dsh-llm-verifier 0.3.0 → 0.3.1

package/lib/index.js

@@ -1313,6 +1313,17 @@ function canonicalTodoSnapshots(index) {
 * A checkpoint already renders the todo/team snapshot these tools wrote, so their
 * own result repeats it while displacing the real work output that came just before
 * them.
+*
+* `present` belongs here for the same reason: declaring deliverables produces no
+* independent output, and it is always the LAST call of a turn. A live session ran
+* typecheck + the full suite, committed, then presented — and every stop boundary
+* showed the judge "Latest observed tool output at routing time (run_code):
+* presented: 8" instead of the verification run one call earlier. Following its own
+* rule ("a state without real verification should not exceed K"), the judge capped
+* the newest checkpoint at exactly K = 52.6% against a 0.8 threshold on four
+* consecutive routes, so each one steered "continue the unfinished work" for work
+* that was finished and verified. The `deliverables/presented` event still records
+* what was presented; the tool result adds nothing the judge can grade.
 */
 const BOOKKEEPING_TOOLS = /* @__PURE__ */ new Set([
 	"todo_write",
@@ -1322,7 +1333,8 @@ const BOOKKEEPING_TOOLS = /* @__PURE__ */ new Set([
 	"interrupt_agent",
 	"list_agents",
 	"exit_plan_mode",
-	"skill"
+	"skill",
+	"present"
 ]);
 /** Whether a wrapper's dispatches were all bookkeeping; undefined/empty means it dispatched nothing. */
 function onlyBookkeepingDispatches(names) {