fix: forward CHOIR_AGENT_TOKEN to codex mcp-stdio via env_vars

.choir/config.toml

@@ -6,19 +6,17 @@ verify_cmd = "moon check --target native && moon test --target native"
 # Per-agent model selection (choir-w5m7). Passthrough to `codex --model <value>`.
 # dev (gpt-5.3-codex-spark) PARKED: spark could not drive the agentic TDD-leaf
 # protocol (mis-sequenced Red->Done, false verify-pass, exit-255 build) on first
-# trial. gpt-5.6-luna PARKED too (2026-07-09): every codex dev leaf spawned
-# under luna had its `choir mcp-stdio` MCP subprocess launch with a completely
-# empty CHOIR_* environment (confirmed via /proc/<pid>/environ), unlike gpt-5.5
-# leaves and unlike root's own Claude-Code-spawned mcp-stdio process, which both
-# correctly inherit CHOIR_AGENT_TOKEN etc. Missing token -> registration
-# silently rejected server-side (serve_authenticate_register_request) but
-# mcp_perform_registration_handshake only checks for a non-empty response, not
-# success, so the bridge proceeds as if authenticated and the next real
-# tools/call (tdd_intent et al) then fails opaquely as "server restarting,
-# retry" -- see choir bead for the write-up. Root cause not fully pinned
-# (whether it's luna specifically or the whole 5.6 family under some codex
-# runtime mode), so reverting dev to the known-good gpt-5.5 baseline rather
-# than guessing at terra. Effort is the real quota lever — see choir-mn0l.
+# trial. gpt-5.6-luna was parked on 2026-07-09 blaming it for codex leaves'
+# `choir mcp-stdio` subprocess launching with an empty CHOIR_* environment;
+# that was disproven the same day — gpt-5.5 leaves had identically scrubbed
+# env. Real cause: codex's MCP stdio launcher builds server environments from
+# scratch (env_clear() + DEFAULT_ENV_VARS whitelist in rmcp-client's
+# create_env_for_mcp_server; shell_environment_policy only affects shell/exec
+# tools, not MCP), and choir's codex MCP entry set no per-server env
+# forwarding (choir-ghl9, fixed via
+# `mcp_servers.choir.env_vars` in codex_mcp_override_args). Model was never at
+# fault; re-tiering dev beyond gpt-5.5 is the operator's call once the fix
+# proves out. Effort is the real quota lever — see choir-mn0l.
 [model.codex]
 default = "gpt-5.5"
 

src/workspace/launch.mbt

@@ -154,12 +154,25 @@ pub fn codex_mcp_override_args(
       ]
   }
   let args_lit = toml_string_array_literal(elems)
+  // codex's MCP stdio launcher builds each server's environment from scratch
+  // (`env_clear()` + a DEFAULT_ENV_VARS whitelist in rmcp-client's
+  // `create_env_for_mcp_server`; `shell_environment_policy` only governs
+  // shell/exec tools, not MCP), so `choir mcp-stdio` never inherits CHOIR_*
+  // ambiently (choir-ghl9). `env_vars` is codex's names-only forward
+  // list: values come from codex's own environment at spawn time and never
+  // ride this argv (argv leaks into ps/proc; the token is a secret). Identity
+  // and the socket path come from the per-leaf .choir/config.local.toml in
+  // `cwd` plus the --role/--name args above, so the token is the only
+  // env-sourced input registration needs.
+  let env_vars_lit = toml_string_array_literal(["CHOIR_AGENT_TOKEN"])
   let args = [
     "-c",
     "mcp_servers.choir.command=\"choir\"",
     "-c",
     "mcp_servers.choir.args=" + args_lit,
     "-c",
+    "mcp_servers.choir.env_vars=" + env_vars_lit,
+    "-c",
     "mcp_servers.choir.cwd=\"" + escape_toml_string(workspace_path) + "\"",
   ]
   if semble_pilot {

src/workspace/workspace_test.mbt

@@ -825,6 +825,27 @@ test "codex_mcp_override_args is argv-shaped Toml without sh" {
   assert_false(args[3].contains("-lc"))
 }
 
+///|
+test "codex_mcp_override_args forwards registration env by name only" {
+  for role in [@types.Role::TL, @types.Role::Dev, @types.Role::Worker] {
+    let args = @workspace.codex_mcp_override_args(role, "main.leaf", "/proj")
+    // codex does not inherit the shell environment into MCP stdio
+    // subprocesses; the names-only `env_vars` forward list is the one
+    // mechanism that gets CHOIR_AGENT_TOKEN there without the value.
+    assert_true(
+      args.contains("mcp_servers.choir.env_vars=[\"CHOIR_AGENT_TOKEN\"]"),
+    )
+    for arg in args {
+      // The token value must never ride argv: no `env` value-map entry
+      // (`mcp_servers.choir.env.KEY="..."` / `env={...}`) and no
+      // KEY=value assignment for the token anywhere.
+      assert_false(arg.contains("mcp_servers.choir.env."))
+      assert_false(arg.contains("mcp_servers.choir.env="))
+      assert_false(arg.contains("CHOIR_AGENT_TOKEN="))
+    }
+  }
+}
+
 ///|
 test "codex_mcp_override_args includes semble server when pilot is enabled" {
   let args = @workspace.codex_mcp_override_args(