|
| 1 | +# Task completion receipts |
| 2 | + |
| 3 | +MCP clients and servers can exchange rich task and tool data, but a task result is not always the same thing as an auditable completion state. |
| 4 | + |
| 5 | +For safety-sensitive workflows, consider adding a small receipt to the end of a task or tool-driven workflow. A receipt separates what the agent or tool claims from the evidence that supports the claim, the next owner, and any human approval boundary. |
| 6 | + |
| 7 | +## Why this matters |
| 8 | + |
| 9 | +A final task message such as: |
| 10 | + |
| 11 | +```text |
| 12 | +Done. All tests passed. Ready to publish. |
| 13 | +``` |
| 14 | + |
| 15 | +contains multiple operational claims: |
| 16 | + |
| 17 | +- the task is complete |
| 18 | +- tests passed |
| 19 | +- the output is ready for an external action |
| 20 | + |
| 21 | +Those claims may require different evidence, such as tool output, logs, file diffs, test output, trace items, or human approval. |
| 22 | + |
| 23 | +## Minimal receipt shape |
| 24 | + |
| 25 | +```yaml |
| 26 | +status_code: 412 |
| 27 | +status_text: missing_evidence |
| 28 | +summary: "The task claimed completion and test success, but did not attach command output." |
| 29 | +claims: |
| 30 | + - claim: "Task is complete." |
| 31 | + support_status: unverified |
| 32 | + evidence: [] |
| 33 | + required_fix: "Attach the relevant tool result, file diff, log, or trace item." |
| 34 | + - claim: "All tests passed." |
| 35 | + support_status: unsupported |
| 36 | + evidence: [] |
| 37 | + required_fix: "Attach the test command and output, or downgrade the claim." |
| 38 | +next_owner: ProducingAgent |
| 39 | +human_decision_required: false |
| 40 | +``` |
| 41 | +
|
| 42 | +## Safety boundaries |
| 43 | +
|
| 44 | +Receipts are useful when a task result may cross a boundary: |
| 45 | +
|
| 46 | +- publishing or deploying |
| 47 | +- writing to long-term memory |
| 48 | +- sending messages or comments |
| 49 | +- spending money |
| 50 | +- changing external systems |
| 51 | +- reporting tests or verification as complete |
| 52 | +
|
| 53 | +If a boundary requires approval, the receipt should say so explicitly: |
| 54 | +
|
| 55 | +```yaml |
| 56 | +human_decision_required: true |
| 57 | +required_fix: "Require approval before publishing." |
| 58 | +``` |
| 59 | +
|
| 60 | +## Relationship to MCP data |
| 61 | +
|
| 62 | +MCP data can provide evidence for a receipt: |
| 63 | +
|
| 64 | +```text |
| 65 | +tool result -> evidence item |
| 66 | +task state -> status code |
| 67 | +client/server trace -> evidence path |
| 68 | +approval or user confirmation -> human approval evidence |
| 69 | +``` |
| 70 | + |
| 71 | +The receipt does not replace MCP protocol messages. It is a final-state summary that helps humans and downstream systems audit whether a completion claim is supported. |
| 72 | + |
| 73 | +Reference protocol: |
| 74 | + |
| 75 | +```text |
| 76 | +SACP: https://github.com/aDragon0707/sacp |
| 77 | +``` |
0 commit comments