diff --git a/.github/workflows/docker-image.yml b/.github/workflows/docker-image.yml
index bc299fa14..576be4c7c 100644
--- a/.github/workflows/docker-image.yml
+++ b/.github/workflows/docker-image.yml
@@ -36,6 +36,10 @@ on:
 permissions:
   contents: read
   packages: write
+  # `gh release download` is a Releases-API act, and reading releases is
+  # its own permission under GitHub's per-key GITHUB_TOKEN mapping (the
+  # 2026-09-30 publish 403, documented at release.yml's publish-cli).
+  releases: read
 
 concurrency:
   group: docker-image
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 245416667..e779de8ca 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -1031,6 +1031,15 @@ jobs:
     timeout-minutes: 15
     permissions:
       contents: write
+      # RELEASES IS ITS OWN PERMISSION under GitHub's per-key re-mapping of
+      # GITHUB_TOKEN (the fine-grained model, rolling out org by org); the
+      # classic map where release operations rode contents: write is what
+      # softprops' README still describes. The 2026-09-30 desktop cut proved
+      # the difference is live here: POST /releases answered 403 "Resource
+      # not accessible by integration" on a token whose granted scopes
+      # printed Contents: write, while the same token pushed tags minutes
+      # earlier. Declaring it names the permission under either mapping.
+      releases: write
     strategy:
       fail-fast: false
       matrix:
@@ -1210,6 +1219,13 @@ jobs:
       needs.plan.result == 'success' &&
       needs.plan.outputs.desktop_matrix != '[]' &&
       (needs.publish-cli.result == 'success' || needs.publish-cli.result == 'skipped')
+    # stageSidecar FETCHES the CLI release's assets with the job token, and
+    # reading releases is the Releases permission under GitHub's per-key
+    # GITHUB_TOKEN mapping (see publish-cli). Narrowed from the workflow
+    # default because this job pushes nothing itself.
+    permissions:
+      contents: read
+      releases: read
     # One sign+notarize round trip per darwin artifact again — the mcp
     # companion is retired (the server binary serves its own `mcp` subcommand).
     # A desktop shard additionally compiles ~500 Rust crates (its sidecar is
@@ -1653,6 +1669,8 @@ jobs:
     timeout-minutes: 15
     permissions:
       contents: write
+      # The cut that 403'd; see publish-cli for why this is its own key.
+      releases: write
     strategy:
       fail-fast: false
       matrix:
diff --git a/scripts/__tests__/docker-image-chain.test.ts b/scripts/__tests__/docker-image-chain.test.ts
index 32fb0fa2e..90798f9ff 100644
--- a/scripts/__tests__/docker-image-chain.test.ts
+++ b/scripts/__tests__/docker-image-chain.test.ts
@@ -16,6 +16,14 @@ import { join } from "node:path";
 const WORKFLOW = readFileSync(join(import.meta.dir, "../../.github/workflows/docker-image.yml"), "utf8");
 
 describe("docker-image.yml chain", () => {
+  test("the release-asset download declares the Releases permission", () => {
+    // Same per-key GITHUB_TOKEN mapping as release.yml's publish jobs (the
+    // 2026-09-30 403, documented at publish-cli): gh release download is a
+    // Releases-API act and must not assume contents covers it.
+    const m = /^permissions:\n(?:[ \t]+\S.*\n)+/m.exec(WORKFLOW);
+    if (m === null) throw new Error("docker-image.yml: the permissions block is gone (reshaped?)");
+    expect(m[0]).toContain("releases: read");
+  });
   test("triggered by workflow_run on Release, never by the dead on: release", () => {
     expect(WORKFLOW).toMatch(/workflow_run:\s*\n\s*workflows: \["Release"\]/);
     // An `on: release` block would sit at the triggers' indentation level.
diff --git a/scripts/__tests__/release-cut-order.test.ts b/scripts/__tests__/release-cut-order.test.ts
index 4347c5711..f0a03e892 100644
--- a/scripts/__tests__/release-cut-order.test.ts
+++ b/scripts/__tests__/release-cut-order.test.ts
@@ -86,6 +86,16 @@ describe("cut phases: CLI first, desktop after", () => {
     // needs no such clause: every conjunct reads false on a dead plan's
     // empty outputs, so its gate skips on its own.)
     expect(header("build-desktop")).toContain("needs.plan.result == 'success'");
+    // The token-permission lesson of the 2026-09-30 403: under GitHub's
+    // per-key GITHUB_TOKEN mapping, release operations are the RELEASES
+    // permission, not contents. Every job that touches the Releases API
+    // must say so by key, or a server-side mapping change 403s it with the
+    // scope headers still printing Contents: write (which is exactly how
+    // that cut failed).
+    expect(header("publish-cli")).toContain("releases: write");
+    expect(header("publish-desktop")).toContain("releases: write");
+    // And the sidecar fetch reads release assets with the job token.
+    expect(header("build-desktop")).toContain("releases: read");
     // And no header may use always(), which would displace the gate the other
     // way — papering over a FAILED or CANCELLED chain. (The one always() in
     // the build bodies is a STEP guard, un-rooting the container workspace,