humanish 0.70.0 → 0.72.0

package/README.md

@@ -132,7 +132,10 @@ Three things to know before you rely on it:
   app-server thread, Claude Code through one `claude -p` stream-json session.
   A participant that starts every turn cold cannot remember trying the menu
   and tries it again; measured on one lab, that was 188 actions over 90 turns
-  and no finish against 21 actions over 8 turns.
+  and no finish against 21 actions over 8 turns. `HUMANISH_LOCAL_AGENT_ONE_SHOT=1`
+  keeps the cold-start path for the Claude agent as a measurement switch, so
+  "remembers" can be compared against "does not" on your own lab; the trace's
+  `ids.model` says which ran.
 
 ## Telemetry
 

package/dist/cua-actor-lab.js

@@ -31,7 +31,7 @@ import { adapterScoreFailureMessage, applyBrowserAdapterHooks } from "./adapter-
 import { actorRegistry, isCuaActorDescriptor } from "./actor-registry.js";
 import { CHROMIUM_EVIDENCE_HYGIENE_FLAGS, chromiumEvidenceProfilePreferencesJson } from "./browser-evidence-hygiene.js";
 import { DEFAULT_OPENAI_CU_MODEL } from "./openai-responses-cu.js";
-import { detectLocalAgents } from "./local-agent-cli.js";
+import { createLocalAgentProvider, detectLocalAgents } from "./local-agent-cli.js";
 import { startAppServerSession } from "./local-agent-appserver.js";
 import { startClaudeSession } from "./local-agent-claude-session.js";
 import { createDesktopSandbox, loadE2BDesktopModule } from "./e2b-desktop-launch.js";
@@ -1641,12 +1641,27 @@ export async function runCuaLane(spec, deps) {
             else if (deps.localAgent === "claude") {
                 // One session for the whole run, like the codex thread above (#520). The one-shot
                 // provider (createLocalAgentProvider) spawned `claude -p` per turn, and every turn
-                // started with no memory of the last.
-                claudeSession = await startClaudeSession({
-                    ...(spec.reasoningEffort === undefined ? {} : { reasoningEffort: spec.reasoningEffort }),
-                    ...(config.actors[0]?.model === undefined ? {} : { model: config.actors[0].model })
-                });
-                localAgentProvider = claudeSession.provider;
+                // started with no memory of the last. HUMANISH_LOCAL_AGENT_ONE_SHOT=1 keeps that path
+                // reachable as a MEASUREMENT switch: MemTrapBench (2026-08) reports memory frameworks
+                // degrading agent performance by 10-40% on some tasks, so "remembers" has to be measured
+                // against "does not" on the same lab, not assumed. The trace records which one ran.
+                const oneShot = env.HUMANISH_LOCAL_AGENT_ONE_SHOT !== undefined
+                    && env.HUMANISH_LOCAL_AGENT_ONE_SHOT !== ""
+                    && env.HUMANISH_LOCAL_AGENT_ONE_SHOT !== "0";
+                if (oneShot) {
+                    localAgentProvider = createLocalAgentProvider({
+                        agent: "claude",
+                        ...(spec.reasoningEffort === undefined ? {} : { reasoningEffort: spec.reasoningEffort }),
+                        ...(config.actors[0]?.model === undefined ? {} : { model: config.actors[0].model })
+                    });
+                }
+                else {
+                    claudeSession = await startClaudeSession({
+                        ...(spec.reasoningEffort === undefined ? {} : { reasoningEffort: spec.reasoningEffort }),
+                        ...(config.actors[0]?.model === undefined ? {} : { model: config.actors[0].model })
+                    });
+                    localAgentProvider = claudeSession.provider;
+                }
             }
             // World is ready: release the pipeline gate so the remaining lanes may start.
             provisioned = true;