Load worker skills on demand instead of baking bodies

Baked skill bodies cost roughly 38KB per worker prompt; workers now search and load only the skills the brief names.

Author: TheGreatAxios

src/agent/directors/bake-skills.test.ts

@@ -64,13 +64,14 @@ describe("builderPackage", () => {
     expect(p).toContain("`/implement` does not steal planning from `/plan`");
   });
 
-  test("systemPrompt requires baked core constraints and Ponytail prerequisites", () => {
+  test("systemPrompt requires core constraints and Ponytail prerequisites", () => {
     const p = builderPackage.systemPrompt;
     expect(p).toContain("Prerequisites");
     expect(p).toMatch(
       /style, philosophy, native-runtime, idiot-proof, and Ponytail/i,
     );
-    expect(p).toMatch(/use_skill is not mounted/i);
+    expect(p).toMatch(/load each with skill_search \+ use_skill/i);
+    expect(p).not.toMatch(/use_skill is not mounted/i);
     expect(p).toMatch(
       /including their TypeScript conventions when TypeScript is the task surface/i,
     );

src/agent/directors/bake-skills.ts

@@ -37,7 +37,7 @@ You are a disciplined implementer worker (maySpawn:false) — not Critic, not Ex
 
 ## Prerequisites
 
-Before substantial repo work: follow style, philosophy, native-runtime, idiot-proof, and Ponytail (baked; use_skill is not mounted). Follow AGENTS.md and /docs, including their TypeScript conventions when TypeScript is the task surface.
+Before substantial repo work: follow style, philosophy, native-runtime, idiot-proof, and Ponytail — load each with skill_search + use_skill only when the brief needs it. Follow AGENTS.md and /docs, including their TypeScript conventions when TypeScript is the task surface.
 
 ## Plan
 

src/agent/directors/builder/package.test.ts

@@ -110,7 +110,8 @@ describe("criticPackage", () => {
   test("tools.allow is review surface with product writes", () => {
     const allow = criticPackage.tools?.allow ?? [];
     expect(allow).toContain("read_file");
-    expect(allow).not.toContain("use_skill");
+    expect(allow).toContain("skill_search");
+    expect(allow).toContain("use_skill");
     expect(allow).toContain("write_file");
     expect(allow).toContain("edit_file");
     expect(allow).toContain("delete_file");

src/agent/directors/builder/package.ts

@@ -61,7 +61,7 @@ API contract check (blocking when brief specifies signatures):
 - Prefer reading tests/callers; a tiny sync call that would hang on a Promise is evidence.
 - Rank these as blocking, not style nits.
 
-Before substantial review work: follow style, philosophy, native-integration, and idiot-proof (baked; use_skill is not mounted). Read the code under review.
+Before substantial review work: follow style, philosophy, native-integration, and idiot-proof — load each with skill_search + use_skill only when the brief needs it. Read the code under review.
 
 OUT OF LANE → refuse or reclassify under Blockers:
 - implementing fixes (route to builder)

src/agent/directors/critic/package.test.ts

@@ -133,7 +133,8 @@ describe("draperPackage", () => {
   test("tools.allow is review surface with file writes for evidence tests", () => {
     const allow = draperPackage.tools?.allow ?? [];
     expect(allow).toContain("read_file");
-    expect(allow).not.toContain("use_skill");
+    expect(allow).toContain("skill_search");
+    expect(allow).toContain("use_skill");
     expect(allow).toContain("write_file");
     expect(allow).toContain("edit_file");
     expect(allow).toContain("delete_file");

src/agent/directors/critic/package.ts

@@ -78,7 +78,8 @@ describe("emilPackage", () => {
   test("tools.allow is review surface with product writes", () => {
     const allow = emilPackage.tools?.allow ?? [];
     expect(allow).toContain("read_file");
-    expect(allow).not.toContain("use_skill");
+    expect(allow).toContain("skill_search");
+    expect(allow).toContain("use_skill");
     expect(allow).toContain("write_file");
     expect(allow).toContain("edit_file");
     expect(allow).toContain("delete_file");

src/agent/directors/draper/package.test.ts

@@ -25,7 +25,7 @@ PRIMARY INTENT: architecture judgment. Judge approach soundness, constraint owne
 
 You are Greybeard — not a second Skywalker, not Critic (code defects with evidence), not Builder. Your value is architectural judgment, not legwork or implementation.
 
-Follow style and philosophy conventions (baked into this prompt) when reviewing plans or approaches — skills are active constraints, not background docs.
+Follow style and philosophy conventions (load on demand with skill_search + use_skill) when reviewing plans or approaches — skills are active constraints, not background docs.
 
 Your value is analysis, not delegation: reach the judgment yourself with
 targeted reads (read_file, grep) and pointed questions (ask_director)
@@ -44,7 +44,7 @@ Blinders: do not call search_agents to discover the fleet. Do not spawn builder,
 
 Guide quality — advise what good architecture looks like for this change. Do not assert enforcement theater (fake caps, pretend runtime gates, or "must spawn N" rules the harness does not enforce).
 
-Before substantial review work: follow style and philosophy conventions (baked; use_skill is not mounted on workers).
+Before substantial review work: follow style and philosophy conventions — load each with skill_search + use_skill only when the brief needs it.
 
 OUT OF LANE: shipping product code, pedantic style-only nitpicking, being a second primary orchestrator, discovering or dispatching the full fleet.`,
 };