diff --git a/.github/dependabot.yml b/.github/dependabot.yml
new file mode 100644
index 0000000..d37d95f
--- /dev/null
+++ b/.github/dependabot.yml
@@ -0,0 +1,34 @@
+version: 2
+
+# Grouped so a week's updates arrive as one reviewable change per ecosystem
+# rather than one per dependency. The packages under `packages/` share a lock
+# discipline, so pip updates are grouped across the repository.
+updates:
+  - package-ecosystem: pip
+    directory: /
+    schedule:
+      interval: weekly
+    open-pull-requests-limit: 3
+    groups:
+      # Torch decides what the engine can run and how large every artifact
+      # is, so it lands on its own rather than inside a bulk update.
+      torch:
+        patterns:
+          - torch
+          - torchvision
+      python:
+        patterns:
+          - "*"
+        exclude-patterns:
+          - torch
+          - torchvision
+
+  - package-ecosystem: github-actions
+    directory: /
+    schedule:
+      interval: weekly
+    open-pull-requests-limit: 3
+    groups:
+      actions:
+        patterns:
+          - "*"
diff --git a/.github/workflows/agent-issue.yml b/.github/workflows/agent-issue.yml
new file mode 100644
index 0000000..40fd74a
--- /dev/null
+++ b/.github/workflows/agent-issue.yml
@@ -0,0 +1,134 @@
+name: Work an issue
+
+# Turns a labelled issue into a pull request without a person opening a
+# session. Label an issue `agent`, or comment `/agent`, and the work starts.
+#
+# The model is reached through aider, which speaks to any OpenAI-compatible
+# endpoint, so the provider is configuration rather than code:
+#
+#   secret   LLM_API_KEY      the provider key
+#   secret   AUTOMATION_TOKEN a PAT or App token
+#   variable LLM_BASE_URL     e.g. https://openrouter.ai/api/v1
+#   variable LLM_MODEL        the provider's exact model slug
+#
+# `AUTOMATION_TOKEN` is not optional. GitHub does not start workflow runs for
+# events raised by the built-in `GITHUB_TOKEN`, so a pull request opened with
+# it sits with no checks and can never satisfy auto-merge.
+#
+# The agent cannot merge, cannot approve, and cannot push to main: `main`
+# requires a pull request and its checks. That is the whole reason it is
+# reasonable to let a model write here unattended.
+
+on:
+  issues:
+    types: [labeled]
+  issue_comment:
+    types: [created]
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  work:
+    # `issue_comment` fires for pull requests too; `.issue.pull_request` is
+    # absent on a real issue, which is how the two are told apart.
+    if: >
+      (github.event_name == 'issues' &&
+       github.event.label.name == 'agent') ||
+      (github.event_name == 'issue_comment' &&
+       !github.event.issue.pull_request &&
+       startsWith(github.event.comment.body, '/agent'))
+    runs-on: [self-hosted, linux, x64, spikeforge-ci]
+    timeout-minutes: 60
+    steps:
+      - name: Check the configuration exists
+        shell: bash
+        run: |
+          missing=""
+          [ -z "${{ secrets.LLM_API_KEY }}" ] && missing="$missing LLM_API_KEY"
+          [ -z "${{ secrets.AUTOMATION_TOKEN }}" ] && missing="$missing AUTOMATION_TOKEN"
+          [ -z "${{ vars.LLM_BASE_URL }}" ] && missing="$missing LLM_BASE_URL"
+          [ -z "${{ vars.LLM_MODEL }}" ] && missing="$missing LLM_MODEL"
+          if [ -n "$missing" ]; then
+            echo "::error::missing configuration:$missing"
+            exit 1
+          fi
+
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          token: ${{ secrets.AUTOMATION_TOKEN }}
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install the agent
+        run: python -m pip install --upgrade aider-chat
+
+      - name: Write the instructions
+        shell: bash
+        env:
+          ISSUE_TITLE: ${{ github.event.issue.title }}
+          ISSUE_BODY: ${{ github.event.issue.body }}
+        run: |
+          # Through a file, not an argument: an issue body is arbitrary text
+          # and interpolating it into a shell command would let its contents
+          # decide what runs.
+          {
+            echo "Work this issue."
+            echo
+            echo "Title: $ISSUE_TITLE"
+            echo
+            echo "$ISSUE_BODY"
+            echo
+            echo "Read this repository's own guidance first -- rules.md, and"
+            echo "AGENTS.md or CLAUDE.md if present. It is binding."
+            echo
+            echo "Five published distributions and two applications build"
+            echo "against this repository. If you change a version, update"
+            echo "compatibility.json to match: the dashboard pins both its"
+            echo "end-to-end suite and the engine inside the desktop"
+            echo "application to that file, so leaving it behind ships a"
+            echo "mismatched build."
+          } > /tmp/agent-instructions.txt
+          cat /tmp/agent-instructions.txt
+
+      - name: Make the change
+        env:
+          OPENAI_API_KEY: ${{ secrets.LLM_API_KEY }}
+          OPENAI_API_BASE: ${{ vars.LLM_BASE_URL }}
+        run: |
+          set -euo pipefail
+          branch="agent/issue-${{ github.event.issue.number }}"
+          git config user.name "spikeforge-agent"
+          git config user.email "contact@capsizegames.com"
+          git checkout -b "$branch"
+          aider --yes --no-analytics --no-gitignore \
+            --model "openai/${{ vars.LLM_MODEL }}" \
+            --message-file /tmp/agent-instructions.txt
+          if git diff --quiet HEAD; then
+            echo "::error::the agent produced no change"
+            exit 1
+          fi
+
+      - name: Check its work
+        run: |
+          python -m pip install -e ".[dev]" || python -m pip install -e .
+          python -m pytest -x -q
+
+      - name: Open the pull request
+        env:
+          GH_TOKEN: ${{ secrets.AUTOMATION_TOKEN }}
+        run: |
+          set -euo pipefail
+          branch="agent/issue-${{ github.event.issue.number }}"
+          git push --set-upstream origin "$branch"
+          gh pr create --base main --head "$branch" \
+            --title "${{ github.event.issue.title }}" \
+            --body "Closes #${{ github.event.issue.number }}.
+
+          Written by \`${{ vars.LLM_MODEL }}\` from the issue text. Nobody has
+          read this. The test suite passed on the runner before it was opened;
+          the required checks decide whether it merges."
diff --git a/.github/workflows/agent-review.yml b/.github/workflows/agent-review.yml
new file mode 100644
index 0000000..ae7a3f2
--- /dev/null
+++ b/.github/workflows/agent-review.yml
@@ -0,0 +1,59 @@
+name: Review a pull request
+
+# Reviews every pull request. With nobody reading the diff, this is the only
+# thing that looks at a change as a change rather than as a test result.
+#
+# It talks to any OpenAI-compatible chat-completions endpoint -- OpenRouter,
+# DeepInfra, or anything else speaking that protocol -- so the provider and
+# model are configuration, not code:
+#
+#   secret   LLM_API_KEY    the provider key
+#   variable LLM_BASE_URL   e.g. https://openrouter.ai/api/v1
+#   variable LLM_MODEL      the provider's exact model slug
+#
+# It posts a comment. It cannot approve, block, or merge: the required checks
+# decide that.
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+permissions:
+  contents: read
+  pull-requests: write
+
+jobs:
+  review:
+    if: ${{ !github.event.pull_request.draft }}
+    runs-on: [self-hosted, linux, x64, spikeforge-ci]
+    timeout-minutes: 20
+    steps:
+      - name: Skip without a provider
+        id: guard
+        shell: bash
+        run: |
+          if [ -z "${{ secrets.LLM_API_KEY }}" ] \
+            || [ -z "${{ vars.LLM_BASE_URL }}" ] \
+            || [ -z "${{ vars.LLM_MODEL }}" ]; then
+            echo "LLM_API_KEY, LLM_BASE_URL or LLM_MODEL is unset;"
+            echo "no review will be posted."
+            echo "skip=true" >> "$GITHUB_OUTPUT"
+          fi
+
+      - uses: actions/checkout@v4
+        if: steps.guard.outputs.skip != 'true'
+
+      - uses: actions/setup-node@v4
+        if: steps.guard.outputs.skip != 'true'
+        with:
+          node-version: "22"
+
+      - name: Post the review
+        if: steps.guard.outputs.skip != 'true'
+        env:
+          LLM_API_KEY: ${{ secrets.LLM_API_KEY }}
+          LLM_BASE_URL: ${{ vars.LLM_BASE_URL }}
+          LLM_MODEL: ${{ vars.LLM_MODEL }}
+          GH_TOKEN: ${{ github.token }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+        run: node scripts/llm_review.mjs
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index e5c8e22..d85b9cd 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -21,10 +21,16 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
+      # No `cache: pip` anywhere in this file. This fleet's runners keep
+      # `_work` on another drive behind a symlink, and setup-python's
+      # dependency-file glob does not see through it: it reports that no
+      # pyproject.toml exists and fails the job before anything is built,
+      # while all seven of them sit in `packages/`. A self-hosted runner also
+      # keeps `~/.cache/pip` between jobs on its own, so the action's cache
+      # was buying nothing here even when it worked.
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install the core dev extra
         run: |
@@ -56,7 +62,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
-          cache: pip
 
       - name: Install torch CPU wheels
         run: |
@@ -89,7 +94,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install torch CPU wheels
         run: |
@@ -141,7 +145,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install torch CPU wheels
         run: |
@@ -182,7 +185,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install torch CPU wheels
         run: |
@@ -243,7 +245,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install torch CPU wheels
         run: |
@@ -310,7 +311,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install docs dependencies
         run: |
@@ -337,7 +337,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Install the renderer PyPI itself uses
         run: |
@@ -405,7 +404,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-          cache: pip
 
       - name: Build the core wheel with no extras
         run: |
diff --git a/scripts/llm_review.mjs b/scripts/llm_review.mjs
new file mode 100644
index 0000000..3256fdc
--- /dev/null
+++ b/scripts/llm_review.mjs
@@ -0,0 +1,186 @@
+#!/usr/bin/env node
+/**
+ * Review a pull request with an OpenAI-compatible model and post the result.
+ *
+ * Written against the chat-completions shape rather than any one vendor's SDK,
+ * so OpenRouter, DeepInfra, or anything else speaking that protocol works by
+ * changing two variables. Nothing here is specific to a model.
+ *
+ * Configuration, all from the environment:
+ *
+ *   LLM_API_KEY    the provider key (secret)
+ *   LLM_BASE_URL   e.g. https://openrouter.ai/api/v1
+ *   LLM_MODEL      the provider's exact model slug
+ *   GH_TOKEN       token used to read the diff and post the comment
+ *   GITHUB_REPOSITORY, PR_NUMBER
+ *
+ * The model slug is deliberately not defaulted. A wrong default fails at the
+ * provider with an opaque error on somebody else's pull request, and guessing
+ * one is worse than refusing to start.
+ */
+
+import { readFileSync } from "node:fs";
+
+/** Diffs beyond this are truncated; large ones exhaust the context window. */
+const MAX_DIFF_BYTES = 120_000;
+
+const REQUIRED = ["LLM_API_KEY", "LLM_BASE_URL", "LLM_MODEL", "GH_TOKEN"];
+
+export function missingConfig(env) {
+  return REQUIRED.filter((name) => !env[name]);
+}
+
+/**
+ * Cut a diff down to a size the model can read, at a file boundary.
+ *
+ * Truncating mid-hunk hands the model a fragment it may read as the whole
+ * change, so the cut lands on the last `diff --git` before the limit and the
+ * omission is stated in the text.
+ */
+export function truncateDiff(diff, limit = MAX_DIFF_BYTES) {
+  if (diff.length <= limit) return { diff, truncated: false };
+  const head = diff.slice(0, limit);
+  const lastFile = head.lastIndexOf("\ndiff --git ");
+  const cut = lastFile > 0 ? head.slice(0, lastFile) : head;
+  return { diff: cut, truncated: true };
+}
+
+/** The instructions the model is judged against. */
+export function buildPrompt({ number, title, body, diff, truncated, guide }) {
+  const omitted = truncated
+    ? "\n\nThis diff was truncated. Review only what is shown, and say so."
+    : "";
+  return [
+    `Review pull request #${number}: ${title}`,
+    body ? `\nIts description:\n${body}` : "",
+    guide ? `\nThe repository's own guidance is binding:\n${guide}` : "",
+    "\nReport only defects you can point at in the diff. Say plainly when",
+    "you find none; a review that invents work to look useful is worse than",
+    "a short one. Rank what you do find, most serious first.",
+    "\nWeigh hardest the things this repository has shipped broken:",
+    "- a new file under desktop/ missing from the files: list in",
+    "  electron-builder.yml, which silently will not be packaged;",
+    "- a check that passes while the product is broken, such as asserting a",
+    "  server answered rather than that the application works;",
+    "- a startup path that can leave the application running with no window;",
+    "- public-facing copy stating something the code does not do;",
+    "- the hard limits in the guidance: file and function length, `any`,",
+    "  non-null assertions, suppression comments, dead code.",
+    "\nWrite GitHub-flavoured markdown. Do not approve or request changes;",
+    "the required checks decide whether this merges.",
+    `\nThe diff:\n\n${diff}${omitted}`,
+  ].join("\n");
+}
+
+/** Pull the assistant's text out of a chat-completions response. */
+export function extractMessage(payload) {
+  const text = payload?.choices?.[0]?.message?.content;
+  if (typeof text !== "string" || text.trim() === "") {
+    throw new Error(
+      `the model returned no message: ${JSON.stringify(payload).slice(0, 400)}`,
+    );
+  }
+  return text.trim();
+}
+
+async function github(path, { token, accept, method, body }) {
+  const response = await fetch(`https://api.github.com${path}`, {
+    method: method ?? "GET",
+    headers: {
+      Accept: accept ?? "application/vnd.github+json",
+      Authorization: `Bearer ${token}`,
+      "Content-Type": "application/json",
+    },
+    body: body ? JSON.stringify(body) : undefined,
+  });
+  if (!response.ok) {
+    throw new Error(
+      `GitHub ${method ?? "GET"} ${path}: ${response.status} ` +
+        `${await response.text()}`,
+    );
+  }
+  return accept?.includes("diff") ? response.text() : response.json();
+}
+
+async function complete({ baseUrl, key, model, prompt }) {
+  const response = await fetch(`${baseUrl.replace(/\/$/, "")}/chat/completions`, {
+    method: "POST",
+    headers: {
+      Authorization: `Bearer ${key}`,
+      "Content-Type": "application/json",
+    },
+    body: JSON.stringify({
+      model,
+      messages: [{ role: "user", content: prompt }],
+      temperature: 0.2,
+    }),
+  });
+  if (!response.ok) {
+    throw new Error(
+      `${baseUrl}: ${response.status} ${await response.text()}`,
+    );
+  }
+  return extractMessage(await response.json());
+}
+
+/** The repository's guidance, when it has some, as prompt context. */
+function readGuide() {
+  for (const name of ["AGENTS.md", "CLAUDE.md", "rules.md"]) {
+    try {
+      return readFileSync(name, "utf8");
+    } catch {
+      // Try the next one; a repository without any is fine.
+    }
+  }
+  return "";
+}
+
+async function main() {
+  const env = process.env;
+  const missing = missingConfig(env);
+  if (missing.length > 0) {
+    console.error(`missing configuration: ${missing.join(", ")}`);
+    process.exit(78);
+  }
+  const repo = env.GITHUB_REPOSITORY;
+  const number = env.PR_NUMBER;
+  const token = env.GH_TOKEN;
+
+  const pull = await github(`/repos/${repo}/pulls/${number}`, { token });
+  const raw = await github(`/repos/${repo}/pulls/${number}`, {
+    token,
+    accept: "application/vnd.github.v3.diff",
+  });
+  const { diff, truncated } = truncateDiff(raw);
+
+  const review = await complete({
+    baseUrl: env.LLM_BASE_URL,
+    key: env.LLM_API_KEY,
+    model: env.LLM_MODEL,
+    prompt: buildPrompt({
+      number,
+      title: pull.title,
+      body: pull.body,
+      diff,
+      truncated,
+      guide: readGuide(),
+    }),
+  });
+
+  await github(`/repos/${repo}/issues/${number}/comments`, {
+    token,
+    method: "POST",
+    body: {
+      body: `${review}\n\n<sub>Automated review — ${env.LLM_MODEL}. Not an ` +
+        `approval; the required checks decide whether this merges.</sub>`,
+    },
+  });
+  console.log(`posted a review on #${number}`);
+}
+
+if (process.argv[1] && process.argv[1].endsWith("llm_review.mjs")) {
+  main().catch((error) => {
+    console.error(error.message);
+    process.exit(1);
+  });
+}
diff --git a/scripts/llm_review.test.mjs b/scripts/llm_review.test.mjs
new file mode 100644
index 0000000..404a355
--- /dev/null
+++ b/scripts/llm_review.test.mjs
@@ -0,0 +1,106 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import {
+  buildPrompt,
+  extractMessage,
+  missingConfig,
+  truncateDiff,
+} from "./llm_review.mjs";
+
+test("names every missing setting at once", () => {
+  assert.deepEqual(missingConfig({}), [
+    "LLM_API_KEY",
+    "LLM_BASE_URL",
+    "LLM_MODEL",
+    "GH_TOKEN",
+  ]);
+  assert.deepEqual(
+    missingConfig({
+      LLM_API_KEY: "k",
+      LLM_BASE_URL: "u",
+      LLM_MODEL: "m",
+      GH_TOKEN: "t",
+    }),
+    [],
+  );
+});
+
+/**
+ * An empty model slug must count as missing. Passing "" to the provider is a
+ * 400 on somebody else's pull request, reported as a review failure rather
+ * than as the configuration mistake it is.
+ */
+test("treats an empty setting as missing", () => {
+  const missing = missingConfig({
+    LLM_API_KEY: "k",
+    LLM_BASE_URL: "u",
+    LLM_MODEL: "",
+    GH_TOKEN: "t",
+  });
+  assert.deepEqual(missing, ["LLM_MODEL"]);
+});
+
+test("leaves a diff that fits alone", () => {
+  const diff = "diff --git a/a b/a\n+one\n";
+  assert.deepEqual(truncateDiff(diff, 1000), { diff, truncated: false });
+});
+
+/**
+ * The cut has to land on a file boundary. Handing the model half a hunk
+ * invites it to review a fragment as though it were the whole change.
+ */
+test("truncates a long diff at a file boundary", () => {
+  const first = `diff --git a/a b/a\n${"+x\n".repeat(50)}`;
+  const second = `diff --git a/b b/b\n${"+y\n".repeat(50)}`;
+  const { diff, truncated } = truncateDiff(first + second, first.length + 20);
+  assert.equal(truncated, true);
+  assert.equal(diff, first.replace(/\n$/, ""));
+  assert.ok(!diff.includes("b/b"));
+});
+
+test("still truncates when no boundary is reachable", () => {
+  const huge = `diff --git a/a b/a\n${"+x\n".repeat(200)}`;
+  const { diff, truncated } = truncateDiff(huge, 50);
+  assert.equal(truncated, true);
+  assert.equal(diff.length, 50);
+});
+
+test("tells the model when it is seeing a partial diff", () => {
+  const whole = buildPrompt({ number: 1, title: "t", diff: "d" });
+  assert.doesNotMatch(whole, /truncated/);
+  const part = buildPrompt({ number: 1, title: "t", diff: "d", truncated: true });
+  assert.match(part, /truncated/);
+});
+
+test("carries the repository's guidance into the prompt", () => {
+  const prompt = buildPrompt({
+    number: 7,
+    title: "Add a thing",
+    diff: "d",
+    guide: "NEVER use any",
+  });
+  assert.match(prompt, /#7: Add a thing/);
+  assert.match(prompt, /NEVER use any/);
+});
+
+test("reads the assistant message", () => {
+  const payload = { choices: [{ message: { content: "  looks fine  " } }] };
+  assert.equal(extractMessage(payload), "looks fine");
+});
+
+/**
+ * A provider that answers 200 with an error body, or with an empty message,
+ * must not post an empty review that reads like approval.
+ */
+test("refuses an empty or malformed completion", () => {
+  assert.throws(() => extractMessage({}), /no message/);
+  assert.throws(
+    () => extractMessage({ choices: [{ message: { content: "   " } }] }),
+    /no message/,
+  );
+  assert.throws(
+    () => extractMessage({ error: { message: "bad model" } }),
+    /no message/,
+  );
+});
diff --git a/tests/test_serving_bundle.py b/tests/test_serving_bundle.py
index c50e0c5..3643411 100644
--- a/tests/test_serving_bundle.py
+++ b/tests/test_serving_bundle.py
@@ -148,7 +148,12 @@ def test_rebuild_is_bit_exact_in_a_subprocess(tmp_path: Any) -> None:
         env=dict(os.environ, PYTHONPATH=root),
         capture_output=True,
         text=True,
-        timeout=180,
+        # Generous because this is a cold Python process importing Torch, and
+        # the runners share one workstation with every other project's CI. At
+        # 180s it timed out under load and reported a bit-exactness failure
+        # that had not happened, which is the worst kind of red: it looks like
+        # the thing the test is named after.
+        timeout=600,
     )
     assert result.returncode == 0, result.stderr
     assert result.stdout.strip() == "ok"