@windyroad/itil 2.1.2 → 2.1.3

package/.claude-plugin/plugin.json

@@ -497,5 +497,5 @@
     }
   },
   "name": "wr-itil",
-  "version": "2.1.2"
+  "version": "2.1.3"
 }

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "wr-itil",
-  "version": "2.1.2",
+  "version": "2.1.3",
   "description": "ITIL problem-management workflows for AI coding agents",
   "author": {
     "name": "Windy Road Technology",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@windyroad/itil",
-  "version": "2.1.2",
+  "version": "2.1.3",
   "description": "ITIL-aligned IT service management for Claude Code and Codex",
   "bin": {
     "windyroad-itil": "./bin/install.mjs"

package/scripts/evaluate-relevance.sh

@@ -6,7 +6,9 @@
 # evidence shapes per evidence-based-relevance-close-pass-for-the-problem-backlog-phase-1-file-no-longer-exists-phase-2-adr-shipped-confirmed-named-skill-exists-self-marker-in-body-driver-child-closed-phase-1-false-positive-fixes-architecture-rule (Phase 1 + Phase 2):
 #
 #   Shape 1 — file-no-longer-exists  (Phase 1, original)
+#             Elided prose references (`.../`) are not paths.
 #   Shape 2 — ADR-shipped with `human-oversight: confirmed`  (Phase 2)
+#             CORROBORATING-ONLY — never a clean CLOSE on its own.
 #   Shape 3 — named-skill-or-feature-exists  (Phase 2)
 #   Shape 4 — self-marker-in-body (line-anchored)  (Phase 2)
 #   Shape 5 — driver-child-ticket-closed  (Phase 2)
@@ -244,9 +246,19 @@ record_shape() {
 # candidate. Detects state-suffix / sibling-file / rename to avoid Phase 1
 # false-positives (agent-defers-mitigation-selection-to-user-during-active-incident-surfaces-mitigation-choice-as-user-authority-when-skill-contract-empowers-agent-driven-reversible-mitigations-problem/phase-3-f9-wr-itil-plugin-maturity-list-in-suite-display-shim-reads-installed-plugins-plugin-json-maturity-field-emits-ndjson-per-surface-rollup-per-plugin-problem/rfc-first-trace-invariant-not-enforced-fixes-start-without-rfc-story-map-or-jtbd-trace-problem); on detection routes to KEEP-WITH-NOTE.
 
+# The character class admits `.` and `/`, so an ELIDED reference in prose
+# (`packages/itil/hooks/lib/.../detectors.sh`, `docs/decisions/044-....md`)
+# extracts as a path, is found absent, and is counted as evidence the file
+# was deleted — concluding a ticket was fixed because its prose was
+# abbreviated. An abbreviation was never a path assertion, so drop any
+# candidate carrying an ASCII ellipsis. The Unicode `…` form needs no
+# filter — it is outside the character class above, so it never extracts
+# as a path in the first place (pinned by a bats case regardless). Fourth
+# member of the Phase 1 false-positive class alongside agent-defers-mitigation-selection-to-user-during-active-incident-surfaces-mitigation-choice-as-user-authority-when-skill-contract-empowers-agent-driven-reversible-mitigations-problem/phase-3-f9-wr-itil-plugin-maturity-list-in-suite-display-shim-reads-installed-plugins-plugin-json-maturity-field-emits-ndjson-per-surface-rollup-per-plugin-problem/rfc-first-trace-invariant-not-enforced-fixes-start-without-rfc-story-map-or-jtbd-trace-problem.
 candidates=$(grep -oE '(packages|docs|\.changeset|src|test|scripts)/[A-Za-z0-9._/-]+\.(md|sh|ts|tsx|js|jsx|json|yml|yaml|bats|py|txt|html)' "$ticket_file" 2>/dev/null \
   | sort -u \
   | grep -v '^docs/problems/' \
+  | grep -v '\.\.\.' \
   || true)
 
 shape1_missing=0
@@ -318,6 +330,12 @@ if [ -n "$adr_refs" ]; then
   done <<< "$adr_refs"
 fi
 if [ -n "$shape2_confirmed" ]; then
+  # Corroborating-only. `human-oversight: confirmed` attests that a human
+  # ratified a DECISION (human-oversight-marker-wr-architect-review-decisions-drain-for-recorded-decisions-architecture-rule), not that anything was built or released
+  # — treating it as delivery evidence is the marker-is-not-an-
+  # implementation-licence conflation human-oversight-marker-wr-architect-review-decisions-drain-for-recorded-decisions-architecture-rule's 2026-05-27 amendment names.
+  # It is recorded so the cite survives (agent-output-grounding-no-fabricated-quantitative-estimates-cite-persist-uncertainty-architecture-rule cumulative evidence), but
+  # the demotion below stops it closing a ticket by itself.
   record_shape "ADR-shipped-confirmed" "ADRs human-oversight-confirmed: ${shape2_confirmed}"
 fi
 
@@ -483,7 +501,27 @@ fi
 # caveat short-tag + one-line prose so the SKILL Step 4.6b template can
 # splice the **Caveat** field directly.
 
-if [ -n "$shapes" ]; then
+# Shape 2 demotion (relevance-close-evaluator-over-fires-a-bare-adr-skill-citation-is-read-as-fix-shipped-so-live-tickets-return-close-candidate-problem Option C, shape 2) — ratification is not
+# delivery. When shape 2 is the ONLY shape that fired, the verdict is
+# demoted from a clean CLOSE-CANDIDATE to CLOSE-CANDIDATE-WITH-CAVEAT so
+# the review-problems Step 4.6 AFK path queues it for the maintainer
+# instead of closing it silently. Measured against origin/main
+# 2026-08-30: 107/120 decision records carried the marker and 91/118
+# open/known-error tickets cited at least one, so the shape fired on ~77%
+# of the live backlog and discriminated almost nothing. relevance-close-evaluator-over-fires-a-bare-adr-skill-citation-is-read-as-fix-shipped-so-live-tickets-return-close-candidate-problem holds the
+# wider shape-2/shape-3 tightening pending a superseding decision; this
+# changes the verdict lane only, not the recorded mechanical check.
+#
+# Evaluated BEFORE the multi-phase block so this tag wins when both would
+# fire — the caveat field is a single structured tag per architect
+# condition C2, and "the evidence is the wrong kind" is the more
+# fundamental reason the verdict cannot be clean.
+if [ "$shapes" = "ADR-shipped-confirmed" ]; then
+  caveat_tag="ratification-is-not-delivery"
+  caveat_msg="only evidence is ADR ratification, which records a human approving a decision — not that a fix shipped; confirm the work actually landed before closing"
+fi
+
+if [ -n "$shapes" ] && [ -z "$caveat_tag" ]; then
   # Multi-phase umbrella detection: unticked checkboxes in the ticket
   # body + at least one shipped-evidence shape match. The shape match
   # itself is the "progress made" signal — unticked tasks indicate
@@ -521,7 +559,7 @@ fi
 # No shape fired — fall back to the legacy KEEP / SKIP routing.
 
 if [ -z "$candidates" ]; then
-  echo "SKIP $basename — no extractable file paths (after self-reference exclusion)"
+  echo "SKIP $basename — no extractable file paths (after self-reference and elided-reference exclusion)"
   exit 2
 fi
 

package/skills/review-problems/SKILL.md

@@ -295,7 +295,7 @@ For each `.open.md` / `.known-error.md` ticket aged ≥ 7 days, evaluate whether
 | Shape | Phase | Mechanical check | Empirical closes (2026-05-31) |
 |---|---|---|---|
 | 1. `file-no-longer-exists` | Phase 1 | grep ticket body for `(packages\|docs\|...)/...\.(md\|sh\|...)`; verify each via `git ls-files --error-unmatch` | 0 of 14 |
-| 2. `ADR-shipped-confirmed` | Phase 2 | grep ticket body for `ADR-NNN`; for each, verify `docs/decisions/<NNN>-*.md` exists AND frontmatter has `human-oversight: confirmed` | 8 of 14 — the "Problem 012: Skill Testing Harness Scope Undefined" problem/the "Problem 015: TDD enforcement does not flag vague Gherkin outcome steps" problem/the "Problem 018: TDD plugin should enforce BDD + Example Mapping principles, with examples traceable to Jobs-To-Be-Done" problem/the "Problem 022: Agents must not fabricate time estimates without grounded data" problem/the "Problem 033: No persistent risk register for ISO 31000 / ISO 27001 compliance" problem/the "Problem 039: Autonomous loops conflate diagnose with implement" problem/the "ADRs accumulate forward-chronology evidence inline (Phase 2 dogfood evidence, amendment history, cross-iter cross-references) — `decisions` bucket dominates context at 41% / 1.3 MiB" problem/the "Reconcile release-cadence text with (lean / release-sooner; appetite is a ceiling, not a trigger) + add the dogfood-location-before-public clause" problem |
+| 2. `ADR-shipped-confirmed` **(corroborating-only)** | Phase 2 | grep ticket body for `ADR-NNN`; for each, verify `docs/decisions/<NNN>-*.md` exists AND frontmatter has `human-oversight: confirmed`. **Never a clean close on its own** — the marker records that a human ratified a DECISION (the ": Human-oversight marker + `/wr-architect:review-decisions` drain for recorded decisions" architecture rule), not that a fix shipped. Shape 2 alone demotes to `CLOSE-CANDIDATE-WITH-CAVEAT` with tag `ratification-is-not-delivery` (the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule reassessment 2026-08-30) | 8 of 14 — the "Problem 012: Skill Testing Harness Scope Undefined" problem/the "Problem 015: TDD enforcement does not flag vague Gherkin outcome steps" problem/the "Problem 018: TDD plugin should enforce BDD + Example Mapping principles, with examples traceable to Jobs-To-Be-Done" problem/the "Problem 022: Agents must not fabricate time estimates without grounded data" problem/the "Problem 033: No persistent risk register for ISO 31000 / ISO 27001 compliance" problem/the "Problem 039: Autonomous loops conflate diagnose with implement" problem/the "ADRs accumulate forward-chronology evidence inline (Phase 2 dogfood evidence, amendment history, cross-iter cross-references) — `decisions` bucket dominates context at 41% / 1.3 MiB" problem/the "Reconcile release-cadence text with (lean / release-sooner; appetite is a ceiling, not a trigger) + add the dogfood-location-before-public clause" problem |
 | 3. `named-skill-or-feature-exists` | Phase 2 | grep for SKILL.md / hook / agent paths + `/wr-<plugin>:<skill>` slash-command refs; verify each via `git ls-files` | 6 of 14 — the "Problem 014: No lightweight aside invocation for governance skills (problems, retros, ADRs)" problem/the "Problem 034: Centralise risk reports for cross-project skill improvement" problem/the "Problem 045: Auto plugin install on user's machine after governance release" problem/the "Problem 079: No inbound sync of upstream-reported problems — reports filed via the intake templates never surface in the local backlog" problem/the "Agent designs schemas with user-asked classification fields when the framework should derive silently OR eliminate the classification entirely — deeper generalisation of derive-first-don't-ask" problem/the "Broaden + rename the `solo-developer` persona → `developer`" problem |
 | 4. `self-marker-in-body` | Phase 2 | line-anchored grep for `Close to (Verifying\|Closed)`, `DONE 2026-`, `## Fix Released` heading, `fix shipped session`, `awaiting K→V`. Pattern MUST anchor to line-start to avoid mid-prose false-positives (architect advisory A2) | explicit in the "Broaden + rename the `solo-developer` persona → `developer`" problem; contributory in the "Problem 033: No persistent risk register for ISO 31000 / ISO 27001 compliance" problem |
 | 5. `driver-child-ticket-closed` | Phase 2 | parse `## Related` for `P<NNN>` refs; check if any are in `docs/problems/closed/`. Suppressed when child names an unbuilt SKILL/agent path (future work, not stale; architect advisory A1) | contributory in several closes |
@@ -325,13 +325,13 @@ Exit-code routing (one verdict line per ticket on stdout):
 | Exit | Stdout prefix | Action |
 |------|--------------|--------|
 | 0 | `CLOSE-CANDIDATE <basename> — shapes: <comma-list> — <per-shape cite>; ...` | Auto-close branch (4.6b). |
-| 0 | `CLOSE-CANDIDATE-WITH-CAVEAT <basename> — shapes: <comma-list> — caveat: <short-tag>: <one-line> — cites: ...` | Surface-batch-confirm branch (4.6b-with-caveat); the caveat short-tag + one-line splices verbatim into the audit section's **Caveat** field per architect condition C2. |
+| 0 | `CLOSE-CANDIDATE-WITH-CAVEAT <basename> — shapes: <comma-list> — caveat: <short-tag>: <one-line> — cites: ...` | Surface-batch-confirm branch — **never the auto-close branch**. Under AFK, defer per 4.6a step 3: surface it, do not `git mv` it. The audit-section shape in 4.6b applies only once a maintainer has confirmed the close, and the caveat short-tag + one-line then splices verbatim into that section's **Caveat** field per architect condition C2. |
 | 1 | `KEEP <basename> — <M>/<N> paths still present` | No action; log only. |
 | 1 | `KEEP-WITH-NOTE <basename> — <note>: <evidence>` | Phase 1 false-positive class (state-suffix / sibling-file / rename) OR architect-A1 future-work disambiguation. No action; log only. |
 | 2 | `SKIP <basename> — <reason>` | No action (age gate, no Reported date, no extractable evidence). |
 | 3 | error | Log advisory; do not abort the pass — relevance-close is non-blocking per the Step 4.5 fail-soft precedent. |
 
-**Algorithm (canonical body)**: runs each of the five shape detectors over the ticket body. Multi-shape matches emit cumulatively (corroborating evidence is stronger than first-match-wins per the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule): the `shapes:` field carries a comma-joined list, the trailing fragment carries per-shape cites semicolon-separated. The caveat fires when at least one shape matches AND the body has any unticked checkboxes (multi-phase mixed-progress umbrella class). The verdict is intentionally conservative — tickets with no shape match AND no extractable evidence route to `SKIP`, not auto-close.
+**Algorithm (canonical body)**: runs each of the five shape detectors over the ticket body. Multi-shape matches emit cumulatively (corroborating evidence is stronger than first-match-wins per the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule): the `shapes:` field carries a comma-joined list, the trailing fragment carries per-shape cites semicolon-separated. Two independent triggers raise the caveat, checked in this order: (1) shape 2 is the ONLY shape that matched — tag `ratification-is-not-delivery`, because ADR ratification is not delivery evidence; (2) at least one shape matches AND the body has any unticked checkboxes — tag `multi-phase-mixed-progress`. The caveat field carries a single structured tag (architect condition C2), so trigger (1) wins when both would fire: "the evidence is the wrong kind" is the more fundamental reason the verdict cannot be clean. The verdict is intentionally conservative — tickets with no shape match AND no extractable evidence route to `SKIP`, not auto-close.
 
 **Surface-batch-confirm flow** (the methodology that produced today's 14 closes — codified for repeatable use):
 
@@ -344,7 +344,15 @@ Real-backlog smoke test 2026-05-31 against today's labeled fixtures: the "Proble
 
 #### 4.6b. Auto-close action per CLOSE-CANDIDATE
 
-For each `CLOSE-CANDIDATE` or `CLOSE-CANDIDATE-WITH-CAVEAT` ticket, perform the following BEFORE the `git mv`:
+**Applies to clean `CLOSE-CANDIDATE` verdicts.** A `CLOSE-CANDIDATE-WITH-CAVEAT`
+verdict does NOT enter this branch on its own: under AFK it is surfaced and left
+open per 4.6a step 3, and it reaches the steps below only after a maintainer has
+confirmed the close at the next interactive review — at which point it carries
+the extra **Caveat** field noted in step 1. Closing a caveat verdict
+unattended defeats the demotion that produced it (the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule reassessment
+2026-08-30, shape 2 corroborating-only).
+
+For each such ticket, perform the following BEFORE the `git mv`:
 
 1. Use the `Edit` tool to append a `## Closed as no longer relevant` section to the ticket body (cite + persist + uncertainty per the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule):
 
@@ -400,7 +408,7 @@ The relevance-close pass runs **unconditionally** during AFK orchestration: when
 
 **Worked example (Phase 2 surface-batch-confirm, 2026-05-31)**: 14 closes across 5 batches using shapes 2-5. Each batch surfaced via `AskUserQuestion` (≤ 5 candidates per batch); maintainer confirmed clean closes and routed caveat candidates with explicit caveat acknowledgement (e.g. the "Problem 039: Autonomous loops conflate diagnose with implement" problem `shared-template-not-built`; the "ADRs accumulate forward-chronology evidence inline (Phase 2 dogfood evidence, amendment history, cross-iter cross-references) — `decisions` bucket dominates context at 41% / 1.3 MiB" problem `deep-dive-bloat-remains`). All closures batched into per-batch commits per the "Governance Skills Commit Their Own Completed Work" architecture rule. The 14-fixture labeled set is the regression suite (`packages/itil/scripts/test/evaluate-relevance.bats` covers each shape positive + the architect A1/A2 advisory negatives).
 
-**Cross-references**: the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule (this pass's design ADR, Phase 1 + Phase 2), the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule (grounding, cumulative shape cite + structured caveat field), the "Problem lifecycle — add a Verification Pending status between Known Error and Closed" architecture rule + the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule lifecycle extension (Open|Known Error → Closed bypassing Verifying for no-fix-needed conclusions; the Closed-row entry at `/wr-itil:manage-problem` SKILL.md line 59 names Phase 1 + Phase 2 shapes), the "Plugin-bundled scripts invoked from SKILL.md resolve via `bin/` on `$PATH`" architecture rule (PATH shim), the "Behavioural-tests-default for skill testing" architecture rule (behavioural bats at `packages/itil/scripts/test/evaluate-relevance.bats` — 33/33 GREEN), the "Governance Skills Commit Their Own Completed Work" architecture rule (batched closure commit grain per pass), the "— Decision-Delegation Contract: when agents act on the framework vs ask the user" architecture rule cat 4 + the "Agents over-ask in interactive sessions — conflating mechanical-stages with user-interactive-stages of multi-stage skill contracts (inverse-)" problem (mechanical-stage carve-out: ask per-batch, not per-ticket), the "Problem 057: `git mv` + Edit + `git add` staging-ordering trap drops content edits from the commit" problem (staging trap), the "`/wr-itil:review-problems` has no path to close tickets that are no longer relevant (evidence-based, NOT age-based) — structural outflow gap drives monotonic backlog growth" problem (Phase 1 driver), the "Phase 2 — extend `evaluate-relevance.sh` with 4 more evidence shapes + fix Phase 1 file-no-longer-exists false-positive class" problem (Phase 2 driver).
+**Cross-references**: the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule (this pass's design ADR, Phase 1 + Phase 2), the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule (grounding, cumulative shape cite + structured caveat field), the "Problem lifecycle — add a Verification Pending status between Known Error and Closed" architecture rule + the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule lifecycle extension (Open|Known Error → Closed bypassing Verifying for no-fix-needed conclusions; the Closed-row entry at `/wr-itil:manage-problem` SKILL.md line 59 names Phase 1 + Phase 2 shapes), the "Plugin-bundled scripts invoked from SKILL.md resolve via `bin/` on `$PATH`" architecture rule (PATH shim), the "Behavioural-tests-default for skill testing" architecture rule (behavioural bats at `packages/itil/scripts/test/evaluate-relevance.bats` — 35/35 GREEN), the "Governance Skills Commit Their Own Completed Work" architecture rule (batched closure commit grain per pass), the "— Decision-Delegation Contract: when agents act on the framework vs ask the user" architecture rule cat 4 + the "Agents over-ask in interactive sessions — conflating mechanical-stages with user-interactive-stages of multi-stage skill contracts (inverse-)" problem (mechanical-stage carve-out: ask per-batch, not per-ticket), the "Problem 057: `git mv` + Edit + `git add` staging-ordering trap drops content edits from the commit" problem (staging trap), the "`/wr-itil:review-problems` has no path to close tickets that are no longer relevant (evidence-based, NOT age-based) — structural outflow gap drives monotonic backlog growth" problem (Phase 1 driver), the "Phase 2 — extend `evaluate-relevance.sh` with 4 more evidence shapes + fix Phase 1 file-no-longer-exists false-positive class" problem (Phase 2 driver).
 
 ### 5. Rewrite `docs/problems/README.md`
 

package/skills-codex/review-problems/SKILL.md

@@ -306,7 +306,7 @@ For each `.open.md` / `.known-error.md` ticket aged ≥ 7 days, evaluate whether
 | Shape | Phase | Mechanical check | Empirical closes (2026-05-31) |
 |---|---|---|---|
 | 1. `file-no-longer-exists` | Phase 1 | grep ticket body for `(packages\|docs\|...)/...\.(md\|sh\|...)`; verify each via `git ls-files --error-unmatch` | 0 of 14 |
-| 2. `ADR-shipped-confirmed` | Phase 2 | grep ticket body for `ADR-NNN`; for each, verify `docs/decisions/<NNN>-*.md` exists AND frontmatter has `human-oversight: confirmed` | 8 of 14 — the "Problem 012: Skill Testing Harness Scope Undefined" problem/the "Problem 015: TDD enforcement does not flag vague Gherkin outcome steps" problem/the "Problem 018: TDD plugin should enforce BDD + Example Mapping principles, with examples traceable to Jobs-To-Be-Done" problem/the "Problem 022: Agents must not fabricate time estimates without grounded data" problem/the "Problem 033: No persistent risk register for ISO 31000 / ISO 27001 compliance" problem/the "Problem 039: Autonomous loops conflate diagnose with implement" problem/the "ADRs accumulate forward-chronology evidence inline (Phase 2 dogfood evidence, amendment history, cross-iter cross-references) — `decisions` bucket dominates context at 41% / 1.3 MiB" problem/the "Reconcile release-cadence text with (lean / release-sooner; appetite is a ceiling, not a trigger) + add the dogfood-location-before-public clause" problem |
+| 2. `ADR-shipped-confirmed` **(corroborating-only)** | Phase 2 | grep ticket body for `ADR-NNN`; for each, verify `docs/decisions/<NNN>-*.md` exists AND frontmatter has `human-oversight: confirmed`. **Never a clean close on its own** — the marker records that a human ratified a DECISION (the ": Human-oversight marker + `/wr-architect:review-decisions` drain for recorded decisions" architecture rule), not that a fix shipped. Shape 2 alone demotes to `CLOSE-CANDIDATE-WITH-CAVEAT` with tag `ratification-is-not-delivery` (the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule reassessment 2026-08-30) | 8 of 14 — the "Problem 012: Skill Testing Harness Scope Undefined" problem/the "Problem 015: TDD enforcement does not flag vague Gherkin outcome steps" problem/the "Problem 018: TDD plugin should enforce BDD + Example Mapping principles, with examples traceable to Jobs-To-Be-Done" problem/the "Problem 022: Agents must not fabricate time estimates without grounded data" problem/the "Problem 033: No persistent risk register for ISO 31000 / ISO 27001 compliance" problem/the "Problem 039: Autonomous loops conflate diagnose with implement" problem/the "ADRs accumulate forward-chronology evidence inline (Phase 2 dogfood evidence, amendment history, cross-iter cross-references) — `decisions` bucket dominates context at 41% / 1.3 MiB" problem/the "Reconcile release-cadence text with (lean / release-sooner; appetite is a ceiling, not a trigger) + add the dogfood-location-before-public clause" problem |
 | 3. `named-skill-or-feature-exists` | Phase 2 | grep for SKILL.md / hook / agent paths + `/wr-<plugin>:<skill>` slash-command refs; verify each via `git ls-files` | 6 of 14 — the "Problem 014: No lightweight aside invocation for governance skills (problems, retros, ADRs)" problem/the "Problem 034: Centralise risk reports for cross-project skill improvement" problem/the "Problem 045: Auto plugin install on user's machine after governance release" problem/the "Problem 079: No inbound sync of upstream-reported problems — reports filed via the intake templates never surface in the local backlog" problem/the "Agent designs schemas with user-asked classification fields when the framework should derive silently OR eliminate the classification entirely — deeper generalisation of derive-first-don't-ask" problem/the "Broaden + rename the `solo-developer` persona → `developer`" problem |
 | 4. `self-marker-in-body` | Phase 2 | line-anchored grep for `Close to (Verifying\|Closed)`, `DONE 2026-`, `## Fix Released` heading, `fix shipped session`, `awaiting K→V`. Pattern MUST anchor to line-start to avoid mid-prose false-positives (architect advisory A2) | explicit in the "Broaden + rename the `solo-developer` persona → `developer`" problem; contributory in the "Problem 033: No persistent risk register for ISO 31000 / ISO 27001 compliance" problem |
 | 5. `driver-child-ticket-closed` | Phase 2 | parse `## Related` for `P<NNN>` refs; check if any are in `docs/problems/closed/`. Suppressed when child names an unbuilt SKILL/agent path (future work, not stale; architect advisory A1) | contributory in several closes |
@@ -336,13 +336,13 @@ Exit-code routing (one verdict line per ticket on stdout):
 | Exit | Stdout prefix | Action |
 |------|--------------|--------|
 | 0 | `CLOSE-CANDIDATE <basename> — shapes: <comma-list> — <per-shape cite>; ...` | Auto-close branch (4.6b). |
-| 0 | `CLOSE-CANDIDATE-WITH-CAVEAT <basename> — shapes: <comma-list> — caveat: <short-tag>: <one-line> — cites: ...` | Surface-batch-confirm branch (4.6b-with-caveat); the caveat short-tag + one-line splices verbatim into the audit section's **Caveat** field per architect condition C2. |
+| 0 | `CLOSE-CANDIDATE-WITH-CAVEAT <basename> — shapes: <comma-list> — caveat: <short-tag>: <one-line> — cites: ...` | Surface-batch-confirm branch — **never the auto-close branch**. Under AFK, defer per 4.6a step 3: surface it, do not `git mv` it. The audit-section shape in 4.6b applies only once a maintainer has confirmed the close, and the caveat short-tag + one-line then splices verbatim into that section's **Caveat** field per architect condition C2. |
 | 1 | `KEEP <basename> — <M>/<N> paths still present` | No action; log only. |
 | 1 | `KEEP-WITH-NOTE <basename> — <note>: <evidence>` | Phase 1 false-positive class (state-suffix / sibling-file / rename) OR architect-A1 future-work disambiguation. No action; log only. |
 | 2 | `SKIP <basename> — <reason>` | No action (age gate, no Reported date, no extractable evidence). |
 | 3 | error | Log advisory; do not abort the pass — relevance-close is non-blocking per the Step 4.5 fail-soft precedent. |
 
-**Algorithm (canonical body)**: runs each of the five shape detectors over the ticket body. Multi-shape matches emit cumulatively (corroborating evidence is stronger than first-match-wins per the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule): the `shapes:` field carries a comma-joined list, the trailing fragment carries per-shape cites semicolon-separated. The caveat fires when at least one shape matches AND the body has any unticked checkboxes (multi-phase mixed-progress umbrella class). The verdict is intentionally conservative — tickets with no shape match AND no extractable evidence route to `SKIP`, not auto-close.
+**Algorithm (canonical body)**: runs each of the five shape detectors over the ticket body. Multi-shape matches emit cumulatively (corroborating evidence is stronger than first-match-wins per the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule): the `shapes:` field carries a comma-joined list, the trailing fragment carries per-shape cites semicolon-separated. Two independent triggers raise the caveat, checked in this order: (1) shape 2 is the ONLY shape that matched — tag `ratification-is-not-delivery`, because ADR ratification is not delivery evidence; (2) at least one shape matches AND the body has any unticked checkboxes — tag `multi-phase-mixed-progress`. The caveat field carries a single structured tag (architect condition C2), so trigger (1) wins when both would fire: "the evidence is the wrong kind" is the more fundamental reason the verdict cannot be clean. The verdict is intentionally conservative — tickets with no shape match AND no extractable evidence route to `SKIP`, not auto-close.
 
 **Surface-batch-confirm flow** (the methodology that produced today's 14 closes — codified for repeatable use):
 
@@ -355,7 +355,15 @@ Real-backlog smoke test 2026-05-31 against today's labeled fixtures: the "Proble
 
 #### 4.6b. Auto-close action per CLOSE-CANDIDATE
 
-For each `CLOSE-CANDIDATE` or `CLOSE-CANDIDATE-WITH-CAVEAT` ticket, perform the following BEFORE the `git mv`:
+**Applies to clean `CLOSE-CANDIDATE` verdicts.** A `CLOSE-CANDIDATE-WITH-CAVEAT`
+verdict does NOT enter this branch on its own: under AFK it is surfaced and left
+open per 4.6a step 3, and it reaches the steps below only after a maintainer has
+confirmed the close at the next interactive review — at which point it carries
+the extra **Caveat** field noted in step 1. Closing a caveat verdict
+unattended defeats the demotion that produced it (the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule reassessment
+2026-08-30, shape 2 corroborating-only).
+
+For each such ticket, perform the following BEFORE the `git mv`:
 
 1. Use the `Edit` tool to append a `## Closed as no longer relevant` section to the ticket body (cite + persist + uncertainty per the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule):
 
@@ -411,7 +419,7 @@ The relevance-close pass runs **unconditionally** during AFK orchestration: when
 
 **Worked example (Phase 2 surface-batch-confirm, 2026-05-31)**: 14 closes across 5 batches using shapes 2-5. Each batch surfaced via `request_user_input` (≤ 5 candidates per batch); maintainer confirmed clean closes and routed caveat candidates with explicit caveat acknowledgement (e.g. the "Problem 039: Autonomous loops conflate diagnose with implement" problem `shared-template-not-built`; the "ADRs accumulate forward-chronology evidence inline (Phase 2 dogfood evidence, amendment history, cross-iter cross-references) — `decisions` bucket dominates context at 41% / 1.3 MiB" problem `deep-dive-bloat-remains`). All closures batched into per-batch commits per the "Governance Skills Commit Their Own Completed Work" architecture rule. The 14-fixture labeled set is the regression suite (`<itil-plugin-root>/scripts/test/evaluate-relevance.bats` covers each shape positive + the architect A1/A2 advisory negatives).
 
-**Cross-references**: the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule (this pass's design ADR, Phase 1 + Phase 2), the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule (grounding, cumulative shape cite + structured caveat field), the "Problem lifecycle — add a Verification Pending status between Known Error and Closed" architecture rule + the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule lifecycle extension (Open|Known Error → Closed bypassing Verifying for no-fix-needed conclusions; the Closed-row entry at `/wr-itil:manage-problem` SKILL.md line 59 names Phase 1 + Phase 2 shapes), the "Plugin-bundled scripts invoked from SKILL.md resolve via `bin/` on `$PATH`" architecture rule (PATH shim), the "Behavioural-tests-default for skill testing" architecture rule (behavioural bats at `<itil-plugin-root>/scripts/test/evaluate-relevance.bats` — 33/33 GREEN), the "Governance Skills Commit Their Own Completed Work" architecture rule (batched closure commit grain per pass), the "— Decision-Delegation Contract: when agents act on the framework vs ask the user" architecture rule cat 4 + the "Agents over-ask in interactive sessions — conflating mechanical-stages with user-interactive-stages of multi-stage skill contracts (inverse-)" problem (mechanical-stage carve-out: ask per-batch, not per-ticket), the "Problem 057: `git mv` + Edit + `git add` staging-ordering trap drops content edits from the commit" problem (staging trap), the "`/wr-itil:review-problems` has no path to close tickets that are no longer relevant (evidence-based, NOT age-based) — structural outflow gap drives monotonic backlog growth" problem (Phase 1 driver), the "Phase 2 — extend `evaluate-relevance.sh` with 4 more evidence shapes + fix Phase 1 file-no-longer-exists false-positive class" problem (Phase 2 driver).
+**Cross-references**: the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule (this pass's design ADR, Phase 1 + Phase 2), the "Agent output grounding — no fabricated quantitative estimates, cite + persist + uncertainty" architecture rule (grounding, cumulative shape cite + structured caveat field), the "Problem lifecycle — add a Verification Pending status between Known Error and Closed" architecture rule + the "Evidence-based relevance-close pass for the problem backlog (Phase 1: file-no-longer-exists; Phase 2: ADR-shipped-confirmed + named-skill-exists + self-marker-in-body + driver-child-closed + Phase 1 false-positive fixes)" architecture rule lifecycle extension (Open|Known Error → Closed bypassing Verifying for no-fix-needed conclusions; the Closed-row entry at `/wr-itil:manage-problem` SKILL.md line 59 names Phase 1 + Phase 2 shapes), the "Plugin-bundled scripts invoked from SKILL.md resolve via `bin/` on `$PATH`" architecture rule (PATH shim), the "Behavioural-tests-default for skill testing" architecture rule (behavioural bats at `<itil-plugin-root>/scripts/test/evaluate-relevance.bats` — 35/35 GREEN), the "Governance Skills Commit Their Own Completed Work" architecture rule (batched closure commit grain per pass), the "— Decision-Delegation Contract: when agents act on the framework vs ask the user" architecture rule cat 4 + the "Agents over-ask in interactive sessions — conflating mechanical-stages with user-interactive-stages of multi-stage skill contracts (inverse-)" problem (mechanical-stage carve-out: ask per-batch, not per-ticket), the "Problem 057: `git mv` + Edit + `git add` staging-ordering trap drops content edits from the commit" problem (staging trap), the "`/wr-itil:review-problems` has no path to close tickets that are no longer relevant (evidence-based, NOT age-based) — structural outflow gap drives monotonic backlog growth" problem (Phase 1 driver), the "Phase 2 — extend `evaluate-relevance.sh` with 4 more evidence shapes + fix Phase 1 file-no-longer-exists false-positive class" problem (Phase 2 driver).
 
 ### 5. Rewrite `docs/problems/README.md`