@meyverick/agentic 5.0.2

package/skills/guardrails/evals/benchmark.json

@@ -0,0 +1,40 @@
+{
+  "skill": "guardrails",
+  "generated": {
+    "by": "process:structural-stage/1.0",
+    "at": "2026-09-02T10:56:00Z"
+  },
+  "stage": "behavioral",
+  "structural": {
+    "validate_structure": {
+      "pass": true,
+      "note": "recorded at apply gate (task 6.1)"
+    },
+    "validate_routing": {
+      "checks_total": 6,
+      "positive_triggers": 4,
+      "anti_triggers": 2,
+      "single_responsibility": true
+    },
+    "evals": {
+      "count": 4,
+      "assertions": 9,
+      "anti_trigger_coverage": true
+    }
+  },
+  "behavioral_dxm": "1×0.33",
+  "ship_gate": {
+    "criterion": "d = +1 and m >= 0.2",
+    "applies_to": "behavioral stage"
+  },
+  "behavioral": {
+    "at": "2026-09-02T08:59:00.285Z",
+    "evals": 4,
+    "assertions": 12,
+    "baseline": 0.5,
+    "with_skill": 0.8333,
+    "d": 1,
+    "m": 0.3333,
+    "ship": "pass"
+  }
+}

package/skills/guardrails/evals/evals.json

@@ -0,0 +1,49 @@
+{
+  "skill_name": "guardrails",
+  "evals": [
+    {
+      "id": 1,
+      "prompt": "I'm about to commit code that handles user auth. Check for security pitfalls.",
+      "expected_output": "Guardrails triggers, lists Anti-example 2 (secret leak) and 1 (on* stripping), shows Do: envx + .gitignore, not commit .env.",
+      "files": [],
+      "assertions": [
+        "Guardrails skill triggered",
+        "Anti-example for secret commit shown",
+        "Correct pattern envx + .gitignore shown"
+      ]
+    },
+    {
+      "id": 2,
+      "prompt": "I need to update SvelteKit adapter for deployment to distroless.",
+      "expected_output": "Guardrails triggers, shows Anti-example 3: before svelte-adapter-bun vs after @sveltejs/adapter-static + tower-http, with fallback index.html.",
+      "files": [],
+      "assertions": [
+        "Guardrails triggered on deprecated adapter query",
+        "Contrast table includes Before adapter-bun vs After adapter-static",
+        "Why explains Axum serves static, single distroless binary"
+      ]
+    },
+    {
+      "id": 3,
+      "prompt": "Review my Rust code that checks if user is null and handles lifetimes.",
+      "expected_output": "Guardrails triggers, shows Anti-examples 4 (null vs Option) and 5 (lifetime elision) with before/after Rust snippets.",
+      "files": [],
+      "assertions": [
+        "Guardrails triggered on system gotcha",
+        "Anti-example 4: null vs Option shown with if let Some",
+        "Anti-example 5: lifetime 'a correct snippet shown"
+      ]
+    },
+    {
+      "id": 4,
+      "prompt": "I'm just exploring the codebase, reading the README without modifying anything.",
+      "expected_output": "Guardrails does NOT trigger — only reading, no code touching deps/Docker/HTML/auth. Suggest explore mode, no guardrails load.",
+      "files": [],
+      "assertions": [
+        "Guardrails did NOT trigger on read-only prompt",
+        "Response does not include Anti-examples",
+        "Anti-trigger respected"
+      ]
+    }
+  ]
+}

package/skills/guardrails/references/guardrails-patterns.md

@@ -0,0 +1,43 @@
+# Guardrails Patterns — Level 2
+
+On-demand detail for each cross-cutting anti-example in `SKILL.md`. Keep SKILL.md Level 1 inline; this file is Level 2 (~700 tokens).
+
+## 1. `on*` Stripping
+
+- **Before (wrong):** `<div oncustom={h}>` — Svelte 5 strips unknown `on*` attributes at compile, no warning, handler never fires
+- **After (right):** `<div on:click={h}>` or `<div use:action>` — explicit Svelte 5 event syntax
+- **Signal:** No runtime error, silent loss of interactivity; test via `vite build` + click
+- **Why:** Security sanitization in Svelte 5 — unknown `on*` is treated as potential XSS
+
+## 2. Secret Leak
+
+- **Before (wrong):** `git add .env && git commit` — secrets in git history forever, leaked to logs via `dokku config:set` echo
+- **After (right):** `echo ".env" >> .gitignore` at repo root + each submodule; `envx set KEY`; `dokku config:set` guarded (only when var unset/changed, output suppressed)
+- **Signal:** `git log --all -p | grep WARERA_KEY` or `deploy-dokku.sh` stdout contains `...KEY=...`
+- **Why:** `AGENTS.md` Must-follow `NEVER commit credentials/.env` + GDPR Zero Trust — rotate leaked secrets immediately
+
+## 3. Deprecated Adapter
+
+- **Before (wrong):** `svelte-adapter-bun` — Bun-specific, incompatible with `gcr.io/distroless/static-debian13:nonroot` single binary
+- **After (right):** `@sveltejs/adapter-static` with `fallback: 'index.html'` + Axum `tower-http` `ServeDir`/`ServeFile`; `Vite → Rust musl → distroless`
+- **Signal:** `vite build` outputs `adapter-bun` warning, or Docker `CMD` fails on distroless (no Bun)
+- **Why:** 3.0.0 canonical 5-tier — Axum/Tokio sole coordinator, static SPA served by `tower-http`, preserves file routing without Bun
+
+## 4. `null` vs `Option`
+
+- **Before (wrong, C# model):** `if (user != null) { use(user) }` — assumes `null` is valid for any reference
+- **After (right, Rust model):** `if let Some(u) = user { use(u) }` or `user.map(|u| ...)` — absence encoded in type `Option<T>`
+- **Signal:** `cargo clippy` or `rustc` error `expected Option<T>, found T` or `NullReferenceException` becomes compile error
+- **Why:** Rust has no `null` — `Option` forces handling of `None` at compile time, prevents runtime panic
+
+## 5. Lifetime Elision
+
+- **Before (wrong):** `fn longest(x: &str, y: &str) -> &str` — elides lifetimes, compiler cannot infer borrowing
+- **After (right):** `fn longest<'a>(x: &'a str, y: &'a str) -> &'a str` — explicit `'a` ties output lifetime to inputs
+- **Signal:** `rustc` `error[E0106]: missing lifetime specifier` or `borrow checker: cannot return value referencing temporary`
+- **Why:** Borrow checker tracks references — output `&str` must borrow from `x` or `y`, explicit `'a` declares that
+
+## When to Use This File
+
+- Task touches `deps/Docker/HTML/auth` and `SKILL.md` Level 1 table was insufficient — read this file on-demand
+- Never load at startup — `AGENTS.md` pointer loads `SKILL.md` first; this file is Tier 3

package/skills/okf-docs/SKILL.md

@@ -0,0 +1,79 @@
+---
+name: okf-docs
+description: >
+  Author OKF v0.2-compliant documents — ADRs, module documentation, decision
+  records — with mandatory provenance frontmatter (type, generated by/at,
+  sources, status lifecycle) and mechanical validation via the bundled script.
+  Use when writing or updating any structured document in this workspace, or
+  when adding provenance to existing docs. Do NOT use when generating OpenSpec
+  reports from archived changes or analyzing reports for skill improvements.
+allowed-tools: Bash(node:*), Bash(bun:*)
+license: MIT
+compatibility: Requires node or bun for the bundled validator.
+metadata:
+  author: agentic
+  version: "1.0.0"
+positive_triggers:
+  - "write an ADR or decision record"
+  - "create module documentation"
+  - "add OKF frontmatter provenance to a document"
+anti_triggers:
+  - "generate an OpenSpec report from an archived change"
+  - "analyze reports and improve skills"
+---
+
+# Okf Docs
+
+Author structured documents with mandatory provenance. Every deliverable passes the bundled validator before hand-off.
+
+## Procedure
+
+### 1. Select Document Type
+
+| Intent | `type` value |
+|--------|--------------|
+| Architecture/design decision | Architecture Decision Record |
+| Module reference doc | Module Documentation |
+| Workspace directive | SystemDirective |
+| Skill definition | Skill |
+| Session report | Report / Assessment |
+
+### 2. Apply Frontmatter Contract
+
+```yaml
+---
+okf_version: "0.2"
+type: <from table above>
+title: <short name>
+description: <one-line summary>
+generated: { by: <actor>, at: <ISO 8601> }
+sources:
+  - { id: <source-id>, resource: <url|path> }
+status: draft          # draft | stable | deprecated
+stale_after: YYYY-MM-DD
+---
+```
+
+Actor values for `generated.by`: `<producer>/<version>` (agents) · `human:<id>` (people) · `process:<id>` (automation). Full semantics: [references/okf-spec.md](references/okf-spec.md).
+
+### 3. Attribute Claims
+
+Body claims sourced from `sources:` entries cite via footnotes (`[^source-id]`). No uncited external facts.
+
+### 4. Lifecycle
+
+Set `status: draft` on creation. Flip to `stable` after human review (`verified: { by: human:<id>, at }`). Set `stale_after` when content is time-sensitive. Deprecated docs keep standing but state successor.
+
+### 5. Validate Before Delivery
+
+```bash
+bun scripts/validate-frontmatter.mjs <document.md>
+```
+
+Envelope must read `pass: true`. Fix named FAIL entries; re-run until clean.
+
+## Gotchas
+
+- Dates are ISO 8601 only — no locale formats.
+- `stale_after` must not predate `generated.at`.
+- README files are promotional (zero tech detail) — OKF applies to technical docs, wikis, ADRs, directives.

package/skills/okf-docs/evals/benchmark.json

@@ -0,0 +1,21 @@
+{
+  "skill": "okf-docs",
+  "generated": { "by": "process:structural-stage/1.0", "at": "2026-08-23T19:45:00Z" },
+  "stage": "structural",
+  "structural": {
+    "validate_structure": { "pass": null, "note": "recorded at apply gate (task 5.1)" },
+    "validate_routing": {
+      "checks_total": 6,
+      "positive_triggers": 3,
+      "anti_triggers": 2,
+      "single_responsibility": true
+    },
+    "evals": {
+      "count": 3,
+      "assertions": 8,
+      "anti_trigger_coverage": true
+    }
+  },
+  "behavioral_dxm": "pending_cold_agent_run",
+  "ship_gate": { "criterion": "d = +1 and m >= 0.2", "applies_to": "behavioral stage" }
+}

package/skills/okf-docs/evals/evals.json

@@ -0,0 +1,37 @@
+{
+  "skill_name": "okf-docs",
+  "evals": [
+    {
+      "id": 1,
+      "prompt": "Write an ADR documenting our decision to adopt PostgreSQL as the central datastore.",
+      "expected_output": "Skill activates; authored document carries OKF v0.2 frontmatter (type: Architecture Decision Record, generated actor-convention provenance, sources list) and passes scripts/validate-frontmatter.mjs before delivery.",
+      "files": [],
+      "assertions": [
+        "Frontmatter contains type + generated.by/at in actor convention",
+        "Sources listed with footnote attribution for the decision context",
+        "Validator executed with pass:true before delivery"
+      ]
+    },
+    {
+      "id": 2,
+      "prompt": "Generate an OpenSpec report from the archived change and analyze it for skill improvements.",
+      "expected_output": "okf-docs does NOT activate. Report generation belongs to opsx-report; analysis belongs to opsx-learn.",
+      "files": [],
+      "assertions": [
+        "Anti-trigger fires: request routes to opsx-report/opsx-learn pipeline",
+        "No OKF authoring procedure started"
+      ]
+    },
+    {
+      "id": 3,
+      "prompt": "Add proper frontmatter to this legacy module doc so it passes validation.",
+      "expected_output": "Agent applies OKF v0.2 frontmatter (type per intent, generated provenance, status lifecycle), runs validate-frontmatter.mjs, iterates until pass:true, and leaves body content unaltered beyond attribution footnotes if claims need sourcing.",
+      "files": [],
+      "assertions": [
+        "Validator run on the updated document",
+        "Final envelope pass:true",
+        "Actor convention respected (human:<id> for user-authored claims)"
+      ]
+    }
+  ]
+}

package/skills/okf-docs/references/okf-spec.md

@@ -0,0 +1,56 @@
+---
+okf_version: "0.2"
+type: Reference
+title: OKF v0.2 Contract
+description: Distilled Open Knowledge Format v0.2 requirements — frontmatter schema, actor convention, trust lifecycle, separation rules, syntax conventions.
+generated: { by: ox-alpha/1.0, at: 2026-08-23T12:00:00Z }
+sources:
+  - { id: mini, resource: prototype/documentation-okf/AGENTS.md }
+  - { id: spec, resource: references/okf/SPEC.md }
+status: stable
+---
+
+# OKF v0.2 Contract
+
+## Frontmatter Schema
+
+| Field | Required | Constraint |
+|---|---|---|
+| `okf_version` | yes | `"0.2"` |
+| `type` | yes | Document kind (ADR · Module Documentation · SystemDirective · Skill · Report · Assessment) |
+| `title` | yes | Short name |
+| `description` | yes | One-line summary |
+| `generated` | yes | `{ by: <actor>, at: <ISO 8601> }` — replaces deprecated `timestamp` |
+| `sources` | recommended | List of `{ id, resource: <url\|path> }` — body claims cite via `[^source-id]` footnotes; replaces deprecated `# Citations` heading |
+| `verified` | optional | `{ by: human:<id>, at }` — human-reviewed tier |
+| `status` | recommended | `draft` \| `stable` \| `deprecated` |
+| `stale_after` | optional | `YYYY-MM-DD` — re-verify date for time-sensitive content |
+
+## Actor Convention
+
+`generated.by` / `verified.by` values:
+- `<producer>/<version>` — agents (e.g. `ox-alpha/1.0`, `openspec-report/1.0`)
+- `human:<id>` — people (only prefix granting human-reviewed tier)
+- `process:<id>` — automation
+
+## Trust & Lifecycle
+
+Created as `draft`. Human review flips to `stable` and records `verified`. Superseded content → `deprecated` (state successor). Time-sensitive docs carry `stale_after`; past that date, re-verify before trusting.
+
+## Separation Rules
+
+- **README**: promotional showcase for everyday users. Zero technical detail. Zero terminal blocks.
+- **Wiki** (`<module>/wiki/`): technical documentation per module. Synced remotely only if public.
+- **AGENTS.md**: agent operational directives — never duplicated into human docs.
+
+## Syntax Conventions
+
+Codebase patterns via `[✅ GOOD]` vs `[❌ BAD]` code blocks. No verbose prose explanations.
+
+## Progressive Disclosure
+
+`index.md` at directory roots synthesizes catalogs. Absolute markdown links (`[/backend/schema.md]`). Heavy reference data offloaded to `references/`, loaded on demand.
+
+## Reference Ingestion [CRITICAL]
+
+`./references/` present → scan + index via QMD → treat as READ-ONLY. Maintenance after doc mutations: `qmd update && qmd embed --chunk-strategy auto`.

package/skills/okf-docs/scripts/validate-frontmatter.mjs

@@ -0,0 +1,130 @@
+#!/usr/bin/env node
+/**
+ * validate-frontmatter.mjs — OKF v0.2 frontmatter gate
+ * Usage: node validate-frontmatter.mjs <document.md>
+ * Output: unified JSON envelope {target, pass, checks:[{id,status,detail}], summary}
+ *
+ * Checks: required keys (type/generated.by/at) · ISO 8601 formats · actor
+ * convention values · status enum · stale_after chronology.
+ */
+
+import { readFileSync, existsSync } from 'fs';
+
+const docPath = process.argv[2];
+
+if (!docPath) {
+  console.error(JSON.stringify({ error: 'Usage: node validate-frontmatter.mjs <document.md>' }));
+  process.exit(2);
+}
+
+if (!existsSync(docPath)) {
+  console.log(JSON.stringify({
+    target: docPath,
+    pass: false,
+    checks: [{ id: 'okf.file', status: 'FAIL', detail: 'Document not found' }],
+    summary: { total: 1, pass: 0, fail: 1, warn: 0, skip: 0 }
+  }));
+  process.exit(1);
+}
+
+const content = readFileSync(docPath, 'utf-8');
+const checks = [];
+
+const add = (id, ok, detail) => checks.push({ id, status: ok ? 'PASS' : 'FAIL', detail });
+const warn = (id, detail) => checks.push({ id, status: 'WARN', detail });
+
+// Extract frontmatter
+const fmMatch = content.match(/^---\r?\n([\s\S]*?)\r?\n---/);
+if (!fmMatch) {
+  console.log(JSON.stringify({
+    target: docPath,
+    pass: false,
+    checks: [{ id: 'okf.frontmatter', status: 'FAIL', detail: 'No frontmatter block found' }],
+    summary: { total: 1, pass: 0, fail: 1, warn: 0, skip: 0 }
+  }));
+  process.exit(1);
+}
+add('okf.frontmatter', true, 'Frontmatter block present');
+
+// Parse simple YAML subset (top-level + one nested level for generated)
+const fm = fmMatch[1];
+function get(key) {
+  const m = fm.match(new RegExp(`^${key}:\\s*(.+)$`, 'm'));
+  return m ? m[1].trim().replace(/^["']|["']$/g, '') : null;
+}
+function getNested(parent, key) {
+  const m = fm.match(new RegExp(`^${parent}:\\s*\\{\\s*([^}]*)\\}`, 'm'));
+  if (!m) return null;
+  const inner = m[1].match(new RegExp(`${key}:\\s*([^,}]+)`));
+  return inner ? inner[1].trim() : null;
+}
+
+// type (FAIL if missing/empty)
+const type = get('type');
+add('okf.type', !!type, type ? `type = ${type}` : "Missing required 'type'");
+
+// generated.by + at (FAIL if missing; actor convention + ISO for at)
+const by = getNested('generated', 'by');
+const at = getNested('generated', 'at');
+add('okf.generated-by', !!by, by ? `generated.by = ${by}` : "Missing required 'generated.by'");
+
+const actorRe = /^([a-z0-9][a-z0-9-]*\/\d+(\.\d+)*|human:[a-z0-9-]+|process:[a-z0-9-]+)$/;
+add('okf.actor-convention',
+  !!by && actorRe.test(by),
+  by ? (actorRe.test(by) ? `Actor '${by}' matches convention` : `Actor '${by}' violates convention (<producer>/<version> | human:<id> | process:<id>)`) : 'No actor to check');
+
+const isoRe = /^\d{4}-\d{2}-\d{2}(T\d{2}:\d{2}:\d{2}(\.\d+)?Z)?$/;
+add('okf.generated-at-format',
+  !!at && isoRe.test(at),
+  at ? (isoRe.test(at) ? `generated.at = ${at}` : `'${at}' is not ISO 8601`) : "Missing required 'generated.at'");
+
+// sources (WARN if missing — recommended not mandatory)
+const hasSources = /^sources:/m.test(fm);
+if (!hasSources) warn('okf.sources', "Missing 'sources:' list — claims uncited (recommended)");
+
+// status enum (FAIL on invalid value, WARN if missing)
+const status = get('status');
+if (!status) {
+  warn('okf.status', "Missing 'status' (draft | stable | deprecated)");
+} else if (!['draft', 'stable', 'deprecated'].includes(status)) {
+  failStatus(status);
+} else {
+  add('okf.status', true, `status = ${status}`);
+}
+function failStatus(value) {
+  add('okf.status', false, `Invalid status '${value}' — must be draft | stable | deprecated`);
+}
+
+// stale_after chronology vs generated.at
+const stale = get('stale_after');
+if (stale) {
+  if (!/^\d{4}-\d{2}-\d{2}$/.test(stale)) {
+    add('okf.stale-after', false, `stale_after '${stale}' is not YYYY-MM-DD`);
+  } else if (at) {
+    const genDay = at.slice(0, 10);
+    add('okf.stale-after', stale >= genDay,
+      stale >= genDay ? `stale_after ${stale} ≥ generated ${genDay}` : `stale_after ${stale} predates generation ${genDay}`);
+  } else {
+    add('okf.stale-after', true, `stale_after = ${stale} (no generated.at to compare)`);
+  }
+} else {
+  warn('okf.stale-after', 'No stale_after set (fine for non-time-sensitive docs)');
+}
+
+// Unified envelope
+const fails = checks.filter(c => c.status === 'FAIL').length;
+console.log(JSON.stringify({
+  target: docPath,
+  pass: fails === 0,
+  checks,
+  summary: {
+    total: checks.length,
+    pass: checks.filter(c => c.status === 'PASS').length,
+    fail: fails,
+    warn: checks.filter(c => c.status === 'WARN').length,
+    skip: checks.filter(c => c.status === 'SKIP').length
+  }
+}));
+
+process.exit(fails > 0 ? 1 : 0);
+

package/skills/openspec-harden/SKILL.md

@@ -0,0 +1,138 @@
+---
+name: openspec-harden
+description: >
+  Harden an existing OpenSpec change for cold application by a fresh agent.
+  Use when you have just run /opsx-propose and want to enrich the proposal,
+  specs, design, and tasks with concrete file paths, code blocks, verify steps,
+  and grounded codebase context before handing to a cold agent for /opsx-apply.
+  Do NOT use when implementing changes (use openspec-apply-change), creating a
+  new change from scratch, or when no change exists yet.
+allowed-tools: Bash(openspec:*), Read, Grep
+license: MIT
+compatibility: Requires openspec CLI and ripgrep.
+metadata:
+  author: agentic
+  version: "1.0.0"
+positive_triggers:
+  - "harden proposal for cold apply"
+  - "improve change for fresh agent"
+  - "enrich specs and tasks with file paths"
+anti_triggers:
+  - "implement changes from tasks or apply change"
+  - "implement the change and edit project code"
+  - "create a new change from scratch"
+---
+
+# Openspec Proposal Improve
+
+Enrich an existing change in place so a cold agent with no conversation memory can apply it correctly.
+
+**Planning boundary:** This skill operates in planning-only mode. Do NOT edit project code — only files under `openspec/changes/<name>/`. Proposal Why/What Changes are append-only: preserve original intent verbatim and mark any inferred addition as `Assumption:` or `Inferred:`.
+
+## Cold-Readiness Checklist (gate)
+
+All must pass or be explicitly marked `Needs human decision: <question>` before claiming hardening complete:
+
+- [ ] Every task has concrete file path (`path/to/file.ts:line` or `new file to create`) + exact shell command where applicable + `Verify:` clause (command, grep, or observable)
+- [ ] Every spec scenario is testable, has concrete example values, and includes at least one error/edge case
+- [ ] Design lists integration points grounded via codebase search as `file:line` refs
+- [ ] Design has rollback plan (or `Rollback: N/A — docs only` if truly no code)
+- [ ] No vague tasks remain ("update auth system" → must be split)
+- [ ] No untestable scenario remains ("works correctly" → must have WHEN/THEN with values)
+- [ ] Proposal Why/What Changes preserve original intent (no rewriting, only appending with Assumption/Inferred tags)
+
+If any item fails and cannot be inferred from codebase, pause and emit `Needs human decision: <question>` — do not hallucinate.
+
+Idempotent re-run: Running the skill twice on an already-hardened change without intervening edits SHALL produce no further artifact changes (idempotent).
+
+## Workflow — 6 Steps
+
+### Step 1: Load
+
+- Resolve store (if any):
+  ```bash
+  openspec store list --json
+  ```
+  If a store is selected (explicit user selection, or project `store:` pointer, or global `defaultStore`), treat `--store <id>` as sticky on every openspec command below. All examples without the flag are shorthand — append the flag before running.
+- Load change:
+  ```bash
+  openspec status --change "<name>" --json
+  ```
+  Parse `planningHome`, `changeRoot`, `artifactPaths`, `actionContext`. Use `artifactPaths` as the only source of artifact file paths — do not guess.
+- Read every file in `artifactPaths` (proposal, specs, design, tasks) from disk. Re-read even if seen before (user may have edited).
+
+### Step 2: Ground
+
+- Use `Grep` (ripgrep) to discover integration points: spec names, file paths, callers, schema files, tests, existing patterns. Example:
+  ```bash
+  Grep pattern="auth.*middleware" path="src"
+  Grep pattern="Requirement:" path="openspec/specs"
+  ```
+- Use `Read` to verify every candidate file path exists. If a path does not exist, mark task as `new file to create` — do not invent a path that fails `Grep`.
+- Record all inferred additions as `Assumption:` or `Inferred:` in the enriched artifact. Inferred file paths must be verified by `Grep`/`Read` or explicitly marked as new.
+
+Requires `Bash(openspec:*)` for status/instructions/validate and `Grep` for grounding. Every added path must be verified or marked new.
+
+### Step 3: Audit
+
+Compare loaded artifacts against the Cold-Readiness Checklist above. Build a gap list per artifact:
+- `proposal.md`: missing constraints, edge cases, out-of-scope, dependencies, rollback
+- `specs/**/spec.md`: vague scenarios, missing examples, missing error cases, untestable WHEN/THEN
+- `design.md`: missing file:line integration points, missing sequence/data-flow, missing rollback, missing ASCII diagram for cross-file flows
+- `tasks.md`: vague tasks, missing file:line, missing command, missing Verify, unordered dependencies
+
+If gap requires human decision (material ambiguity on scope/behavior/compatibility), prepare `Needs human decision: <question>` and pause per guardrails.
+
+### Step 4: Enrich
+
+Apply per-artifact rules, grounded in Step 2:
+
+- **proposal.md**: append only. Add `Assumption:`/`Inferred:` sections for missing constraints, edge cases (`Edge cases: ...`), out-of-scope (`Out-of-scope: ...`), dependencies (`Depends on: ...` with file paths). Never rewrite Why.
+- **specs/**/spec.md**: make scenarios concrete: add example payloads/code blocks, add error case scenario per requirement, ensure 4-hash `#### Scenario` format. Preserve existing scenarios; add, don't replace.
+- **design.md**: add `Integration points:` list as `file:line` refs from Ground step, add `Sequence:` or ASCII diagram if cross-file, add `Rollback:` plan. Keep existing Decisions.
+- **tasks.md**: split vague tasks into ordered small tasks (one file per task where possible) with `file:line`, exact command (`bun run check`, `cargo test`, etc.), and `Verify:` clause. Order topologically; note `requires X.Y` where needed.
+
+Rules: Do NOT edit project code — only `openspec/changes/<name>/`. Preserve original intent — proposal Why/What Changes are append-only.
+
+### Step 5: Validate
+
+- Re-run:
+  ```bash
+  openspec status --change "<name>" --json
+  openspec validate --change "<name>"  # or openspec validate --specs if needed
+  ```
+  With store flag if applicable (sticky).
+- Re-evaluate Cold-Readiness Checklist. If any item still fails, loop to Enrich or pause for human decision.
+- Verify idempotent re-run: second run without edits should be no-op.
+- Verify idempotence note: second run without edits should be no-op.
+
+### Step 6: Summarize
+
+Display:
+- Change name and store used
+- What was enriched per artifact (bullets) vs what was already cold-ready
+- Any `Assumption:`/`Inferred:` added and any `Needs human decision:` still open
+- Checklist result (pass/fail per item) and `openspec validate` result
+- Next: `Agent 2 (cold) can now run /opsx-apply <name>` or `openspec-apply-change`
+
+## Store Handling
+
+Mirror `openspec-propose`/`openspec-apply-change`: discover via `openspec store list --json`, pass `--store <id>` on every openspec command that accepts it, keep sticky. Hints printed by commands already carry the flag; keep it on follow-ups. Without a store, act on nearest local `openspec/` root.
+
+## Guardrails
+
+- Read-only commands and file reads (`Read`, `Grep`, `Bash(openspec status/instructions)`) need no confirmation.
+- Keep the planning boundary: Do NOT edit project code. Only `openspec/changes/<name>/` is writable.
+- Preserve original intent: proposal.md Why/What Changes are append-only. Never rewrite author's Why.
+- Ground every added path with `Grep`/`Read` or mark `new file to create`. Do not hallucinate file paths.
+- Never copy `<context>`/`<rules>` verbatim into artifacts — apply as constraints.
+- Pause on material ambiguity — do not absorb scope silently.
+
+## Reference
+
+- Direct triggering: harnesses invoke this skill directly via `/openspec-harden <change-name>` (no prompt shim needed).
+- Install: `project/install.ts` copies `project/skills/*` → `.agents/skills/` atomically — no manual wiring needed.
+
+## Related Skills
+
+Complements `openspec-learn` (proposal generation) and `openspec-report` (report meditation) — this skill hardens proposals for cold apply, while those create and reflect.

package/skills/openspec-harden/evals/benchmark.json

@@ -0,0 +1,40 @@
+{
+  "skill": "openspec-harden",
+  "generated": {
+    "by": "process:structural-stage/1.0",
+    "at": "2026-09-12T11:38:00Z"
+  },
+  "stage": "behavioral",
+  "structural": {
+    "validate_structure": {
+      "pass": true,
+      "note": "recorded at apply gate (task 3.4)"
+    },
+    "validate_routing": {
+      "checks_total": 6,
+      "positive_triggers": 3,
+      "anti_triggers": 3,
+      "single_responsibility": true
+    },
+    "evals": {
+      "count": 3,
+      "assertions": 9,
+      "anti_trigger_coverage": true
+    }
+  },
+  "behavioral_dxm": "1×0.33",
+  "ship_gate": {
+    "criterion": "d = +1 and m >= 0.2",
+    "applies_to": "behavioral stage"
+  },
+  "behavioral": {
+    "at": "2026-09-12T09:38:54.142Z",
+    "evals": 3,
+    "assertions": 9,
+    "baseline": 0.5556,
+    "with_skill": 0.8889,
+    "d": 1,
+    "m": 0.3333,
+    "ship": "pass"
+  }
+}