ReturnReady: Check the Return Handoff

Check the return handoff before a package moves

A phone answer can conflict with written return instructions, and a later correction can invalidate an earlier matching answer. ReturnReady puts those sources side by side so an operator can see what still needs clarification.

It checks five details: authorization, destination, shipping payer, deadline, and refund or replacement terms. A matching score never authorizes shipment.

Current demonstration · Merged contribution PR #408 · Reviewed source and setup · Executed correction checks

Upstream status, verified September 13

PR #408 was merged by the maintainer on September 11, 2026, at 09:20 UTC. The merge commit is a6035f235b27a6b1307c91b7f2dca5d161f3de85. The maintainer's September 11 superseding review states that no current Must Fix remains. This replaces the earlier pending-re-review status.

Maintainer acceptance is not a claim that every older fork, cached object or previously published media file was erased. It also does not establish competition eligibility, a successful return enquiry or an award.

Review corrections, September 10

The repository contribution is v0.4.1 with wholly synthetic example conversations, outcomes, dates, durations and identifiers. Earlier call-derived documentation and operational evidence directories were removed from the contribution, and its unique PR history was replaced on September 10 with one correctly named commit. The correction-test source remains pinned to f0dba156f2d30745aa4a69b022e14dfaa87569d4; the maintainer subsequently resolved the catalog conflict and merged the contribution.

Phone inputs require ASCII E.164 without surrounding whitespace. The actual live HTTP boundary rechecks syntax and blocks reserved fictional destinations. The UK full-number fixture is in Ofcom's reserved London drama range and is used only by injected request spies and loopback tests. Syntax validation does not establish subscriber identity, provider availability or permission.

This is a source correction, not a replacement of the already hosted presentation. The existing v3 video still includes selected redacted metadata from the earlier blocked test. It is not relabelled wholly synthetic. The repository's synthetic-only tests are not evidence that a successful real enquiry occurred.

A complete review workflow

Load written instructions and a call result. Inspect each comparison against its original recipient quotation. When a destination changes later, ReturnReady retains that later statement, removes the earlier destination from the supported-match count, and creates a source-linked clarification question.

Download the review with its input fingerprints, original evidence and unresolved questions. Save a session to continue later. Reopening recomputes the review instead of trusting a saved approval or making an old observation fresh. Editing inputs disables stale exports; an invalid import leaves the current review intact.

The comparison is conservative English lexical review, not general semantic contradiction detection. Caller speech and generated summaries cannot replace quoted recipient evidence.

How CALL-E is used

An agent host uses CALL-E's supported MCP planning and confirmation flow, retains the actual run ID, and reads that same run after it ends. ReturnReady imports the finished result through its data view or offline CLI. The operator supplies the independently expected run ID, original observation time and explicit permission review.

Wrong runs, multiple attempts, unsupported transcript formats and inconsistent receipts are rejected. Unknown or unavailable permission, an explicit provider error, a blocked next step or a recognized refusal produces a hold: no transcript or extracted claims enter the normalized review and no retry is requested.

For a permitted result, the importer supports the observed timestamped BOT/USER transcript format. It does not invent return fields. Separately reviewed annotations must include an exact quotation and turn index. The file import is operator-mediated, not an authenticated provider fetch.

Import format, CLI and limitations

Demonstrated behavior and limits

The earlier support-authorized hotline attempt did not grant the required permission or provide return instructions. Completion flags were not treated as usable return evidence. The caller continued scenario questions despite its stop instruction; this remains a limitation, not a successful consent-control demonstration. No further call was made as part of these corrections or this status update.

The post-call importer cannot terminate an active call or erase provider recordings. No hotline transcript, recording, phone number or provider identifier is included in the current contribution. The historical video is not a merchant pilot, a successful return enquiry or proof that Nigerian calling works.

The separate optional REST adapter is not live-verified. Its explicit live flag, specific-recipient authorization, number confirmation, recipient-local hours and SQLite no-retry ledger remain separate from the MCP importer. Only unstarted previews can be cancelled by this application.

Run and inspect it

Use Python 3.11 or newer. Open apps/python/returnready in the pinned source, run python returnready.py, and open the printed localhost address. Default mode needs no API key and places no calls.

The recorded correction run passed 147 Python unit/HTTP checks, comprising 130 retained assertions and 17 new input/transport checks. Both actual localhost browser suites passed: 18 existing review workflows and 15 finished-result workflows. Repository validation and the branch-name check passed. All call-like test inputs were wholly synthetic and no provider requests were made. These are historical execution results, not new tests run for the September 13 status update. They establish tested software behavior, not user adoption, general accuracy or live-service compatibility.

Presentation and provenance

The linked video is Joseph Ayanda's simplified v3 upload. It uses enlarged app detail views, recorded application footage and paced stock neural narration. Synthetic examples and the historical redacted negative-path metadata remain distinguished. Edited footage and held detail views are explanatory, not an uninterrupted real-time session.

No purchases, refund initiation, courier booking, fee acceptance or automatic shipment authorization are performed. Fingerprints identify supplied bytes, not merchant identity, permission or truth. Recording and publication permissions are separate.

The next validation is an authorized return-information trial and operator usability testing. No customer benefit, paying user or prize is claimed. Original work by Joseph Ayanda, developed with substantial AI assistance under the MIT license. The existing CALL-E entry 1173962 remains submitted with the confirmed account details and the video link supplied by Joseph. The upstream contribution is merged; competition eligibility and awards remain decisions of the organizers.

Built With

• call-e
• css
• html5
• javascript
• playwright
• python
• sqlite