archilyzer

commit cd0ffcddd55c037370572899081ef6f407a99249
parent d7a2445848574eb70106bbbe3465bfa94b035a71
Author: I Mean I'm Just Saying <imeanimjustsaying@kiwifarms.st>
Date:   Sat, 26 Sep 2026 14:27:09 -0400

mcp: the ask and sweep plans send clip media through fetch_clip, never a hand-run yt-dlp

One step, shared by both builders: after the citation step in ask, before
Finish in the sweep. It passes the plan's corpus as source (the Rumble
mapping reads it), offers full: true for a whole recording, and says the
README's yt-dlp command is the operator's fallback, not the agent's.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Diffstat:
M
mcp/src/instructions.test.ts
 | 
26
++++++++++++++++++++++++++
M
mcp/src/instructions.ts
 | 
23
+++++++++++++++++++++++

2 files changed, 49 insertions(+), 0 deletions(-)
diff --git a/mcp/src/instructions.test.ts b/mcp/src/instructions.test.ts
@@ -188,3 +188,29 @@ test("the sweep reads operator corrections from a sibling manifest before the fi
   assert.ok(at > 0 && batch > at, `corrections step at ${at}, batch step at ${batch}`);
   assert.match(text, /never re-asserted/);
 });
+
+// ─── Media for a citation ───
+
+test("both builders send clip media through fetch_clip, never a hand-run yt-dlp", () => {
+  for (const text of [sweep("coffee channels=c"), ask("coffee")]) {
+    assert.match(text, /`fetch_clip`/);
+    assert.match(text, /NEVER run yt-dlp yourself/);
+    // The corpus rides along, so a Rumble embed id maps to the editor's dir.
+    assert.match(text, /pass source: "remote:https:\/\/site\.example" so a Rumble id/);
+    // The whole recording is on offer, but the window is the default.
+    assert.match(text, /full: true when the ask genuinely needs the whole recording/);
+    // Only the operator falls back to yt-dlp.
+    assert.match(text, /the operator's fallback, not yours/);
+  }
+  // In the sweep, before Finish: a report that needs a clip asks while the
+  // citation is in hand, not after the summary is written.
+  const s = sweep("coffee channels=c");
+  const clip = s.indexOf("Media for a cited moment");
+  const finish = s.indexOf("**Finish.**");
+  assert.ok(clip > 0 && finish > clip, `clip step at ${clip}, Finish at ${finish}`);
+  // In the ask, after the answer's citation rule.
+  const a = ask("coffee");
+  const cite = a.indexOf("Answer with citations");
+  const askClip = a.indexOf("Media for a cited moment");
+  assert.ok(cite > 0 && askClip > cite, `citations at ${cite}, clip step at ${askClip}`);
+});
diff --git a/mcp/src/instructions.ts b/mcp/src/instructions.ts
@@ -140,6 +140,25 @@ function batchStep(req: PromptRequest, ctx: PlanContext, reportPath: string): st
   );
 }
 
+// Media for a citation goes through the editor's fetch job, never a yt-dlp the
+// agent runs itself (the operator's rule: every fetch through Archilyzer). One
+// step, shared by both builders so the two cannot drift. `wait_seconds` and
+// `full` stay un-backticked: they are arguments, not tools.
+function clipStep(ctx: PlanContext): string {
+  return (
+    `**Media for a cited moment — through the editor only.** When I ask for ` +
+    `the clip behind a citation (or a report needs one), call \`fetch_clip\` ` +
+    `with that citation's channel slug, video id, start/end in seconds (or ` +
+    `mm:ss) and a one-line reason (or full: true when the ask genuinely needs ` +
+    `the whole recording — a window is the default), and pass ` +
+    `source: "${ctx.corpus}" so a Rumble id resolves to the right directory. ` +
+    `NEVER run yt-dlp yourself, in any form. If the tool reports no editor is ` +
+    `configured, say so and stop — the README's yt-dlp command is the ` +
+    `operator's fallback, not yours. If it returns queued, call it again with ` +
+    `the job it names.`
+  );
+}
+
 // The full sweep instructions.
 export function buildSweepInstructions(
   req: PromptRequest,
@@ -236,6 +255,8 @@ export function buildSweepInstructions(
 
   steps.push(batchStep(req, ctx, reportPath));
 
+  steps.push(clipStep(ctx));
+
   steps.push(
     `**Finish.** Work to the end of the worklist, then write a summary ` +
       `section: the scope swept, **how many of the N you actually read**, ` +
@@ -329,6 +350,8 @@ export function buildAskInstructions(
       `corpus does NOT show — an absence of matches is a finding, not a gap ` +
       `to paper over.`,
   );
+
+  steps.push(clipStep(ctx));
   }
 
   const numbered = steps.map((s, i) => `${i + 1}. ${s}`).join("\n\n");