diff --git a/apps/streamdeck/package.json b/apps/streamdeck/package.json
index 10da5d802..f3b4c2195 100644
--- a/apps/streamdeck/package.json
+++ b/apps/streamdeck/package.json
@@ -8,9 +8,9 @@
     "build": "node esbuild.js",
     "watch": "node esbuild.js --watch",
     "check-types": "tsc --noEmit",
-    "validate": "streamdeck validate com.cluesmith.codev.sdPlugin",
+    "validate": "node scripts/validate.mjs",
     "pack": "streamdeck pack com.cluesmith.codev.sdPlugin --output dist --force",
-    "package": "node esbuild.js && rm -f com.cluesmith.codev.sdPlugin/bin/plugin.js.map && rm -rf com.cluesmith.codev.sdPlugin/logs && find com.cluesmith.codev.sdPlugin -name .DS_Store -delete && streamdeck validate com.cluesmith.codev.sdPlugin && streamdeck pack com.cluesmith.codev.sdPlugin --output dist --force",
+    "package": "node esbuild.js && rm -f com.cluesmith.codev.sdPlugin/bin/plugin.js.map && rm -rf com.cluesmith.codev.sdPlugin/logs && find com.cluesmith.codev.sdPlugin -name .DS_Store -delete && node scripts/validate.mjs && streamdeck pack com.cluesmith.codev.sdPlugin --output dist --force",
     "test": "vitest run"
   },
   "dependencies": {
diff --git a/apps/streamdeck/scripts/validate.d.mts b/apps/streamdeck/scripts/validate.d.mts
new file mode 100644
index 000000000..30dd686c5
--- /dev/null
+++ b/apps/streamdeck/scripts/validate.d.mts
@@ -0,0 +1,27 @@
+// Type declarations for validate.mjs so its exported, unit-tested core resolves under tsc
+// (NodeNext maps a `./validate.mjs` import to this `./validate.d.mts`) without a suppression.
+
+/** One run of the validate command: its exit code and combined stdout+stderr. */
+export interface RunResult {
+  code: number;
+  output: string;
+}
+
+/** A finished attempt loop: the last run's result plus how many attempts it took. */
+export interface BackoffResult extends RunResult {
+  attempts: number;
+}
+
+export interface BackoffOptions {
+  run: (attempt: number) => Promise<RunResult>;
+  attempts?: number;
+  baseBackoffMs?: number;
+  isTransient?: (output: string) => boolean;
+  sleep?: (ms: number) => Promise<unknown>;
+  log?: (message: string) => void;
+}
+
+export const TRANSIENT_SIGNATURES: string[];
+export const DEFAULTS: { attempts: number; baseBackoffMs: number };
+export function isTransientError(output: string): boolean;
+export function runWithBackoff(options?: BackoffOptions): Promise<BackoffResult>;
diff --git a/apps/streamdeck/scripts/validate.mjs b/apps/streamdeck/scripts/validate.mjs
new file mode 100644
index 000000000..fbb88c5fa
--- /dev/null
+++ b/apps/streamdeck/scripts/validate.mjs
@@ -0,0 +1,124 @@
+// Run `streamdeck validate` with bounded re-attempts + backoff around transient network errors (#1436).
+//
+// WHY: the Elgato CLI's `manifestUrlsExist` validation rule does a live HEAD request to the
+// manifest's top-level `URL` (ours is https://github.com/cluesmith/codev). Its catch block turns
+// only `ENOTFOUND` into a graceful validation error; ANY other fetch failure (UND_ERR_SOCKET,
+// ECONNRESET, "fetch failed", …) is rethrown and crashes the whole `validate` run. That put the
+// network on CI's pass/fail path and flaked unrelated PRs (#1432, #1434) with no code defect.
+//
+// FIX: run the WHOLE validate command again a few times with exponential backoff, but ONLY when
+// the failure output matches a transient-network signature. Real validation errors fail fast on
+// the first attempt (no masking, no wasted backoff). `--no-update-check` does NOT help here — it
+// only gates the separate schema-update fetch, not this URL-reachability probe.
+//
+// The core loop is exported and unit-tested (see src/__tests__/validate.test.ts);
+// `main()` only wires it to the real child process. Mirrors scripts/render-action-icons.mjs.
+
+import { execFile } from 'node:child_process';
+import { dirname, join } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const HERE = dirname(fileURLToPath(import.meta.url));
+const PLUGIN_DIR = 'com.cluesmith.codev.sdPlugin';
+
+// Case-insensitive signatures of a transient network failure worth another attempt. These are the
+// error shapes the CLI rethrows from a failed `fetch(URL, { method: 'HEAD' })`. Each is a specific
+// error code/phrase, not a loose word, so a plugin description that merely mentions "network"
+// can't trigger a false re-run.
+//
+// EAI_AGAIN (temporary DNS failure) IS here but ENOTFOUND (permanent "host doesn't exist") is
+// NOT — and that asymmetry is deliberate: the CLI already converts ENOTFOUND into a graceful
+// "must be resolvable" validation error rather than rethrowing it, so a genuinely bad URL fails
+// loudly on attempt 1, while a transient DNS blip gets another attempt.
+export const TRANSIENT_SIGNATURES = [
+  'UND_ERR_SOCKET',
+  'UND_ERR_CONNECT_TIMEOUT',
+  'fetch failed',
+  'ECONNRESET',
+  'ECONNREFUSED',
+  'ETIMEDOUT',
+  'EAI_AGAIN',
+  'ENETUNREACH',
+  'ENETDOWN',
+  'socket hang up',
+];
+
+export const DEFAULTS = { attempts: 3, baseBackoffMs: 1000 };
+
+/** Does this combined stdout+stderr look like a transient network failure? */
+export function isTransientError(output) {
+  const haystack = String(output ?? '').toLowerCase();
+  return TRANSIENT_SIGNATURES.some((sig) => haystack.includes(sig.toLowerCase()));
+}
+
+const defaultSleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
+
+/**
+ * Bounded attempt loop. `run()` resolves to { code, output }. We run again only when a non-zero
+ * result is transient AND attempts remain; otherwise we return the last result (fail fast on real
+ * errors, and still surface the failure after exhausting transient attempts). Injectable `run`,
+ * `sleep`, and `log` keep this deterministic under test.
+ */
+export async function runWithBackoff({
+  run,
+  attempts = DEFAULTS.attempts,
+  baseBackoffMs = DEFAULTS.baseBackoffMs,
+  isTransient = isTransientError,
+  sleep = defaultSleep,
+  log = () => {},
+} = {}) {
+  let last;
+  for (let attempt = 1; attempt <= attempts; attempt++) {
+    last = await run(attempt);
+    if (last.code === 0) {
+      return { ...last, attempts: attempt };
+    }
+    const transient = isTransient(last.output);
+    const hasMore = attempt < attempts;
+    if (!transient || !hasMore) {
+      return { ...last, attempts: attempt };
+    }
+    const backoff = baseBackoffMs * 2 ** (attempt - 1);
+    log(
+      `streamdeck validate: transient network error on attempt ${attempt}/${attempts}; ` +
+        `trying again in ${backoff}ms…`,
+    );
+    await sleep(backoff);
+  }
+  return { ...last, attempts };
+}
+
+/** Spawn `streamdeck validate <plugin>` once, capturing combined output. */
+function runValidateOnce() {
+  return new Promise((resolve) => {
+    const child = execFile(
+      'streamdeck',
+      ['validate', PLUGIN_DIR],
+      { cwd: join(HERE, '..'), encoding: 'utf-8', maxBuffer: 10 * 1024 * 1024 },
+      (error, stdout, stderr) => {
+        let output = `${stdout ?? ''}${stderr ?? ''}`;
+        // A spawn failure (e.g. the CLI isn't on PATH) yields empty stdio; surface the error
+        // message so CI shows *why* rather than an exit 1 with no diagnostic. ENOENT and the like
+        // aren't in TRANSIENT_SIGNATURES, so appending it won't trigger a spurious re-run.
+        if (error && output.trim() === '') {
+          output = `${error.message}\n`;
+        }
+        resolve({ code: error ? (error.code ?? 1) : 0, output });
+      },
+    );
+    // The execFile callback already receives spawn errors; this handler just prevents an
+    // unhandled 'error' event from crashing the process before the callback resolves.
+    child.on('error', () => {});
+  });
+}
+
+async function main() {
+  const result = await runWithBackoff({ run: runValidateOnce, log: (m) => console.warn(m) });
+  process.stdout.write(result.output);
+  process.exit(result.code === 0 ? 0 : 1);
+}
+
+// Only run when invoked as a script, not when imported by the test.
+if (process.argv[1] && fileURLToPath(import.meta.url) === process.argv[1]) {
+  main();
+}
diff --git a/apps/streamdeck/src/__tests__/validate.test.ts b/apps/streamdeck/src/__tests__/validate.test.ts
new file mode 100644
index 000000000..5300b2b61
--- /dev/null
+++ b/apps/streamdeck/src/__tests__/validate.test.ts
@@ -0,0 +1,87 @@
+import { describe, it, expect, vi } from 'vitest';
+import {
+  runWithBackoff,
+  isTransientError,
+  TRANSIENT_SIGNATURES,
+} from '../../scripts/validate.mjs';
+
+/**
+ * #1436: `streamdeck validate`'s `manifestUrlsExist` rule does a live HEAD request to the
+ * manifest's `URL`. Anything other than ENOTFOUND (UND_ERR_SOCKET, "fetch failed", …) is rethrown
+ * and crashes the whole validate run, flaking unrelated PRs' CI. The fix wraps the invocation in a
+ * bounded loop that runs again ONLY on transient network failures. These tests pin that contract:
+ * they fail against a single-shot implementation and pass with the backoff loop.
+ */
+
+const ok = { code: 0, output: 'Validation successful.' };
+const socketFail = { code: 1, output: 'validate failed\nTypeError: fetch failed\n  UND_ERR_SOCKET' };
+const realFail = {
+  code: 1,
+  output: 'manifest.json\n  error: Actions must not be empty\n1 error',
+};
+
+// No real waiting under test.
+const noSleep = () => Promise.resolve();
+
+describe('isTransientError', () => {
+  it('matches the observed socket errors', () => {
+    expect(isTransientError('TypeError: fetch failed\n  UND_ERR_SOCKET')).toBe(true);
+    expect(isTransientError('read ECONNRESET')).toBe(true);
+    expect(isTransientError('connect ETIMEDOUT')).toBe(true);
+  });
+
+  it('does not match a real validation failure or ENOTFOUND', () => {
+    expect(isTransientError('error: Actions must not be empty')).toBe(false);
+    // ENOTFOUND is reported by the CLI as a normal "must be resolvable" error, not a crash.
+    expect(isTransientError('URL must be resolvable (ENOTFOUND)')).toBe(false);
+  });
+
+  it('is case-insensitive across every declared signature', () => {
+    for (const sig of TRANSIENT_SIGNATURES) {
+      expect(isTransientError(`prefix ${sig.toUpperCase()} suffix`)).toBe(true);
+      expect(isTransientError(`prefix ${sig.toLowerCase()} suffix`)).toBe(true);
+    }
+  });
+});
+
+describe('runWithBackoff', () => {
+  it('recovers from a transient failure then a success (fails with a single attempt)', async () => {
+    const run = vi.fn().mockResolvedValueOnce(socketFail).mockResolvedValueOnce(ok);
+    const result = await runWithBackoff({ run, sleep: noSleep });
+    expect(run).toHaveBeenCalledTimes(2);
+    expect(result.code).toBe(0);
+    expect(result.attempts).toBe(2);
+  });
+
+  it('runs up to the attempt cap on repeated transient failures, then surfaces the failure', async () => {
+    const run = vi.fn().mockResolvedValue(socketFail);
+    const result = await runWithBackoff({ run, attempts: 3, sleep: noSleep });
+    expect(run).toHaveBeenCalledTimes(3);
+    expect(result.code).toBe(1);
+    expect(result.attempts).toBe(3);
+  });
+
+  it('fails fast on a real validation error without a second attempt', async () => {
+    const run = vi.fn().mockResolvedValue(realFail);
+    const result = await runWithBackoff({ run, attempts: 3, sleep: noSleep });
+    expect(run).toHaveBeenCalledTimes(1);
+    expect(result.code).toBe(1);
+  });
+
+  it('applies exponential backoff between transient attempts', async () => {
+    const run = vi.fn().mockResolvedValue(socketFail);
+    const sleep = vi.fn().mockResolvedValue(undefined);
+    await runWithBackoff({ run, attempts: 3, baseBackoffMs: 1000, sleep });
+    // Two waits between three attempts: 1000ms then 2000ms.
+    expect(sleep.mock.calls.map((c) => c[0])).toEqual([1000, 2000]);
+  });
+
+  it('returns immediately on first-attempt success', async () => {
+    const run = vi.fn().mockResolvedValue(ok);
+    const sleep = vi.fn().mockResolvedValue(undefined);
+    const result = await runWithBackoff({ run, sleep });
+    expect(run).toHaveBeenCalledTimes(1);
+    expect(sleep).not.toHaveBeenCalled();
+    expect(result.code).toBe(0);
+  });
+});
diff --git a/codev/projects/bugfix-1436-ci-streamdeck-validate-step-fl/status.yaml b/codev/projects/bugfix-1436-ci-streamdeck-validate-step-fl/status.yaml
new file mode 100644
index 000000000..e907374eb
--- /dev/null
+++ b/codev/projects/bugfix-1436-ci-streamdeck-validate-step-fl/status.yaml
@@ -0,0 +1,22 @@
+id: bugfix-1436
+title: ci-streamdeck-validate-step-fl
+protocol: bugfix
+phase: pr
+plan_phases: []
+current_plan_phase: null
+gates:
+  pr:
+    status: approved
+    requested_at: '2026-08-14T00:40:33.414Z'
+    approved_at: '2026-08-15T02:53:15.320Z'
+iteration: 1
+build_complete: false
+history: []
+started_at: '2026-08-14T00:27:39.607Z'
+updated_at: '2026-08-15T02:53:15.321Z'
+pr_ready_for_human: false
+pr_history:
+  - phase: pr
+    pr_number: 1451
+    branch: builder/bugfix-1436
+    created_at: '2026-08-14T00:40:49.690Z'
diff --git a/codev/state/bugfix-1436_thread.md b/codev/state/bugfix-1436_thread.md
new file mode 100644
index 000000000..816b417a0
--- /dev/null
+++ b/codev/state/bugfix-1436_thread.md
@@ -0,0 +1,78 @@
+# bugfix-1436 — streamdeck validate CI flake on transient network errors
+
+Issue #1436. BUGFIX protocol, strict mode.
+
+## Investigate (iter 1)
+
+### Root cause (traced from source, not assumed)
+
+The flake is NOT a schema-update fetch as the issue title guessed. The exact failure
+path is the Elgato CLI validation rule `manifestUrlsExist`
+(`@elgato/cli` dist, rule from `src/validation/plugin/rules/manifest-urls-exist.ts`):
+
+```js
+const { status } = await fetch(url.value, { method: "HEAD" });   // url = manifest top-level "URL"
+...
+} catch (err) {
+  if (err.cause?.code === "ENOTFOUND") {
+    this.addError(..., "must be resolvable", url);   // graceful validation error
+  } else {
+    throw err;   // <-- ANY other fetch error (UND_ERR_SOCKET, ECONNRESET, ETIMEDOUT,
+                  //     "fetch failed") is RETHROWN → crashes `streamdeck validate` → CI job fails
+  }
+}
+```
+
+Our manifest declares `"URL": "https://github.com/cluesmith/codev"`
+(`com.cluesmith.codev.sdPlugin/manifest.json:10`). `streamdeck validate` does a live
+HEAD request to that URL every run. A transient socket error (not ENOTFOUND) is
+rethrown unhandled and fails the whole validate step — exactly the observed
+`UND_ERR_SOCKET` / `fetch failed` on PRs #1432, #1434, with no code defect.
+
+### Offline fallback verified NOT viable
+CLI `--help` empirically checked: `streamdeck validate` has `--no-update-check`
+("Disables updating schemas") and `--force-update-check`. But those only gate the
+SCHEMA update; they do NOT disable the manifest-URL reachability probe. So offline /
+`--no-update-check` would NOT remove this flake. Retry is the correct and only clean fix.
+Schemas themselves are bundled locally via `@elgato/schemas`.
+
+### Fix (architect preference: bounded retry with backoff)
+Wrap the `streamdeck validate` invocation in a bounded retry (3 attempts, exp backoff),
+retrying ONLY on transient/network error signatures; fail fast on real validation errors.
+Place a small testable helper `apps/streamdeck/scripts/validate.mjs` (mirrors
+the existing `scripts/render-action-icons.mjs` + matching vitest test pattern) and point
+the `validate` npm script at it. Both CI workflows call `pnpm validate`
+(test.yml:113, sdk-canary.yml:57), so this fixes the flake at both sites. Local `package`
+script inlines `streamdeck validate` — swap that one call for the helper too (no script
+restructuring).
+
+Regression test: simulate a transient failure then success against the retry helper;
+fails without retry (single-shot throws), passes with it. Real validation errors must
+NOT be retried.
+
+Scope: << 300 LOC. Fits BUGFIX.
+
+## Fix + PR (iter 1)
+
+Implemented `apps/streamdeck/scripts/validate.mjs` (retry core exported +
+unit-tested), wired `package.json` `validate` + inline `package` call to it. 170
+streamdeck tests pass (11 new). Verified end-to-end: happy path exit 0, real error
+fails fast (exit 1), transient retries in unit tests, regression test fails without fix.
+Committed 84fd5e785. Porch checks (build, tests) passed. PR #1451 opened (Fixes #1436).
+
+CMAP: first run failed to auto-detect project from worktree ("Multiple projects found");
+re-ran with `--issue 1436 --project-id bugfix-1436` — both resolve PR #1451 correctly.
+Awaiting all three verdicts before notifying architect + `porch done` (pr gate).
+
+## CMAP verdicts + review fixes
+
+- gemini=APPROVE (HIGH), codex=APPROVE (HIGH), claude=COMMENT (HIGH, non-blocking nits).
+- Addressed claude's substantive points:
+  - Dropped the over-broad `'network'` substring from TRANSIENT_SIGNATURES (could false-retry
+    a plugin description mentioning "network"); added specific `ENETUNREACH`/`ENETDOWN`.
+  - Spawn failure now surfaces `error.message` in output instead of exit-1-with-empty-output.
+  - Sharpened the comment on the deliberate EAI_AGAIN (retry) vs ENOTFOUND (fail-fast) asymmetry.
+  - Corrected PR body test count: 8 new tests (170 suite total), not 11.
+- Left as-is (correct by design): ENOTFOUND exclusion — the CLI already reports it as a graceful
+  "must be resolvable" error, so we must not retry a genuinely bad URL.
+- 170 tests still pass after the fixes.