test(agents): run agents through npx -y idx-mcp, as users configure it

- agent-smoke `--npx <spec>` registers `npx -y <spec>` (the published package
  or an `npm pack` tarball) with a fresh npm cache, instead of the local build,
  so each agent's own npx handling and a cold install are tested.
- `nix develop .#agents-fhs` (Linux) runs agents in an FHS sandbox: Hermes
  replaces a bare `npx` with its own generic-Linux Node, which NixOS cannot
  load. `agent-fhs <cmd>` runs anything inside it.
- Hermes: only the idx toolset (with its terminal it repaired a broken
  server mid-run and passed), the Node toolchain fetched before the first
  case (`hermes -z` waits ~15s for MCP tools), JSON config.yaml (as profiles
  write it), linked plugins for plugin model providers, and a copy of its
  mcp-stderr.log per run.
- A failed run names its cause inline, since output is deleted without --keep.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WB13omP6h7YeyJx7GNtUnF

docs/SMOKE.md

@@ -86,8 +86,19 @@ and agent behavior varies between runs, so it is never part of the gate.
 nix develop .#agents --command tests/agents/agent-smoke.sh            # every agent found
 nix develop .#agents --command tests/agents/agent-smoke.sh --agents claude,codex --runs 3
 tests/agents/agent-smoke.sh --live --case shareholders --keep          # real data
+tests/agents/agent-smoke.sh --npx idx-mcp                              # the published npm package
+tests/agents/agent-smoke.sh --npx "$(cd npm/idx-mcp && realpath "$(npm pack --silent)")"  # this checkout's wrapper
 ```
 
+`--npx <spec>` registers `npx -y <spec>` the way users configure it, so each
+agent's own `npx` handling and a cold npm install are part of the test. It
+runs the release binary, so it implies `--live`; `IDX_BINARY_URL` swaps in a
+local one. Hermes Agent replaces a bare `npx` with its own generic-Linux
+Node, which NixOS cannot load; on NixOS use `nix develop .#agents-fhs`, which
+runs every agent in an FHS sandbox (and `agent-fhs hermes` for an interactive
+session). Hermes logs a server's stderr to `$HERMES_HOME/logs/mcp-stderr.log`;
+each run keeps a copy next to its output as `<case>.<run>.mcp-stderr`.
+
 - Each agent gets an isolated config that registers only the local server;
   your own agent settings and MCP servers are not loaded or changed. Logins
   are copied from the agent's home (Codex `auth.json`, pi and Hermes auth
@@ -95,7 +106,8 @@ tests/agents/agent-smoke.sh --live --case shareholders --keep          # real da
   output but still deletes every copied agent config.
 - Tools are auto-approved for the run (Claude `dontAsk` with
   `mcp__idx__*`, Codex `default_tools_approval_mode="approve"`); Claude and pi
-  also get no built-in tools, so they can only answer through MCP.
+  also get no built-in tools, and Hermes only the `idx` toolset, so they
+  can only answer through MCP.
 - `IDX_AGENT_MODEL_<AGENT>` picks a model (for example
   `IDX_AGENT_MODEL_CLAUDE=sonnet`). `IDX_AGENT_EXEC_PREFIX` prefixes every
   agent command, for hosts that inject API keys through a wrapper.

flake.nix

@@ -160,8 +160,27 @@
             ./scripts/build-ownership-snapshot.sh
           ];
         };
+        # `nix develop .#agents-fhs` (Linux): the agents shell, with every agent
+        # run inside an FHS sandbox. Hermes runs a bare `npx` with its own
+        # generic-Linux Node, which NixOS cannot load, so testing the
+        # `npx -y idx-mcp` setup users have needs this on NixOS.
+        # `agent-fhs <cmd>` runs any command (e.g. an interactive `hermes`) inside it.
+        agentsFhsShell =
+          let
+            agentFhs = pkgs.buildFHSEnv {
+              name = "agent-fhs";
+              targetPkgs = p: [ p.stdenv.cc.cc.lib p.zlib p.openssl ];
+              runScript = pkgs.writeShellScript "agent-fhs-run" ''exec "''${@:-bash}"'';
+            };
           in
-      {
+          self.devShells.${system}.agents.overrideAttrs (old: {
+            shellHook = (old.shellHook or "") + ''
+              export PATH="${agentFhs}/bin:$PATH"
+              export IDX_AGENT_EXEC_PREFIX="''${IDX_AGENT_EXEC_PREFIX:+$IDX_AGENT_EXEC_PREFIX }agent-fhs"
+            '';
+          });
+      in
+      lib.recursiveUpdate {
         packages.default = idxPackage;
         packages.idx-mcp = mcpPackage;
         packages.ownership-publisher = ownershipPublisher;
@@ -352,6 +371,6 @@
             RUST_BACKTRACE = "1";
           };
         };
-      }
+      } (lib.optionalAttrs pkgs.stdenv.isLinux { devShells.agents-fhs = agentsFhsShell; })
     );
 }

tests/agents/agent-smoke.sh

@@ -20,6 +20,9 @@ Options:
   --runs <n>        Runs per agent and case (default: 1)
   --live            Real upstream data and the published ownership snapshot
                     instead of mock fixtures; answer text is not checked
+  --npx <spec>      Register `npx -y <spec>` as users do, instead of the local
+                    build: `idx-mcp` (published) or an `npm pack` tarball.
+                    Runs the release binary, so it implies --live
   --keep            Keep the work directory (raw agent output, results.jsonl)
   --help            Show this help
 
@@ -29,6 +32,8 @@ Environment:
   IDX_AGENT_MODEL_<AGENT> Model override, e.g. IDX_AGENT_MODEL_CLAUDE=sonnet,
                           IDX_AGENT_MODEL_OPENCODE=opencode-go/glm-5.3
   IDX_AGENT_TIMEOUT       Seconds per agent run (default: 300)
+  IDX_BINARY_URL          With --npx: binary the npm package installs instead
+                          of the release download (see docs/NPM_DISTRIBUTION.md)
 
 Exit status: 0 when every agent passed every case at least once, 1 when some
 agent failed a case in all runs, 2 on usage or setup errors.
@@ -43,6 +48,7 @@ only_cases=()
 runs=1
 live=0
 keep=0
+npx_spec=
 timeout_secs="${IDX_AGENT_TIMEOUT:-300}"
 read -r -a exec_prefix <<< "${IDX_AGENT_EXEC_PREFIX:-}"
 
@@ -52,6 +58,7 @@ while [[ $# -gt 0 ]]; do
         --case) only_cases+=("${2:?}"); shift 2 ;;
         --runs) runs="${2:?}"; shift 2 ;;
         --live) live=1; shift ;;
+        --npx) npx_spec="${2:?}"; live=1; shift 2 ;;
         --keep) keep=1; shift ;;
         --help|-h) usage; exit 0 ;;
         *) usage >&2; exit 2 ;;
@@ -63,7 +70,7 @@ if ((BASH_VERSINFO[0] < 4 || (BASH_VERSINFO[0] == 4 && BASH_VERSINFO[1] < 4)));
     echo "agent-smoke: bash 4.4+ is required (run it inside nix develop)" >&2
     exit 2
 fi
-for tool in jq cargo timeout; do
+for tool in jq cargo timeout ${npx_spec:+npx}; do
     command -v "$tool" >/dev/null || { echo "agent-smoke: $tool is required" >&2; exit 2; }
 done
 
@@ -118,16 +125,36 @@ else
     done
 fi
 
-# One launcher for every agent: isolated data dirs, and the idx-core crate
-# as working directory (the mock provider reads tests/fixtures from there).
-launcher="$work/idx-mcp"
-{
+# One server definition for every agent, as {command, args, env}.
+if [[ -n "$npx_spec" ]]; then
+    # What users configure, so each agent's own npx handling is exercised
+    # (Hermes swaps in its managed Node). A fresh npm cache makes the first
+    # agent run a cold install; agents may filter env, so pass it explicitly.
+    export npm_config_cache="$work/npm-cache"
+    server_env=("${data_env[@]}" "npm_config_cache=$npm_config_cache")
+    [[ -z "${IDX_BINARY_URL:-}" ]] || server_env+=("IDX_BINARY_URL=$IDX_BINARY_URL")
+    # npx runs a tarball only as a package to install: `npx -y -p file:<tgz> idx-mcp`.
+    if [[ "$npx_spec" == *.tgz && ! -f "$npx_spec" ]]; then
+        echo "agent-smoke: tarball not found: $npx_spec" >&2; exit 2
+    fi
+    npx_args=(-y "$npx_spec")
+    [[ ! -f "$npx_spec" ]] || npx_args=(-y -p "file:$(realpath "$npx_spec")" idx-mcp)
+    server_json="$(jq -nc --argjson args "$(printf '%s\n' "${npx_args[@]}" | jq -Rnc '[inputs]')" '{command: "npx", args: $args,
+        env: ($ARGS.positional | map(capture("^(?<key>[^=]+)=(?<value>.*)$")) | from_entries)}' \
+        --args "${server_env[@]}")"
+else
+    # A launcher with isolated data dirs, and the idx-core crate as working
+    # directory (the mock provider reads tests/fixtures from there).
+    launcher="$work/idx-mcp"
+    {
         echo '#!/usr/bin/env bash'
         printf 'export %q\n' "${data_env[@]}"
         printf 'cd %q\n' "$repo_root/crates/idx-core"
         printf 'exec %q "$@"\n' "$repo_root/target/debug/idx-mcp"
-} > "$launcher"
-chmod +x "$launcher"
+    } > "$launcher"
+    chmod +x "$launcher"
+    server_json="$(jq -nc --arg cmd "$launcher" '{command: $cmd, args: [], env: {}}')"
+fi
 
 # --- per-agent setup, run, and tool-call extraction ---------------------------
 # Each run_<agent> writes raw output to $out.raw, the answer to $out.answer, and
@@ -137,7 +164,7 @@ chmod +x "$launcher"
 copy_if_present() { [[ -f "$1" ]] && cp "$1" "$2/"; return 0; }
 
 setup_claude() {
-    printf '{"mcpServers":{"idx":{"type":"stdio","command":"%s","args":[]}}}\n' "$launcher" > "$work/claude-mcp.json"
+    jq -n --argjson s "$server_json" '{mcpServers: {idx: ({type: "stdio"} + $s)}}' > "$work/claude-mcp.json"
 }
 run_claude() {
     local prompt="$1" out="$2" model=()
@@ -158,11 +185,15 @@ setup_codex() {
     copy_if_present "${CODEX_HOME:-$HOME/.codex}/auth.json" "$work/codex"
 }
 run_codex() {
-    local prompt="$1" out="$2" model=()
+    local prompt="$1" out="$2" model=() server=()
     [[ -n "${IDX_AGENT_MODEL_CODEX:-}" ]] && model=(-m "$IDX_AGENT_MODEL_CODEX")
+    # -c values are TOML; JSON strings and arrays of them are valid TOML.
+    mapfile -t server < <(jq -r '"mcp_servers.idx.command=\(.command | tojson)",
+        "mcp_servers.idx.args=\(.args | tojson)",
+        "mcp_servers.idx.env={\(.env | to_entries | map("\(.key)=\(.value | tojson)") | join(","))}"' <<< "$server_json")
     CODEX_HOME="$work/codex" "${exec_prefix[@]}" codex exec --json --ephemeral \
         --skip-git-repo-check -s read-only "${model[@]}" \
-        -c "mcp_servers.idx.command=\"$launcher\"" -c 'mcp_servers.idx.args=[]' \
+        -c "${server[0]}" -c "${server[1]}" -c "${server[2]}" \
         -c 'mcp_servers.idx.default_tools_approval_mode="approve"' \
         "$prompt" < /dev/null > "$out.raw"
     jq -c 'select(.type=="item.completed" and .item.type=="mcp_tool_call" and .item.server=="idx")
@@ -175,7 +206,7 @@ setup_pi() {
     mkdir -p "$work/pi"
     for file in auth.json models.json settings.json; do copy_if_present "$src/$file" "$work/pi"; done
     # "direct" exposes each MCP tool by name (the default "codemode" wraps them).
-    jq -n --arg cmd "$launcher" '{mcpServers: {idx: {command: $cmd, exposure: "direct"}}}' > "$work/pi/mcp.json"
+    jq -n --argjson s "$server_json" '{mcpServers: {idx: ($s + {exposure: "direct"})}}' > "$work/pi/mcp.json"
 }
 run_pi() {
     local prompt="$1" out="$2" model=()
@@ -193,17 +224,42 @@ setup_hermes() {
     mkdir -p "$work/hermes"
     copy_if_present "$src/auth.json" "$work/hermes"
     copy_if_present "$src/.env" "$work/hermes"
+    # Model providers can be user plugins. A symlink: cleanup removes only the link.
+    [[ ! -d "$src/plugins" ]] || ln -s "$src/plugins" "$work/hermes/plugins"
+    local server
+    server="$(jq -c '. + {enabled: true}' <<< "$server_json")"
+    if jq -e 'type == "object"' "$src/config.yaml" >/dev/null 2>&1; then
+        # Hermes also writes config.yaml as JSON (e.g. in profiles).
+        jq --argjson s "$server" '{model, mcp_servers: {idx: $s}} | del(.. | nulls)' \
+            "$src/config.yaml" > "$work/hermes/config.yaml"
+    else
         {
             # The top-level `model:` block, up to the next top-level key.
             [[ -f "$src/config.yaml" ]] && awk '/^[^[:space:]#]/ { in_model = /^model:/ } in_model' "$src/config.yaml"
-        printf 'mcp_servers:\n  idx:\n    command: %s\n    enabled: true\n' "$launcher"
+            # JSON is valid YAML.
+            printf 'mcp_servers:\n  idx: %s\n' "$server"
         } > "$work/hermes/config.yaml"
+    fi
+    # A bare `npx` runs on Hermes' own Node, fetched into a fresh HERMES_HOME
+    # on first use (~20s); `hermes -z` waits only ~15s for MCP tools, so fetch
+    # it first or the first case runs without them.
+    if [[ -n "$npx_spec" ]]; then
+        HERMES_HOME="$work/hermes" "${exec_prefix[@]}" hermes pm install npm >/dev/null \
+            || { echo "agent-smoke: hermes pm install npm failed" >&2; exit 2; }
+    fi
 }
 run_hermes() {
-    local prompt="$1" out="$2" model=() session
+    local prompt="$1" out="$2" model=() session status=0
     [[ -n "${IDX_AGENT_MODEL_HERMES:-}" ]] && model=(-m "$IDX_AGENT_MODEL_HERMES")
-    HERMES_HOME="$work/hermes" "${exec_prefix[@]}" hermes -z "$prompt" "${model[@]}" \
-        --usage-file "$out.usage.json" < /dev/null > "$out.answer"
+    # Only the idx toolset: with its terminal, Hermes can debug and repair a
+    # broken server mid-run and pass for the wrong reason.
+    HERMES_HOME="$work/hermes" "${exec_prefix[@]}" hermes -z "$prompt" "${model[@]}" -t idx \
+        --usage-file "$out.usage.json" < /dev/null > "$out.answer" || status=$?
+    # Hermes reports a dead server only as "Connection closed"; its stderr is here.
+    if [[ -f "$work/hermes/logs/mcp-stderr.log" ]]; then
+        cp "$work/hermes/logs/mcp-stderr.log" "$out.mcp-stderr"
+    fi
+    ((status == 0)) || return "$status"
     # -z prints only the answer; tool calls come from the session export.
     session="$(jq -r '.session_id // empty' "$out.usage.json")"
     [[ -n "$session" ]] || return 1
@@ -219,8 +275,9 @@ run_hermes() {
 
 setup_opencode() {
     mkdir -p "$work/opencode/cwd"
-    jq -n --arg cmd "$launcher" \
-        '{"$schema": "https://opencode.ai/config.json", mcp: {idx: {type: "local", command: [$cmd], enabled: true}}}' \
+    jq -n --argjson s "$server_json" \
+        '{"$schema": "https://opencode.ai/config.json",
+          mcp: {idx: {type: "local", command: ([$s.command] + $s.args), environment: $s.env, enabled: true}}}' \
         > "$work/opencode/cwd/opencode.json"
 }
 run_opencode() {
@@ -284,12 +341,15 @@ for agent in "${agents[@]}"; do
             mkdir -p "$(dirname "$out")"
             : > "$out.calls"; : > "$out.answer"
             started=$SECONDS
-            if timeout "$timeout_secs" bash -c "set -eo pipefail; $(declare -f "run_$agent"); $(declare -p work launcher exec_prefix); run_$agent \"\$1\" \"\$2\"" _ "$prompt" "$out" 2> "$out.stderr"; then
+            if timeout "$timeout_secs" bash -c "set -eo pipefail; $(declare -f "run_$agent"); $(declare -p work server_json exec_prefix); run_$agent \"\$1\" \"\$2\"" _ "$prompt" "$out" 2> "$out.stderr"; then
                 # Malformed agent output fails this run, not the whole matrix.
                 failures="$(check_run "$case_json" "$out")" \
                     || failures="${failures:+$failures$'\n'}could not check the agent output (see $out.raw)"
             else
-                failures="agent exited with status $? (see $out.stderr)"
+                status=$?
+                # Name the cause here: without --keep the output is deleted.
+                reason="$(grep -m1 -v '^[[:space:]]*$' "$out.answer" || tail -n1 "$out.stderr" || true)"
+                failures="agent exited with status $status${reason:+: ${reason:0:300}}"
             fi
             status=pass
             [[ -z "$failures" ]] || status=fail