diff --git a/config/scripts/generate-bundled-skill-guides.test.mjs b/config/scripts/generate-bundled-skill-guides.test.mjs
index d2743d319d3..8c0a2a782c8 100644
--- a/config/scripts/generate-bundled-skill-guides.test.mjs
+++ b/config/scripts/generate-bundled-skill-guides.test.mjs
@@ -1,5 +1,5 @@
 import { execFile } from 'node:child_process'
-import { cp, mkdir, mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'
+import { cp, mkdir, mkdtemp, readFile, readdir, rm, writeFile } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
 import path from 'node:path'
 import { promisify } from 'node:util'
@@ -351,3 +351,55 @@ describe('bundled skill guide generator', () => {
     await expect(buildArtifacts(root)).rejects.toThrow('Guide reference is empty')
   })
 })
+
+// Why generalized: `orchestration-skill-guidance.test.mjs` pins this both-directions routing for
+// orchestration alone. Any guide that grows a `references/` directory needs the same contract, or a
+// reference can ship unroutable or a gate can route a file that does not exist.
+describe('guide reference routing', () => {
+  async function guidesWithReferences() {
+    const guideRoot = path.join(projectDir, 'skill-guides')
+    const entries = await readdir(guideRoot, { withFileTypes: true })
+    const owners = []
+    for (const entry of entries.filter((candidate) => candidate.isDirectory())) {
+      const referenceRoot = path.join(guideRoot, entry.name, 'references')
+      const shipped = await readdir(referenceRoot).catch(() => null)
+      if (shipped === null) {
+        continue
+      }
+      owners.push({
+        name: entry.name,
+        referenceRoot,
+        shipped: shipped.filter((file) => file.endsWith('.md')).sort()
+      })
+    }
+    return owners
+  }
+
+  it('routes every shipped reference from its own guide, in both directions', async () => {
+    const owners = await guidesWithReferences()
+    // A vacuous loop would pass forever; orchestration is the guide that owns references today.
+    expect(owners.map((owner) => owner.name)).toContain('orchestration')
+
+    const mismatches = []
+    for (const owner of owners) {
+      const guidePath = path.join(projectDir, 'skill-guides', `${owner.name}.md`)
+      const guide = await readFile(guidePath, 'utf8').catch(() => null)
+      if (guide === null) {
+        mismatches.push(`${owner.name}: references/ exists with no ${owner.name}.md beside it`)
+        continue
+      }
+      const routed = [
+        ...new Set([...guide.matchAll(/`references\/([^`]+\.md)`/gu)].map((match) => match[1]))
+      ].sort()
+      const unshipped = routed.filter((file) => !owner.shipped.includes(file))
+      const unrouted = owner.shipped.filter((file) => !routed.includes(file))
+      if (unshipped.length > 0) {
+        mismatches.push(`${owner.name}: routes missing ${unshipped}`)
+      }
+      if (unrouted.length > 0) {
+        mismatches.push(`${owner.name}: ships unrouted ${unrouted}`)
+      }
+    }
+    expect(mismatches).toEqual([])
+  })
+})
diff --git a/config/scripts/skill-description-length.test.mjs b/config/scripts/skill-description-length.test.mjs
index e7a9db79541..b39af4b6da5 100644
--- a/config/scripts/skill-description-length.test.mjs
+++ b/config/scripts/skill-description-length.test.mjs
@@ -7,6 +7,10 @@ const skillsDir = resolve(import.meta.dirname, '../../skills')
 // Why: the Agent Skills spec caps `description` at 1024 chars and conforming installers
 // reject the whole skill (#17935); the frontmatter is what the installer parses, so check it.
 const MAX_DESCRIPTION_LENGTH = 1024
+// Why raw, not backtick-stripped: NVIDIA SkillEvaluator rejects `<tag>` in a description as a
+// schema error, and Cowork's validator parses descriptions as HTML and fails the whole plugin
+// silently (compound-engineering #602). Neither honors backticks, so placeholders belong in the body.
+const ANGLE_BRACKET_TOKEN = /<[A-Za-z][\w.-]*>/u
 
 function readDescription(skillName) {
   const skillMarkdown = readFileSync(join(skillsDir, skillName, 'SKILL.md'), 'utf8')
@@ -36,4 +40,13 @@ describe('bundled skill descriptions', () => {
       `${name}: description is ${description.length} chars`
     ).toBeLessThanOrEqual(MAX_DESCRIPTION_LENGTH)
   })
+
+  it.each(skillNames)('%s keeps angle-bracket placeholders out of its description', (name) => {
+    const token = ANGLE_BRACKET_TOKEN.exec(readDescription(name) ?? '')
+
+    expect(
+      token?.[0],
+      `${name}: rephrase or move "${token?.[0] ?? ''}" into the skill body`
+    ).toBeUndefined()
+  })
 })
diff --git a/config/scripts/skill-guide-size-budget.test.mjs b/config/scripts/skill-guide-size-budget.test.mjs
new file mode 100644
index 00000000000..13ee7d7f186
--- /dev/null
+++ b/config/scripts/skill-guide-size-budget.test.mjs
@@ -0,0 +1,74 @@
+import { readdirSync, readFileSync } from 'node:fs'
+import { join, resolve } from 'node:path'
+import { describe, expect, it } from 'vitest'
+
+const guideRoot = resolve(import.meta.dirname, '../../skill-guides')
+
+/**
+ * Provenance: the Agent Skills spec's "keep your main SKILL.md under 500 lines" is an explicit
+ * recommendation, not a limit, and nothing rejects a longer guide. 300 is the tighter bound this
+ * repo already practices — six of eight guides sit under it, and `orchestration.md` holds 201 lines
+ * by routing detail into `references/`, which is the restructure this budget is meant to push.
+ * A line count is not a token count; treat a green run as a shape check, not a context-budget proof.
+ */
+const MAX_GUIDE_LINES = 300
+
+/**
+ * Guides that already exceed the bound, with the size they may not grow past. Recorded sizes are a
+ * ratchet ceiling, not a target: shrink them freely and delete the entry once the guide fits.
+ * A name may leave this set. A name may never join it — split the guide into `references/` instead.
+ */
+const OVER_BUDGET = new Map([
+  ['orca-cli', 424],
+  ['orca-per-workspace-env', 794]
+])
+
+/** Matches `wc -l`: a trailing newline ends the last line rather than starting a new one. */
+function lineCount(contents) {
+  const lines = contents.split(/\r?\n/u)
+  return lines.at(-1) === '' ? lines.length - 1 : lines.length
+}
+
+function guideSizes() {
+  return new Map(
+    readdirSync(guideRoot, { withFileTypes: true })
+      .filter((entry) => entry.isFile() && entry.name.endsWith('.md'))
+      .map((entry) => [
+        entry.name.replace(/\.md$/u, ''),
+        lineCount(readFileSync(join(guideRoot, entry.name), 'utf8'))
+      ])
+  )
+}
+
+describe('always-loaded skill guide size budget', () => {
+  const sizes = guideSizes()
+
+  it('measures every shipped guide', () => {
+    expect(sizes.size).toBeGreaterThanOrEqual(8)
+    expect(sizes.get('orchestration')).toBeGreaterThan(0)
+  })
+
+  it('keeps every guide outside OVER_BUDGET under the bound', () => {
+    const violations = [...sizes]
+      .filter(([name, size]) => size > MAX_GUIDE_LINES && !OVER_BUDGET.has(name))
+      .map(([name, size]) => `${name}: ${size} lines > ${MAX_GUIDE_LINES}`)
+
+    expect(violations).toEqual([])
+  })
+
+  it('never lets an OVER_BUDGET guide grow past its recorded size', () => {
+    const grown = [...OVER_BUDGET]
+      .filter(([name, ceiling]) => (sizes.get(name) ?? 0) > ceiling)
+      .map(([name, ceiling]) => `${name}: ${sizes.get(name)} lines > recorded ${ceiling}`)
+
+    expect(grown).toEqual([])
+  })
+
+  it('drops OVER_BUDGET entries that now fit, so the set only ratchets down', () => {
+    const stale = [...OVER_BUDGET.keys()].filter(
+      (name) => !sizes.has(name) || (sizes.get(name) ?? 0) <= MAX_GUIDE_LINES
+    )
+
+    expect(stale).toEqual([])
+  })
+})
diff --git a/src/cli/skill-guide-cli-parity.test.ts b/src/cli/skill-guide-cli-parity.test.ts
new file mode 100644
index 00000000000..11a2ad15c8d
--- /dev/null
+++ b/src/cli/skill-guide-cli-parity.test.ts
@@ -0,0 +1,177 @@
+import { readdirSync, readFileSync } from 'node:fs'
+import { join, relative, resolve } from 'node:path'
+import { describe, expect, it } from 'vitest'
+import { CLI_GLOBAL_FLAGS } from '../shared/cli-argument-boundary'
+import { specPaths } from './command-spec'
+import { COMMAND_SPECS } from './specs'
+
+// Why: a guide is the version-matched surface for the binary that shipped it, so a command
+// path or flag it names must exist in COMMAND_SPECS. `orca emulator camera --webcam` was
+// documented for months without ever existing (#16904 review C1).
+
+// Why __dirname: it works under both Vitest and the CommonJS tsc emit that build:cli type-checks
+// this file against; import.meta.dirname does not (TS1470).
+const projectDir = resolve(__dirname, '..', '..')
+const guideRoot = join(projectDir, 'skill-guides')
+const MAX_COMMAND_DEPTH = 3
+
+type Invocation = { file: string; line: number; text: string }
+
+function guideFiles(directory: string): string[] {
+  return readdirSync(directory, { withFileTypes: true }).flatMap((entry) => {
+    const full = join(directory, entry.name)
+    if (entry.isDirectory()) {
+      return guideFiles(full)
+    }
+    return entry.isFile() && entry.name.endsWith('.md') ? [full] : []
+  })
+}
+
+/**
+ * The invocation span is the command text only — never the surrounding prose or table cell.
+ * `skill-guides/orca-emulator.md` describes serve-sim's own `--detach` in a Notes column beside
+ * an `ORCA ...` cell, and that is correct prose a line-scoped check would flag.
+ */
+function invocationSpans(contents: string, file: string): Invocation[] {
+  const found: Invocation[] = []
+  let inFence = false
+  contents.split(/\r?\n/u).forEach((line, index) => {
+    if (/^\s*(?:```|~~~)/u.test(line)) {
+      inFence = !inFence
+      return
+    }
+    const spans = inFence ? [line] : [...line.matchAll(/`([^`]+)`/gu)].map((match) => match[1])
+    for (const span of spans) {
+      const starts = [...span.matchAll(/\bORCA\b/gu)].map((match) => match.index)
+      starts.forEach((start, position) => {
+        found.push({
+          file,
+          line: index + 1,
+          text: span.slice(start, starts[position + 1] ?? span.length).trim()
+        })
+      })
+    }
+  })
+  return found
+}
+
+/** Blank out quoted values so a nested `--model` inside `--command "codex --model ..."` is not read as a flag. */
+function maskQuotedValues(text: string): string {
+  let masked = ''
+  let quote: string | null = null
+  for (const character of text) {
+    if (quote) {
+      masked += character === quote ? character : ' '
+      if (character === quote) {
+        quote = null
+      }
+    } else if (character === '"' || character === "'") {
+      quote = character
+      masked += character
+    } else {
+      masked += character
+    }
+  }
+  return masked
+}
+
+const specByPath = new Map<string, (typeof COMMAND_SPECS)[number]>()
+const pathPrefixes = new Set<string>()
+for (const spec of COMMAND_SPECS) {
+  for (const path of specPaths(spec)) {
+    specByPath.set(path.join(' '), spec)
+    for (let length = 1; length < path.length; length += 1) {
+      pathPrefixes.add(path.slice(0, length).join(' '))
+    }
+  }
+}
+
+function longestKnownPrefix(tokens: string[]): string | null {
+  for (let length = tokens.length; length >= 1; length -= 1) {
+    const candidate = tokens.slice(0, length).join(' ')
+    if (specByPath.has(candidate) || pathPrefixes.has(candidate)) {
+      return candidate
+    }
+  }
+  return null
+}
+
+function allowedFlagsFor(prefix: string): Set<string> {
+  const exact = specByPath.get(prefix)
+  const flags = new Set<string>(CLI_GLOBAL_FLAGS)
+  const specs = exact
+    ? [exact]
+    : COMMAND_SPECS.filter((spec) =>
+        specPaths(spec).some((path) => path.join(' ').startsWith(`${prefix} `))
+      )
+  for (const spec of specs) {
+    for (const flag of spec.allowedFlags) {
+      flags.add(flag)
+    }
+  }
+  return flags
+}
+
+function describeFailure(invocation: Invocation, detail: string): string {
+  const location = `${relative(projectDir, invocation.file)}:${invocation.line}`
+  return `${location}: ${detail}\n    ${invocation.text}`
+}
+
+function parityFailures(invocation: Invocation): string[] {
+  const masked = maskQuotedValues(invocation.text).replace(/\s#.*$/u, '')
+  const tokens: string[] = []
+  for (const token of masked.slice('ORCA'.length).trim().split(/\s+/u)) {
+    if (!/^[a-z][a-z0-9-]*$/u.test(token) || tokens.length === MAX_COMMAND_DEPTH) {
+      break
+    }
+    tokens.push(token)
+  }
+  if (tokens.length === 0) {
+    return []
+  }
+
+  const failures: string[] = []
+  let command: string | null = null
+  for (let length = tokens.length; length >= 1 && command === null; length -= 1) {
+    const candidate = tokens.slice(0, length).join(' ')
+    if (specByPath.has(candidate)) {
+      command = candidate
+    }
+  }
+  if (command === null) {
+    // An incomplete reference like `ORCA emulator ...` names a real prefix and pins no flags.
+    if (pathPrefixes.has(tokens.join(' '))) {
+      return failures
+    }
+    failures.push(
+      describeFailure(invocation, `no COMMAND_SPECS path or alias for "${tokens.join(' ')}"`)
+    )
+    command = longestKnownPrefix(tokens)
+    if (command === null) {
+      return failures
+    }
+  }
+
+  const allowed = allowedFlagsFor(command)
+  for (const match of masked.matchAll(/--([a-z][a-z0-9-]*)/gu)) {
+    if (!allowed.has(match[1])) {
+      failures.push(describeFailure(invocation, `--${match[1]} is not a flag of "${command}"`))
+    }
+  }
+  return failures
+}
+
+describe('skill guides only name commands and flags the CLI defines', () => {
+  const invocations = guideFiles(guideRoot).flatMap((file) =>
+    invocationSpans(readFileSync(file, 'utf8'), file)
+  )
+
+  it('extracts invocations from every guide and reference', () => {
+    expect(invocations.length).toBeGreaterThan(150)
+    expect(new Set(invocations.map((invocation) => invocation.file)).size).toBeGreaterThan(8)
+  })
+
+  it('resolves every ORCA invocation against COMMAND_SPECS', () => {
+    expect(invocations.flatMap(parityFailures)).toEqual([])
+  })
+})