skillpacks 0.1.0

package/packs/agentic-skills-bench/claude/benchmark-test-skill/SKILL.md

@@ -0,0 +1,151 @@
+---
+name: benchmark-test-skill
+description: Run verify and benchmark tests for one agentic-skills skill, producing pass-rate, latency, cost, and consistency metrics
+type: execution
+version: v0.1
+argument-hint: "<skill name>"
+---
+
+# Benchmark Test Skill
+
+Invoke as `/benchmark-test-skill <skill>`.
+
+Use this skill when the user wants to benchmark-test a skill defined in this repository. The trailing argument is the skill under test, not a mode for that skill. For example, `/benchmark-test-skill design-system` tests the `design-system` skill with the harness; it does not run `design-system` against the current app or website.
+
+Run the agentic-skills test harness verification gate followed by the benchmark extension for a single skill. By default, benchmark both Claude and Codex runners and report them separately.
+
+Produce deterministic benchmark evidence only. It should hand off to `/benchmark-agent-review <skill>` as a separate step when the user needs subjective ergonomic judgment or remediation planning for the generated skill outputs.
+
+## Input
+
+- Required: one skill name, such as `design-system`.
+- If no skill name is provided, ask the user which skill to benchmark-test.
+- When invoked as `/benchmark-test-skill <skill>`, resolve `benchmark-test-skill` as the active command first, including this project-local pack path, before treating the trailing argument as the skill under test.
+- Do not reinterpret the trailing argument as the active workflow unless the benchmark-test-skill command cannot be found after checking project-local packs.
+
+## Execution
+
+Run commands from `/Users/georgele/projects/tools/agentic-skills/tests`.
+
+### Step 0 - Command Resolution Guard
+
+Confirm the active workflow is this `packs/agentic-skills-bench` skill. The requested target skill is data for the benchmark harness, not a command to execute directly.
+
+- For command-like invocations, preserve the leading command and route by the pack command first.
+- If a target such as `design-system`, `run`, or `ship` is provided, never run that skill directly as the benchmark action.
+- If command resolution is ambiguous, stop and report the ambiguity instead of running the target skill.
+
+### Step 1 - Eligibility Preflight
+
+Before running verify, check whether the requested skill is a repository skill known to the benchmark harness:
+
+```bash
+pnpm bench --list-skills
+```
+
+- If `<SKILL>` is not listed, stop immediately and report `unknown skill: <SKILL>`.
+- List the known repository skills from the command output.
+- Do not run `pnpm verify` or `pnpm bench` for unknown skills.
+- Read and report the listed coverage status for `<SKILL>`: `custom`, `generic`, or `blocked`.
+- Skills with custom layer4 setups use skill-specific fixtures and hard assertions. Some setups also include deterministic output-quality rubrics.
+- Skills without custom layer4 setups use the harness generic smoke benchmark. Treat that as invocation/compliance evidence, not deep domain-quality evidence.
+- If the row is `blocked`, stop before verify and bench. Report the blocked reason and next command from the list output.
+- If the row is `generic`, continue only as generic smoke evidence and route missing custom coverage to `/targeted-skill-builder <SKILL> benchmark coverage`.
+
+### Step 2 - Verify
+
+```bash
+pnpm verify --skill <SKILL>
+```
+
+- Expect layer1 to pass and layer2 to pass when target-specific tests exist.
+- Treat layer1 as the static harness-contract gate, including basic benchmark setup alignment checks such as expected next-route handoffs, runner command conventions, output file expectations, and quality-rubric facts against the current skill contract.
+- If layer1 fails because a benchmark setup is misaligned with the skill contract, classify it as a harness/benchmark coverage defect and route to `/targeted-skill-builder <SKILL> benchmark failure`; do not spend agent budget on `pnpm bench`.
+- If layer2 reports no tests matched `<SKILL>`, treat that layer as skipped and continue to the benchmark step. Record the skip clearly because generic benchmark coverage is weaker than target-specific layer2 verification.
+- Record pass/fail and wall time per layer.
+- If verify fails, stop and report the failure. Do not run the benchmark step.
+
+### Step 3 - Bench
+
+Run only if verify passes:
+
+```bash
+pnpm bench --skill <SKILL> --agent both --runs 3 --chunk-size 3 --pause 0
+```
+
+- Use 3 iterations by default.
+- Use `--agent both` by default. Only use `--agent claude` or `--agent codex` when the user explicitly asks to isolate one runner.
+- The expected budget is about $1 per run for the current design-system variant; report actual cost from the benchmark output when available.
+- The bench system persists raw data to `tests/benchmarks/runs/<skill>-<agent>-<sessionId>/` and generates `report.json`.
+- Treat rate limits, quota exhaustion, and similar runner-capacity errors as infrastructure-blocked runs, not skill failures. Report them separately from evaluated pass rate.
+- If recent same-skill benchmark reports or triage reports show repeated same-family benchmark false negatives, do not recommend another blind rerun as the next step. Report the recurrence and route to `/targeted-skill-builder <SKILL> benchmark repeated false-negative generalization` so the harness/setup gets a family-level semantic evaluator, fixture set, or infrastructure classifier instead of another one-off tolerance patch.
+
+### Step 4 - Report
+
+Write results to `benchmark/test-<SKILL>-<YYYY-MM-DD>.md` at the repository root. Use the current date.
+
+Populate the report from `report.json` and verify the output includes:
+
+- verify table with layer status and wall time
+- agent name, evaluated pass rate, blocked-run count, and Wilson 95% confidence interval
+- failed assertions, if any
+- output-quality score summary when the setup defines a quality evaluator, including threshold failures, critical failures, and lowest-scoring criteria when present
+- infrastructure-blocked runs, if any
+- latency p50, p95, and p99
+- cost per run and total cost
+- mean pairwise similarity and outlier count
+- raw session path
+- recommended next route, using a literal label such as `Recommended next command:` or `Recommended next skill:`
+
+After writing the report, verify the file exists and contains the benchmark target, agent rows, pass-rate or blocked-run data, latency, cost, and raw session path. If any required report field is missing, treat the workflow as incomplete and fix the report before marking the benchmark done.
+
+## Output
+
+Print a concise benchmark summary:
+
+- verify pass/fail
+- hard assertion pass rate
+- output-quality score when present, labeled as an additional rubric score rather than a statistical confidence measure
+- infrastructure-blocked count, if any
+- p50 latency
+- total cost
+- report path
+- next review handoff when subjective output-quality judgment or remediation planning is still needed
+
+The Markdown report and the final assistant response must both include a literal next-route label accepted by the harness, such as `Recommended next skill: /benchmark-agent-review <skill>` or `Recommended next command: /ship`.
+
+## Alignment Page
+
+When this skill produces durable deliverables (research, specs, plans, reports, prototypes, or any document output), build a full-depth HTML alignment page following `ALIGNMENT-PAGE.md` in this skill's directory. Output: `alignment/benchmark-test-skill-{topic}.html`.
+
+## Constraints
+
+- Do not audit or benchmark the app, website, docs, or product surface unless the user explicitly asks for a separate website/product benchmark workflow.
+- Do not run `pnpm bench` when `pnpm verify` fails.
+- Do not fabricate benchmark metrics. Use the command output and `report.json`.
+- Do not present quality score as a replacement for hard assertion pass rate, or present a small benchmark run as statistically definitive.
+- Do not omit the final next-step route. Completion output must include either `Recommended next skill: <command>` or the two-line pair `**Next work:** <specific task or "none">` and `**Recommended next command:** <one command or route>`.
+- Do not create or modify GitHub Actions workflows.
+
+## Next-Step Routing
+
+If the skill is unknown to the repository, recommend checking the skill name or creating the skill first with `/create-agentic-skill`.
+
+If the skill has blocked benchmark coverage, recommend the row's `next_command`.
+
+If the skill only has generic smoke benchmark coverage or otherwise lacks custom domain-quality assertions, recommend `/targeted-skill-builder <skill> benchmark coverage`.
+
+If the skill fails verification, hard benchmark assertions, or configured quality thresholds, recommend `/session-triage <skill> benchmark failure`.
+
+If benchmark runs are blocked only by rate limits or quota exhaustion, recommend re-running `/benchmark-test-skill <skill>` after the reset instead of treating the skill as failed.
+
+If the failure pattern has already been triaged as a repeated same-family benchmark false negative, recommend `/targeted-skill-builder <skill> benchmark repeated false-negative generalization` instead of another `/benchmark-test-skill <skill>` rerun.
+
+If evaluated benchmark runs completed and subjective output-quality review or remediation planning has not yet been performed, recommend `/benchmark-agent-review <skill>` as the next separate step.
+
+If the skill passes, the report is written, and no subjective review is needed or the separate `/benchmark-agent-review <skill>` step has already been completed, recommend `/ship`.
+
+## Default Shipping Contract
+
+Follow the shared shipping contract convention in CLAUDE.md.
+

package/packs/agentic-skills-bench/claude/benchmark-test-skill/archive/v0.0/SKILL.md

@@ -0,0 +1,167 @@
+---
+name: benchmark-test-skill
+description: Run verify and benchmark tests for one agentic-skills skill, producing pass-rate, latency, cost, and consistency metrics
+type: execution
+version: v0.0
+argument-hint: "<skill name>"
+---
+
+# Benchmark Test Skill
+
+Invoke as `/benchmark-test-skill <skill>`.
+
+Use this skill when the user wants to benchmark-test a skill defined in this repository. The trailing argument is the skill under test, not a mode for that skill. For example, `/benchmark-test-skill design-system` tests the `design-system` skill with the harness; it does not run `design-system` against the current app or website.
+
+This skill runs the agentic-skills test harness verification gate followed by the benchmark extension for a single skill. By default, benchmark both Claude and Codex runners and report them separately.
+
+This skill produces deterministic benchmark evidence only. It should hand off to `/benchmark-agent-review <skill>` as a separate step when the user needs subjective ergonomic judgment or remediation planning for the generated skill outputs.
+
+## Input
+
+- Required: one skill name, such as `design-system`.
+- If no skill name is provided, ask the user which skill to benchmark-test.
+- When invoked as `/benchmark-test-skill <skill>`, resolve `benchmark-test-skill` as the active command first, including this project-local pack path, before treating the trailing argument as the skill under test.
+- Do not reinterpret the trailing argument as the active workflow unless the benchmark-test-skill command cannot be found after checking project-local packs.
+
+## Execution
+
+Run commands from `/Users/georgele/projects/tools/agentic-skills/tests`.
+
+### Step 0 - Command Resolution Guard
+
+Confirm the active workflow is this `packs/agentic-skills-bench` skill. The requested target skill is data for the benchmark harness, not a command to execute directly.
+
+- For command-like invocations, preserve the leading command and route by the pack command first.
+- If a target such as `design-system`, `run`, or `ship` is provided, never run that skill directly as the benchmark action.
+- If command resolution is ambiguous, stop and report the ambiguity instead of running the target skill.
+
+### Step 1 - Eligibility Preflight
+
+Before running verify, check whether the requested skill is a repository skill known to the benchmark harness:
+
+```bash
+pnpm bench --list-skills
+```
+
+- If `<SKILL>` is not listed, stop immediately and report `unknown skill: <SKILL>`.
+- List the known repository skills from the command output.
+- Do not run `pnpm verify` or `pnpm bench` for unknown skills.
+- Read and report the listed coverage status for `<SKILL>`: `custom`, `generic`, or `blocked`.
+- Skills with custom layer4 setups use skill-specific fixtures and hard assertions. Some setups also include deterministic output-quality rubrics.
+- Skills without custom layer4 setups use the harness generic smoke benchmark. Treat that as invocation/compliance evidence, not deep domain-quality evidence.
+- If the row is `blocked`, stop before verify and bench. Report the blocked reason and next command from the list output.
+- If the row is `generic`, continue only as generic smoke evidence and route missing custom coverage to `/targeted-skill-builder <SKILL> benchmark coverage`.
+
+### Step 2 - Verify
+
+```bash
+pnpm verify --skill <SKILL>
+```
+
+- Expect layer1 to pass and layer2 to pass when target-specific tests exist.
+- Treat layer1 as the static harness-contract gate, including basic benchmark setup alignment checks such as expected next-route handoffs, runner command conventions, output file expectations, and quality-rubric facts against the current skill contract.
+- If layer1 fails because a benchmark setup is misaligned with the skill contract, classify it as a harness/benchmark coverage defect and route to `/targeted-skill-builder <SKILL> benchmark failure`; do not spend agent budget on `pnpm bench`.
+- If layer2 reports no tests matched `<SKILL>`, treat that layer as skipped and continue to the benchmark step. Record the skip clearly because generic benchmark coverage is weaker than target-specific layer2 verification.
+- Record pass/fail and wall time per layer.
+- If verify fails, stop and report the failure. Do not run the benchmark step.
+
+### Step 3 - Bench
+
+Run only if verify passes:
+
+```bash
+pnpm bench --skill <SKILL> --agent both --runs 3 --chunk-size 3 --pause 0
+```
+
+- Use 3 iterations by default.
+- Use `--agent both` by default. Only use `--agent claude` or `--agent codex` when the user explicitly asks to isolate one runner.
+- The expected budget is about $1 per run for the current design-system variant; report actual cost from the benchmark output when available.
+- The bench system persists raw data to `tests/benchmarks/runs/<skill>-<agent>-<sessionId>/` and generates `report.json`.
+- Treat rate limits, quota exhaustion, and similar runner-capacity errors as infrastructure-blocked runs, not skill failures. Report them separately from evaluated pass rate.
+- If recent same-skill benchmark reports or triage reports show repeated same-family benchmark false negatives, do not recommend another blind rerun as the next step. Report the recurrence and route to `/targeted-skill-builder <SKILL> benchmark repeated false-negative generalization` so the harness/setup gets a family-level semantic evaluator, fixture set, or infrastructure classifier instead of another one-off tolerance patch.
+
+### Step 4 - Report
+
+Write results to `benchmark/test-<SKILL>-<YYYY-MM-DD>.md` at the repository root. Use the current date.
+
+Populate the report from `report.json` and verify the output includes:
+
+- verify table with layer status and wall time
+- agent name, evaluated pass rate, blocked-run count, and Wilson 95% confidence interval
+- failed assertions, if any
+- output-quality score summary when the setup defines a quality evaluator, including threshold failures, critical failures, and lowest-scoring criteria when present
+- infrastructure-blocked runs, if any
+- latency p50, p95, and p99
+- cost per run and total cost
+- mean pairwise similarity and outlier count
+- raw session path
+- recommended next route, using a literal label such as `Recommended next command:` or `Recommended next skill:`
+
+After writing the report, verify the file exists and contains the benchmark target, agent rows, pass-rate or blocked-run data, latency, cost, and raw session path. If any required report field is missing, treat the workflow as incomplete and fix the report before marking the benchmark done.
+
+## Output
+
+Print a concise benchmark summary:
+
+- verify pass/fail
+- hard assertion pass rate
+- output-quality score when present, labeled as an additional rubric score rather than a statistical confidence measure
+- infrastructure-blocked count, if any
+- p50 latency
+- total cost
+- report path
+- next review handoff when subjective output-quality judgment or remediation planning is still needed
+
+The Markdown report and the final assistant response must both include a literal next-route label accepted by the harness, such as `Recommended next skill: /benchmark-agent-review <skill>` or `Recommended next command: /ship`.
+
+## Alignment Page
+
+When this skill produces durable deliverables (research, specs, plans, reports, prototypes, or any document output), build a full-depth HTML alignment page at `alignment/benchmark-test-skill-{topic}.html`. Use a normalized topic slug derived from the app, feature, research subject, report subject, or output filename.
+
+**Full content requirement.** The alignment page must contain the complete content of every proposed markdown deliverable -- every section, every finding, every detail, every list item. It is a thorough interactive review document, not a summary. Render the full deliverable content in clean, readable HTML with appropriate hierarchy, styling, and navigation. If the skill writes multiple scoped deliverables in one run, build one alignment page that contains all deliverables with anchor-linked navigation. Durable tracker artifacts, such as `research/assumption-tracker.md`, remain canonical markdown outputs but must also be fully rendered into the alignment page before approval.
+
+**Dark-mode styling.** Use a dark color scheme by default. Base CSS variables: `--bg: #0d1117; --surface: #161b22; --border: #30363d; --text: #c9d1d9; --text-muted: #8b949e; --accent: #58a6ff; --green: #3fb950; --red: #f85149; --orange: #d29922; --purple: #bc8cff;`. Apply `background: var(--bg); color: var(--text);` on body. Use `--surface` for cards, nav, and table headers. Use `--border` for all borders. Use `--purple` for question blocks and gate headings. Use `--accent` for links and section headings. Keep headings `color: #fff` or `var(--accent)` for hierarchy. Question block backgrounds should use `#1c2333`.
+
+**Alignment gates.** Treat gates as explicit review sections inside the HTML page. A gate blocks finalization until its required inline questions are answered and compiled into YAML. Include every gate that applies to the skill output, and include these gate types whenever relevant: evidence coverage, assumptions/confidence, scope/non-goals, candidate/verdict decisions, artifact destination, proposed file changes, coverage checkpoint, and post-approval route.
+
+**Report-only research gates.** For report-only or pre-approval research skills, the alignment page must explicitly contain evidence coverage, assumptions/confidence, recommended path, proposed file changes, and approval gates before any canonical research, spec, or task file is created or updated.
+
+
+**Required inline questions.** Each gate must contain at least one required inline question placed directly under the content it governs, inside a visually distinct question block. Each question must use radio-button inputs and include two standing options after the skill-generated choices: "Other / None of the above" backed by a multi-line text box for free-form input, and "Need clarification" backed by an optional notes box where the user can explain what is unclear. When any radio option other than "Other" or "Need clarification" is selected, show an optional "Additional notes" text box beneath it so the user can qualify their choice. Generate questions based on what genuinely needs user input -- do not add filler questions. Do not create a separate bottom "Decisions & Clarifications" section.
+
+**Gate YAML contract.** At the bottom of the page, include a "Compile Answers" button that aggregates answers from all inline gate questions throughout the page, including free-text notes. The button remains disabled until every required question has a selection, shows a count of remaining unanswered questions, and scrolls to the first unanswered question if clicked early. When every question is answered, generate a structured YAML block with one item per gate answer using this stable shape: `section`, `gate_type`, `status` (`answered`, `other`, or `needs-clarification`), `answer`, optional `notes`, and optional `target_artifact` or `target_path` when the gate controls file output. After successful compilation, automatically attempt to copy the YAML to the clipboard with the Clipboard API, display copy status, and display the YAML in a read-only textarea with an explicit "Copy YAML" button. The copy button must retry clipboard copy when supported and fall back to selecting the textarea contents when clipboard access is unavailable or blocked.
+
+**Pre-approval stop.** Before user approval, the next action is review of the HTML alignment page, not downstream routing. Ask the user to review the page and provide the compiled YAML answers. Do not include `Recommended next skill`, `Recommended next command`, or downstream routing language until after compiled YAML has been provided and the approved artifacts have been written or updated.
+
+**Diff highlighting on updates.** When the agent updates an existing alignment page after receiving compiled answers, highlight what changed since the previous version. The agent chooses inline annotation or side-by-side layout per situation.
+
+**Archiving.** Before replacing an existing alignment page, archive it to `docs/history/archive/YYYY-MM-DD/HHMMSS/alignment/benchmark-test-skill-{topic}.html`.
+
+**Browser open.** Attempt to open the resulting HTML page in the browser and report whether the open succeeded or was blocked. A blocked browser-open attempt does not make the skill fail when the files were written correctly.
+
+## Constraints
+
+- Do not audit or benchmark the app, website, docs, or product surface unless the user explicitly asks for a separate website/product benchmark workflow.
+- Do not run `pnpm bench` when `pnpm verify` fails.
+- Do not fabricate benchmark metrics. Use the command output and `report.json`.
+- Do not present quality score as a replacement for hard assertion pass rate, or present a small benchmark run as statistically definitive.
+- Do not omit the final next-step route. Completion output must include either `Recommended next skill: <command>` or the two-line pair `**Next work:** <specific task or "none">` and `**Recommended next command:** <one command or route>`.
+- Do not create or modify GitHub Actions workflows.
+
+## Next-Step Routing
+
+If the skill is unknown to the repository, recommend checking the skill name or creating the skill first with `/create-agentic-skill`.
+
+If the skill has blocked benchmark coverage, recommend the row's `next_command`.
+
+If the skill only has generic smoke benchmark coverage or otherwise lacks custom domain-quality assertions, recommend `/targeted-skill-builder <skill> benchmark coverage`.
+
+If the skill fails verification, hard benchmark assertions, or configured quality thresholds, recommend `/session-triage <skill> benchmark failure`.
+
+If benchmark runs are blocked only by rate limits or quota exhaustion, recommend re-running `/benchmark-test-skill <skill>` after the reset instead of treating the skill as failed.
+
+If the failure pattern has already been triaged as a repeated same-family benchmark false negative, recommend `/targeted-skill-builder <skill> benchmark repeated false-negative generalization` instead of another `/benchmark-test-skill <skill>` rerun.
+
+If evaluated benchmark runs completed and subjective output-quality review or remediation planning has not yet been performed, recommend `/benchmark-agent-review <skill>` as the next separate step.
+
+If the skill passes, the report is written, and no subjective review is needed or the separate `/benchmark-agent-review <skill>` step has already been completed, recommend `/ship`.

package/packs/agentic-skills-bench/codex/benchmark-agent-review/ALIGNMENT-PAGE.md

@@ -0,0 +1,86 @@
+# Alignment Page — benchmark-agent-review
+
+When this skill produces durable deliverables (research, specs, plans, reports, prototypes, or any document output), build a full-depth HTML alignment page at `alignment/benchmark-agent-review-{topic}.html`. Use a normalized topic slug derived from the app, feature, research subject, report subject, or output filename.
+
+**Visual rendering tier.** Each alignment page declares a rendering tier:
+- `document` (default) — text, tables, approval gates. No charts or interactive elements.
+- `visual` — adds inline charts, diagrams, and spatial layouts rendered via self-contained Canvas/SVG. Data is embedded as inline JSON. No external fetches.
+- `prototype` — adds functional interactive components (clickable flows, live CSS previews, drag interactions). Prototypes are ephemeral previews, not production code.
+
+The tier is set per skill via `visual_tier` in SKILL.md frontmatter (or detected by the upgrade script). The bundled ALIGNMENT-PAGE.md receives tier-appropriate rendering guidance from the generator.
+
+**Alignment lifecycle.** Alignment pages have three lifecycle states. `review` is the draft/pre-approval page: it contains the complete deliverable preview, evidence, required inline gate questions, section feedback controls, feedback-only YAML, and final compile controls. `confirmed` is the post-approval page after the agent has consumed final compiled YAML, made any approved user-requested edits, and written or updated the approved canonical artifacts. A confirmed page is finished and current for the completed alignment cycle, but it remains amendable when later research changes or extends the conclusions. `amended` is a future revision of a confirmed page: archive the previous confirmed page first, mark what changed, preserve the prior approval record, and highlight the new evidence, decisions, or scope that changed the earlier conclusion.
+
+**Central alignment index.** When a research-producing skill writes or updates an HTML alignment page, also maintain the repository's central alignment index. Look for `alignment/index.html` at the repository root; if the target repo documents a different central alignment-index path, use that path and report it. If the index does not exist, create it as a self-contained HTML page that works from `file://`, links to every active `alignment/*.html` page except itself, and includes the newly written page. Every index entry must include the page's alignment date in a muted meta span after the link, formatted `YYYY-MM-DD`, for example `<a href="page.html">Title</a> <span class="meta">2026-06-09</span>`. The alignment date is the review or confirmed date written inside the page itself; new pages use today's date. If the index exists, update it instead of blindly appending: preserve valid existing entries, metadata, and entry dates; remove duplicate entries for the same page path; keep links repo-relative; and include the new or updated page. When updating an existing index, add a date only to entries that are missing one, deriving it from the page's internal alignment date, or from file history when the page has no internal date. If a `new` marker is used, combine it with the date as `<span class="meta">new · YYYY-MM-DD</span>`, and remove stale `new` markers from entries that are no longer the latest addition.
+
+**Alignment index categories.** The index groups alignment pages into five categories. Each alignment page belongs to exactly one category. When writing or updating the index, assign each page to a category and render the index grouped by category in the order listed below.
+
+1. **Research** (`research`) — market research, competitive analysis, investigative reports. Skill prefixes: `devtool-positioning`, `devtool-adoption`, `devtool-monetization`, `devtool-user-map`, `devtool-workflow`, `devtool-dx-journey`, `devtool-integration-map`, `deep-research`, `repo-glossary`.
+2. **Product Design & Spec** (`product-design`) — product scoping, idea briefs, design plans, specifications. Skill prefixes: `idea-scope-brief`, `fork-idea-branch`, `animation-design-planner`, `skills-showcase`, `skills-inventory`.
+3. **Utility & Maintenance** (`utility`) — diagnostics, triage, incident investigation, infrastructure maintenance. Skill prefixes: `investigate`, `session-triage`, `prompt-history-backfill`.
+4. **QA & Meta-Skill Improvement** (`qa-meta`) — code review, benchmarking, skill building/fixing, expert review. Skill prefixes: `expert-review`, `benchmark`, `targeted-skill-builder`, `devtool-docs-audit`.
+5. **Ops & Session Analysis** (`ops-analysis`) — operational intelligence, session history analysis, cost/token/workflow analysis. Skill prefixes: `analyze-sessions`, `canonical-workflow`, `run-batch`.
+
+**Category assignment rule.** Match the skill name that produced the alignment page against the prefix lists above; use the first matching category. When no prefix matches, read the page title and summary to assign the best-fit category. When the category is ambiguous, default to **Research**.
+
+**Visual tier badge.** When writing or updating the alignment index, display a visual tier badge next to each linked page: `[doc]` for document tier, `[visual]` for visual tier (pages with charts/diagrams), `[proto]` for prototype tier (interactive pages). Read the tier from the alignment page's `data-visual-tier` attribute on the `<html>` element. When the attribute is missing, default to `[doc]`.
+
+**Category data attribute.** Set `data-alignment-category` on the `<html>` element of each alignment page with the category slug (`research`, `product-design`, `utility`, `qa-meta`, or `ops-analysis`). Also set `data-visual-tier` on the `<html>` element with the rendering tier (`document`, `visual`, or `prototype`). When an existing alignment page lacks the attribute, infer the category from the skill-name prefix or page content and add it on the next update. The index should read this attribute when available; when it is missing, apply the prefix-matching rule to assign a category for display.
+
+**Category and product-path grouping.** In product-line or product-path repositories, identify product from `research/.progress.yaml` `product_paths[]`, active `research/{slug}/` or `specs/{slug}/` scopes, page metadata, page title, or the alignment page path. Group index entries by category first, then by product/product path within each category, with the repo's existing secondary sort preserved when clear. In flat repositories, group by category, then preserve the existing index organization within each category or use a clear newest-first or title order when creating a new index.
+
+**Research quality contract.** For research-producing outputs, do not synthesize research findings, recommendations, candidate rankings, or working packets until the user approves the research scope from the `review` alignment page's final compiled YAML. Before scope approval, do only minimal discovery needed to propose the research scope, source plan, assumptions, output paths, and approval questions; label any observed files or source availability as scope evidence, not findings. After scope approval, perform the synthesized research and separate and label `claims` (what the report concludes), `evidence` (source, repo artifact or file path, quote or observation, date, and confidence), `inference` (why that evidence supports the claim), `assumptions` (what remains unproven), and `decision impact` (what the user should approve, reject, or correct). Do not collapse evidence and inference into unsupported summary prose.
+
+**Embed prohibition.** Do not use `<object>`, `<iframe>`, `<embed>`, or "open working packet" / "view full document" links as the primary rendering of any proposed deliverable, evidence matrix, interview log, or manifest entry. The alignment page HTML must contain the rendered content directly. Embedded or linked external documents are permitted only as supplemental references alongside the fully rendered content already present in the page. **Exception:** Self-contained inline `<canvas>` and `<svg>` elements are permitted for charts and diagrams within the same HTML file.
+
+**Visual accessibility contract.** Every chart or diagram in a `visual` or `prototype` tier alignment page must include: (1) a `<table>` or text fallback rendering the same data, (2) an `aria-label` attribute on every `<canvas>` and `<svg>` element describing the chart type and data summary, and (3) a "View as table" toggle button that swaps between the chart and its table fallback. The toggle must default to showing the chart, with the table hidden via a `hidden` attribute that is toggled by the button.
+
+**Inline chart contract.** Visual and prototype tier pages must inline all chart rendering code as a `<script>` block within the HTML file. No external CDN references, relative file imports, or `<script src="...">` tags for chart libraries. Use the alignment chart snippet library (`scripts/alignment-chart-snippets.js`) as the reference implementation — copy the needed functions directly into the page's `<script>` block. Charts must read CSS variables (`--chart-1` through `--chart-8`) for colors and adapt to the dark-mode palette.
+
+**No context loss rule.** The alignment page must contain the complete content of every proposed markdown deliverable -- every section, every finding, every detail, every list item -- plus the decision-relevant substance from search logs, interview logs, source notes, repo scans, and approval notes. If a fact, source, caveat, uncertainty, alternative, rejected or lower-confidence finding, or decision rationale appears in a proposed deliverable or research log, it must either appear in the HTML page or be explicitly linked from the exact section that depends on it. The page is a thorough interactive review document, not a summary. Render the full deliverable content in clean, readable HTML with appropriate hierarchy, styling, and navigation. If the skill writes multiple scoped deliverables in one run, build one alignment page that contains all deliverables with anchor-linked navigation. Durable tracker artifacts, such as `research/assumption-tracker.md`, remain canonical markdown outputs but must also be fully rendered into the alignment page before approval.
+
+**Page layout contract.** After the page title and short summary, include a top-of-page "Table of Contents" section with anchor links to the major review sections and the bottom compile section. Keep the Table of Contents in normal document flow. Do not use a sidebar, side rail, drawer, split-shell layout, or sticky navigation for the Table of Contents unless the user explicitly asks for that layout. Do not place compile, copy, feedback, or answer controls in a sticky or fixed bottom banner/footer. Bottom compile controls must appear as ordinary content in a bottom compile section, so they scroll with the page and do not cover content at high zoom.
+
+**Research translation requirements.** For research outputs, the HTML page must include an evidence matrix, confidence/assumption register, alternatives considered, rejected or lower-confidence findings, source coverage gaps, and downstream implications. The evidence matrix must map each major claim to source or repo evidence, inference, confidence, assumption status, and decision impact. The confidence/assumption register must show which conclusions are evidence-backed, which are provisional, and what evidence would change them.
+
+**Dark-mode styling.** Use a dark color scheme by default. Base CSS variables: `--bg: #0d1117; --surface: #161b22; --border: #30363d; --text: #c9d1d9; --text-muted: #8b949e; --accent: #58a6ff; --green: #3fb950; --red: #f85149; --orange: #d29922; --purple: #bc8cff;`. Data visualization palette (for `visual` and `prototype` tier pages): `--chart-1: #58a6ff; --chart-2: #3fb950; --chart-3: #d29922; --chart-4: #bc8cff; --chart-5: #f85149; --chart-6: #79c0ff; --chart-7: #f0883e; --chart-8: #a5d6ff;`. Apply `background: var(--bg); color: var(--text);` on body. Use `--surface` for cards, the top Table of Contents, and table headers. Use `--border` for all borders. Use `--purple` for question blocks and gate headings. Use `--accent` for links and section headings. Keep headings `color: #fff` or `var(--accent)` for hierarchy. Question block backgrounds should use `#1c2333`.
+
+**Responsive layout.** Include `<meta name="viewport" content="width=device-width, initial-scale=1">` in the head. Use at most two breakpoints: `@media (max-width: 860px)` for tablet and `@media (max-width: 560px)` for phone. Wrap every `<table>` in a `<div style="overflow-x: auto">` so tables scroll horizontally instead of breaking layout. All interactive elements (buttons, radio labels, feedback controls) must have a minimum touch target of `44px` height and at least `10px 16px` padding. Use `rem` for font sizes and respect the user's base font size — do not set a fixed `font-size` on `html` or `body`. Flex containers that place a heading beside controls (`display: flex; justify-content: space-between`) must include `flex-wrap: wrap; gap: 8px` so controls drop below the heading on narrow screens. Grid layouts must use `minmax(min(210px, 100%), 1fr)` instead of bare `minmax(210px, 1fr)` to prevent horizontal overflow when the viewport is narrower than the minimum. Set `max-width: 100%; overflow-wrap: break-word;` on `main` and any fixed-width container so no element forces a horizontal scrollbar at any zoom level.
+
+**Alignment gates.** In `review` pages, treat gates as explicit review sections inside the HTML page. A gate blocks finalization only while the page is in `review`, until its required inline questions are answered and compiled into YAML. Include every gate that applies to the skill output, and include these gate types whenever relevant: evidence coverage, assumptions/confidence, scope/non-goals, candidate/verdict decisions, artifact destination, proposed file changes, coverage checkpoint, and post-approval route. In `confirmed` pages, preserve the gate decisions as read-only approval records instead of active approval controls.
+
+**Output gate semantics and de-duplication.** The `artifact destination` gate approves the durable or review artifact location, such as the alignment page, working packet, report, spec, plan, prototype, or document path. The `proposed file changes` gate approves downstream mutation scope, timing, and the allowed file set after review. Keep these as separate approval concepts. If both gate concepts ask only the same path-destination question, render one combined visual section such as `Artifact Destination & Proposed File Changes` instead of duplicating path-only sections. If the decisions differ, render separate gates. Do not duplicate YAML entries solely because one visual section title names both concepts: emit separate YAML items only when the page asked separate decisions, and keep each item's `gate_type` equal to the decision it controls (`artifact destination` or `proposed file changes`).
+
+**Research completeness gate.** For research outputs, include a research completeness gate with inline questions asking whether the evidence is sufficient for the recommendation, which claims need more support, and whether missing context could change the recommendation. Place these questions directly under the evidence matrix or recommendation section they govern.
+
+**Source coverage expectations.** For web research, organize source coverage by category rather than citation list alone; use categories such as competitors, pricing, user sentiment, positioning, integrations, and recent activity when relevant to the topic. For repo or codebase research, include file/path evidence and clearly distinguish observed code facts from inferred product, workflow, or user conclusions.
+
+**Report-only research gates.** For report-only or pre-approval research skills, the first `review` alignment page must explicitly contain proposed research scope, source plan, assumptions/confidence, recommended output path, proposed file changes, and approval gates before synthesized research begins and before any working packet, canonical research, spec, or task file is created or updated.
+
+**Staged research workflow.** Heavy report-first research uses a scope-first three-stage approval workflow. Stage 1 is minimal scope discovery only: inspect enough repository, user, and source context to propose research scope, source plan, assumptions, output paths, and approval questions. Build the `review` alignment page before synthesized research. The page must render the proposed scope, available source categories, known context, assumptions/confidence, proposed working-packet and canonical output paths, and research-scope approval gates. Stop for final compiled YAML approval of the research scope. Do not synthesize findings, rank candidates, recommend a path, or create working packets, canonical research, spec, or task files in Stage 1. Stage 2 starts only after final compiled YAML approves the research scope with no unresolved `needs-clarification`, unresolved `down` feedback, or other unresolved negative feedback. Perform the synthesized research, run required source/code checks, and write only a non-canonical working packet. Flat mode uses `research/_working/preliminary-benchmark-agent-review-research.md`; product-path mode uses `research/{slug}/_working/preliminary-benchmark-agent-review-research.md`. Existing raw evidence or search-log outputs may remain as supporting evidence when the skill contract already requires them, but synthesized deliverables stay in the working packet. Update the same `review` alignment page so it renders the full preliminary packet, evidence matrix, assumptions/confidence register, source coverage gaps, proposed canonical file changes, and artifact approval gates. Feedback-only YAML revises the working packet and review page, then remains in Stage 2. Stage 3 consumes final compiled YAML for artifact approval only when it has no unresolved `needs-clarification`, unresolved `down` feedback, or other unresolved negative feedback. Apply approved edits first, archive the working packet to `docs/history/archive/YYYY-MM-DD/HHMMSS/<original-working-path>`, remove the active working packet, write the approved canonical artifacts to the unchanged output paths, and convert the page to `confirmed` with the approval record preserved.
+
+**Glossary Additions gate.** When this research introduces domain-specific terms, acronyms, or concept definitions that a reader outside the project would not know, include a `## Glossary Additions` section in the alignment page. Render a table of proposed terms with columns: Term, Definition, Source (`benchmark-agent-review`), Category (business/tooling/workflow/technical/domain), and per-term Approve/Edit/Reject/Flag radio controls. Only user-approved terms are appended to the target glossary (`research/glossary.md` or `research/{slug}/glossary.md` for scoped paths) during the confirmed-page write step. When the skill operates in a product-path scope, ask whether each term belongs in the parent or scoped glossary; default to the scoped glossary.
+
+**Required inline questions.** In `review` pages, each gate must contain at least one required inline question placed directly under the content it governs, inside a visually distinct question block. Each question must use radio-button inputs and include two standing options after the skill-generated choices: "Other / None of the above" backed by a multi-line text box for free-form input, and "Need clarification" backed by an optional notes box where the user can explain what is unclear. When any radio option other than "Other" or "Need clarification" is selected, show an optional "Additional notes" text box beneath it so the user can qualify their choice. Generate questions based on what genuinely needs user input -- do not add filler questions. Do not create a separate bottom "Decisions & Clarifications" section. Confirmed pages must not keep these as required input controls; render the answered decisions as read-only records.
+
+**Section feedback controls.** Every major section of the page (each deliverable section, evidence matrix, confidence register, gate, and any anchor-linked top-level heading) carries a lightweight section-feedback control near its heading: three mutually exclusive choices -- "emphasize" (add more weight to this section or to the point named in the notes), thumbs down (reject or flag a concern), and "clarification needed". Selecting any one reveals a multi-line section-feedback textarea placed directly under or beside that section's emphasize/down/clarify controls; deselecting hides the textarea and any local feedback YAML output for that section. This textarea is separate from required gate-question text inputs: even when the same section has gate questions with their own text boxes, selecting emphasize/down/clarify must still reveal the section-feedback textarea near the feedback controls. Use `--green` for active "emphasize", `--red` for the active thumbs-down, and `--orange` for active "clarification needed", with muted/inactive states otherwise. These controls are optional for final approval and do not replace required gate questions. They also power the separate feedback-only YAML path so the user can send emphasis requests, concerns, or clarification requests before answering every required gate question. When a section contains nested subsections (h3 or deeper), duplicate the section-feedback control at the bottom of the section, just before the closing `</section>` tag. The bottom copy mirrors the top copy — same `data-section`, same id with a `-bottom` suffix, same button set. Sections without nested subsections keep only the top control. Pair top and bottom controls by matching `data-section` attribute. When a feedback button is clicked on one copy, mirror the active state, `has-feedback` class, and `dataset.feedback` value to the paired copy. Textarea content should sync on compile — whichever textarea has content takes precedence (prefer the one the user last typed in). Both copies' local compile/copy buttons produce the same YAML. Use the top copy's `id` (without `-bottom`) as the canonical feedback id for YAML output.
+
+**Feedback-only YAML contract.** Provide feedback-only YAML in two places: locally under each selected section-feedback textarea, and globally in the bottom compile section. When a section-feedback control is selected, show local "Compile Feedback YAML" and "Copy YAML" controls plus a read-only YAML textarea directly under that section's feedback textarea. The local feedback compile generates YAML with `alignment_page: alignment/benchmark-agent-review-{topic}.html`, `feedback_status: revision-request`, `approval_status: not-approved`, `unanswered_required_questions`, and a `section_feedback` list containing the single selected section-feedback entry for that local control. The bottom "Compile Feedback YAML" control generates the same YAML shape but aggregates every selected section-feedback entry on the page. Enable both local and bottom feedback compile controls as soon as at least one section-feedback control is set, even if required inline gate questions are unanswered. Populate `alignment_page` from the known repo-relative output path used to write the HTML page, not from the page title, browser URL, window location, or any display label. Each feedback entry uses the shape `section`, `feedback` (`emphasize`, `down`, or `needs-clarification`), optional `notes` from that section's feedback textarea, and `requested_agent_action` (`add-weight-to-section`, `investigate-and-revise`, or `clarify-before-approval`). For `emphasize` feedback, the YAML must make clear that the agent should add more weight, prominence, evidence, detail, or recommendation emphasis to that section or to the specific point named in the notes; this is a revision request, not approval of the section as-is. For `down` and `needs-clarification` feedback, the YAML must make clear that the agent should evaluate the feedback, investigate further when needed, and contextually amend the HTML alignment page before asking again for final approval answers. Automatically attempt to copy feedback YAML to the clipboard, display local or bottom copy status, display it in the matching read-only textarea, and provide a "Copy YAML" button with the same retry and fallback behavior as final gate YAML. Do not render the bottom feedback compile controls as a sticky or fixed banner.
+
+**Gate YAML contract.** In `review` pages, at the bottom of the page, include an ordinary in-flow compile section with a "Compile Feedback YAML" button for selected section feedback and a separate "Compile Answers" button for final approval answers. The bottom compile section must not be sticky, fixed, floating, or styled as a persistent banner. The "Compile Answers" button aggregates answers from all inline gate questions throughout the page, including free-text notes. The final approval button remains disabled until every required question has a selection, shows a count of remaining unanswered questions, and scrolls to the first unanswered question if clicked early. When every question is answered, generate a structured YAML block with `alignment_page: alignment/benchmark-agent-review-{topic}.html`, `approval_status: ready-for-agent-review`, one item per gate answer using this stable shape: `section`, `gate_type`, `status` (`answered`, `other`, or `needs-clarification`), `answer`, optional `notes`, and optional `target_artifact` or `target_path` when the gate controls file output. Populate `alignment_page` from the known repo-relative output path used to write the HTML page, not from the page title, browser URL, window location, or any display label. The same final compiled output also includes any section-feedback controls that the user set, using the `section_feedback` shape from the feedback-only YAML contract. After successful compilation, automatically attempt to copy the YAML to the clipboard with the Clipboard API, display copy status, and display the YAML in a read-only textarea with an explicit "Copy YAML" button. The copy button must retry clipboard copy when supported and fall back to selecting the textarea contents when clipboard access is unavailable or blocked.
+
+**Confirmed page contract.** After final approval has been applied, convert the active alignment page from `review` to `confirmed`. Include a visible status block with `alignment_status: confirmed`, confirmation date, confirmed artifact paths, approval source summary, and explicit finished but amendable language: the page is current for the completed alignment cycle, and later research can amend it only by archiving the prior confirmed page and highlighting the amendment. Remove required gate-question controls, the final "Compile Answers" button, unanswered-question counters, disabled approval buttons, and approval-blocking language. Preserve the full research and approval record: answered decisions, user requests, evidence matrix, assumptions/confidence register, source gaps, proposed file changes, final compiled YAML or faithful approval summary, and what changed during confirmation. Keep research caveats visible; `confirmed` means approved/current, not immutable or permanently true.
+
+**Pre-approval stop.** While an alignment page is in `review`, the next action is review of the HTML alignment page, not downstream routing. Ask the user to review the page and provide either local or bottom feedback-only YAML for emphasis requests, concerns, or clarification, or final compiled YAML answers when ready. Do not require the user to answer every gate before sending emphasis requests, negative feedback, or clarification needs. When feedback-only YAML is provided, treat it as a revision request: evaluate the feedback, investigate further when needed, archive and amend the HTML page contextually, highlight the changes, and ask again for review. Do not include `Recommended next skill`, `Recommended next command`, or downstream routing language until after final compiled YAML has been provided and the approved artifacts have been written or updated.
+
+**After approval handling.** When final compiled YAML is provided, inspect it before confirming the page. If it contains approvals plus user-requested edits, first update the page, research, and proposed canonical artifacts to satisfy the request; then write or update the approved canonical artifacts and confirm the page. If it contains `needs-clarification`, unresolved `down` feedback, or any unresolved negative feedback, revise the page and return it to `review` instead of marking it `confirmed`. Before replacing any existing alignment page, archive it to `docs/history/archive/YYYY-MM-DD/HHMMSS/alignment/benchmark-agent-review-{topic}.html`. Future amendments to confirmed pages follow the same archive-first rule and must show what changed from the prior confirmed record.
+
+**Diff highlighting on updates.** When the agent updates an existing alignment page after receiving feedback, compiled answers, confirmation edits, or later amendment evidence, the rendered HTML must visibly indicate what changed since the previous version. Highlight changed sections or blocks in-place with a `Changed`/`Updated` badge, distinct border or background, `<ins>`/`<del>` markup, or a side-by-side before/after comparison when that is clearer. A top-level change summary is useful but not sufficient by itself unless the changed content is also visibly marked where the reader reviews it. Remove or refresh stale change markers on later updates so the highlights describe the current revision only.
+
+**Archiving.** Before replacing an existing alignment page, archive it to `docs/history/archive/YYYY-MM-DD/HHMMSS/alignment/benchmark-agent-review-{topic}.html`.
+
+**Brief Me (text-to-speech).** Every alignment page must include a `<script src="../scripts/alignment-tts-kokoro.js"></script>` tag before `</body>`. The script adds a "Brief Me" button that reads the page aloud section-by-section using Kokoro TTS (an 82M-parameter model running client-side via WebGPU/WASM from `kokoro-js`). The transport bar includes play/pause/stop/skip controls, a voice selector (persisted in localStorage), and speed selection. On first click the voice model downloads (~92MB quantized on WASM; ~326MB fp32 on WebGPU). Weights persist between visits via Cache API origin storage, and this works on `file://` origins in Chrome — including `file://wsl.localhost` UNC paths (verified 2026-06-09 with an instrumented probe in Windows Chrome: `caches` is available, entries survive reloads and browser restarts, and Chrome shares a single `file://` origin across all local pages, so one download serves every alignment page). Repeat visits pay model initialization only, not re-download. If `caches` is undefined (non-Chromium or locked-down browsers), the model re-downloads each visit; serving pages from a local HTTP origin (e.g. `node scripts/serve-alignment.mjs` in the repo root, then `http://localhost:8907/alignment/<page>.html`) guarantees caching there — note `http://localhost` is a separate origin from `file://`, so the first localhost visit downloads once even when the `file://` cache is already warm. After the first successful load on a machine, the script preloads the model during browser idle time on later page loads (gated by a `tts-kokoro-used` localStorage flag, so users who never click pay nothing). Text is prepared for speech before synthesis: block boundaries (headings, paragraphs, list items, table cells) get sentence punctuation so the model does not read them as run-ons, and speech-hostile symbols are normalized (`~92MB` reads as "approximately 92MB", `v0.3` as "version 0.3", em-dashes as comma pauses, `e.g.`/`i.e.`/`vs.` expanded). Synthesis is pipelined with small chunks: the first audio chunk is kept short (~250 chars) for fast time-to-first-audio, later chunks stay small (~300 chars, cut at sentence or clause boundaries) with up to two chunks synthesizing ahead during playback, and the next section's opening chunk is prefetched while the current section's last chunk plays, so chunk and section boundaries are gapless. If Kokoro fails to load (old browser, network error), it falls back to the Web Speech API automatically. It skips gate controls, compile UI, feedback textareas, and stat grids. Charts and tables use narrative summaries instead of raw data: containers with a `data-tts-narrative` attribute have the narrative read aloud; containers without the attribute are silently skipped. To inject TTS into existing pages, run `node scripts/inject-tts.mjs` (idempotent). Use `--force` to replace the existing TTS block with an updated version. Do not inline the TTS script; always use the `src` tag. Do not use `type="module"` — module scripts are blocked by CORS on `file://` URLs.
+
+**TTS narrative contract.** Every `.chart-container` and `.table-wrap` element should carry a `data-tts-narrative` attribute with a 1-3 sentence human-readable summary of what the chart or table shows. Chart functions in `scripts/alignment-chart-snippets.js` auto-generate a baseline narrative from their data; author-provided `data-tts-narrative` attributes take precedence. When authoring alignment pages manually, add `data-tts-narrative` to chart containers and table wrappers so that Brief Me reads a meaningful summary instead of skipping the element.
+
+**Browser open.** After writing the resulting HTML page, run `node scripts/open-html-page.mjs alignment/benchmark-agent-review-{topic}.html --browser auto` from the repository root, replacing the path with the actual page path when needed. Report the script status (`focused`, `opened`, `fallback-opened`, `blocked`, or `failed`) in the handoff. Continue when opening is `blocked` if the files were written and verified correctly; a blocked browser-open attempt does not make the skill fail.

package/packs/agentic-skills-bench/codex/benchmark-agent-review/CHANGELOG.md

@@ -0,0 +1,9 @@
+# Changelog
+
+## v0.0
+
+- Archived previous skill contract.
+
+## v0.1 - 2026-05-25
+
+- Added research-quality alignment requirements covering claim/evidence/inference separation, no-context-loss HTML translation, evidence matrices, confidence/assumption registers, source coverage, and research completeness gates.

package/packs/agentic-skills-bench/codex/benchmark-agent-review/SKILL.md

@@ -0,0 +1,118 @@
+---
+name: benchmark-agent-review
+description: Review persisted benchmark run outputs with one or more agent graders and report subjective ergonomic quality separately from deterministic benchmark scores
+type: analysis
+version: v0.1
+argument-hint: "<skill name or run path> [--reviewers codex,claude] [--runs N]"
+---
+
+# Benchmark Agent Review
+
+Invoke as `$benchmark-agent-review <skill-or-run-path>`.
+
+Use this skill after `$benchmark-test-skill <skill>` when the deterministic benchmark passed but the user wants agent judgment on whether the generated artifacts are actually excellent, ergonomic, and useful for the next caller.
+
+Act as a follow-up review workflow. It does not replace hard benchmark assertions, deterministic output-quality rubrics, or verify/bench commands.
+
+The primary object of review is the generated skill output, not the benchmark harness. Treat hard assertions and deterministic output-quality scores as context and triage signals only. Lead with whether each retained output is excellent, good, usable, weak, or failing under the agent-review rubric; discuss deterministic rubric tightening only after output-quality findings, and only when it would help future triage surface the same output-quality issue.
+
+## Input
+
+- Required: a benchmark target skill name such as `run`, or a raw run directory such as `tests/benchmarks/runs/run-codex-47e0dd54/`.
+- Optional: `--reviewers codex,claude` to request named reviewer families when both outputs are available.
+- Optional: `--runs N` to request multiple independent review passes. Default to 3 when practical; use 1 when only one reviewer pass is available in the active environment.
+
+## Process
+
+1. Resolve benchmark evidence:
+   - If the input is a run directory, inspect that directory.
+   - If the input is a skill name, find the newest matching directories under `tests/benchmarks/runs/<skill>-*/`.
+   - Prefer the latest Claude and Codex run directories when both exist.
+   - Read `report.md`, `report.json`, each `run-*.json`, and any persisted generated artifact content visible in stdout/stderr or file snapshots.
+   - If generated artifact content is not fully available, state that limitation and grade only the retained evidence.
+
+2. Extract outputs under review:
+   - Identify the artifact path the benchmark expected, such as `run-plan.md` or `benchmark/test-*.md`.
+   - Extract the generated artifact text when available.
+   - Preserve runner identity, run index, hard assertions, deterministic quality score, and infrastructure-blocked status.
+   - Exclude infrastructure-blocked runs from subjective scoring.
+
+3. Build the review packet:
+   - Include the original benchmark prompt.
+   - Include the benchmark fixture facts.
+   - Include the output artifact text or retained summary.
+   - Include hard assertions and deterministic quality scores for context.
+   - Include the target skill contract only when needed to judge ergonomics.
+
+4. Grade each evaluated output with this rubric:
+   - **Task selection clarity**: the selected work is unambiguous and traceable to fixture evidence.
+   - **Implementation specificity**: the output gives a next caller concrete enough steps to act without redoing basic discovery.
+   - **Validation strength**: proposed checks prove behavior or artifact quality, not just file existence or generic completion.
+   - **Scope control**: the output respects the fixture and user constraints without becoming timid or incomplete.
+   - **Next-route ergonomics**: the next command is correct for the runner mode and explains what it will do.
+   - **No invented facts**: the output avoids unsupported services, files, metrics, commands, deploys, or repository claims.
+   - **Residual-risk awareness**: the output names meaningful uncertainty, missing evidence, or follow-up risk when relevant.
+
+5. Score:
+   - Use a 0-100 integer score per reviewed output.
+   - Treat 90-100 as excellent, 80-89 as good, 70-79 as usable but meaningfully incomplete, 60-69 as weak, and below 60 as failing human review.
+   - When multiple review passes are available, report median score, score range, and common findings.
+   - Do not blend subjective scores into hard assertion pass rate or deterministic quality score.
+
+6. Optional multi-reviewer handling:
+   - Use subagents only when the active Codex tool instructions permit them and the user explicitly requested multiple agent review passes.
+   - Assign each reviewer one independent grading pass over the same review packet.
+   - Do not let reviewers edit repository files.
+   - Synthesize reviewer findings into a normalized score table.
+
+7. Write the report:
+   - Create `benchmark/review-<SKILL>-<YYYY-MM-DD>.md` when the input resolves to a skill.
+   - Create `benchmark/review-<RUN-DIR-NAME>-<YYYY-MM-DD>.md` when the input is one run directory.
+   - If the review is exploratory and the user asks for chat-only output, do not write a file.
+
+8. Build the remediation handoff:
+   - Convert every material weakness into a remediation target instead of stopping at broad advice.
+   - Classify each target as target-skill contract, benchmark rubric, retained-evidence gap, harness/setup issue, or one-off run behavior.
+   - Name the exact owner file, skill contract, benchmark setup, or report artifact when known; when the exact file is not proven, name the narrowest known owner surface and state the lookup needed to confirm it.
+   - Propose the exact contract, rubric, fixture, or evidence-capture behavior to add or tighten; avoid vague changes such as "update the skill" without naming the behavior that changes.
+   - Include the validation command or contract-lint assertion that would prove the issue is fixed; a focused fixture rerun is acceptable only when it names the expected assertion or artifact-quality behavior.
+   - When retained artifact text contains placeholder risk, monitoring, validation, or known-unknown sections such as `Not captured`, `Not specified`, `TBD`, `None`, or `N/A`, the remediation must identify the owner target that should reject or repair that placeholder and the validation check that would fail before the fix.
+   - Choose one definitive next route from the highest-impact verified remediation; do not leave the next caller to choose among generic options.
+
+## Output
+
+Report:
+
+- Source benchmark report paths.
+- Reviewed run directories and run indexes.
+- Hard assertion pass rate and deterministic output-quality score from the benchmark report.
+- Output-quality verdict against the agent-review rubric, explicitly focused on the generated skill artifacts rather than the benchmark's ease or strictness.
+- Agent-review score table with reviewer, runner, run index, score, and grade band.
+- Median subjective score and score range when multiple scores exist.
+- Common strengths.
+- Common weaknesses.
+- Remediation table with finding, classification, owner target, proposed change, validation check, and route.
+- Remediation rows must be implementation-ready: each material finding needs a concrete owner target, a proposed behavior change, and a validation check or command. Broad rows like "tighten the rubric" or "update the skill" are incomplete unless they also name the owning file/surface, exact behavior, and proof.
+- Optional deterministic-rubric notes only when the retained output-quality findings show the deterministic rubric failed to surface a meaningful issue or produced misleading context.
+- **Next work:** the one definitive remediation selected from the remediation table, or no follow-up when all evaluated outputs are excellent and no meaningful issue remains.
+- **Recommended next command:** one command derived from that remediation, usually `$targeted-skill-builder <skill> <specific output-quality gap>`, `$targeted-skill-builder <benchmark setup or reviewed skill> <specific rubric gap>`, `$session-triage <skill> benchmark review`, or `$ship` only when no remediation is needed.
+
+## Alignment Page
+
+When this skill produces durable deliverables (research, specs, plans, reports, prototypes, or any document output), build a full-depth HTML alignment page following `ALIGNMENT-PAGE.md` in this skill's directory. Output: `alignment/benchmark-agent-review-{topic}.html`.
+
+## Constraints
+
+- Do not re-run `$benchmark-test-skill` unless the requested benchmark artifacts are missing or stale and the user explicitly asks for a fresh benchmark.
+- Do not present subjective agent-review scores as statistically definitive.
+- Do not merge subjective scores into deterministic output-quality scores.
+- Do not frame benchmark pass/fail laxness as the primary problem when the task is to judge skill output quality. If the benchmark intentionally passes weak-but-compliant outputs, grade the output quality directly and use deterministic-rubric notes only as supporting context.
+- Do not grade infrastructure-blocked runs as skill outputs.
+- Do not fabricate missing artifact content. State when only stdout summaries, assertions, or quality results are available.
+- Do not collapse multiple material weaknesses into a vague handoff such as "tighten the rubric"; the remediation table must preserve the responsible target, exact behavior change, and validation check for each issue.
+- Do not create or modify GitHub Actions workflows.
+
+## Default Shipping Contract
+
+Follow the shared shipping contract convention in CLAUDE.md.
+