fix(skill): stop hardcoding invoke_agent in load-skill override-forwarding prose

docs/api-reference/veryfront/agent.md

@@ -498,7 +498,7 @@ Input delivered to a hosted agent-service detached execution callback.
 | `LOAD_SKILL_CONTINUE_SAME_TURN`                                 | Loading a skill is not doing the work; the turn continues.                      | [source](https://github.com/veryfront/veryfront-code/blob/main/src/skill/load-skill-policy.ts#L30)                  |
 | `LOAD_SKILL_CONTINUE_SAME_TURN_NOW`                             | Shared load skill continue same turn now value.                                 | [source](https://github.com/veryfront/veryfront-code/blob/main/src/agent/conversation/delegation-policy.ts#L27)     |
 | `LOAD_SKILL_DELEGATION_THRESHOLD`                               | Shared load skill delegation threshold value.                                   | [source](https://github.com/veryfront/veryfront-code/blob/main/src/agent/conversation/delegation-policy.ts#L31)     |
-| `LOAD_SKILL_OVERRIDE_FORWARDING`                                | A skill's model/thinking/maxSteps only take effect if the caller forwards them. | [source](https://github.com/veryfront/veryfront-code/blob/main/src/skill/load-skill-policy.ts#L41)                  |
+| `LOAD_SKILL_OVERRIDE_FORWARDING`                                | A skill's model/thinking/maxSteps only take effect if the caller forwards them. | [source](https://github.com/veryfront/veryfront-code/blob/main/src/skill/load-skill-policy.ts#L49)                  |
 | `LOAD_SKILL_ROOT_OWNERSHIP`                                     | Shared load skill root ownership value.                                         | [source](https://github.com/veryfront/veryfront-code/blob/main/src/agent/conversation/delegation-policy.ts#L29)     |
 | `MAX_RUNTIME_SKILL_PROMPT_ENTRIES`                              | Maximum value for runtime skill prompt entries.                                 | [source](https://github.com/veryfront/veryfront-code/blob/main/src/agent/runtime/skill-prompt.ts#L12)               |
 | `NO_DELEGATION_NARRATION_UNLESS_ASKED`                          | Shared no delegation narration unless asked value.                              | [source](https://github.com/veryfront/veryfront-code/blob/main/src/agent/conversation/delegation-policy.ts#L20)     |

src/skill/load-skill-policy.ts

@@ -37,9 +37,17 @@ export const LOAD_SKILL_CONTINUE_SAME_TURN = "Continue the same turn after calli
  * those fields. `createLoadSkillTool` returns `{ skillId, instructions,
  * references, scripts }`, so telling a factory-built agent to forward returned
  * overrides would name fields its `load_skill` never produces.
+ *
+ * The `invoke_agent` condition is load-bearing, not incidental. Overrides are
+ * applied only to that tool (`applySkillDelegationOverridesToToolInput` returns
+ * its input unchanged for any other tool name), and scoped `agent_<id>`
+ * delegates accept only `{ input }` (`AgentToolInput`), so they cannot carry
+ * model/thinking/maxSteps at all. Phrasing the clause conditionally keeps it
+ * accurate on runs where `invoke_agent` is absent, without needing a dynamic
+ * description.
  */
 export const LOAD_SKILL_OVERRIDE_FORWARDING =
-  "Pass through any returned model, thinking, or maxSteps overrides to invoke_agent when delegating.";
+  "If invoke_agent is available, pass through any returned model, thinking, or maxSteps overrides when delegating to it.";
 
 /**
  * The behavioural contract every `load_skill` tool description must state,

src/skill/tools.test.ts

@@ -737,4 +737,21 @@ describe("load_skill orchestration contract", () => {
       false,
     );
   });
+
+  it("states override forwarding conditionally, because it only works for invoke_agent", () => {
+    // Some runs expose only scoped delegate tools (`agent_<id>`) and no
+    // `invoke_agent`. Naming `invoke_agent` unconditionally points at a tool
+    // that is absent; generalising to "the available delegation tool" is worse,
+    // because scoped delegates CANNOT carry overrides — `AgentToolInput` is
+    // `{ input: string }`, and applySkillDelegationOverridesToToolInput returns
+    // its input unchanged for any tool other than invoke_agent.
+    //
+    // So the clause must be conditional: a no-op when invoke_agent is absent,
+    // accurate when it is present. veryfront/veryfront-issue-inbox#411.
+    assertStringIncludes(LOAD_SKILL_OVERRIDE_FORWARDING, "If invoke_agent is available");
+    assertEquals(
+      LOAD_SKILL_OVERRIDE_FORWARDING.includes("the available delegation tool"),
+      false,
+    );
+  });
 });