Skip to content
Technical Analysis
The approval is spent. Was the effect delivered?
Research

The approval is spent. Was the effect delivered?

Isolated probes share one consumed-approval state across success, pre-effect failure and post-effect exception. Replay control and delivery evidence require separate accounts.

2026-10-29中文 →

The approval is spent. Was the effect delivered? ​

An approval has been consumed and retrying it is rejected. Yet the business effect might never have occurred—or it might have occurred without a successful outcome record. Systems that approve agents' external operations need different responses to those outcomes. A single “consumed” state cannot distinguish them.

The same-day Research Object's isolated probes locate the distinction in an execution order: using an approval once and successfully delivering a business effect are different guarantees. The architectural question is which evidence establishes approved identity, consumption, delivered outcome and authority to recover.

One consumed state, three execution outcomes ​

The pinned ASU gateway methods validate an approval, consume it, call the executor and record the effect after normal return. Our internal experiment uses the already completed Research Object: extracted methods and actual conditional update SQL, with an in-memory database, synthetic effect descriptor, fixed Boolean receipt oracle and fake executor. It performs no remote business operation. Pinned gateway

All eight controls matched expectations. Three execution paths expose the central difference. The environment was Python 3.12.9 with SQLite 3.45.3. All three paths leave the approval consumed. “Memory effects” below count mutations in the fake executor; “successful records” are experimental records written after normal return. Neither is proof of delivery by a remote service.

Synthetic pathFirst → retry responseEffects / records
Normal return200 → 4031 / 1
Exception before effect500 → 4030 / 0
Exception after effect; reply lost500 → 4031 / 0

Consumption is identical while effects and records differ. The last two paths even have identical first and retry response codes: those signals alone cannot tell the caller whether nothing happened or acknowledgement was lost after an effect. These are internally injected synthetic failures, not observed production incidents.

Five other controls cover changed tool, changed arguments, invalid receipt, pending approval and expired consumption. None invokes the executor. The rejection paths matter: this is not a demonstration of arbitrary modification or allowed replay.

The interval between consumption and recording ​

The examined implementation checks session and user, terminal approval status, stored and resubmitted effect digests, and freshly resolved raw arguments. Receipt verification precedes the conditional approved-and-unexpired database update. Execution receives the stored descriptor. Digest binding answers whether execution concerns the approved object. Gateway and consumption implementation

Our interpretation follows the observed order: consumption commits before execution, while effect recording depends on normal return. Whether an approval can be reused and whether the external effect succeeded are separate questions. Rejecting reuse addresses the former; it does not supply the latter's missing outcome evidence. An execution error does not automatically imply zero business effect.

This does not justify simply moving consumption after execution. Early consumption can prioritize replay prevention. Delayed consumption would have to address concurrency and repeated effects anew. A real executor might already provide idempotency—repeated submissions under one key produce one business effect—or result queries by identity. We did not audit full executors and make no universal claim of effect loss.

A protocol requirement is not a deployment result ​

Microsoft's Agent Governance Toolkit ADR 0030 requires consumption to share an execution transaction or idempotency boundary. It also requires execution-time revalidation of the digest, expiry, policy and approval-chain versions, and authenticated approvers. An individual allow vote cannot itself execute an action. Action-bound approval ADR

This independently authored source provides a design requirement. It is not an independent reproduction of our probe or a measured deployment advantage. The comparison is useful because it names the transaction or idempotency responsibility beyond single-use credentials. A requirement's existence and evidence that a running system satisfies it are different things.

Entry-point coverage also does not follow automatically from effect binding. OpenHands documentation explicitly assigns direct-tool-call safeguards to the host caller, while new conversation actions consult confirmation policy. The same-day Research Object is used only as a boundary comparison: an approval constraint matters on paths that traverse it. This alleges no vulnerability and does not place OpenHands in the ASU paper's audit. Official security guide

Which fact is recovery missing? ​

Governed execution should distinguish the approved effect, consumed credential, confirmed business outcome and authority to start recovery. The first three concern different evidence objects. The fourth concerns permission and cannot be inferred from an error code.

After an effect with lost acknowledgement, “try again” may duplicate the business effect, while “assume success” may hide failure. This article tests no general recovery solution. Its Research Object supports a direction for further verification: connect approval identity to executor idempotency keys and queryable results, then specify who may reconcile, compensate or request renewed approval when the response is ambiguous. These are research questions, not demonstrated capabilities.

What these probes do not establish ​

The public code is the authors' primary implementation material. Our eight controls are internal isolated experiments without independent third-party reproduction. Descriptor hashing and canonicalization use synthetic substitutes, and receipt verification uses a fixed Boolean oracle. We did not test actual signatures, informed human consent, trusted rendering or credential enrollment. In-memory exceptions are not a complete distributed-crash experiment.

The fetched source revision is 056742b. The paper's attack matrix identifies another experimental working tree with uncommitted modifications. We did not recreate that tree, replay formal proofs or reproduce the paper's large experiments. The paper and formal artifacts remain author-provided material; publication, review, citation and implementation status do not establish general validity.

The supportable conclusion ends at this evidence boundary: single-use consumption can constrain credential reuse, while successful delivery needs separate evidence. What remains to verify is which outcomes can be queried and who may act on ambiguity, rather than renaming “consumed” as “completed.”

Last updated: