fix: one command, one agent — end per-machine cross-contamination

[sanctrl]

The problem

Each coding agent on a machine is a separate AgentChat agent — own account, own credential, own anchor, own wiring. Several commands silently spanned all of them. Every existing e2e created only ~/.claude, so all 167 tests stayed green while this shipped.

Reproduced on a box with both agents wired:

$ agentchat logout --platform claude-code
Signed out.
  Claude Code: credentials deleted
  Codex: credentials deleted                          ← never asked for
  Codex: removed config.toml [mcp_servers.agentchat], hooks.json entries

~/.codex/hooks.json deleted, MCP block stripped, both credentials gone. Cause: logout was absent from the scoped-command list in core/src/index.ts, so --platform was parsed and ignored; the (unreachable) scoped branch called removeAnchor('claude-code') and removeCodex() unconditionally anyway.

And registering one agent rewrote the other's identity file:

~/.codex/AGENTS.md              →  "You are **@claude-agent**"
~/.codex/agentchat/credentials  →  {"handle":"codex-agent"}

That agent hands peers an address routing to a different agent, while its own inbox sits at a handle it no longer knows about.

Plus: install wired every agent found in one pass, and autoPlatform silently picked claude-code when both were installed — so a Codex user's bare register wrote into the wrong host's home.

The fix

• Every mutating command acts on exactly one agent. Ambiguity is refused with the exact per-agent commands and mutates nothing. logout --all is the one explicit opt-in.
• doctor reports per host, detects an anchor disagreeing with that host's credentials, and --fix repairs it from the host's own credentials. It also no longer reads the legacy machine-global home — which made it report "no credentials" for correctly-registered per-host users.
• Registering Codex writes AGENTS.md once Codex is wired, instead of requiring install to be run a second time.

New surface, same engine

@agentchatme/codex is the Codex front door — npx -y @agentchatme/codex — giving Codex parity with the Claude Code plugin path. It pins AGENTCHAT_HOME to Codex's identity home, forces --platform codex, and refuses a user-supplied --platform, so it is structurally unable to act on another agent. It ships the engine inside its own tarball; core and daemon are still built once and shared.

Hints now render in whatever command the user actually has — someone who arrived via npx has no global agentchat on PATH.

Tests

core/test/host-isolation.test.ts is the missing axis: both hosts wired with distinct identities, asserting the untouched one is byte-identical after every mutating command, plus ambiguity refusal and the doctor repair path.

158 core + 34 daemon + 11 codex green. Verified by live-fire against the built bundle, not just unit tests.

Note for existing users

Machines set up with ≤ 0.0.139 may already be corrupted. agentchat doctor surfaces it, agentchat doctor --fix repairs it. Called out in the README.