Multi-architect routing: builder→architect messages misrouted to 'main' when worktree has its own state.db

[waleedkadous]

Summary

Builder→architect messages are silently misrouted to main (instead of the spawning sibling architect) when run inside a builder worktree that has its own .agent-farm/state.db. The affinity primitive shipped in v3.0.5 is bypassed end-to-end for sibling-architect workspaces — every sibling architect (ob-refine in the reporting case) gets cut out of its own builders' messages.

Repro

In a workspace with two architects registered:

• main (terminal fe6b1cd8...)
• ob-refine (terminal 48ad2d87...)

ob-refine spawned builder-bugfix-1599. The builder's .agent-farm/state.db row correctly records:

builder-bugfix-1599 | implementing | spawned_by_architect=ob-refine | terminal=513db860...

The builder then ran afx send architect "..." from inside .builders/bugfix-1599/. Tower log:

[2026-05-19T22:01:13.318Z] [INFO] Message sent: bugfix-1599 → architect (terminal fe6b1cd8...)

Terminal fe6b1cd8... is main, not ob-refine. Wrong terminal.

Root cause

The sender logged is bugfix-1599, not the canonical builder-bugfix-1599. That mismatch breaks the affinity lookup downstream.

detectCurrentBuilderId() (packages/codev/src/agent-farm/commands/send.ts:48) does:

const worktreeDirName = match[1];                 // 'bugfix-1599'
const state = loadState();                        // <-- singleton DB
const builder = state.builders.find(b => /* worktree path match */);
return builder ? builder.id : worktreeDirName;    // fallback when miss

loadState() uses the singleton getDb(), which resolves stateDir via findWorkspaceRoot(). When CWD is .builders/<id>/, findWorkspaceRoot() returns the worktree path, not the workspace path, because the worktree is itself a full git checkout with its own codev/ (config.ts:79). So getDb() opens:

/Users/.../shannon/.builders/bugfix-1599/.agent-farm/state.db

This file exists (created on first DB open) and is empty:

$ sqlite3 .../bugfix-1599/.agent-farm/state.db "SELECT COUNT(*) FROM builders;"
0

So the state.builders.find(...) returns undefined, and detectCurrentBuilderId falls back to the worktree directory name bugfix-1599. Tower receives from='bugfix-1599'.

In tower-messages.ts:318, lookupBuilderSpawningArchitect('bugfix-1599', '/Users/.../shannon') returns undefined because no row exists with id='bugfix-1599' — the row is keyed builder-bugfix-1599. With spawningArchitect === undefined, the routing skips the builder-context branch (line 320-342) and falls through to the non-builder-sender branch at line 344-347:

const terminalId =
  entry.architects.get(DEFAULT_ARCHITECT_NAME) ?? entry.architects.values().next().value!;
return { terminalId, workspacePath, agent: 'architect' };

→ Routes to main. The spawning architect is silently bypassed.

Why this matters

Affinity routing was the headline feature of v3.0.5/v3.0.6 multi-architect. This bug means every sibling-architect workspace loses messages from its builders today. The porch wake-up path (gate-approved, pr-ready notifications) and the builder's own architect-bound updates all hit the wrong terminal. The data is correct in state.db, the dashboard is correct, the in-memory architect registry is correct — only the sender-ID resolution at send time is wrong.

Fix sketch

detectCurrentBuilderId should open the workspace's state.db directly, not the singleton (which resolves to the worktree's). The correct pattern already exists in state.ts:380 (lookupBuilderSpawningArchitect):

const dbPath = path.join(workspacePath, '.agent-farm', 'state.db');
if (!existsSync(dbPath)) return undefined;
const wsDb = new Database(dbPath, { readonly: true });
try {
  // ... lookup ...
} finally {
  wsDb.close();
}

Apply the same pattern in detectCurrentBuilderId: from CWD .builders/<id>/, the workspace root is two levels up; open that state.db (readonly), find the builder by worktree-path match, return its canonical id. The fallback to worktreeDirName should remain as a safety net but should never be hit in practice.

Add a regression test: spawn-fixture, builder in worktree with empty worktree-level state.db, assert detectCurrentBuilderId() returns the canonical builder-<X> ID by reading the workspace's state.db.

Verification

Manual repro on shannon@2026-05-19:

• ob-refine spawned bugfix-1599
• builder's afx send architect → routed to main terminal fe6b1cd8...
• Expected: routed to ob-refine terminal 48ad2d87...

Severity

High — this silently breaks the multi-architect feature for every adopter (Shannon in production). User-visible symptom is "my sibling architect never hears from its builders." No data loss, but the feature is non-functional.

Affected versions

v3.0.5 onward (anything with sibling-architect support).

[waleedkadous]

On it! Working on a fix now.