0030 — Engine-delegated wire and projection consumers

TL;DR. Both ends of the wire run the same terminal engine (libghostty), so the wire carries only identity, lifecycle, transport, opaque terminal bytes, and L3 metadata; every structured surface — screen state, semantic events, panes, layouts, “run and wait” — is a consumer-side projection of that shared engine, never a wire tier. This rejects the gRPC/cells-on-wire agent design, dissolves the L2 collection tier into L3 grouping plus a single atomic L1 batch operation (KILL_TERMINALS), names phux-web as the reference projection pattern, and reaffirms ADR-0017 as the constraint that keeps the reference TUI a pure consumer.

Status: Accepted Date: 2026-06-06

Context

A docs-wide audit surfaced a cluster of contradictions that all reduce to one unresolved question: what is allowed to be a wire tier, and what must be a consumer-side projection of the shared engine? The symptoms:

• Session/collection lifecycle verbs (CREATE_SESSION, KILL_COLLECTION, RENAME_SESSION) ship today as L1 commands in phux-protocol’s Command enum, while ADR-0017 says session vocabulary is banished from the wire and ADR-0015 reserves the collection lifecycle for L2.
• Two agent surfaces disagree. The live wire realizes agent needs as L1 commands (GET_SCREEN, ROUTE_INPUT, GET_TERMINAL_STATE) plus an AgentEvent push frame; the unbuilt docs/spec/L2_AGENT_PROTOCOL.md (marked stability: scratch) prescribes a gRPC+JSON transport carrying structured cells and a different event taxonomy.
• Two SDK docs disagree: a hand-rolled L1 codec versus a gRPC/tonic structured-state service.
• The encoding spec prescribes field-tagged TLV; every message body is positional. This is acknowledged divergence, not a settled choice.

These are not four bugs. They are four places where structure that belongs to a consumer’s view of the terminal has either leaked onto the wire or been proposed for it. Without a stated principle, each gets re-litigated in isolation and the wire accretes product opinions one verb at a time — the failure mode ADR-0015 and ADR-0017 already named for sessions and layout, now recurring for agent state.

ADR-0013 established that both ends run libghostty and the wire carries VT bytes. ADR-0018 generalized that to lazy state synchronization of engine state. ADR-0022 framed every consumer as a different projection of one source-of-truth Terminal. This ADR makes the shared premise of all three normative and uses it to settle the audit.

Decision

1. The engine is delegated; structure is always a projection

phux does not own terminal semantics. libghostty owns the mapping from bytes to a grid of cells and from input atoms to bytes, and both ends of the wire run that same engine (ADR-0013). phux never re-encodes terminal state into a second representation on the wire.

It follows that any structured view of a terminal — a cell grid, a semantic command-boundary stream, a pane tree, a layout, a “run a command and collect its output” result — is computed by a consumer from the engine it already runs. Structure is a projection, not a transmission. Putting a structured terminal representation on the wire would re-create the tmux re-parse liability one layer up: a second model that can drift from the engine and degrade under capability mismatch, which is precisely the cost ADR-0013 paid to remove.

2. The wire’s job, stated as a closed list

The wire carries exactly: terminal identity (ADR-0016); terminal lifecycle, including an atomic multi-terminal batch operation; transport framing and capability negotiation; opaque terminal bytes (output, snapshot, input atoms per ADR-0024); and L3 metadata the server stores without interpreting. It does not carry structured screen state, semantic event taxonomies as a normative type system, panes, layouts, or command-runner results. Those are consumer projections.

The existing L1 agent affordances are read in this light. GET_SCREEN and GET_TERMINAL_STATE return engine-derived snapshots a consumer could also compute locally; they are a convenience for consumers that have not yet adopted the carry-your-own-engine pattern (point 4), not a license to make structured state a normative wire contract. New structured surfaces SHALL NOT be added to the wire; they belong in the projection.

3. The gRPC/cells-on-wire agent design is rejected

docs/spec/L2_AGENT_PROTOCOL.md proposes a separate gRPC+JSON transport that puts structured cells and a parallel event enum on the wire. It is rejected and superseded by this ADR. It violates point 1 (structure on the wire), it forks the codec and event taxonomy from the live L1 surface, and it contradicts ADR-0022’s “agents are a projection, the CLI + JSON schema is the contract.” The structured agent surface (cells, command results, semantic events) is a local projection over the shared engine, exposed to agents through the CLI and its versioned JSON schema, not a wire service.

4. Consumers are peers; phux-web is the reference projection

The TUI, the browser client, and the agent surface are peers — none is protocol-privileged (ADR-0017). The reference pattern for a consumer is carry your own engine and project locally: phux-web (ADR-0025) runs ghostty-vt.wasm, speaks the exact wire codec (ADR-0024), and computes its rendered view from engine state it owns. The agent SDK SHOULD follow this pattern — run the engine, project to structured state locally — rather than become a gRPC structured-state service. The wire stays identical across all three; only the projection differs.

5. Group lifecycle is L3 metadata plus an atomic L1 batch op; no L2 tier

The one thing a consumer-side projection genuinely cannot do is an atomic group operation — kill a bundle of terminals such that no observer sees a partial state. That, and only that, is irreducible — and it is a single L1 operation, not a tier. The wire gains KILL_TERMINALS { ids: [TerminalId] }, applied atomically under the server’s existing single Mutex<ServerState> lock (one lock acquisition, all-or-nothing for a local server; cross-host atomicity is out of scope). There is no L2 collection tier: group membership and names are L3 metadata plus client logic, and sessions, windows, panes, and layouts remain L3 conventions plus client logic (ADR-0017, ADR-0019, ADR-0027).

The current code is wrong here: CREATE_SESSION, KILL_COLLECTION, and RENAME_SESSION ride as L1 commands, putting session vocabulary and collection lifecycle in the substrate tier. They decompose and are removed: create is SPAWN_TERMINAL(s) plus L3 metadata, rename is an L3 metadata SET on a name key, and kill-group is the new KILL_TERMINALS { ids }. The CollectionId plumbing is retired where it existed only to serve those verbs; where it is entangled with SPAWN_TERMINAL or L3 metadata scoping it may remain as a documented opaque grouping key (not a lifecycle tier) with the remnant noted for a follow-up bead. The migration is tracked as a code task; until it lands the docs state the target and flag the divergence inline.

6. The reference TUI is the wedge, held pure by ADR-0017

The reference TUI is the daily-driver adoption surface that bootstraps a population of terminals-on-the-wire, and it is worth heavy product investment. Its differentiator is the wire itself — attach/detach, remoting, and humans and their agents sharing the same live terminals — not local splits, so it is not merely a second local multiplexer. The constraint that keeps the wedge from corrupting the platform is ADR-0017: the TUI gets no protocol-level standing, and its needs land as L3 conventions and client logic, never as new wire surface. Investing in the TUI as a product and holding it as a pure consumer are not in tension; the leaked L1 session verbs (point 5) are the current breach of that line, and the thin or unpublished wire/agent/web docs are the current gap.

Rationale

The delegation principle is what gives phux its central property: because the engine is shared and never re-encoded, the wire cannot introduce a second terminal model that drifts or degrades. Every time structure has been proposed for the wire — layout in an early draft of ADR-0015, session verbs in ADR-0021, cells-on-wire in the agent spec — the same argument retires it: the structure is recoverable from the shared engine, so transmitting it adds a drift surface and a conformance tax on consumers that do not want it, while buying nothing the projection lacks.

Group lifecycle is the lone exception because atomicity is not recoverable from a projection: a client tearing down N terminals one at a time exposes intermediate states and races a concurrent observer. A server-side atomic operation is the cheapest correct answer — and the cheapest form of that answer is a single L1 verb (KILL_TERMINALS { ids }) under the existing server lock, not a whole tier. Atomicity earns one op; it does not earn L2.

Naming phux-web as the reference pattern turns an abstract principle into a copyable shape: a consumer that wants structure runs the engine and reads it, exactly as the browser client already does in shipping code.

Tradeoffs

• A structured agent surface costs each agent consumer an engine. Running libghostty to project structured state is more work than reading cells off a gRPC stream. We accept it: it is the same cost the browser client already pays, and it removes the drift surface a structured wire would add.
• Removing the leaked verbs is a wire-affecting change that today’s code has already shipped at L1. Decomposing them into SPAWN_TERMINAL + L3 metadata + KILL_TERMINALS (and retiring the CollectionId plumbing that only served them) is real migration work, not a doc edit, and bumps PROTOCOL_VERSION 0.2.0 → 0.3.0; the docs carry an inline divergence marker until it lands.
• The encoding question is left open here. Whether message bodies migrate from positional to field-tagged TLV is orthogonal to tiering and is not decided by this ADR; it remains tracked separately. Naming it avoids the reader inferring that “engine-delegated wire” settled the codec shape.
• Heavy TUI investment under a no-privilege constraint means TUI features must be expressible as L3 conventions and client logic. That is a real design discipline, occasionally more work than a bespoke wire message would be, and it is the price of keeping the substrate product-agnostic.

Alternatives

(B) Dissolve L2 entirely into L3 metadata plus an L1 batch-kill — ADOPTED. Drop the Collection tier; represent grouping as L3 metadata (a set of TerminalIds under a well-known key) and provide atomicity through a single L1 batch operation (KILL_TERMINALS { ids }) that the server applies as one unit under its existing lock. Adopted because it is the most faithful expression of this ADR’s own principle — atomicity is the lone exception, so it earns exactly one op, not a tier — and because it is better for federation: there is a single federated identity, TerminalId, with no second CollectionId to federate and reconcile. The honest cost is that it pushes membership consistency into client-maintained metadata (no server-enforced membership view) and turns “kill this group” into a client-assembled id list rather than a named server-side entity. We accept that: grouping is presentation, and the one correctness need (atomic teardown) is fully served by the batch op.

(A) Minimal lifecycle-only L2 collection tier — considered, not chosen. Keep L2 as a narrow tier whose entire job is atomic create/kill/membership of a Collection, with sessions/windows/panes/layout as L3 plus client logic. Honest description: it isolates the one operation a projection cannot perform atomically behind a server surface, makes grouping a named server-consistent entity, and leaves ADR-0015’s three-tier model intact. Not chosen because it adds wire surface (a whole tier) for what is one op of irreducible need, and it introduces a second federated identity (CollectionId) to carry, reconcile, and version alongside TerminalId — weight that option B avoids while still delivering atomic teardown.

(C) Adopt the gRPC/cells-on-wire agent protocol. Build the agent surface as the separate structured-state service L2_AGENT_PROTOCOL.md describes. Rejected per Decision point 3: it puts structure on the wire, forks the codec and event taxonomy, and contradicts the projection thesis. Recorded here only to mark it as considered and closed.