Holmes LabAgentic Software Engineering Architect

How can different AI CLI sessions collaborate explicitly?

A review request sent to another CLI has no reply. Should you send it again? Follow one illustrative request to separate delivery, acknowledgement, work completion and decisions under uncertainty.

Case edition Updated

Inspect Bridge and its limits

Case-edition basis: 2026-11-08. The messages and state transitions below are illustrative, not logs from actual sessions. Read the distribution guidance against Agent Bridge 0.1.0-alpha.1 public documentation. This editorial change runs no new Claude/Codex/AGY round trip or model call, and does not turn bounded tests recorded in public documentation into support for every environment.

The case: ask for one review, not implementation

Suppose sender A asks receiver B to review a greeting-function change without editing it. A shared directory does not establish the recipient, conversation or authority. First verify actual enrolled sessions, canonical workspaces and the allowed route.

The task is one review response. A message does not create permission to edit files, install, deploy or use additional models. For a real trial, the human first bounds request count, elapsed time, cost and stop conditions.

Read delivery and work state separately

The table compares evidence at different layers. completed is a work-receipt state, not a guarantee that automatically follows transport acceptance. Showing a successful send API as a completed review reports work that has not been established.

Different facts to observe for one request
Layer / stateWhat it establishesWhat remains unproven
Delivery · queuedThe request is queuedThat the receiver read it
Delivery · transport_acceptedThe transport accepted itAgent acknowledgement or completed work
Delivery · agent_acknowledgedThe receiver acknowledgement is recordedThat the review is complete
Work · completedA separate completion receipt existsCorrectness or independent verification
Delivery · delivery_unknownDelivery may have happened; the outcome is uncertainConfirmed nondelivery or safe redelivery

An illustrative message and receipt sequence

This is illustrative notation for reading, not a Bridge API request format or execution log. A real receiver reads bridge_inbox and acknowledges with bridge_ack. Check work completion separately through bridge_work_receipt and the result itself.

Illustrative scenario · not a live delivery log

A → B: Review trim/default behavior. Do not edit files.
delivery: transport_accepted
B reads inbox and acknowledges: agent_acknowledged
B reports: normal name preserved; padded/empty tests reviewed
work receipt: completed
A checks the report and its supporting evidence.

Decide under uncertainty before retrying

No reply does not establish nondelivery. If delivery happened and only the response was lost, another send can duplicate work. Inspect receipts and the receiver session first, then determine whether the official reconcile procedure is needed.

Consider a determination such as confirmed_not_sent only with evidence of nondelivery. It can trigger redelivery; never select it merely to continue a workflow. If the operator cannot establish what happened, stop and report the uncertainty.

What to inspect when the outcome is uncertain
ObservationNext decision
Only transport_accepted existsInspect receiver inbox and acknowledgement
Acknowledgement exists, but no review resultInspect work state and the separate completion report
delivery_unknownInspect receiver evidence without automatic redelivery

Check the selected distribution’s limits

The existing first-use guide is based on 0.1.0-alpha.1, with macOS, Node-version and license restrictions. For an actual installation, compare the selected release’s official README with installed help. Do not borrow Windows/Linux or managed AGY support claims from another release.

Bounded round trips recorded in public documentation do not establish longevity, every crash boundary or isolation across real session pairs. Bridge does not promise exactly-once remote execution. This article is neither a new compatibility certification nor a release announcement.

Authority stays separate when using Kit

Kit connects change intent, specifications and verification; Bridge handles explicit message delivery. A received message does not approve a specification or activate hooks. The receiving task still follows the project’s actual human permissions and specification workflow.

A checklist before the first round trip

Start with one greeting containing no execution instructions. Record enrollment, allowed routes and acknowledgement evidence, then follow the selected guide’s cleanup steps such as pause/stop. Never publish actual session IDs, credentials or private paths.

  • Are the actual enrolled receiver and workspace route verified?
  • Did the human bound requests, time, cost and stop conditions?
  • Are transport acceptance and acknowledgement recorded separately?
  • Are the completion receipt and result evidence checked?
  • Are uncertain deliveries kept out of automatic retries?

Sources and verification

This case edition expands an earlier introduction. Source versions and the evidence basis above define its scope.

Share this reference

Copy a link or the prepared introduction. Nothing is posted automatically.

Introduction draft for an external post
How can different AI CLI sessions collaborate explicitly?

A review request sent to another CLI has no reply. Should you send it again? Follow one illustrative request to separate delivery, acknowledgement, work completion and decisions under uncertainty.

https://holmes-lab.web.app/en/notes/cli-agent-handoff/

Continue from here