Workflow: social-media; step: execute-required-saas-connect-post; item: verify-saas-discovery-and-execution.

The retained server log (agent_go/logs/server_debug.log, 2026-09-06 IST) shows two overlapping executions using the same runtime identity: msgseq-iteration-0-default-default-step-11-execute-required-saas-connect-post.

• 23:28:23: the first execution registers its working directory and guard.
• 23:30:18: its replacement registers the same runtime identity and directory.
• 23:31:38: the cancelled first execution finishes unwinding, closes its agent, and reaches message-sequence cleanup. Its execution ID ends in 1788717503412367000.
• 23:31:44: the replacement's shell requests begin failing with missing cwd. Repeated requests continue failing through 23:32:28.
• 23:32:38: the replacement execution, ending in 1788717618256685000, fails.

The cwd was not absent at creation: it was registered repeatedly before the failure. The configuration error reports the symptom; shared session ownership explains why it disappeared. The cleanup function clears shell configuration and releases CLI resources by that identity. The failure is reproduced using the production cleanup function in a local unit test.

Allocate a unique UUID suffix for each new message-sequence runtime. Keep the readable run/group/step prefix for diagnostics. Reuse the ID only from the sequence's live in-memory runtime across its turns. Do not reclaim a serialized runtime ID when reconstructing conversation history after execution/restart.

The old execution now releases only its own shell guard, cwd, environment, managed DB capability, and CLI runtime. No workspace-root fallback or additional file/DB permission was introduced. Existing conversation/history metadata and cost recording paths are unchanged; runtime_session_id records the new owner.

• Before the change, TestMessageSequenceOldRuntimeCleanupPreservesReplacement failed with the replacement shell config becoming nil.
• After the change, message-sequence, working-directory, and direct-learning tests passed.
• Regression coverage verifies late cleanup preserves all replacement shell configuration, including blocked paths and DB capability; the completed runtime is still cleaned up.
• Resume coverage verifies saved history cannot reclaim the previous runtime, while subsequent turns reuse the live replacement runtime.
• Focused race-detector tests for runtime ownership, resume, stable turn identity, and working-directory propagation passed. git diff --check passed.

No AgentWorks server was started/restarted and no live workflow was retried. The fix is local source pending deployment. This does not prove that no public X action occurred earlier in either run. That requires authoritative action receipts or external read-back. Nor does it serialize all shared workflow artifact writes: this patch isolates runtime/session ownership, not execution of the same business action in parallel. Repeated agent shell retries cannot repair missing server-side session state.