[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M75FJT77H2Q3JVWWYVJ4ZP07":3,"skill-same:01M75FJT77H2Q3JVWWYVJ4ZP07":43},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":13,"repositoryHtmlUrl":26,"repositoryHomepageUrl":6,"ownerHtmlUrl":27,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":28,"install":29,"resources":42},"01M75FJT77H2Q3JVWWYVJ4ZP07","workflow-retrospective",null,"Measure how the SDD workflow itself performed (rework cycles, blocked tasks, review tickets, quality-gate failures) and propose human-approved improvements to project workflow files.","aharada54914\u002Fsdd-forge","aharada54914","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F40755043?v=4","01M75FFC4V7WXZF0PK9RYWHNGF","01M75FFC5YJJ11B7MDCRVWB9Z6",0,"MIT","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Fplugins\u002Fsdd-quality-loop\u002Fskills\u002Fworkflow-retrospective\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: workflow-retrospective\ndescription: Measure how the SDD workflow itself performed (rework cycles, blocked tasks, review tickets, quality-gate failures) and propose human-approved improvements to project workflow files.\ndisable-model-invocation: false\nuser-invocable: false\n---\n\n# Workflow Retrospective\n\nObserve and improve the SDD workflow process itself, not the application code.\n\n## Invocation\n\nCodex:\n```\nUse the workflow-retrospective skill for specs\u002F\u003Cfeature>\n```\n\nClaude Code — internal skill, not directly invocable. Reached by following\nthis file from its entry command:\n\n```\n\u002Fsdd-ship:ship --retro specs\u002F\u003Cfeature>\u002Ftasks.md\n```\n\n## Metrics Collection (read-only)\n\nRead the following sources; do not modify them:\n\n- `reports\u002Fimplementation\u002F` — one report per task cycle\n- `reports\u002Fquality-gate\u002F` — one report per quality-gate run\n- `docs\u002Freview-tickets\u002F` — all YAML tickets for the feature\n- `git log --oneline` — commit history scoped to the feature path\n- `reports\u002Fspec-review\u002F` — spec-review-contract.json files for each feature\n- `reports\u002Ftask-review\u002F` — task-review-contract.json files for each feature\n- `reports\u002Fimpl-review\u002F` — impl-review-contract.json files for each feature\n- previous `reports\u002Fretrospective\u002F*.md` — comparison baseline and prior WFI\n  verification results\n- `reports\u002Fruns\u002FRUN-*.json` — machine-readable run records from prior runs\n  (attribution baseline for WFI verification)\n- `docs\u002Fworkflow-improvements\u002FWFI-*.md` — Applied WFIs with open verification\n  horizons\n- `docs\u002Fworkflow-improvements\u002Fretention-checklist.md` — recurrence conditions\n  for previously Verified WFIs\n\nFor each task derive:\n\n- **QG Cycles** — count of quality-gate reports for that task\n- **Task Attempts — read `Task Attempt Count`** from the latest\n  current-schema implementation report for the task.\n- **Review Rounds — count independent review rounds** recorded for the task\n  across its implementation and review evidence.\n- **Quality-Gate Runs — count quality-gate reports** for the task, including\n  failed and blocked runs rather than only the final passing run.\n- **Model Escalations — count complete escalation transitions** whose prior\n  tier, next tier, failure class, attempt number, and reason are all recorded.\n- **Blocked Count** — count of `Blocked` decisions across those reports\n- **Tickets** — count of review tickets by severity (`critical`, `major`, `minor`)\n- **Auto-fixed** — count of tickets where `auto_fix_allowed: true` and\n  `status: resolved`\n- **Outcome** — final task status (`Done` or still open)\n\nLegacy implementation reports without these additive fields contribute `N\u002FA`;\ndo not infer or fabricate missing attempt, review-round, or escalation values.\n\n### Deterministic artifact rules\n\nApply these rules before calculating the table. All paths below are canonical\nrepository-relative paths, and `\u003Cfeature>` and `T-NNN` must exactly match the\nrequested feature and task:\n\n1. **Implementation attempts.** The authoritative candidates are\n   `reports\u002Fimplementation\u002F\u003Cfeature>\u002FT-NNN.md` and\n   `reports\u002Fimplementation\u002F\u003Cfeature>\u002FT-NNN-attempt-\u003Cpositive integer>.md`.\n   The report heading task ID and its `Run ID` must agree with the path and be\n   non-empty. De-duplicate candidates with the same `(task ID, Run ID)` by\n   retaining the lexicographically smallest canonical path. Select the current\n   task report by the greatest numeric `Task Attempt Count`; break a tie by the\n   lexicographically greatest `Run ID`, then the lexicographically smallest\n   canonical path. The selected report supplies the cumulative Task Attempts\n   value. Schema-less legacy candidates sort below current-schema candidates\n   and contribute `N\u002FA` only when no current-schema candidate remains.\n2. **Independent review rounds.** The authoritative artifacts are\n   `reports\u002Fimplementation\u002F\u003Cfeature>\u002FT-NNN-review-\u003Cpositive integer>.md`.\n   Associate by the path task ID and require the report's `Task` identity to\n   match. The suffix is the round number. De-duplicate on\n   `(task ID, round number)` by retaining the lexicographically smallest\n   canonical path, then count the retained rounds in numeric round order.\n3. **Quality-gate runs.** The authoritative artifacts are Markdown files under\n   `reports\u002Fquality-gate\u002F` with exactly one task-identity line — the\n   canonical `Task ID: T-NNN` (quality-report template, WFI-020), or the\n   legacy `Task: T-NNN` form in reports that predate it — and one non-empty\n   `Run ID`. Associate only by that exact task identity.\n   De-duplicate on `(task ID, Run ID)` by retaining the lexicographically\n   smallest canonical path. Count retained reports, including PASS, FAIL, and\n   BLOCKED, ordered by `Run ID` and then canonical path.\n4. **Model escalations.** Read complete transition records from the retained\n   implementation, independent-review, and quality-gate artifacts. Associate\n   each record with the artifact's validated task identity. De-duplicate the\n   same transition across evidence sources on\n   `(task ID, escalation attempt number, prior tier, next tier, failure class)`.\n   Retain the record from the lexicographically smallest canonical path and\n   order by numeric escalation attempt, prior tier, next tier, failure class,\n   then path. Conflicting reasons for one de-duplication key make Data\n   Completeness `Blocked`; do not choose or count either conflicting record.\n5. **Invalid or ambiguous evidence.** A candidate with a mismatched identity,\n   malformed positive integer, duplicate identity with conflicting metric\n   values, or missing required association field is excluded and recorded in\n   the report's data-completeness explanation. Such evidence must never be\n   repaired from filenames, timestamps, chat history, or filesystem iteration\n   order.\n\nAlso derive dataset quality indicators:\n\n- **Sample Size** — number of tasks, review contracts, quality-gate reports, and\n  review tickets used.\n- **Data Completeness** — Complete when all expected report roots exist for the\n  feature; Partial when an optional source is missing; Blocked when required task\n  or quality-gate evidence is absent.\n- **Confidence** — High for recurring patterns across at least three tasks or\n  two independent evidence types; Medium for recurring patterns across two\n  tasks; Low for single-task observations. Low-confidence observations may be\n  reported but must not create a WFI.\n\n### Domain-drift metrics (when `domain\u002F` exists)\n\nWhen the project carries a `domain\u002F` directory, extend the metric roll-up\nwith domain-drift counts sourced only from `check-domain-conformance`\nfindings already recorded in `reports\u002Fquality-gate\u002F*.md` (the retained\nreports selected by rule 3 above). Do not re-run `check-domain-conformance`\nand do not read `domain\u002F` artifacts directly — this is a read-only rollup\nover already-recorded quality-gate text, not a new evidence-collection path.\nWhen `domain\u002F` is absent, skip this subsection entirely and omit the Domain\nDrift Metrics table from the report (do not emit a zero-filled table).\n\n1. Within each retained quality-gate report's text, locate every\n   `check-domain-conformance WARN (\u003Cn> finding(s)):` or\n   `check-domain-conformance FAILED (\u003Cn> finding(s)):` block (the exact\n   output format produced by\n   `plugins\u002Fsdd-quality-loop\u002Fscripts\u002Fcheck-domain-conformance.{sh,ps1}`) and\n   collect its `- \u003Cfinding text>` lines.\n2. Classify each finding line by matching its text against the fixed set of\n   messages the script can produce:\n   - **Term deviation** — lines matching `unrecognized term '...'`, or\n     `aggregate reference '...' not found in domain-contract.json\n     aggregates`, or `aggregate reference '...' has no domain\u002Faggregates\u002F\n     ....md card`. These all report a name\u002Fvocabulary the domain contract\n     does not recognize.\n   - **Boundary violation** — lines matching `Bounded-Context '...' not\n     found in domain-contract.json`, or `Bounded-Context lists two contexts\n     (...) with no declared relation in context map`. These report an\n     undeclared context or an undeclared cross-context relation (AC-015).\n   - A line matching neither pattern (e.g. a `requirements.md not found: ...`\n     \u002F `design.md not found: ...` input error) is not a drift finding; do not\n     count it in either bucket.\n3. **Term-Deviation Count** — total term-deviation finding lines across all\n   retained quality-gate reports for the feature in this period.\n4. **Boundary-Violation Count** — total boundary-violation finding lines\n   across the same reports.\n5. **Domain-Drift Trend** — compare this period's combined\n   (Term-Deviation Count + Boundary-Violation Count) against the previous\n   retrospective's combined total (from its Domain Drift Metrics table, when\n   one exists). Report `N\u002FA` when no previous retrospective recorded this\n   metric. A sustained increase across two or more consecutive retrospectives\n   is a friction pattern candidate under the Improvement Loop (Section 1),\n   subject to the same two-task recurrence and confidence rules as any other\n   friction pattern — never draft a WFI from a single period's counts alone.\n\nDe-duplicate on the same `(task ID, Run ID)` quality-gate report selection\nalready established in rule 3 of the Deterministic artifact rules above; a\nreport excluded there (e.g. a duplicate `Run ID`) does not contribute\ndomain-drift findings either.\n\nFor spec-review, task-review, and impl-review metrics, scan the contract files\nand derive:\n\n- **spec_review_rounds_per_feature** — for each feature, the round number of the\n  final passing spec-review-contract.json (rounds consumed to reach PASS).\n  If no PASS contract exists, record the maximum round reached.\n- **spec_review_blocked_rate** — percentage of features where the\n  spec-review-loop reached BLOCKED state (verdict == BLOCKED in the final\n  attempt's last round).\n\n- **task_review_rounds_per_feature** — for each feature, the round number of the\n  final passing task-review-contract.json (rounds consumed to reach PASS).\n  If no PASS contract exists, record the maximum round reached.\n- **task_review_blocked_rate** — percentage of features where the task-review-loop\n  reached BLOCKED state (verdict == BLOCKED in the final attempt's last round).\n- **impl_review_rounds_per_feature** — for each feature, the round number of the\n  final passing impl-review-contract.json.\n- **impl_review_blocked_rate** — percentage of features where the impl-review-loop\n  reached BLOCKED state.\n- **impl_review_legacy_design_rate** — percentage of features where\n  `legacy_design: true` in at least one impl-review-contract.json.\n\n## Applied-WFI Verification (horizon and retention)\n\nRun these two checks on every retrospective, before drafting any new WFI.\nThey are what makes self-improvement measurable instead of impressionistic:\nnever classify a WFI from the feeling that \"the run went smoother\" — only from\nrun records and report counts.\n\n**Horizon check.** For every WFI with `Status: Applied`:\n\n1. Read its `Verification Metric` block (`Target-Metric`, `Expected-Direction`,\n   `Horizon`, `Baseline`, `Target`).\n2. Compare the baseline against the current run records (`reports\u002Fruns\u002F`) and\n   this retrospective's metrics. Report counts, not percentages, when fewer\n   than 20 runs exist.\n3. Classify:\n   - Target met → set `Status: Verified`, append the comparison to `## Result`,\n     and add the WFI's recurrence condition to\n     `docs\u002Fworkflow-improvements\u002Fretention-checklist.md`.\n   - Direction improved, target unmet, horizon still open → leave Applied and\n     record `Needs-Followup` in `## Result`.\n   - Horizon expired without the target being met, or the metric worsened →\n     set `Status: Rejected`, append the evidence to `## Result`, and present\n     the WFI's `Rollback-Plan` to the human for approval. Do not execute the\n     rollback without human approval.\n\n**Retention check.** Read `docs\u002Fworkflow-improvements\u002Fretention-checklist.md`\n(if absent, skip and note that in the report). For each entry, check whether the\nrecurrence condition matches any evidence in this period. On recurrence:\nset the source WFI's `Status: Regressed`, record the evidence in its\n`## Result`, remove the checklist entry, and treat the recurrence as a friction\npattern candidate for a follow-up WFI. Retention answers the question most\nloops forget: do previously fixed failure modes stay fixed?\n\nBefore crediting or blaming any WFI, read the transcripts or reports behind the\nchanged metric: grading artifacts can masquerade as effects.\n\n## Output\n\nGenerate `reports\u002Fretrospective\u002F\u003Ctimestamp>.md` using the structure below.\nFill every section; do not leave any `{{}}` field blank.\n\n```markdown\n# Retrospective Report\n\n## Header\n\n| Field | Value |\n|---|---|\n| Feature | {{feature}} |\n| Period | {{period_start}} – {{period_end}} |\n| Generated | {{generated_timestamp}} |\n| Sample Size | {{task_count}} tasks, {{review_contract_count}} review contracts, {{qg_report_count}} QG reports, {{ticket_count}} tickets |\n| Data Completeness | {{complete_partial_blocked}} |\n| Confidence | {{high_medium_low_with_reason}} |\n\n## Metrics\n\n| Task | Task Attempts | Review Rounds | Quality-Gate Runs | Model Escalations | Blocked Count | Tickets (C\u002FM\u002FMin) | Outcome |\n|---|---|---|---|---|---|---|---|\n| {{task_id}} | {{task_attempts}} | {{review_rounds}} | {{quality_gate_runs}} | {{model_escalations}} | {{blocked_count}} | {{critical}}\u002F{{major}}\u002F{{minor}} | {{outcome}} |\n| **Total** | | | | | | | |\n\n_C = Critical, M = Major, Min = Minor_\n\n## Domain Drift Metrics\n\n_Include this section only when `domain\u002F` exists; omit entirely otherwise._\n\n| Metric | This Period | Previous Period | Trend |\n|---|---|---|---|\n| Term-Deviation Count | {{term_deviation_count}} | {{prev_term_deviation_count}} | {{trend}} |\n| Boundary-Violation Count | {{boundary_violation_count}} | {{prev_boundary_violation_count}} | {{trend}} |\n| Combined Domain-Drift Count | {{combined_drift_count}} | {{prev_combined_drift_count}} | {{trend}} |\n\nCounts are sourced only from `check-domain-conformance` findings already\nrecorded in the retained `reports\u002Fquality-gate\u002F*.md` reports for this\nfeature; see the Domain-drift metrics rules above. A sustained increase\nacross two or more consecutive retrospectives is a friction-pattern\ncandidate (see Friction Patterns \u002F Improvement Loop below), not an automatic\nWFI trigger.\n\n## Friction Patterns\n\nPatterns observed across two or more tasks in this period.\n\n### {{pattern_id}}: {{pattern_title}}\n\n- **Evidence:** {{affected_tasks_and_ticket_ids}}\n- **Frequency:** {{occurrence_count}} occurrences\n- **Phase:** {{workflow_phase}}\n- **Confidence:** {{high_medium_low}}\n- **Do Not Overfit:** {{why_this_is_not_a_single_task_exception}}\n\n## Proposed Improvements\n\n| WFI-ID | Status | Problem | Target File(s) |\n|---|---|---|---|\n| {{wfi_id}} | Draft | {{problem_summary}} | {{target_files}} |\n\n## Improvement Verification Plan\n\n| WFI-ID | Expected Effect Metric | Baseline | Target | Next Checkpoint |\n|---|---|---|---|---|\n| {{wfi_id}} | {{metric}} | {{baseline}} | {{target}} | {{next_feature_or_date}} |\n\n## Review Gate Metrics\n\n| Feature | Spec Review Rounds | Spec Review Verdict | Task Review Rounds | Task Review Verdict | Impl Review Rounds | Impl Review Verdict | Legacy Design |\n|---|---|---|---|---|---|---|---|\n| {{feature}} | {{spec_review_rounds}} | {{spec_review_verdict}} | {{task_review_rounds}} | {{task_review_verdict}} | {{impl_review_rounds}} | {{impl_review_verdict}} | {{legacy_design}} |\n\n## Comparison With Previous Retrospective\n\n| Metric | Previous | This Period | Trend |\n|---|---|---|---|\n| Avg QG Cycles per Task | {{prev_avg_qg}} | {{curr_avg_qg}} | {{trend}} |\n| Avg Task Attempts | {{prev_task_attempts}} | {{curr_task_attempts}} | {{trend}} |\n| Avg Review Rounds | {{prev_review_rounds}} | {{curr_review_rounds}} | {{trend}} |\n| Avg Quality-Gate Runs | {{prev_quality_gate_runs}} | {{curr_quality_gate_runs}} | {{trend}} |\n| Total Model Escalations | {{prev_model_escalations}} | {{curr_model_escalations}} | {{trend}} |\n| Total Blocked Count | {{prev_blocked}} | {{curr_blocked}} | {{trend}} |\n| Total Review Tickets | {{prev_tickets}} | {{curr_tickets}} | {{trend}} |\n| Auto-fix Rate | {{prev_autofix_pct}} | {{curr_autofix_pct}} | {{trend}} |\n| Avg Spec Review Rounds | {{prev_spec_review_rounds}} | {{curr_spec_review_rounds}} | {{trend}} |\n| Spec Review Blocked Rate | {{prev_spec_review_blocked}} | {{curr_spec_review_blocked}} | {{trend}} |\n| Avg Task Review Rounds | {{prev_task_review_rounds}} | {{curr_task_review_rounds}} | {{trend}} |\n| Task Review Blocked Rate | {{prev_task_review_blocked}} | {{curr_task_review_blocked}} | {{trend}} |\n| Avg Impl Review Rounds | {{prev_impl_review_rounds}} | {{curr_impl_review_rounds}} | {{trend}} |\n| Impl Review Blocked Rate | {{prev_impl_review_blocked}} | {{curr_impl_review_blocked}} | {{trend}} |\n| Impl Legacy Design Rate | {{prev_legacy_design_rate}} | {{curr_legacy_design_rate}} | {{trend}} |\n| Repeat Finding Rate | {{prev_repeat_finding_rate}} | {{curr_repeat_finding_rate}} | {{trend}} |\n| WFI Verification Rate | {{prev_wfi_verification_rate}} | {{curr_wfi_verification_rate}} | {{trend}} |\n\n_If no previous retrospective exists, mark all \"Previous\" cells as N\u002FA._\n\n## Applied WFI Horizon Check\n\n| WFI-ID | Target-Metric | Baseline | Target | Current | Horizon | Classification |\n|---|---|---|---|---|---|---|\n| {{wfi_id}} | {{target_metric}} | {{baseline}} | {{target}} | {{current}} | {{horizon}} | {{verified_needs_followup_rejected}} |\n\n_List every WFI that was `Status: Applied` at the start of this retrospective.\nIf none, state \"No applied WFIs with open horizons.\"_\n\n## Retention Check\n\n| Source WFI | Recurrence Condition | Recurred? | Evidence |\n|---|---|---|---|\n| {{wfi_id}} | {{condition}} | {{yes_no}} | {{evidence_or_dash}} |\n\n_One row per entry in retention-checklist.md. If the checklist is absent or\nempty, state that explicitly._\n```\n\n## Run Record (machine-readable)\n\nAfter writing the retrospective report, emit the machine-readable run record:\n\n```\nscripts\u002Femit-run-record.sh \u003Cfeature> --track \u003Cfull|lite> \\\n  --model-main \u003Cmodel id used for implementation> \\\n  --model-reviewers \u003Cmodel id used for review agents> \\\n  --plugin-version \u003Cinstalled sdd plugin version>\n```\n\n(or `emit-run-record.ps1` with the equivalent named parameters; if the project\nhas no `scripts\u002F` copy, run it from the installed sdd-quality-loop plugin's\n`scripts\u002F` directory).\n\nThe script deterministically counts tasks, first-try gate passes, gate reports,\nblocked reports, and review tickets, and records the currently Applied WFIs —\ndo not compute these numbers yourself and do not edit the emitted JSON. Pass\nthe model ids and track as arguments; they are per-run confound metadata that\nfuture attribution analysis cannot recover after the fact. If a model version\nchanged since the previous run record, note that in the retrospective report:\nmetric shifts across a model change must not be attributed to WFIs.\n\n## Improvement Loop\n\n1. **Identify friction.** Flag patterns that recur across at least two tasks:\n   - Same `type` of review ticket appears repeatedly.\n   - A phase produces `Blocked` more than once.\n   - Auto-fix rate drops below 50 % for a ticket type.\n   - A prior WFI's expected-effect metric does not improve by its next\n     checkpoint.\n\n   Do not draft a WFI from a single-task observation. Record it under Friction\n   Patterns with Low confidence only when it may become relevant later.\n\n1.5. **Classify the WFI.** Before drafting, walk the Section 1 flowchart of\n   `plugins\u002Fsdd-quality-loop\u002Freferences\u002Fwfi-category-guide.md` to set the\n   scope axis (`Category`), then set the mechanism axis (`Mechanism`) from\n   Section 5:\n\n   - **`measurement`** (checked first): the change touches graders, gate\n     thresholds, retrospective\u002Faudit logic, or run-record definitions. Forces\n     `Meta-Change: true` and the strict audit lane (category guide Section 5).\n   - **`human-process`**: the change alters approval policy, escalation rules,\n     or what humans review.\n   - **`plugin-improvement`**: friction evidence comes from the \"Review Gate\n     Metrics\" table (`spec_review_rounds`, `spec_review_blocked_rate`,\n     `impl_review_rounds`, `task_review_blocked_rate`,\n     `impl_review_blocked_rate`, `impl_review_legacy_design_rate`) or involves\n     cross-plugin handoff transitions (design review → task decomposition →\n     implementation flow). These WFIs are expressed in generic workflow terms\n     (see Section 2 of the category guide) and will be tracked as GitHub Issues.\n   - **`app-dev-efficiency`**: all other friction patterns (task sizing, test\n     coverage gaps, spec quality, project-specific recurring ticket types). These\n     WFIs use project-specific concrete language (feature slugs, task IDs, RT-IDs).\n\n   `Mechanism` is one of `instructions | memory | tools | architecture |\n   model-routing`. Set `Meta-Change: true` whenever the proposed change touches\n   anything that measures the workflow, regardless of Category.\n\n   For `plugin-improvement` WFIs: read Section 2 of `wfi-category-guide.md` and\n   apply all term substitutions to Root Cause Hypothesis, Proposed Change, and\n   Expected Effect. Problem Evidence may cite raw metric names from the retrospective.\n\n1.75. **Run the why-why analysis (なぜなぜ分析).** Before writing any Root Cause\n   Hypothesis, walk a 5-Whys chain from the observed friction down to a root\n   cause the Proposed Change can actually fix:\n\n   - Row 1 asks why the friction from Problem Evidence occurred; each\n     subsequent row asks why the previous row's answer occurred (chain rule).\n   - Go at least 3 levels deep, 5 when the chain supports it. Stop only when\n     the answer names a **controllable process or mechanism cause** — never at\n     a restated symptom, at \"human\u002Fagent error\", or at a cause outside the\n     workflow's control. Do not pad the chain past the root cause just to\n     reach 5 rows.\n   - Back each row with evidence (retrospective table row, RT-ID, BL-ID,\n     report path, or file:line). Mark rows you cannot corroborate as\n     `(hypothesis)`.\n   - The final row's answer becomes the `## Root Cause Hypothesis`, and the\n     `## Proposed Change` must act on that terminal cause — if the change\n     only treats an intermediate \"why\", either deepen the change or shorten\n     the chain to what the evidence supports.\n\n   Record the chain in the WFI's `## Why-Why Analysis` section (structure\n   below). The audit cycle rejects chains that restate symptoms, skip links,\n   or terminate at blame instead of mechanism (WHY-CHAIN-VALID check).\n\n2. **Draft a WFI.** For each identified friction, create\n   `docs\u002Fworkflow-improvements\u002FWFI-NNN.md` with `Status: Draft` using the\n   structure below. Increment NNN from the highest existing WFI number (start\n   at 001 if none exist). If the repository also runs an automated\n   self-improvement routine (e.g. it has `.github\u002Fself-improvement-prompt.md`),\n   check open issues labeled `self-improvement` first; if one covers the same\n   theme, reference its issue number in the WFI instead of duplicating the work.\n\n   ```markdown\n   # Workflow Improvement\n\n   ## WFI-ID\n\n   {{wfi_id}}\n\n   ## Status\n\n   Status: Draft\n\n   \u003C!-- Allowed values: Draft | Approved | Applied | Verified | Rejected | Regressed -->\n   \u003C!-- Only a human may set status to Approved. The AI sets Draft, Applied,        -->\n   \u003C!-- Verified, Rejected, and Regressed.                                           -->\n\n   ## Category\n\n   Category: {{plugin-improvement|app-dev-efficiency|human-process|measurement}}\n\n   \u003C!-- Scope axis — walk wfi-category-guide.md §1 flowchart.                        -->\n   \u003C!-- plugin-improvement: use generic workflow terms (wfi-category-guide.md §2). -->\n   \u003C!-- app-dev-efficiency: use project-specific detail (feature slug, task IDs).   -->\n   \u003C!-- measurement: forces Meta-Change: true (strict audit lane, guide §5).        -->\n\n   ## Mechanism\n\n   Mechanism: {{instructions|memory|tools|architecture|model-routing}}\n\n   \u003C!-- Mechanism axis — what kind of thing changes (wfi-category-guide.md §5). -->\n\n   ## Meta-Change\n\n   Meta-Change: {{true|false}}\n\n   \u003C!-- true when the change touches graders, thresholds, retrospective\u002Faudit     -->\n   \u003C!-- logic, or run-record definitions (wfi-category-guide.md §5).              -->\n\n   ## GitHub-Issue\n\n   GitHub-Issue: {{N\u002FA for app-dev-efficiency | Pending for plugin-improvement}}\n\n   \u003C!-- Populated by wfi-audit-cycle after gh issue create (plugin-improvement only). -->\n\n   ## Audit-Status\n\n   Audit-Status: Not-Started\n\n   \u003C!-- wfi-audit-cycle advances this through Cycle-1-In-Progress →              -->\n   \u003C!-- Cycle-2-In-Progress → Human-Pending before human review.                  -->\n\n   ## Problem Evidence\n\n   {{metrics_and_ticket_references}}\n\n   \u003C!-- Quote specific BL-IDs, RT-IDs, or retrospective table rows that show the friction. -->\n   \u003C!-- For plugin-improvement: raw metric field names are acceptable here.        -->\n\n   ## Why-Why Analysis\n\n   | # | Why (question) | Because (answer) | Evidence |\n   |---|---|---|---|\n   | 1 | Why did {{observed_friction}} occur? | {{because_1}} | {{evidence_ref_1}} |\n   | 2 | Why did {{because_1}} occur? | {{because_2}} | {{evidence_ref_2}} |\n   | 3 | Why did {{because_2}} occur? | {{because_3}} | {{evidence_ref_3}} |\n\n   \u003C!-- 5-Whys chain from step 1.75. At least 3 levels; each Because is the      -->\n   \u003C!-- next row's Why subject; final Because = Root Cause Hypothesis mechanism. -->\n   \u003C!-- Never terminate at a restated symptom, \"human\u002Fagent error\", or an        -->\n   \u003C!-- uncontrollable cause. Cite evidence per row; mark unproven rows           -->\n   \u003C!-- \"(hypothesis)\". Why\u002FBecause prose follows the category language rules;   -->\n   \u003C!-- the Evidence column may cite raw metric names like Problem Evidence.     -->\n\n   ## Root Cause Hypothesis\n\n   {{root_cause}}\n\n   \u003C!-- Must be the terminal cause of the ## Why-Why Analysis chain.             -->\n   \u003C!-- plugin-improvement: use generic workflow terms only (wfi-category-guide.md §2). -->\n   \u003C!-- app-dev-efficiency: name the specific feature\u002Ftask\u002Fticket driving the issue. -->\n\n   ## Proposed Change\n\n   | Target File | Change Description |\n   |---|---|\n   | {{file_path}} | {{change_description}} |\n\n   \u003C!-- Target files must be project-side workflow files only (AGENTS.md, CLAUDE.md,    -->\n   \u003C!-- specs\u002F templates, task-splitting guidelines). Plugin files must not be listed.  -->\n\n   ## Expected Effect\n\n   {{expected_effect}}\n\n   \u003C!-- State the metric(s) expected to improve and by how much.                       -->\n   \u003C!-- plugin-improvement: use generic metric names from wfi-category-guide.md §2.   -->\n   \u003C!-- app-dev-efficiency: use project-specific metric names with concrete targets.   -->\n\n   ## Verification Metric\n\n   Target-Metric: {{run_record_metric_key}}\n   Expected-Direction: {{increase|decrease}}\n   Horizon: {{within_next_N_runs_or_date}}\n   Baseline: {{baseline_count}}\n   Target: {{target_count}}\n\n   \u003C!-- Target-Metric names one run-record key (reports\u002Fruns\u002FRUN-*.json) or a       -->\n   \u003C!-- retrospective table column. Horizon is binding: expiry without the target    -->\n   \u003C!-- met → Status: Rejected + rollback proposal. Use counts, not percentages.     -->\n\n   ## Verification Plan\n\n   {{verification_plan}}\n\n   \u003C!-- Describe how the next task cycle will confirm the improvement. -->\n   \u003C!-- Reference the specific metric rows from the retrospective template. -->\n\n   ## Rollback-Plan\n\n   {{rollback_plan}}\n\n   \u003C!-- Exact files\u002Fsections changed and how to revert (normally: git revert of  -->\n   \u003C!-- the commit whose message contains this WFI-ID). Required before Approved. -->\n\n   ## Result\n\n   Pending\n\n   \u003C!-- Fill after the next task cycle completes. Append a comparison table. -->\n   ```\n\n2.5. **Trigger audit cycle.** Immediately after creating the WFI Draft, invoke\n   the `wfi-audit-cycle` skill for every new WFI regardless of category:\n\n   Claude Code:\n   ```\n   \u002Fsdd-quality-loop:wfi-audit-cycle WFI-NNN\n   ```\n   Codex:\n   ```\n   Use the wfi-audit-cycle skill for WFI-NNN\n   ```\n\n   The audit cycle runs 2 independent review cycles (Cycle 1: proposal quality;\n   Cycle 2: impact and risk), revises the WFI between cycles, and sets\n   `Audit-Status: Human-Pending` when complete.\n\n   For `plugin-improvement` WFIs, `wfi-audit-cycle` also creates a GitHub Issue\n   after both cycles complete.\n\n   Record the WFI in the retrospective report under \"Proposed Improvements\" only\n   after the audit cycle completes (i.e., after `Audit-Status: Human-Pending` is\n   set). Use these status labels in the table:\n\n   | Category | Status label in table |\n   |---|---|\n   | `plugin-improvement` | `Human-Pending (audited, issue created)` |\n   | `app-dev-efficiency` | `Human-Pending (audited)` |\n\n3. **Await human Approved.** Do not apply any improvement until a human sets\n   the WFI `status` to `Approved`.  Record the pending WFI references in the\n   retrospective report under \"Proposed Improvements\".\n\n4. **Apply approved improvements.** Modify only project-side workflow files:\n   - `AGENTS.md` or `CLAUDE.md` in the project root\n   - `specs\u002F` template files or task-splitting guidelines\n   - Task granularity guidance documents\n\n   Do not modify installed plugin files (skills, references, templates inside\n   `plugins\u002F`).\n\n   Always include the WFI ID (`WFI-NNN`) in the commit message and any PR that\n   applies an approved improvement.  Automated maintenance routines treat\n   `WFI-`-referenced changes as protected and will not revert them; changes\n   without the marker lose that protection. The WFI-ID-in-commit rule is also\n   what makes the `Rollback-Plan` executable: reverting the WFI is reverting\n   that commit.\n\n   **Adoption hygiene (attribution discipline):** apply at most one WFI (or one\n   explicitly labeled batch) per verification window, and never mid-run.\n   Simultaneous unlabeled adoptions make it impossible to attribute a metric\n   shift to any single change.\n\n5. **Verify effect.** After the next task cycle completes, re-run metrics\n   collection (the Applied-WFI Verification section above) and append a\n   `Result` section to the WFI document. Compare the `Verification Metric`\n   baseline and target with the current run records. Classify the result as:\n   - `Verified`: target met without introducing a worse repeated friction.\n     Register the recurrence condition in\n     `docs\u002Fworkflow-improvements\u002Fretention-checklist.md`.\n   - `Needs-Followup`: direction improved but target was not met and the\n     horizon is still open.\n   - `Rejected`: metric worsened, the horizon expired without the target, or\n     the proposal caused a new repeated failure. Present the `Rollback-Plan`\n     to the human.\n   - `Regressed` (set by the retention check, possibly many cycles later): a\n     previously Verified WFI's failure mode recurred.\n\n   This mirrors continuous eval practice: keep the successful regression signal,\n   promote recurring misses into new WFI candidates, and avoid reactive changes\n   from one-off failures. A model-version change between baseline and\n   checkpoint invalidates the comparison — note it and extend the horizon\n   instead of classifying.\n\n## Sudo Mode\n\nA valid `SDD_SUDO` flag does **not** bypass WFI approval. Setting a WFI `status`\nto `Approved` changes the SDD workflow itself (governance), so it remains a\nhuman-only action even under sudo: continue to await human `Approved` before\napplying any improvement. The WFI records status as an inline `Status: \u003Cvalue>`\nfield; a hook guard denies any agent edit introducing `Status: Approved` in a\n`docs\u002Fworkflow-improvements\u002FWFI-*.md` file, and is never bypassed by sudo.\nRead-only metrics collection runs as normal. See\n`plugins\u002Fsdd-quality-loop\u002Freferences\u002Fsudo-mode-policy.md`.\n\n## Boundaries\n\n- Do not modify application code.\n- Do not change task status fields.\n- Do not commit or push unless the human explicitly requests it.\n- Do not create, resolve, or modify review tickets.\n- Do not invoke `quality-gate` or `fix-by-review-ticket`.\n","# Workflow Retrospective\n\nObserve and improve the SDD workflow process itself, not the application code.\n\n## Invocation\n\nCodex:\n```\nUse the workflow-retrospective skill for specs\u002F\u003Cfeature>\n```\n\nClaude Code — internal skill, not directly invocable. Reached by following\nthis file from its entry command:\n\n```\n\u002Fsdd-ship:ship --retro specs\u002F\u003Cfeature>\u002Ftasks.md\n```\n\n## Metrics Collection (read-only)\n\nRead the following sources; do not modify them:\n\n- `reports\u002Fimplementation\u002F` — one report per task cycle\n- `reports\u002Fquality-gate\u002F` — one report per quality-gate run\n- `docs\u002Freview-tickets\u002F` — all YAML tickets for the feature\n- `git log --oneline` — commit history scoped to the feature path\n- `reports\u002Fspec-review\u002F` — spec-review-contract.json files for each feature\n- `reports\u002Ftask-review\u002F` — task-review-contract.json files for each feature\n- `reports\u002Fimpl-review\u002F` — impl-review-contract.json files for each feature\n- previous `reports\u002Fretrospective\u002F*.md` — comparison baseline and prior WFI\n  verification results\n- `reports\u002Fruns\u002FRUN-*.json` — machine-readable run records from prior runs\n  (attribution baseline for WFI verification)\n- `docs\u002Fworkflow-improvements\u002FWFI-*.md` — Applied WFIs with open verification\n  horizons\n- `docs\u002Fworkflow-improvements\u002Fretention-checklist.md` — recurrence conditions\n  for previously Verified WFIs\n\nFor each task derive:\n\n- **QG Cycles** — count of quality-gate reports for that task\n- **Task Attempts — read `Task Attempt Count`** from the latest\n  current-schema implementation report for the task.\n- **Review Rounds — count independent review rounds** recorded for the task\n  across its implementation and review evidence.\n- **Quality-Gate Runs — count quality-gate reports** for the task, including\n  failed and blocked runs rather than only the final passing run.\n- **Model Escalations — count complete escalation transitions** whose prior\n  tier, next tier, failure class, attempt number, and reason are all recorded.\n- **Blocked Count** — count of `Blocked` decisions across those reports\n- **Tickets** — count of review tickets by severity (`critical`, `major`, `minor`)\n- **Auto-fixed** — count of tickets where `auto_fix_allowed: true` and\n  `status: resolved`\n- **Outcome** — final task status (`Done` or still open)\n\nLegacy implementation reports without these additive fields contribute `N\u002FA`;\ndo not infer or fabricate missing attempt, review-round, or escalation values.\n\n### Deterministic artifact rules\n\nApply these rules before calculating the table. All paths below are canonical\nrepository-relative paths, and `\u003Cfeature>` and `T-NNN` must exactly match the\nrequested feature and task:\n\n1. **Implementation attempts.** The authoritative candidates are\n   `reports\u002Fimplementation\u002F\u003Cfeature>\u002FT-NNN.md` and\n   `reports\u002Fimplementation\u002F\u003Cfeature>\u002FT-NNN-attempt-\u003Cpositive integer>.md`.\n   The report heading task ID and its `Run ID` must agree with the path and be\n   non-empty. De-duplicate candidates with the same `(task ID, Run ID)` by\n   retaining the lexicographically smallest canonical path. Select the current\n   task report by the greatest numeric `Task Attempt Count`; break a tie by the\n   lexicographically greatest `Run ID`, then the lexicographically smallest\n   canonical path. The selected report supplies the cumulative Task Attempts\n   value. Schema-less legacy candidates sort below current-schema candidates\n   and contribute `N\u002FA` only when no current-schema candidate remains.\n2. **Independent review rounds.** The authoritative artifacts are\n   `reports\u002Fimplementation\u002F\u003Cfeature>\u002FT-NNN-review-\u003Cpositive integer>.md`.\n   Associate by the path task ID and require the report's `Task` identity to\n   match. The suffix is the round number. De-duplicate on\n   `(task ID, round number)` by retaining the lexicographically smallest\n   canonical path, then count the retained rounds in numeric round order.\n3. **Quality-gate runs.** The authoritative artifacts are Markdown files under\n   `reports\u002Fquality-gate\u002F` with exactly one task-identity line — the\n   canonical `Task ID: T-NNN` (quality-report template, WFI-020), or the\n   legacy `Task: T-NNN` form in reports that predate it — and one non-empty\n   `Run ID`. Associate only by that exact task identity.\n   De-duplicate on `(task ID, Run ID)` by retaining the lexicographically\n   smallest canonical path. Count retained reports, including PASS, FAIL, and\n   BLOCKED, ordered by `Run ID` and then canonical path.\n4. **Model escalations.** Read complete transition records from the retained\n   implementation, independent-review, and quality-gate artifacts. Associate\n   each record with the artifact's validated task identity. De-duplicate the\n   same transition across evidence sources on\n   `(task ID, escalation attempt number, prior tier, next tier, failure class)`.\n   Retain the record from the lexicographically smallest canonical path and\n   order by numeric escalation attempt, prior tier, next tier, failure class,\n   then path. Conflicting reasons for one de-duplication key make Data\n   Completeness `Blocked`; do not choose or count either conflicting record.\n5. **Invalid or ambiguous evidence.** A candidate with a mismatched identity,\n   malformed positive integer, duplicate identity with conflicting metric\n   values, or missing required association field is excluded and recorded in\n   the report's data-completeness explanation. Such evidence must never be\n   repaired from filenames, timestamps, chat history, or filesystem iteration\n   order.\n\nAlso derive dataset quality indicators:\n\n- **Sample Size** — number of tasks, review contracts, quality-gate reports, and\n  review tickets used.\n- **Data Completeness** — Complete when all expected report roots exist for the\n  feature; Partial when an optional source is missing; Blocked when required task\n  or quality-gate evidence is absent.\n- **Confidence** — High for recurring patterns across at least three tasks or\n  two independent evidence types; Medium for recurring patterns across two\n  tasks; Low for single-task observations. Low-confidence observations may be\n  reported but must not create a WFI.\n\n### Domain-drift metrics (when `domain\u002F` exists)\n\nWhen the project carries a `domain\u002F` directory, extend the metric roll-up\nwith domain-drift counts sourced only from `check-domain-conformance`\nfindings already recorded in `reports\u002Fquality-gate\u002F*.md` (the retained\nreports selected by rule 3 above). Do not re-run `check-domain-conformance`\nand do not read `domain\u002F` artifacts directly — this is a read-only rollup\nover already-recorded quality-gate text, not a new evidence-collection path.\nWhen `domain\u002F` is absent, skip this subsection entirely and omit the Domain\nDrift Metrics table from the report (do not emit a zero-filled table).\n\n1. Within each retained quality-gate report's text, locate every\n   `check-domain-conformance WARN (\u003Cn> finding(s)):` or\n   `check-domain-conformance FAILED (\u003Cn> finding(s)):` block (the exact\n   output format produced by\n   `plugins\u002Fsdd-quality-loop\u002Fscripts\u002Fcheck-domain-conformance.{sh,ps1}`) and\n   collect its `- \u003Cfinding text>` lines.\n2. Classify each finding line by matching its text against the fixed set of\n   messages the script can produce:\n   - **Term deviation** — lines matching `unrecognized term '...'`, or\n     `aggregate reference '...' not found in domain-contract.json\n     aggregates`, or `aggregate reference '...' has no domain\u002Faggregates\u002F\n     ....md card`. These all report a name\u002Fvocabulary the domain contract\n     does not recognize.\n   - **Boundary violation** — lines matching `Bounded-Context '...' not\n     found in domain-contract.json`, or `Bounded-Context lists two contexts\n     (...) with no declared relation in context map`. These report an\n     undeclared context or an undeclared cross-context relation (AC-015).\n   - A line matching neither pattern (e.g. a `requirements.md not found: ...`\n     \u002F `design.md not found: ...` input error) is not a drift finding; do not\n     count it in either bucket.\n3. **Term-Deviation Count** — total term-deviation finding lines across all\n   retained quality-gate reports for the feature in this period.\n4. **Boundary-Violation Count** — total boundary-violation finding lines\n   across the same reports.\n5. **Domain-Drift Trend** — compare this period's combined\n   (Term-Deviation Count + Boundary-Violation Count) against the previous\n   retrospective's combined total (from its Domain Drift Metrics table, when\n   one exists). Report `N\u002FA` when no previous retrospective recorded this\n   metric. A sustained increase across two or more consecutive retrospectives\n   is a friction pattern candidate under the Improvement Loop (Section 1),\n   subject to the same two-task recurrence and confidence rules as any other\n   friction pattern — never draft a WFI from a single period's counts alone.\n\nDe-duplicate on the same `(task ID, Run ID)` quality-gate report selection\nalready established in rule 3 of the Deterministic artifact rules above; a\nreport excluded there (e.g. a duplicate `Run ID`) does not contribute\ndomain-drift findings either.\n\nFor spec-review, task-review, and impl-review metrics, scan the contract files\nand derive:\n\n- **spec_review_rounds_per_feature** — for each feature, the round number of the\n  final passing spec-review-contract.json (rounds consumed to reach PASS).\n  If no PASS contract exists, record the maximum round reached.\n- **spec_review_blocked_rate** — percentage of features where the\n  spec-review-loop reached BLOCKED state (verdict == BLOCKED in the final\n  attempt's last round).\n\n- **task_review_rounds_per_feature** — for each feature, the round number of the\n  final passing task-review-contract.json (rounds consumed to reach PASS).\n  If no PASS contract exists, record the maximum round reached.\n- **task_review_blocked_rate** — percentage of features where the task-review-loop\n  reached BLOCKED state (verdict == BLOCKED in the final attempt's last round).\n- **impl_review_rounds_per_feature** — for each feature, the round number of the\n  final passing impl-review-contract.json.\n- **impl_review_blocked_rate** — percentage of features where the impl-review-loop\n  reached BLOCKED state.\n- **impl_review_legacy_design_rate** — percentage of features where\n  `legacy_design: true` in at least one impl-review-contract.json.\n\n## Applied-WFI Verification (horizon and retention)\n\nRun these two checks on every retrospective, before drafting any new WFI.\nThey are what makes self-improvement measurable instead of impressionistic:\nnever classify a WFI from the feeling that \"the run went smoother\" — only from\nrun records and report counts.\n\n**Horizon check.** For every WFI with `Status: Applied`:\n\n1. Read its `Verification Metric` block (`Target-Metric`, `Expected-Direction`,\n   `Horizon`, `Baseline`, `Target`).\n2. Compare the baseline against the current run records (`reports\u002Fruns\u002F`) and\n   this retrospective's metrics. Report counts, not percentages, when fewer\n   than 20 runs exist.\n3. Classify:\n   - Target met → set `Status: Verified`, append the comparison to `## Result`,\n     and add the WFI's recurrence condition to\n     `docs\u002Fworkflow-improvements\u002Fretention-checklist.md`.\n   - Direction improved, target unmet, horizon still open → leave Applied and\n     record `Needs-Followup` in `## Result`.\n   - Horizon expired without the target being met, or the metric worsened →\n     set `Status: Rejected`, append the evidence to `## Result`, and present\n     the WFI's `Rollback-Plan` to the human for approval. Do not execute the\n     rollback without human approval.\n\n**Retention check.** Read `docs\u002Fworkflow-improvements\u002Fretention-checklist.md`\n(if absent, skip and note that in the report). For each entry, check whether the\nrecurrence condition matches any evidence in this period. On recurrence:\nset the source WFI's `Status: Regressed`, record the evidence in its\n`## Result`, remove the checklist entry, and treat the recurrence as a friction\npattern candidate for a follow-up WFI. Retention answers the question most\nloops forget: do previously fixed failure modes stay fixed?\n\nBefore crediting or blaming any WFI, read the transcripts or reports behind the\nchanged metric: grading artifacts can masquerade as effects.\n\n## Output\n\nGenerate `reports\u002Fretrospective\u002F\u003Ctimestamp>.md` using the structure below.\nFill every section; do not leave any `{{}}` field blank.\n\n```markdown\n# Retrospective Report\n\n## Header\n\n| Field | Value |\n|---|---|\n| Feature | {{feature}} |\n| Period | {{period_start}} – {{period_end}} |\n| Generated | {{generated_timestamp}} |\n| Sample Size | {{task_count}} tasks, {{review_contract_count}} review contracts, {{qg_report_count}} QG reports, {{ticket_count}} tickets |\n| Data Completeness | {{complete_partial_blocked}} |\n| Confidence | {{high_medium_low_with_reason}} |\n\n## Metrics\n\n| Task | Task Attempts | Review Rounds | Quality-Gate Runs | Model Escalations | Blocked Count | Tickets (C\u002FM\u002FMin) | Outcome |\n|---|---|---|---|---|---|---|---|\n| {{task_id}} | {{task_attempts}} | {{review_rounds}} | {{quality_gate_runs}} | {{model_escalations}} | {{blocked_count}} | {{critical}}\u002F{{major}}\u002F{{minor}} | {{outcome}} |\n| **Total** | | | | | | | |\n\n_C = Critical, M = Major, Min = Minor_\n\n## Domain Drift Metrics\n\n_Include this section only when `domain\u002F` exists; omit entirely otherwise._\n\n| Metric | This Period | Previous Period | Trend |\n|---|---|---|---|\n| Term-Deviation Count | {{term_deviation_count}} | {{prev_term_deviation_count}} | {{trend}} |\n| Boundary-Violation Count | {{boundary_violation_count}} | {{prev_boundary_violation_count}} | {{trend}} |\n| Combined Domain-Drift Count | {{combined_drift_count}} | {{prev_combined_drift_count}} | {{trend}} |\n\nCounts are sourced only from `check-domain-conformance` findings already\nrecorded in the retained `reports\u002Fquality-gate\u002F*.md` reports for this\nfeature; see the Domain-drift metrics rules above. A sustained increase\nacross two or more consecutive retrospectives is a friction-pattern\ncandidate (see Friction Patterns \u002F Improvement Loop below), not an automatic\nWFI trigger.\n\n## Friction Patterns\n\nPatterns observed across two or more tasks in this period.\n\n### {{pattern_id}}: {{pattern_title}}\n\n- **Evidence:** {{affected_tasks_and_ticket_ids}}\n- **Frequency:** {{occurrence_count}} occurrences\n- **Phase:** {{workflow_phase}}\n- **Confidence:** {{high_medium_low}}\n- **Do Not Overfit:** {{why_this_is_not_a_single_task_exception}}\n\n## Proposed Improvements\n\n| WFI-ID | Status | Problem | Target File(s) |\n|---|---|---|---|\n| {{wfi_id}} | Draft | {{problem_summary}} | {{target_files}} |\n\n## Improvement Verification Plan\n\n| WFI-ID | Expected Effect Metric | Baseline | Target | Next Checkpoint |\n|---|---|---|---|---|\n| {{wfi_id}} | {{metric}} | {{baseline}} | {{target}} | {{next_feature_or_date}} |\n\n## Review Gate Metrics\n\n| Feature | Spec Review Rounds | Spec Review Verdict | Task Review Rounds | Task Review Verdict | Impl Review Rounds | Impl Review Verdict | Legacy Design |\n|---|---|---|---|---|---|---|---|\n| {{feature}} | {{spec_review_rounds}} | {{spec_review_verdict}} | {{task_review_rounds}} | {{task_review_verdict}} | {{impl_review_rounds}} | {{impl_review_verdict}} | {{legacy_design}} |\n\n## Comparison With Previous Retrospective\n\n| Metric | Previous | This Period | Trend |\n|---|---|---|---|\n| Avg QG Cycles per Task | {{prev_avg_qg}} | {{curr_avg_qg}} | {{trend}} |\n| Avg Task Attempts | {{prev_task_attempts}} | {{curr_task_attempts}} | {{trend}} |\n| Avg Review Rounds | {{prev_review_rounds}} | {{curr_review_rounds}} | {{trend}} |\n| Avg Quality-Gate Runs | {{prev_quality_gate_runs}} | {{curr_quality_gate_runs}} | {{trend}} |\n| Total Model Escalations | {{prev_model_escalations}} | {{curr_model_escalations}} | {{trend}} |\n| Total Blocked Count | {{prev_blocked}} | {{curr_blocked}} | {{trend}} |\n| Total Review Tickets | {{prev_tickets}} | {{curr_tickets}} | {{trend}} |\n| Auto-fix Rate | {{prev_autofix_pct}} | {{curr_autofix_pct}} | {{trend}} |\n| Avg Spec Review Rounds | {{prev_spec_review_rounds}} | {{curr_spec_review_rounds}} | {{trend}} |\n| Spec Review Blocked Rate | {{prev_spec_review_blocked}} | {{curr_spec_review_blocked}} | {{trend}} |\n| Avg Task Review Rounds | {{prev_task_review_rounds}} | {{curr_task_review_rounds}} | {{trend}} |\n| Task Review Blocked Rate | {{prev_task_review_blocked}} | {{curr_task_review_blocked}} | {{trend}} |\n| Avg Impl Review Rounds | {{prev_impl_review_rounds}} | {{curr_impl_review_rounds}} | {{trend}} |\n| Impl Review Blocked Rate | {{prev_impl_review_blocked}} | {{curr_impl_review_blocked}} | {{trend}} |\n| Impl Legacy Design Rate | {{prev_legacy_design_rate}} | {{curr_legacy_design_rate}} | {{trend}} |\n| Repeat Finding Rate | {{prev_repeat_finding_rate}} | {{curr_repeat_finding_rate}} | {{trend}} |\n| WFI Verification Rate | {{prev_wfi_verification_rate}} | {{curr_wfi_verification_rate}} | {{trend}} |\n\n_If no previous retrospective exists, mark all \"Previous\" cells as N\u002FA._\n\n## Applied WFI Horizon Check\n\n| WFI-ID | Target-Metric | Baseline | Target | Current | Horizon | Classification |\n|---|---|---|---|---|---|---|\n| {{wfi_id}} | {{target_metric}} | {{baseline}} | {{target}} | {{current}} | {{horizon}} | {{verified_needs_followup_rejected}} |\n\n_List every WFI that was `Status: Applied` at the start of this retrospective.\nIf none, state \"No applied WFIs with open horizons.\"_\n\n## Retention Check\n\n| Source WFI | Recurrence Condition | Recurred? | Evidence |\n|---|---|---|---|\n| {{wfi_id}} | {{condition}} | {{yes_no}} | {{evidence_or_dash}} |\n\n_One row per entry in retention-checklist.md. If the checklist is absent or\nempty, state that explicitly._\n```\n\n## Run Record (machine-readable)\n\nAfter writing the retrospective report, emit the machine-readable run record:\n\n```\nscripts\u002Femit-run-record.sh \u003Cfeature> --track \u003Cfull|lite> \\\n  --model-main \u003Cmodel id used for implementation> \\\n  --model-reviewers \u003Cmodel id used for review agents> \\\n  --plugin-version \u003Cinstalled sdd plugin version>\n```\n\n(or `emit-run-record.ps1` with the equivalent named parameters; if the project\nhas no `scripts\u002F` copy, run it from the installed sdd-quality-loop plugin's\n`scripts\u002F` directory).\n\nThe script deterministically counts tasks, first-try gate passes, gate reports,\nblocked reports, and review tickets, and records the currently Applied WFIs —\ndo not compute these numbers yourself and do not edit the emitted JSON. Pass\nthe model ids and track as arguments; they are per-run confound metadata that\nfuture attribution analysis cannot recover after the fact. If a model version\nchanged since the previous run record, note that in the retrospective report:\nmetric shifts across a model change must not be attributed to WFIs.\n\n## Improvement Loop\n\n1. **Identify friction.** Flag patterns that recur across at least two tasks:\n   - Same `type` of review ticket appears repeatedly.\n   - A phase produces `Blocked` more than once.\n   - Auto-fix rate drops below 50 % for a ticket type.\n   - A prior WFI's expected-effect metric does not improve by its next\n     checkpoint.\n\n   Do not draft a WFI from a single-task observation. Record it under Friction\n   Patterns with Low confidence only when it may become relevant later.\n\n1.5. **Classify the WFI.** Before drafting, walk the Section 1 flowchart of\n   `plugins\u002Fsdd-quality-loop\u002Freferences\u002Fwfi-category-guide.md` to set the\n   scope axis (`Category`), then set the mechanism axis (`Mechanism`) from\n   Section 5:\n\n   - **`measurement`** (checked first): the change touches graders, gate\n     thresholds, retrospective\u002Faudit logic, or run-record definitions. Forces\n     `Meta-Change: true` and the strict audit lane (category guide Section 5).\n   - **`human-process`**: the change alters approval policy, escalation rules,\n     or what humans review.\n   - **`plugin-improvement`**: friction evidence comes from the \"Review Gate\n     Metrics\" table (`spec_review_rounds`, `spec_review_blocked_rate`,\n     `impl_review_rounds`, `task_review_blocked_rate`,\n     `impl_review_blocked_rate`, `impl_review_legacy_design_rate`) or involves\n     cross-plugin handoff transitions (design review → task decomposition →\n     implementation flow). These WFIs are expressed in generic workflow terms\n     (see Section 2 of the category guide) and will be tracked as GitHub Issues.\n   - **`app-dev-efficiency`**: all other friction patterns (task sizing, test\n     coverage gaps, spec quality, project-specific recurring ticket types). These\n     WFIs use project-specific concrete language (feature slugs, task IDs, RT-IDs).\n\n   `Mechanism` is one of `instructions | memory | tools | architecture |\n   model-routing`. Set `Meta-Change: true` whenever the proposed change touches\n   anything that measures the workflow, regardless of Category.\n\n   For `plugin-improvement` WFIs: read Section 2 of `wfi-category-guide.md` and\n   apply all term substitutions to Root Cause Hypothesis, Proposed Change, and\n   Expected Effect. Problem Evidence may cite raw metric names from the retrospective.\n\n1.75. **Run the why-why analysis (なぜなぜ分析).** Before writing any Root Cause\n   Hypothesis, walk a 5-Whys chain from the observed friction down to a root\n   cause the Proposed Change can actually fix:\n\n   - Row 1 asks why the friction from Problem Evidence occurred; each\n     subsequent row asks why the previous row's answer occurred (chain rule).\n   - Go at least 3 levels deep, 5 when the chain supports it. Stop only when\n     the answer names a **controllable process or mechanism cause** — never at\n     a restated symptom, at \"human\u002Fagent error\", or at a cause outside the\n     workflow's control. Do not pad the chain past the root cause just to\n     reach 5 rows.\n   - Back each row with evidence (retrospective table row, RT-ID, BL-ID,\n     report path, or file:line). Mark rows you cannot corroborate as\n     `(hypothesis)`.\n   - The final row's answer becomes the `## Root Cause Hypothesis`, and the\n     `## Proposed Change` must act on that terminal cause — if the change\n     only treats an intermediate \"why\", either deepen the change or shorten\n     the chain to what the evidence supports.\n\n   Record the chain in the WFI's `## Why-Why Analysis` section (structure\n   below). The audit cycle rejects chains that restate symptoms, skip links,\n   or terminate at blame instead of mechanism (WHY-CHAIN-VALID check).\n\n2. **Draft a WFI.** For each identified friction, create\n   `docs\u002Fworkflow-improvements\u002FWFI-NNN.md` with `Status: Draft` using the\n   structure below. Increment NNN from the highest existing WFI number (start\n   at 001 if none exist). If the repository also runs an automated\n   self-improvement routine (e.g. it has `.github\u002Fself-improvement-prompt.md`),\n   check open issues labeled `self-improvement` first; if one covers the same\n   theme, reference its issue number in the WFI instead of duplicating the work.\n\n   ```markdown\n   # Workflow Improvement\n\n   ## WFI-ID\n\n   {{wfi_id}}\n\n   ## Status\n\n   Status: Draft\n\n   \u003C!-- Allowed values: Draft | Approved | Applied | Verified | Rejected | Regressed -->\n   \u003C!-- Only a human may set status to Approved. The AI sets Draft, Applied,        -->\n   \u003C!-- Verified, Rejected, and Regressed.                                           -->\n\n   ## Category\n\n   Category: {{plugin-improvement|app-dev-efficiency|human-process|measurement}}\n\n   \u003C!-- Scope axis — walk wfi-category-guide.md §1 flowchart.                        -->\n   \u003C!-- plugin-improvement: use generic workflow terms (wfi-category-guide.md §2). -->\n   \u003C!-- app-dev-efficiency: use project-specific detail (feature slug, task IDs).   -->\n   \u003C!-- measurement: forces Meta-Change: true (strict audit lane, guide §5).        -->\n\n   ## Mechanism\n\n   Mechanism: {{instructions|memory|tools|architecture|model-routing}}\n\n   \u003C!-- Mechanism axis — what kind of thing changes (wfi-category-guide.md §5). -->\n\n   ## Meta-Change\n\n   Meta-Change: {{true|false}}\n\n   \u003C!-- true when the change touches graders, thresholds, retrospective\u002Faudit     -->\n   \u003C!-- logic, or run-record definitions (wfi-category-guide.md §5).              -->\n\n   ## GitHub-Issue\n\n   GitHub-Issue: {{N\u002FA for app-dev-efficiency | Pending for plugin-improvement}}\n\n   \u003C!-- Populated by wfi-audit-cycle after gh issue create (plugin-improvement only). -->\n\n   ## Audit-Status\n\n   Audit-Status: Not-Started\n\n   \u003C!-- wfi-audit-cycle advances this through Cycle-1-In-Progress →              -->\n   \u003C!-- Cycle-2-In-Progress → Human-Pending before human review.                  -->\n\n   ## Problem Evidence\n\n   {{metrics_and_ticket_references}}\n\n   \u003C!-- Quote specific BL-IDs, RT-IDs, or retrospective table rows that show the friction. -->\n   \u003C!-- For plugin-improvement: raw metric field names are acceptable here.        -->\n\n   ## Why-Why Analysis\n\n   | # | Why (question) | Because (answer) | Evidence |\n   |---|---|---|---|\n   | 1 | Why did {{observed_friction}} occur? | {{because_1}} | {{evidence_ref_1}} |\n   | 2 | Why did {{because_1}} occur? | {{because_2}} | {{evidence_ref_2}} |\n   | 3 | Why did {{because_2}} occur? | {{because_3}} | {{evidence_ref_3}} |\n\n   \u003C!-- 5-Whys chain from step 1.75. At least 3 levels; each Because is the      -->\n   \u003C!-- next row's Why subject; final Because = Root Cause Hypothesis mechanism. -->\n   \u003C!-- Never terminate at a restated symptom, \"human\u002Fagent error\", or an        -->\n   \u003C!-- uncontrollable cause. Cite evidence per row; mark unproven rows           -->\n   \u003C!-- \"(hypothesis)\". Why\u002FBecause prose follows the category language rules;   -->\n   \u003C!-- the Evidence column may cite raw metric names like Problem Evidence.     -->\n\n   ## Root Cause Hypothesis\n\n   {{root_cause}}\n\n   \u003C!-- Must be the terminal cause of the ## Why-Why Analysis chain.             -->\n   \u003C!-- plugin-improvement: use generic workflow terms only (wfi-category-guide.md §2). -->\n   \u003C!-- app-dev-efficiency: name the specific feature\u002Ftask\u002Fticket driving the issue. -->\n\n   ## Proposed Change\n\n   | Target File | Change Description |\n   |---|---|\n   | {{file_path}} | {{change_description}} |\n\n   \u003C!-- Target files must be project-side workflow files only (AGENTS.md, CLAUDE.md,    -->\n   \u003C!-- specs\u002F templates, task-splitting guidelines). Plugin files must not be listed.  -->\n\n   ## Expected Effect\n\n   {{expected_effect}}\n\n   \u003C!-- State the metric(s) expected to improve and by how much.                       -->\n   \u003C!-- plugin-improvement: use generic metric names from wfi-category-guide.md §2.   -->\n   \u003C!-- app-dev-efficiency: use project-specific metric names with concrete targets.   -->\n\n   ## Verification Metric\n\n   Target-Metric: {{run_record_metric_key}}\n   Expected-Direction: {{increase|decrease}}\n   Horizon: {{within_next_N_runs_or_date}}\n   Baseline: {{baseline_count}}\n   Target: {{target_count}}\n\n   \u003C!-- Target-Metric names one run-record key (reports\u002Fruns\u002FRUN-*.json) or a       -->\n   \u003C!-- retrospective table column. Horizon is binding: expiry without the target    -->\n   \u003C!-- met → Status: Rejected + rollback proposal. Use counts, not percentages.     -->\n\n   ## Verification Plan\n\n   {{verification_plan}}\n\n   \u003C!-- Describe how the next task cycle will confirm the improvement. -->\n   \u003C!-- Reference the specific metric rows from the retrospective template. -->\n\n   ## Rollback-Plan\n\n   {{rollback_plan}}\n\n   \u003C!-- Exact files\u002Fsections changed and how to revert (normally: git revert of  -->\n   \u003C!-- the commit whose message contains this WFI-ID). Required before Approved. -->\n\n   ## Result\n\n   Pending\n\n   \u003C!-- Fill after the next task cycle completes. Append a comparison table. -->\n   ```\n\n2.5. **Trigger audit cycle.** Immediately after creating the WFI Draft, invoke\n   the `wfi-audit-cycle` skill for every new WFI regardless of category:\n\n   Claude Code:\n   ```\n   \u002Fsdd-quality-loop:wfi-audit-cycle WFI-NNN\n   ```\n   Codex:\n   ```\n   Use the wfi-audit-cycle skill for WFI-NNN\n   ```\n\n   The audit cycle runs 2 independent review cycles (Cycle 1: proposal quality;\n   Cycle 2: impact and risk), revises the WFI between cycles, and sets\n   `Audit-Status: Human-Pending` when complete.\n\n   For `plugin-improvement` WFIs, `wfi-audit-cycle` also creates a GitHub Issue\n   after both cycles complete.\n\n   Record the WFI in the retrospective report under \"Proposed Improvements\" only\n   after the audit cycle completes (i.e., after `Audit-Status: Human-Pending` is\n   set). Use these status labels in the table:\n\n   | Category | Status label in table |\n   |---|---|\n   | `plugin-improvement` | `Human-Pending (audited, issue created)` |\n   | `app-dev-efficiency` | `Human-Pending (audited)` |\n\n3. **Await human Approved.** Do not apply any improvement until a human sets\n   the WFI `status` to `Approved`.  Record the pending WFI references in the\n   retrospective report under \"Proposed Improvements\".\n\n4. **Apply approved improvements.** Modify only project-side workflow files:\n   - `AGENTS.md` or `CLAUDE.md` in the project root\n   - `specs\u002F` template files or task-splitting guidelines\n   - Task granularity guidance documents\n\n   Do not modify installed plugin files (skills, references, templates inside\n   `plugins\u002F`).\n\n   Always include the WFI ID (`WFI-NNN`) in the commit message and any PR that\n   applies an approved improvement.  Automated maintenance routines treat\n   `WFI-`-referenced changes as protected and will not revert them; changes\n   without the marker lose that protection. The WFI-ID-in-commit rule is also\n   what makes the `Rollback-Plan` executable: reverting the WFI is reverting\n   that commit.\n\n   **Adoption hygiene (attribution discipline):** apply at most one WFI (or one\n   explicitly labeled batch) per verification window, and never mid-run.\n   Simultaneous unlabeled adoptions make it impossible to attribute a metric\n   shift to any single change.\n\n5. **Verify effect.** After the next task cycle completes, re-run metrics\n   collection (the Applied-WFI Verification section above) and append a\n   `Result` section to the WFI document. Compare the `Verification Metric`\n   baseline and target with the current run records. Classify the result as:\n   - `Verified`: target met without introducing a worse repeated friction.\n     Register the recurrence condition in\n     `docs\u002Fworkflow-improvements\u002Fretention-checklist.md`.\n   - `Needs-Followup`: direction improved but target was not met and the\n     horizon is still open.\n   - `Rejected`: metric worsened, the horizon expired without the target, or\n     the proposal caused a new repeated failure. Present the `Rollback-Plan`\n     to the human.\n   - `Regressed` (set by the retention check, possibly many cycles later): a\n     previously Verified WFI's failure mode recurred.\n\n   This mirrors continuous eval practice: keep the successful regression signal,\n   promote recurring misses into new WFI candidates, and avoid reactive changes\n   from one-off failures. A model-version change between baseline and\n   checkpoint invalidates the comparison — note it and extend the horizon\n   instead of classifying.\n\n## Sudo Mode\n\nA valid `SDD_SUDO` flag does **not** bypass WFI approval. Setting a WFI `status`\nto `Approved` changes the SDD workflow itself (governance), so it remains a\nhuman-only action even under sudo: continue to await human `Approved` before\napplying any improvement. The WFI records status as an inline `Status: \u003Cvalue>`\nfield; a hook guard denies any agent edit introducing `Status: Approved` in a\n`docs\u002Fworkflow-improvements\u002FWFI-*.md` file, and is never bypassed by sudo.\nRead-only metrics collection runs as normal. See\n`plugins\u002Fsdd-quality-loop\u002Freferences\u002Fsudo-mode-policy.md`.\n\n## Boundaries\n\n- Do not modify application code.\n- Do not change task status fields.\n- Do not commit or push unless the human explicitly requests it.\n- Do not create, resolve, or modify review tickets.\n- Do not invoke `quality-gate` or `fix-by-review-ticket`.\n",{"name":5,"description":7,"user-invocable":19,"disable-model-invocation":19},"main","0f4e3268b5403855008ae089d3c1611430a41987","plugins\u002Fsdd-quality-loop\u002Fskills\u002Fworkflow-retrospective","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge","https:\u002F\u002Fgithub.com\u002Faharada54914","2026-11-15 02:49:38",{"source":8,"skillName":5,"commands":30,"prompt":40,"zipAvailable":41},[31,34,37],{"tool":32,"command":33},"npx","npx skills add aharada54914\u002Fsdd-forge --skill workflow-retrospective",{"tool":35,"command":36},"bunx","bunx skills add aharada54914\u002Fsdd-forge --skill workflow-retrospective",{"tool":38,"command":39},"pnpm","pnpm dlx skills add aharada54914\u002Fsdd-forge --skill workflow-retrospective","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Fplugins\u002Fsdd-quality-loop\u002Fskills\u002Fworkflow-retrospective\u002FSKILL.md",true,[],[44,52,59,66,73,80],{"skillNo":45,"name":46,"nameZh":6,"description":47,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":48,"categories":49,"keywords":50,"publishedAt":51},"01M7DY6PWMGRM0GFNBXAE2RP3K","status","Show a read-only Jira-style six-view SDD dashboard in the initiating parent's Sites or Claude Artifacts, with source freshness and failure visibility.","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Fplugins\u002Fsdd-ship\u002Fskills\u002Fstatus\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":53,"name":54,"nameZh":6,"description":55,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":18},"01M75FN16WRENJ2A9CXZCXVKBQ","verify-review-reservation","Recheck the bounded identity-ledger reservation invariant after lock, hash-check or publication changes.","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fproject-skills\u002Fformal-review-reservation\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":18},"01M75FMW90YZDCB7QNJ8K5KGKG","aws-cdk","Authors, deploys, and troubleshoots AWS infrastructure using CDK with TypeScript or Python. Covers best practices, stack architecture, and construct patterns. Applies when writing CDK constructs, bootstrapping environments, running cdk deploy\u002Fsynth\u002Fdiff, fixing CDK or CloudFormation errors, planning stack structure, importing existing resources, resolving drift, or refactoring stacks without resource replacement.","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcloud-orchestration\u002Faws-bundle\u002Fupstream\u002Fplugins\u002Faws-core\u002Fskills\u002Faws-cdk\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":18},"01M75FMSTFNT7EQSZY81T4WRHN","quality-gate","Independently verify an Implementation Complete task with a Default-FAIL verification contract, deterministic checks, and an isolated critical reviewer, then decide whether it is Done.","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Fspecs\u002Fsdd-domain-multitarget\u002Fverification\u002FT-003\u002Fattempt10-20261016\u002Fbaseline\u002Fplugins\u002Fsdd-quality-loop\u002Fskills\u002Fquality-gate\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":18},"01M75FMPM8FDE4NQ6FFY7KB83Y","ship","Implement approved SDD tasks through the quality gate to Done — the second of the two-command workflow. Run after \u002Fsdd-bootstrap:bootstrap and human task approval.","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Fspecs\u002Fquality-loop-fixes\u002Fhuman-copy\u002Fplugins\u002Fsdd-ship\u002Fskills\u002Fship\u002FSKILL.md",[],[],{"skillNo":81,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":18},"01M75FMK3G0EPXW77XK8ED3F2E","https:\u002F\u002Fgithub.com\u002Faharada54914\u002Fsdd-forge\u002Fblob\u002Fmain\u002Fspecs\u002Fmcp-readonly-preflight\u002Fhuman-copy\u002Fplugins\u002Fsdd-ship\u002Fskills\u002Fship\u002FSKILL.md",[],[]]