fleetd/docs/CB-401-Peer-Launcher-SPI.md

CB-401 — Peer Launcher SPI (Stage 4: pluggable peers)

Status: design note (feature branch feature/peer-launcher-spi) Stage: 4 — makes the bridge scalable to heterogeneous peers (Claude Code, Codex, …) without the core learning any one peer's environment.

1. Why

claude-bridge is a communication bus between heterogeneous AI agents — its stable surface is the protocol (fleet_spawn / send / poll / reply / ask / list / stop / status), and that surface should stay provider-neutral. Today the daemon can only materialize one kind of peer: an off-subscription Claude Code CLI over herdr. Everything specific to how that peer is set up (ANTHROPIC_BASE_URL, the subscription guard, --mcp-config/system-prompt flags, claude-* naming, git-token injection) is baked directly into the core spawn path.

The goal of CB-401 is a seam — a PeerLauncher SPI — so that "how to bring a peer of kind X to life" lives in a swappable adapter the bus delegates to, while the bus itself owns only transport, session/turn lifecycle, and routing. This both unlocks a second peer kind (Codex, a human, another Claude) and retroactively gives the CB-301-ext / CB-302 environment features a principled home (an adapter) instead of sitting in core.

Non-goal for CB-401: dynamic/external plugin loading (arbitrary jars). That is Stage C and carries a security model of its own (§7). CB-401 delivers a first-party, in-tree, config-selected SPI with exactly one implementation, proving the seam.

2. The boundary

flowchart LR
    subgraph core["BRIDGE CORE — provider-neutral"]
        proto["Protocol verbs<br/>spawn / send / poll / reply / ask / list / stop"]
        sm["SessionManager<br/>FSM · registry · roster · lifecycle"]
        msg["Message store / routing"]
        life["Lifecycle limits<br/>idle_ttl · context_cap · drain"]
    end
    subgraph adapters["PEER ADAPTERS — env-specific"]
        cc["ClaudeCodeLauncher<br/>(today's WorkerService)"]
        cx["CodexLauncher<br/>(future, CB-402)"]
        hu["HumanLauncher<br/>(future)"]
    end
    sm -->|"delegates spawn/release"| spi{{"PeerLauncher SPI"}}
    spi --> cc
    spi --> cx
    spi --> hu
    cc -.->|"herdr transport"| herdr["herdr (terminal multiplexer)"]

Figure 1 — the core delegates peer materialization to a launcher chosen by profile; the core never learns a peer's env.

3. Coupling audit (as-built, main @ 0efb65c)

Where Claude/herdr specifics actually live today:

Concern	Location	Verdict
ANTHROPIC_BASE_URL / ANTHROPIC_MODEL / CLAUDE_CONFIG_DIR / ANTHROPIC_AUTH_TOKEN env	WorkerService.spawn	→ adapter
SubscriptionGuard.assertWorker(baseUrl) (billing boundary)	WorkerService.spawn → guard	→ adapter (it guards an ANTHROPIC_* concept)
--mcp-config + --append-system-prompt REPLY_CHARTER (Claude Code CLI flags)	WorkerService.argvWithBridge	→ adapter
claude-<profile>-<nonce>-<seq> naming, WORKER_NAME regex, orphan reap (CB-117)	WorkerService	→ adapter (naming is a herdr-label detail)
GITEA_TOKEN / GITEA_HOST injection (CB-302 checkpoint)	WorkerService.spawn	→ adapter + a capability (§6)
tab/pane placement, worker space, tab labels	WorkerService.spawnInTab/spawnAsPane via herdr WorkspaceControl	→ adapter (herdr transport detail)
FleetConfig.Worker profile shape (baseUrl, model, configDir, …)	config	mostly adapter-shaped — see §5
FSM, registry, roster, reapIdle/drainAll/contextCap, rosterView	SessionManager	stays core
turn/completion detection (TurnListener, CompletionResolver, StatusPoller, WorkerPresence)	inject/	stays core, but reads herdr terminal output → transport-coupled (§4b)
message store & routing	msg/	stays core
Agent / paneId handle	herdr/	generalize — see §4a

Finding: the extraction is tractable because ~90% of the coupling is already funnelled through one class (WorkerService). Renaming/adapting it to ClaudeCodeLauncher implements PeerLauncher and having SessionManager depend on the interface is the bulk of Stage A.

4. Two friction points

4a. paneId is a herdr handle, not a peer-neutral id

SessionManager keys its registry by paneId, MCP/REST route by paneId, and WorkerSession stores it. paneId is a herdr pane handle — meaningless for a peer that isn't a herdr pane.

Decision: introduce an opaque PeerHandle the launcher returns. It carries a launcher-assigned id (the registry/routing key) plus launcher-private coordinates (for herdr: paneId, tabId, terminalId). Stage A keeps id == paneId for the Claude adapter so nothing downstream changes value, but the type stops being "a herdr pane" — the core routes on PeerHandle.id().

classDiagram
    class PeerLauncher {
        <<interface>>
        +Set~Capability~ capabilities()
        +PeerHandle spawn(SpawnRequest req)
        +void release(PeerHandle h)
        +String effectiveCwd(SpawnRequest req)
        +List~String~ parityOverlay(String profile)
        +int reapOrphans()
    }
    class PeerHandle {
        <<interface>>
        +String id()
    }
    class ClaudeCodeLauncher {
        herdr AgentControl/WorkspaceControl
        SubscriptionGuard
    }
    PeerLauncher <|.. ClaudeCodeLauncher
    ClaudeCodeLauncher ..> PeerHandle : returns

Figure 2 — the SPI the core sees. ClaudeCodeLauncher is today's WorkerService, adapted.

4b. Turn/completion detection reads herdr output

inject/ (turn listener, completion resolver, status poller, presence) infers turn boundaries from herdr terminal scraping. That is genuinely peer-transport-specific — a Codex peer would signal turns differently. For CB-401 this stays in core (it's the Claude/herdr transport's detector), but §6's capability model is what lets a future non-herdr peer bring its own turn-signalling without the core assuming terminal scraping. Out of scope for Stage A; noted so the SPI doesn't accidentally hard-wire "turns come from herdr".

5. Config shape

FleetConfig.Worker is Claude-shaped (baseUrl, model, configDir, tokenEnv). Rather than break existing YAML, CB-401 keeps workers: exactly as-is and treats those fields as the ClaudeCodeLauncher's profile schema. A future peer kind adds a kind: discriminator (default "claude-code") selecting the launcher; unknown-kind → clear config error. No migration of existing configs. (Jackson already ignores unknown keys, so adding kind is backward-safe.)

sequenceDiagram
    participant MCP as fleet_spawn (MCP/REST)
    participant SM as SessionManager
    participant L as PeerLauncher (by profile.kind)
    participant T as transport (herdr)
    MCP->>SM: acquire(profile, cwd, owner)
    SM->>L: spawn(SpawnRequest)
    L->>L: build env + guard + argv (adapter-private)
    L->>T: start(name, argv, env, cwd)
    T-->>L: handle (paneId…)
    L-->>SM: PeerHandle(id)
    SM->>SM: register session keyed by handle.id()
    SM-->>MCP: session view

Figure 3 — spawn delegation. The core's acquire is unchanged in shape; only the thing it calls becomes an interface.

6. Capabilities

Peers are not uniform. Let each launcher declare a capability set; the protocol is the union and degrades gracefully when a launcher lacks one:

Capability	Meaning	Claude Code	Codex (likely)	Human
MID_TURN_ASK	supports fleet_ask rendezvous	✓	?	✗
SELF_PR	can open its own PR at checkpoint (CB-302)	✓ (opt-in token)	?	✗
WORKTREE	can run in a provisioned git worktree	✓	✓	✗
ORPHAN_REAP	spawner can reconcile orphaned peers on boot	✓	?	✗

A verb invoked against a peer that lacks the capability returns a clean "unsupported for this peer" rather than a crash. This keeps the protocol honest as peers diversify and prevents the core from assuming "every peer is a Claude in a worktree" (the drift signal from the identity note).

7. Staging & the Stage-C security gate

flowchart TD
    A["Stage A — CB-401<br/>extract PeerLauncher SPI in-tree<br/>ClaudeCodeLauncher = adapted WorkerService<br/>one impl, config-selected"] --> B["Stage B — CB-402+<br/>2nd in-tree adapter (Codex/human)<br/>proves the SPI held"]
    B --> C["Stage C<br/>dynamic external plugin loading<br/>ServiceLoader / jar discovery"]
    C -.requires.-> G["Trust & capability model<br/>what env/tokens a plugin may inject"]
    classDef gate fill:#b7791f,stroke:#7b341e,color:#ffffff;
    class G gate

Figure 4 — deliver A now; B when a real second peer exists; C only if third parties must ship adapters, and only behind a trust model.

Security note (Stage C, not now): a launcher runs at daemon privilege and touches process spawning and env/token injection into peers — the most sensitive path in the system. "Extra plugins" must mean first-party, in-tree, config-selected for the foreseeable future. A third-party plugin that can inject env into a peer needs a genuine trust/capability model before it may exist. Regardless of stage, the primary remains the merge/verify gate — self-reports over the bus are messages, not verified facts.

8. Stage A scope (this branch, delegation-ready)

Deliverable for CB-401 Stage A — mechanical, behaviour-preserving:

1. PeerLauncher interface + PeerHandle (opaque id) + SpawnRequest (profile, requestedCwd, callerCwd) + Capability enum, new package dev.ltms.fleet.peer.
2. ClaudeCodeLauncher implements PeerLauncher = today's WorkerService, adapted: spawn(...) returns a PeerHandle (id = paneId), capabilities() declares MID_TURN_ASK, SELF_PR(when token), WORKTREE, ORPHAN_REAP.
3. SessionManager depends on PeerLauncher, not WorkerService concretely; routing keys on PeerHandle.id() (== paneId today, so zero value change).
4. Fleetd.main wires the concrete ClaudeCodeLauncher behind the interface.
5. No behaviour change, no config change. Full green gate: ide_sync → ide_diagnostics (0 errors/0 warnings) → mvn clean install with MVN_EXIT captured (no masking pipe). All existing tests pass unchanged; add tests only for the new PeerHandle indirection.

Explicitly out of scope for Stage A: kind: config discriminator, any second adapter, capability enforcement at the verb layer (declare only), touching inject/ turn detection, dynamic loading.