yoi/work-items/open/20260601-132955-tui-peer-pod-handshake-command/artifacts/implementation-report.md

37 lines
3.0 KiB
Markdown

# Implementation report: peer Pod handshake command
Date: 2026-06-02
## Investigation
The current boundaries are documented in `artifacts/investigation-summary.md`. No escalation blocker was found. The main concern identified was avoiding reuse of spawned-child state (`SpawnedPodRegistry`, delegated scope, output cursors, and child completion semantics) for peer communication; the implementation therefore adds separate peer metadata and a separate peer send tool.
## Implemented behavior
- Added reciprocal peer metadata to `PodMetadata` as `peers`, separate from `spawned_children` and `reclaimed_children`.
- Added protocol `Method::RegisterPeer { name }` and `Event::PeerRegistered { result }`.
- Added controller handling for `RegisterPeer`, idle/paused only, validating an existing target Pod and rejecting self-handshakes.
- Added `PodDiscovery::register_peer` that persists both metadata directions and restores the exact prior source-side peer state on ordinary second-side write failure.
- Extended `ListPods` visibility to include `VisibilityReason::Peer`; a successful handshake makes both Pods see each other as `peer` through Pod metadata.
- Added `SendToPeerPod` as a distinct LLM tool. It only sends to visible live peer Pods, delivers `Method::Notify` with a source label, and does not use child delegation, output cursors, parent ownership, or child completion notifications.
- Added TUI command `:peer <pod-name>` for idle attached Pods. Success is reported through a transient actionbar notice when the controller returns `PeerRegistered`.
- Documented peer semantics in `docs/design/pod-session-state.md` and added prompt guidance that peer Pods are not spawned children.
## Reviewer blocker fixes
- `SendToPeerPod` now reuses the existing one-shot Pod socket client path (`connect_and_send`), which drains connect-time `Alert` / `Snapshot` traffic before writing `Notify` and returns an error if method delivery fails.
- Added a regression test where the target socket emits an alert and snapshot before reading the peer `Notify`, proving the peer send drains the prelude and still delivers the message.
- Registration failure rollback now restores the exact prior source peer list instead of unconditionally removing `source -> target`; a target-side injected failure test verifies a pre-existing source relation is preserved.
- Wording now describes `:peer` as metadata-level reciprocal registration rather than live target-controller consent, and documents that `SendToPeerPod` fails for non-live peers instead of auto-restoring them.
## Tests and validation run
- `cargo test -p protocol -p pod-store -p pod -p tui --lib`
- `./tickets.sh doctor`
- `git diff --check`
- `nix build .#yoi`
## Notes
The two-file reciprocal metadata update is not crash-transactional because the existing Pod metadata store has no multi-record transaction boundary. The implementation avoids successful replies with one-sided state for normal validation/write failures by restoring the exact prior source-side peer list if the reciprocal write fails.