Inspect what the stories claim.

Reading the visual narration

Each event has three narrated actions. Playback spends six illustrative seconds on an event by default; this is pacing for explanation, not agent latency. First, Then and Finally let you read the actions manually. Worktree A and B are example isolated checkouts; the host and agents create and inspect them, not the animation.

The envelope follows a message path independently of the work card. The handoff example assumes an available managed TMUX path with the intended receiver verified; the durable handoff can also be read by polling the ledger. Transport arrival does not transfer work ownership. The pending artifact stays with Claude through acknowledgment and moves into Codex’s review only after acceptance.

Why these situations?

A focused qualitative review found recurring questions about handoffs, overlapping edits and shared runtime resources. This is a selected set of discussions, not a representative frequency study. A person describing a problem is not evidence they tested Rally.

• Reddit: Handover between CC and Codex (April 24, 2026). The original author asks how to carry work between tools when switching. Replies also argue that a concise handoff file can be sufficient.
• Reddit: If you run multiple AI agents on the same repo, how do you stop them stepping on each other? (August 24, 2026). The original author describes separate worktrees but difficulty testing against a shared running app. Automated thread summaries and advertised performance claims were excluded from the evidence.
• Codex issue: Automatically isolate and coordinate concurrent writes (August 6, 2026). The requester describes the burden of coordinating local chats, writing scopes and merge order. This is a user report, not a reproduced defect in a current release.
• Hacker News: Is anyone else drowning in terminal tabs running AI coding agents? The discussion includes coordination concerns and a simpler alternative: one feature per worktree, integrating dependencies in order. It is qualitative context, not comparative product validation.

Private community feedback also informed the choice of scenarios. Individual messages and identities are not republished here. The recovery story is an important protocol edge case, not a claim that it is one of the most common complaints.

Overlapping work

Rally records intent and claims. An overlapping exclusive claim can be refused inside Rally. The default edit hook remains advisory; a claim is not an operating-system lock. Separate worktrees allow isolated changes, followed by deliberate integration and checking.

Pinned implementation documentation: claims and enforcement boundaries. A named integration owner in the story is the agents’ agreement for that task.

Handoff and review

The sender identifies a receiver, assignment and evidence. Delivery does not establish receiver acceptance. The receiving session must respond; completion requires a separate result. A changed revision needs its own checks.

Pinned handoff and launch contract. The second review cycle in the replay is compressed into one event and labeled accordingly.

Shared database or service

Rally supports typed service and port claims. A competing claim is refused at claim time. An agent that acts on the database without claiming it is outside that check. Automatic file hooks do not inspect SQL or provision isolated environments.

Pinned resource-scope boundary. This is a hypothetical scenario, not a reconstructed incident. Both sessions claim the same service identifier before using it; the host executes and verifies the actual operations.

Stalled sessions

Stale presence means a check-in is old. It does not prove a process stopped or authorize takeover. The story requires inspection, preserved work and an explicit decision before new ownership.

Pinned command semantics. The simulation never treats the passage of time as permission. Its final step assumes ownership was successfully resolved. Historical audits found failed recovery commands and manual intervention; an unsuccessful recovery must remain paused.

Practice evidence

Evidence was inspected locally on September 13, 2026. Public commit links allow source inspection where the commits are available remotely. Historical agent reports are distinguished from checks rerun for this draft.

Persona Lab: the UI needs the data contract

Initial commit 10a1271 (May 27, 2026) contains the UI scaffold, architecture, data-contract documentation and persona schema. The contemporaneous UI-agent feedback says Codex’s contract handoff was actionable. These files were inspected in the commit.

The same feedback describes pending acknowledgments failing to surface and a human switching terminals. The original May handoff receipts were not available in the current ledger cache. This is a preserved development example with reported handoff benefit, not proof of unattended execution or measured time savings.

Atomize AI: preserve a lane during integration

The September 7, 2026 coordination chain includes a targeted request for the login lane’s revision and ownership status, a response referencing that request, the integrator’s follow-up, and a later owner verification. The commits were checked locally: 0ef70b953 is an ancestor of integrated revision 8c072da69.

This establishes preserved work and a recorded handoff. The execution host and Build Loop performed the integration. It does not prove that Rally merged the code, prevented a particular deletion, or deployed the result. The historical report includes a failing check and narrower reruns; it is not described here as a fully green suite.

Ambient Agent: a review changes the plan

On September 3, 2026, review records returned a “revise” verdict on a branch-consolidation plan. The review identified a validation condition that could not pass and preservation evidence that could not be reproduced. A response linked back to the review task before corrective commit 178d665.

The commit was inspected and addresses both named issues. That supports “review changed the plan before proceeding.” It does not measure loss avoided, authenticate historical model identities, or prove the subsequent runtime behavior. Detailed internal branch inventories and runbooks are not published in this example.

Improving the plugin while using it

Rally’s September 5 implementation report described a hook change; a September 6 review challenged its test-count claim. Both are preserved in the local Rally ledger. Source commits exist, and the focused channel-audience contract test was rerun for this draft: 20 passed, 0 failed.

Route the before-write advisory to the model · Update the renderer tests.

This demonstrates inspectable implementation and review records plus a current focused check. It does not establish that Rally caused the code fix, that every historical count was accurate, or that every host interaction works.

Building the Astro explanation

The local Astro room contains scoped simulation claims, review handoffs and integration reports. The preserved Session Lanes implementation exists in local commit 54f3dd9; its predecessor is bf2fa14. These artifacts were inspected when preparing this draft.

The simulation uses source from the Rally repository to explain its rules. Reading another repo’s source and using Rally in both projects does not by itself prove automatic cross-repo handoff, shared authority or linked rooms.

Room isolation and optional cross-repo discovery. Each repo retains its own coordination history; linked worktrees of one repo share a room.

Launching Claude or Gemini

The source documents managed launch adapters for Claude, Codex, OpenCode and Gemini. In this review, rally run claude --name story-proof --dry-run --json and the equivalent Gemini command returned launch plans with a session identity, worktree and tmux backend.

The returned delivery grade was unverified_transport. These dry runs did not start agents or prove authentication, accepted instructions or completed work. A coding host executes the task after launch.

Supported runtime documentation. This review used the local CLI build 0.2.7+ad9c910 and source snapshot 89be57d; those are distinct evidence surfaces.

Rally release: a hold survives an incomplete fix

July 31, 2026 records show a Claude-authored hold, Codex acknowledgment, a challenge to candidate 36a2789, release of the store-file claim and a new Codex claim. The next audit request pinned fix 4b28c0e and its tree. Both hold resolutions precede the local v0.1.7 tag at 10a4b05.

Selected durable log payloads were cross-checked against the SQLite projection, and the storage fix and tag ancestry were inspected. Historical stress results and CI success remain participant reports, not reruns. One acceptance reference was absent from the inspected projection; the summary names the hold explicitly. Pending injections did not always receive replies. Tool labels are not proof of distinct people or models.

Inspect the storage repair. This demonstrates a recorded release hold and ownership handoff, not a measured outage prevented.

Atomize: a warning can be wrong

September 3, 2026 records show an initial taxonomy-conversion concern retracted after the reviewer read the design. A subsequent concern addressed old application writers encountering a new constraint. Commit 49d2a3ebc removes the bulk update and permits legacy plus canonical values during rollout.

The selected durable records and exact migration diff were inspected. The resolution reviewed an uncommitted file; byte identity with the later commit was not established. Production queries and backfill outcomes were not rerun. An independent audit initially mixed this chain with a separate release-attribution review; that error was removed here. This story is about revising a concern using evidence, not proof of avoided data loss.