following amendment has already started, because it then belongs to that one. A bare range such as AM-1..AM-4 never opens a row. For a SAFEGUARD amendment (a warning/guard/fallback), the named test must prove the safeguard actually TRIGGERS on a real input — not merely that its code path exists (a structurally-dead safeguard passes an existence test and never fires in production).'
 const AMENDMENT_GATE = 'AMENDMENT GATE (P2): do NOT judge this yourself — RUN the check and report what it says. Via Bash run EXACTLY `' + DZ + ' amendment-check --slug ' + SLUG + ' --json` (add `--feature-dir ' + FDIR + '` if the slug does not resolve from your CWD). Parse the JSON and report `amendments: {outcome, counts, reasons}` in your return object. outcome `pass` or `skip` clears the gate; `fail` is a HIGH gap and every reason must be quoted verbatim into the QE report; `not-established` means the check could not be run or the grammar matched nothing — that is NEVER a pass, report it as inconclusive with the tool error. Empty stdout, a crash, or a missing `dz` is `not-established`, not a clean gate. This check proves each amendment RESOLVES to a real test; it does NOT prove the test discriminates — vacuity stays with the discrimination gate above. ' +
   'IO-ON-PURE-PATH + FIXTURE-SWAP HUNT (P5): in the test diff, hunt for replacements of broken/unbound fixtures with healthy ones — the old fixture was probably a NEGATIVE CONTROL proving a path was I/O-free; each such swap requires a compensating negative resource-down test. If the code diff adds I/O (DB/network/file) to a previously-pure path — especially startup/lifespan/health — require a negative resource-down test (broken/unbound resource → the path degrades per its declared contract: fail-open for advisory, explicit fail-fast for load-bearing). Missing → HIGH gap.'
 
@@ -2064,7 +2125,7 @@ if (isMplus) {
   designThunks.push(() => subCheckpoint('qcsd', qcsdPromptText, qcsdOpts, ['03.5_ideation_report.md'], () => designStage(qcsdPromptText, qcsdOpts, FDIR + '/03.5_ideation_report.md', 'qcsd')))
   designPairMeta.push({ role: 'design:qcsd', prompt: qcsdPromptText, model: modelLabel(qcsdOpts), lessons: [] })
   const archExtra = isLplus ? ' Also ' + FDIR + '/04_domain_model.md (DDD).' : ''
-  const archPromptText = (isLplus ? 'Step 4 (DDD) + ' : '') + 'Step 5 (Architecture) of /feature-adr for "' + DESC + '" (' + SLUG + '). READ the code. Write ' + FDIR + '/05_architecture.md (components, data flow, integration points, the emit/merge/wiring shape).' + archExtra + ' Return wrote[] + summary.' + WRITE_DISCIPLINE
+  const archPromptText = (isLplus ? 'Step 4 (DDD) + ' : '') + 'Step 5 (Architecture) of /feature-adr for "' + DESC + '" (' + SLUG + '). READ the code. Write ' + FDIR + '/05_architecture.md (components, data flow, integration points, the emit/merge/wiring shape). It MUST carry a section headed exactly "Observability" answering how anyone would know this feature is working once it ships: what it logs, what it counts, what a failure looks like from outside, and who would notice. If the feature genuinely emits nothing at runtime — a pure refactor, a CI-only gate — say "nothing to observe" and why; that is a complete answer, not a gap. What is not acceptable is leaving the question unanswered.' + archExtra + ' Return wrote[] + summary.' + WRITE_DISCIPLINE
   designThunks.push(() => subCheckpoint('architecture', archPromptText, archOpts, ['05_architecture.md'].concat(isLplus ? ['04_domain_model.md'] : []), () => designStage(archPromptText, archOpts, FDIR + '/05_architecture.md', 'architecture')))
   designPairMeta.push({ role: 'design:architecture', prompt: archPromptText, model: modelLabel(archOpts), lessons: [] })
 }
@@ -2295,7 +2356,7 @@ async function runChallengePanel(planRel, plannerName) {
   if (!authorIsCodex) {
     // author=Claude → Codex adversary (cross-family). Compact prompt: Codex reads the files itself (no 24k
     // brief inlined). safeCodexAgent is the honest exec path; null/invalid ⇒ loud Claude fallback below.
-    const cx = await safeCodexAgent('You are a FRESH adversarial reviewer of an implementation plan you did NOT write. Read these files: ' + planRel + ' , architecture/vision.md , architecture/degradations.md (relative to repo ' + REPO + '). BREAK the plan, do not confirm it. Answer C1 arch-anti-cement (deviating from a pattern in the degradations registry is NOT a finding), C2 prod-ready, C3 test sufficiency+honesty both ways, C4 overengineering, C5 silent decisions, C6 runtime consistency, C7 scope>1.5x, C8 executability. Output ONLY minified JSON {"findings":[{"c","severity":"P0|P1|P2","title","why","where"}],"summary"}.', { label: 'challenge:codex-adversary', phase: 'Plan' })
+    const cx = await safeCodexAgent('You are a FRESH adversarial reviewer of an implementation plan you did NOT write. Read these files: ' + planRel + ' , architecture/vision.md , architecture/testing.md , architecture/map.json , architecture/degradations.md (relative to repo ' + REPO + '). BREAK the plan, do not confirm it. Answer C1 arch-anti-cement (deviating from a pattern in the degradations registry is NOT a finding), C2 prod-ready, C3 test sufficiency+honesty both ways, C4 overengineering, C5 silent decisions, C6 runtime consistency, C7 scope>1.5x, C8 executability. Output ONLY minified JSON {"findings":[{"c","severity":"P0|P1|P2","title","why","where"}],"summary"}.', { label: 'challenge:codex-adversary', phase: 'Plan' })
     if (cx) { try { verdict = sanitizeChallengeVerdict(JSON.parse(String(cx).replace(/^[^{]*/, '').replace(/[^}]*$/, ''))) } catch { verdict = null } }
     if (!verdict) { log('Challenge panel: Codex adversary unavailable/unparseable/invalid — falling back to a FRESH Claude panel (NOT cross-family; run `dz challenge` + codex manually for a cross-family pass)'); adversary = 'claude-fallback' }
   }
@@ -2355,7 +2416,7 @@ if (plan) {
 // plan straight into Step 7. The gate is forced NOT-ESTABLISHED without probing the tree at all.
 let planGate = { verdict: 'not-established', exit: null, reason: 'plan-stage-null', output: 'The Step-6 plan stage returned no result for THIS run (agent died, or produced nothing). Any 06_implementation_plan.md present on disk belongs to an earlier run and cannot vouch for this one, so the gate refuses without reading it.' }
 if (plan) {
-  const planGateOut = await agent('Run EXACTLY this shell snippet via your Bash tool, as ONE command, and return its stdout VERBATIM, nothing else — do not summarize it, do not judge the plan yourself, do not omit the K2_GATE_SCRIPT / K2_GATE_TRIED lines or the trailing K2_EXIT line:\n' + planCompletenessGateCmd(REPO, 'features/' + SLUG, tier, { gateScript: GATE_SCRIPT_ARG }), { label: 'plan:k2-gate', phase: 'Plan', effort: 'low' })
+  const planGateOut = await agent('Run EXACTLY this shell snippet via your Bash tool, as ONE command, and return its stdout VERBATIM, nothing else — do not summarize it, do not judge the plan yourself, do not omit the K2_GATE_SCRIPT / K2_GATE_TRIED lines or the trailing K2_EXIT line:\n' + planCompletenessGateCmd(REPO, 'features/' + SLUG, tier, { gateScript: GATE_SCRIPT_ARG, workspace: WS === null ? undefined : WS }), { label: 'plan:k2-gate', phase: 'Plan', effort: 'low' })
   planGate = parsePlanGateVerdict(planGateOut)
 }
 log('K2 plan-completeness gate: ' + planGate.verdict + ' (exit=' + (planGate.exit === null ? 'unknown' : planGate.exit) + ', reason=' + planGate.reason + ')')
@@ -2424,7 +2485,14 @@ if (QE_SCOPE === 'uncommitted') {
   const baseCmd = changeSetProbeCmd({ scope: 'uncommitted', paths: preCodeTargets, quote: shq })
   if (baseCmd) {
     const baseOut = await agent('Run EXACTLY this via Bash from ' + REPO + ' and return its stdout VERBATIM with NO commentary: cd ' + shq(REPO) + ' && ' + baseCmd, { label: 'qe:baseline-hash', phase: 'Code', effort: 'low' })
-    if (baseOut !== null && baseOut !== undefined) preCodeBaseline = parseHashProbe(String(baseOut), preCodeTargets)
+    // An EMPTY relay reply is not a measurement. parseHashProbe seeds every declared path with null
+    // and returns a valid-looking snapshot, so a failed baseline would later compare null -> hash for
+    // every target and hand Mode B a change set of files Step 7 never touched (cross-family review of
+    // this restoration, P1). A probe that produced no parsable hash line leaves the baseline NULL —
+    // the established not-measured signal that the log below already names and that changedFromHashes
+    // turns into "scope NOT ESTABLISHED", which is never a pass.
+    const baseText = (baseOut === null || baseOut === undefined) ? '' : String(baseOut)
+    if (/^[0-9a-f]{64}\s+\S/m.test(baseText)) preCodeBaseline = parseHashProbe(baseText, preCodeTargets)
   }
   if (preCodeBaseline === null) log('QE: pre-code baseline NOT captured — the change set will read as unmeasured, and a scoped review will refuse rather than guess')
 }