feat(dispatch): slash-command skill invocation + .middle/prompt.md dispatch brief

.claude/skills/implementing-github-issues/SKILL.md

@@ -8,6 +8,10 @@ allowed-tools: Bash(gh:*), Bash(git:*), Bash(pnpm:*), Bash(mkdir:*), Read, Write
 
 End-to-end workflow for taking a GitHub issue from "assigned" to "PR open with verification evidence, marked ready for human review." All phases of one issue land on **one branch** and **one PR**; the PR is the long-lasting context for the workstream.
 
+## Dispatch brief (read first)
+
+Before anything else, check for `.middle/prompt.md` in the working directory. If it exists, it is **middle's dispatch brief** — read it and honor it. It carries the operating framing for an autonomous run (work continuously, don't stop to ask questions you can resolve yourself, terminal state is PR-ready), plus any per-dispatch operator notes (plan changes, scope adjustments, "skip X"). Operator notes in the brief override the defaults below. If the file is absent, you're being run interactively by a human — proceed normally.
+
 ## Core principles
 
 **The code shows WHAT. The PR explains WHY.** Code comments are reserved for non-obvious constraints. Reasoning, alternatives considered, and tradeoffs go in the **decisions log** (`planning/issues/<num>/decisions.md`), then get distilled into PR review comments and the PR description.

packages/adapters/claude/src/prompt.ts

@@ -1,21 +1,26 @@
 /**
  * The literal text `send-keys` carries into a tmux session to start or continue
- * the agent. `send-keys` cannot cleanly carry a multi-line prompt — embedded
- * newlines submit early — so the full prompt lives on disk and this returns a
- * one-line `@`-reference that force-includes it. A single `@` prefixes the
- * whole relative path.
+ * the agent.
+ *
+ * - `initial`: a slash command that force-invokes the implementing skill on the
+ *   Epic. The skill reads `.middle/prompt.md` (the dispatch brief) itself, so a
+ *   single one-line submission both starts the skill and delivers the brief —
+ *   which matters because the Phase 1 workflow only drives one turn.
+ * - `resume` / `answer`: an `@`-reference that force-includes the on-disk brief
+ *   (multi-line context the agent reloads). Used by the fuller multi-turn
+ *   workflow; `send-keys` can't carry multi-line text, hence the file pointer.
  */
 export function buildPromptText(opts: {
   promptFile: string;
   kind: "initial" | "resume" | "answer";
+  epicNumber: number;
 }): string {
-  const ref = `@${opts.promptFile}`;
   switch (opts.kind) {
     case "initial":
-      return ref;
+      return `/implementing-github-issues implement #${opts.epicNumber}`;
     case "resume":
-      return `Resuming this workstream — re-read the linked context and continue. ${ref}`;
+      return `Resuming this workstream — re-read the linked context and continue. @${opts.promptFile}`;
     case "answer":
-      return `A human answered your open question — read the answer and continue. ${ref}`;
+      return `A human answered your open question — read the answer and continue. @${opts.promptFile}`;
   }
 }

packages/adapters/claude/test/adapter.test.ts

@@ -56,16 +56,21 @@ describe("buildLaunchCommand", () => {
 });
 
 describe("buildPromptText", () => {
-  test("initial returns the bare @-reference one-liner", () => {
+  test("initial force-invokes the skill via slash command on the epic", () => {
     expect(
-      claudeAdapter.buildPromptText({ promptFile: ".middle/prompt.md", kind: "initial" }),
-    ).toBe("@.middle/prompt.md");
+      claudeAdapter.buildPromptText({
+        promptFile: ".middle/prompt.md",
+        kind: "initial",
+        epicNumber: 14,
+      }),
+    ).toBe("/implementing-github-issues implement #14");
   });
 
   test("resume frames the @-reference as a continuation", () => {
     const text = claudeAdapter.buildPromptText({
       promptFile: ".middle/resume.md",
       kind: "resume",
+      epicNumber: 14,
     });
     expect(text).toContain("@.middle/resume.md");
     expect(text.toLowerCase()).toContain("resum");
@@ -75,6 +80,7 @@ describe("buildPromptText", () => {
     const text = claudeAdapter.buildPromptText({
       promptFile: ".middle/answer.md",
       kind: "answer",
+      epicNumber: 14,
     });
     expect(text).toContain("@.middle/answer.md");
     expect(text.toLowerCase()).toContain("answer");

packages/core/src/adapter.ts

@@ -21,11 +21,14 @@ export interface AgentAdapter {
 
   /**
    * The literal text to send-keys into the session to start or continue the
-   * agent — includes the `@`-reference to the on-disk prompt file.
+   * agent. `initial` invokes the implementing skill directly (a slash command);
+   * the skill reads the on-disk dispatch brief. `resume`/`answer` carry the
+   * `@`-reference to the brief for the fuller multi-turn workflow.
    */
   buildPromptText(opts: {
     promptFile: string; // path, relative to the worktree
     kind: "initial" | "resume" | "answer";
+    epicNumber: number; // the dispatched Epic/issue number
   }): string;
 
   /** Put the ready session into auto mode — a launch flag or post-ready keystrokes. */

packages/dispatcher/src/workflows/implementation.ts

@@ -68,10 +68,11 @@ function sessionNameFor(input: ImplementationInput): string {
 }
 
 /**
- * Write a plan-style placeholder `.middle/prompt.md` into the worktree if one
- * is not already present. An operator (or Phase 3+ `mm init` / Phase 7's
- * recommender) can override by committing a real prompt in the source repo —
- * the worktree inherits it and this writer leaves it alone.
+ * Write the default dispatch brief to `.middle/prompt.md` if one isn't already
+ * present. The skill (invoked by the slash command) reads this file as its
+ * operating brief — framing, not "use the skill" (the slash command already
+ * did that). An operator-supplied brief (committed in the repo, or written by a
+ * future `mm dispatch --note` / the recommender) is left untouched.
  */
 function ensurePromptFile(worktreePath: string, epicNumber: number): void {
   const middleDir = join(worktreePath, ".middle");
@@ -80,18 +81,21 @@ function ensurePromptFile(worktreePath: string, epicNumber: number): void {
   mkdirSync(middleDir, { recursive: true });
   writeFileSync(
     promptPath,
-    `# middle dispatch — Epic #${epicNumber}
+    `# middle dispatch brief — Epic #${epicNumber}
 
-You are dispatched by middle (the autonomous GitHub-issue dispatcher) to work
-on Epic #${epicNumber} in this repository.
+You are running autonomously under middle. There is no human watching in real
+time. Operating rules for this dispatch:
 
-Use the \`implementing-github-issues\` skill — it is available in this worktree
-at \`.claude/skills/implementing-github-issues/SKILL.md\`. Invoke it via the
-Skill tool with name \`implementing-github-issues\` and input \`implement #${epicNumber}\`.
+- Work through every phase continuously. The mechanical verification gates are
+  the gates between phases — do not pause for confirmation between them.
+- Do not stop to ask questions you can resolve yourself. Pause only if you are
+  genuinely blocked: ambiguous acceptance criteria, or a decision needing more
+  candidate forks than the complexity ceiling.
+- The terminal state is: every phase verified, the PR marked ready for review,
+  and the reviewer's brief posted on both the Epic and the PR. Then stop.
 
-If the skill is not present in this worktree (the target repo has not been
-\`mm init\`'d and is not middle's own dogfood checkout), write a brief
-explanation to \`.middle/failed.json\` as \`{ "reason": "<message>" }\` and stop.
+## Operator notes for this dispatch
+(none)
 `,
   );
 }
@@ -230,6 +234,7 @@ export function createImplementationWorkflow(
       const promptText = adapter.buildPromptText({
         promptFile: ".middle/prompt.md",
         kind: "initial",
+        epicNumber: ctx.input.epicNumber,
       });
       console.error(`${tag} sending prompt: "${promptText}"`);
       await deps.tmux.sendText(sessionName, promptText);