mirror of
https://github.com/0xrsydn/idx-cli.git
synced 2026-11-07 09:45:36 +00:00
687c9ad83f
From a Codex review of the harness: - Drop target/debug from PATH in --npx mode (the dev shell adds it), so a missing package bin cannot fall through to the local build, and fail an agent whose npm cache holds no installed idx-mcp binary. - The live `changes` case must diff the two latest releases, read from the synced snapshot, not any two dates. - Hermes: `hermes -z` waits ~15s for MCP tools regardless of connect_timeout, so a slow cold install lost the first case. Warm up with `hermes mcp test idx`, which installs through Hermes' own npx and honors connect_timeout (now 300s). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
432 lines
22 KiB
Bash
Executable file
432 lines
22 KiB
Bash
Executable file
#!/usr/bin/env bash
|
|
# Agent smoke test: runs real coding agents headless against a local idx-mcp
|
|
# build and checks which MCP tools they call. Opt-in: it needs agent logins
|
|
# and model credits, and agent behavior is nondeterministic. It is not part
|
|
# of `nix flake check`. See docs/SMOKE.md.
|
|
set -Eeuo pipefail
|
|
|
|
usage() {
|
|
cat <<'EOF'
|
|
Usage: tests/agents/agent-smoke.sh [options]
|
|
|
|
Runs each case in tests/agents/cases.json through each agent, with an
|
|
isolated agent config that registers only the local idx-mcp, and checks the
|
|
tools called (and, with mock data, key facts in the answer).
|
|
|
|
Options:
|
|
--agents <list> Comma-separated: claude,codex,pi,hermes,opencode
|
|
(default: every one found on PATH)
|
|
--case <id> Run only this case (repeatable)
|
|
--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
|
|
|
|
Environment:
|
|
IDX_AGENT_EXEC_PREFIX Command prefix for every agent, e.g. a secrets
|
|
wrapper: "with-ai-secrets ai-tools"
|
|
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.
|
|
EOF
|
|
}
|
|
|
|
repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
|
|
cases_file="$repo_root/tests/agents/cases.json"
|
|
all_agents=(claude codex pi hermes opencode)
|
|
agents=()
|
|
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:-}"
|
|
|
|
while [[ $# -gt 0 ]]; do
|
|
case "$1" in
|
|
--agents) IFS=, read -r -a agents <<< "${2:?}"; shift 2 ;;
|
|
--case) only_cases+=("${2:?}"); shift 2 ;;
|
|
--runs) runs="${2:?}"; shift 2 ;;
|
|
--live) live=1; shift ;;
|
|
--npx) npx_spec="${2:?}"; live=1; shift 2 ;;
|
|
--keep) keep=1; shift ;;
|
|
--help|-h) usage; exit 0 ;;
|
|
*) usage >&2; exit 2 ;;
|
|
esac
|
|
done
|
|
|
|
[[ "$runs" =~ ^[1-9][0-9]*$ ]] || { echo "--runs must be a positive integer" >&2; exit 2; }
|
|
if ((BASH_VERSINFO[0] < 4 || (BASH_VERSINFO[0] == 4 && BASH_VERSINFO[1] < 4))); then
|
|
echo "agent-smoke: bash 4.4+ is required (run it inside nix develop)" >&2
|
|
exit 2
|
|
fi
|
|
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
|
|
|
|
if [[ ${#agents[@]} -eq 0 ]]; then
|
|
for agent in "${all_agents[@]}"; do
|
|
command -v "$agent" >/dev/null && agents+=("$agent")
|
|
done
|
|
fi
|
|
[[ ${#agents[@]} -gt 0 ]] || { echo "agent-smoke: no agent CLI found on PATH" >&2; exit 2; }
|
|
for agent in "${agents[@]}"; do
|
|
[[ " ${all_agents[*]} " == *" $agent "* ]] || { echo "unknown agent: $agent" >&2; exit 2; }
|
|
command -v "$agent" >/dev/null || { echo "agent-smoke: $agent not found on PATH" >&2; exit 2; }
|
|
done
|
|
|
|
work="$(mktemp -d "${TMPDIR:-/tmp}/idx-agent-smoke.XXXXXX")"
|
|
cleanup() {
|
|
if [[ "$keep" == 1 ]]; then
|
|
# Never leave copied agent logins or settings (which may hold API keys)
|
|
# behind; each run's raw output is already under runs/.
|
|
rm -rf -- "$work/codex" "$work/pi" "$work/hermes" "$work/opencode"
|
|
echo "agent-smoke: kept $work (copied agent configs removed)"
|
|
elif [[ "$work" == */idx-agent-smoke.* ]]; then
|
|
rm -rf -- "$work"
|
|
fi
|
|
}
|
|
trap cleanup EXIT
|
|
|
|
# --- idx-mcp under test ----------------------------------------------------
|
|
|
|
echo "agent-smoke: building idx-mcp ($([[ $live == 1 ]] && echo live || echo mock) data)"
|
|
features=()
|
|
[[ "$live" == 1 ]] || features=(--features idx-core/testing)
|
|
(cd "$repo_root" && cargo build -q --locked -p idx-cli -p idx-mcp "${features[@]}")
|
|
idx="$repo_root/target/debug/idx"
|
|
|
|
data_env=(
|
|
"XDG_CONFIG_HOME=$work/idx/config"
|
|
"XDG_CACHE_HOME=$work/idx/cache"
|
|
"XDG_DATA_HOME=$work/idx/data"
|
|
)
|
|
latest_release='' previous_release=''
|
|
if [[ "$live" == 1 ]]; then
|
|
env "${data_env[@]}" "$idx" -o json ownership sync >/dev/null
|
|
# Live cases name releases as {latest_release} and {previous_release}.
|
|
mapfile -t releases < <(env "${data_env[@]}" "$idx" -o json ownership releases | jq -r '.[].as_of_date')
|
|
latest_release="${releases[0]:-}" previous_release="${releases[1]:-}"
|
|
else
|
|
data_env+=(IDX_USE_MOCK_PROVIDER=1)
|
|
fixtures="$repo_root/crates/idx-core/tests/fixtures"
|
|
# The importer derives the release from the IDX file name.
|
|
cp "$fixtures/ksei_above1_20260831_excerpt.xlsx" "$work/peng-2026-08-00017-satu-persen.xlsx"
|
|
for file in "$work/peng-2026-08-00017-satu-persen.xlsx" \
|
|
"$fixtures/ksei_balancepos_20260130_excerpt.txt" \
|
|
"$fixtures/ksei_balancepos_20260227_excerpt.txt"; do
|
|
env "${data_env[@]}" "$idx" -o json ownership import --file "$file" >/dev/null
|
|
done
|
|
fi
|
|
|
|
# One server definition for every agent, as {command, args, env}.
|
|
if [[ -n "$npx_spec" ]]; then
|
|
# What users configure, so each agent's own npx handling is exercised
|
|
# (Hermes swaps in its managed Node). Each agent gets a fresh npm cache
|
|
# below, so its first run is a cold install.
|
|
server_env=("${data_env[@]}")
|
|
[[ -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: $npx_spec" >&2; exit 2
|
|
fi
|
|
# In the dev shell target/debug is on PATH; drop it so a missing package
|
|
# bin cannot fall through to the local build.
|
|
PATH="$(tr ':' '\n' <<< "$PATH" | grep -vxF "$repo_root/target/debug" | paste -sd: -)"
|
|
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 "$launcher"
|
|
server_json="$(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
|
|
# one {"name","args"} JSON object per idx tool call to $out.calls, with names
|
|
# normalized to the bare tool name.
|
|
|
|
copy_if_present() { [[ -f "$1" ]] && cp "$1" "$2/"; return 0; }
|
|
|
|
setup_claude() {
|
|
jq -n --argjson s "$server_json" '{mcpServers: {idx: ({type: "stdio"} + $s)}}' > "$work/claude-mcp.json"
|
|
}
|
|
run_claude() {
|
|
local prompt="$1" out="$2" model=()
|
|
[[ -n "${IDX_AGENT_MODEL_CLAUDE:-}" ]] && model=(--model "$IDX_AGENT_MODEL_CLAUDE")
|
|
"${exec_prefix[@]}" claude -p "$prompt" "${model[@]}" \
|
|
--mcp-config "$work/claude-mcp.json" --strict-mcp-config \
|
|
--setting-sources project --no-session-persistence --tools "" \
|
|
--permission-mode dontAsk --allowedTools "mcp__idx__*" \
|
|
--output-format stream-json --verbose --max-turns 10 < /dev/null > "$out.raw"
|
|
jq -c 'select(.type=="assistant") | .message.content[]? | select(.type=="tool_use")
|
|
| select(.name | startswith("mcp__idx__"))
|
|
| {name: (.name | ltrimstr("mcp__idx__")), args: .input}' "$out.raw" > "$out.calls"
|
|
jq -r 'select(.type=="result") | .result // ""' "$out.raw" > "$out.answer"
|
|
}
|
|
|
|
setup_codex() {
|
|
mkdir -p "$work/codex"
|
|
copy_if_present "${CODEX_HOME:-$HOME/.codex}/auth.json" "$work/codex"
|
|
}
|
|
run_codex() {
|
|
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 "${server[0]}" -c "${server[1]}" -c "${server[2]}" \
|
|
-c 'mcp_servers.idx.default_tools_approval_mode="approve"' \
|
|
"$prompt" < /dev/null > "$out.raw"
|
|
jq -c 'select(.type=="item.completed" and .item.type=="mcp_tool_call" and .item.server=="idx")
|
|
| select(.item.status=="completed") | {name: .item.tool, args: .item.arguments}' "$out.raw" > "$out.calls"
|
|
jq -r 'select(.type=="item.completed" and .item.type=="agent_message") | .item.text' "$out.raw" > "$out.answer"
|
|
}
|
|
|
|
setup_pi() {
|
|
local src="${PI_CODING_AGENT_DIR:-$HOME/.pi/agent}"
|
|
mkdir -p "$work/pi"
|
|
for file in auth.json models.json settings.json; do copy_if_present "$src/$file" "$work/pi"; done
|
|
# "direct" exposes each MCP tool by name (the default "codemode" wraps them).
|
|
jq -n --argjson s "$server_json" '{mcpServers: {idx: ($s + {exposure: "direct"})}}' > "$work/pi/mcp.json"
|
|
}
|
|
run_pi() {
|
|
local prompt="$1" out="$2" model=()
|
|
[[ -n "${IDX_AGENT_MODEL_PI:-}" ]] && model=(--model "$IDX_AGENT_MODEL_PI")
|
|
# No built-in tools, so the agent can only reach the data through MCP.
|
|
PI_CODING_AGENT_DIR="$work/pi" "${exec_prefix[@]}" pi -p --mode json --no-session \
|
|
--no-builtin-tools --no-skills --no-context-files "${model[@]}" "$prompt" < /dev/null > "$out.raw"
|
|
jq -c 'select(.type=="tool_execution_start") | select(.toolName | startswith("mcp__idx__"))
|
|
| {name: (.toolName | ltrimstr("mcp__idx__")), args: .args}' "$out.raw" > "$out.calls"
|
|
jq -r 'select(.type=="agent_end") | .messages[-1].content[]? | select(.type=="text") | .text' "$out.raw" > "$out.answer"
|
|
}
|
|
|
|
setup_hermes() {
|
|
local src="${HERMES_HOME:-$HOME/.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
|
|
# A cold `npx` install can outlast the default connect timeout.
|
|
server="$(jq -c '. + {enabled: true, connect_timeout: 300}' <<< "$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' "$src/config.yaml"
|
|
# JSON is valid YAML.
|
|
printf 'mcp_servers:\n idx: %s\n' "$server"
|
|
} > "$work/hermes/config.yaml"
|
|
fi
|
|
# A bare `npx` runs on Hermes' own Node, fetched into a fresh HERMES_HOME
|
|
# on first use (~20s), and the cold package install and binary download
|
|
# come on top. `hermes -z` waits only ~15s for MCP tools whatever
|
|
# connect_timeout says, so fetch Node and install the server first: `mcp
|
|
# test` starts it through Hermes' own npx and waits for connect_timeout.
|
|
if [[ -n "$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; }
|
|
HERMES_HOME="$work/hermes" "${exec_prefix[@]}" hermes mcp test idx < /dev/null > "$work/hermes-warmup.log" 2>&1 \
|
|
|| echo "agent-smoke: hermes mcp test idx failed (see $work/hermes-warmup.log)" >&2
|
|
fi
|
|
}
|
|
run_hermes() {
|
|
local prompt="$1" out="$2" model=() session status=0
|
|
[[ -n "${IDX_AGENT_MODEL_HERMES:-}" ]] && model=(-m "$IDX_AGENT_MODEL_HERMES")
|
|
# Only the idx toolset: with its terminal, Hermes can debug and repair a
|
|
# broken server mid-run and pass for the wrong reason.
|
|
HERMES_HOME="$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 "$session" ]] || return 1
|
|
HERMES_HOME="$work/hermes" hermes sessions export --format jsonl --session-id "$session" \
|
|
"$out.raw" --yes < /dev/null >/dev/null
|
|
# Hermes may call MCP tools directly or through its tool_call wrapper.
|
|
jq -c '.messages[]? | .tool_calls[]? | .function
|
|
| (.arguments | (try fromjson catch {})) as $args
|
|
| if .name == "tool_call" then ($args.calls[]? | {name, args: .arguments}) else {name, args: $args} end
|
|
| select(.name | startswith("mcp__idx__"))
|
|
| .name |= ltrimstr("mcp__idx__")' "$out.raw" > "$out.calls"
|
|
}
|
|
|
|
setup_opencode() {
|
|
local version
|
|
mkdir -p "$work/opencode/cwd"
|
|
version="$("${exec_prefix[@]}" opencode --version 2>/dev/null | grep -oE '[0-9]+\.[0-9]+\.[0-9]+' | head -n1)"
|
|
if [[ "${version%%.*}" -ge 2 ]]; then
|
|
# OpenCode 2 (@opencode/cli). A denied call ends the run, and models
|
|
# use Code Mode's `execute` while the server connects, so deny only
|
|
# the tools that could route around it. --standalone keeps the run
|
|
# off the shared background service.
|
|
jq -n --argjson s "$server_json" \
|
|
'{mcp: {servers: {idx: {type: "local", command: ([$s.command] + $s.args),
|
|
environment: $s.env, codemode: false}}},
|
|
permissions: ((["read", "edit", "glob", "grep", "shell", "subagent", "skill",
|
|
"question", "webfetch", "websearch", "external_directory"]
|
|
| map({action: ., resource: "*", effect: "deny"}))
|
|
+ [{action: "idx_*", resource: "*", effect: "allow"}])}' \
|
|
> "$work/opencode/cwd/opencode.json"
|
|
echo --standalone > "$work/opencode/run-flag"
|
|
else
|
|
jq -n --argjson s "$server_json" \
|
|
'{"$schema": "https://opencode.ai/config.json",
|
|
mcp: {idx: {type: "local", command: ([$s.command] + $s.args), environment: $s.env, enabled: true}},
|
|
permission: {"*": "deny", "idx_*": "allow"}}' \
|
|
> "$work/opencode/cwd/opencode.json"
|
|
echo --pure > "$work/opencode/run-flag"
|
|
fi
|
|
}
|
|
run_opencode() {
|
|
local prompt="$1" out="$2" model=() home="$work/opencode" flag
|
|
[[ -n "${IDX_AGENT_MODEL_OPENCODE:-}" ]] && model=(-m "$IDX_AGENT_MODEL_OPENCODE")
|
|
flag="$(< "$home/run-flag")"
|
|
(cd "$home/cwd" && XDG_CONFIG_HOME="$home/config" XDG_DATA_HOME="$home/data" \
|
|
XDG_CACHE_HOME="$home/cache" XDG_STATE_HOME="$home/state" \
|
|
"${exec_prefix[@]}" opencode run "$flag" --format json "${model[@]}" "$prompt" < /dev/null > "$out.raw")
|
|
jq -c 'select(.type=="tool_use") | .part | select(.tool | startswith("idx_"))
|
|
| select(.state.status=="completed") | {name: (.tool | ltrimstr("idx_")), args: .state.input}' "$out.raw" > "$out.calls"
|
|
jq -r 'select(.type=="text") | .part.text' "$out.raw" > "$out.answer"
|
|
}
|
|
|
|
# --- checks -----------------------------------------------------------------
|
|
|
|
# Prints failure reasons for one run; no output means it passed.
|
|
check_run() {
|
|
local case_json="$1" out="$2"
|
|
jq -rn --argjson case "$case_json" --slurpfile calls "$out.calls" '
|
|
$case.expect_tools[] as $want
|
|
| if $want.any then
|
|
select([$calls[] | .name] | any(. as $n | $want.any | index($n)) | not)
|
|
| "no call to any of \($want.any | join(", "))"
|
|
else
|
|
([$calls[] | select(.name == $want.name)]) as $matches
|
|
| if ($matches | length) == 0 then "no call to \($want.name)"
|
|
elif ($want.args // {}) as $args | ($matches | any(.args | contains($args))) | not
|
|
then "\($want.name) called without \($want.args | tojson): \($matches | map(.args) | tojson)"
|
|
else empty end
|
|
end' || echo "could not parse the agent's tool calls (see $out.raw)"
|
|
if [[ "$live" == 0 ]]; then
|
|
local pattern
|
|
while IFS= read -r pattern; do
|
|
# jq regexes (Oniguruma) support (?i) on every platform, unlike grep -P.
|
|
jq -Rse --arg re "$pattern" 'test($re)' "$out.answer" >/dev/null \
|
|
|| echo "answer does not match /$pattern/"
|
|
done < <(jq -r '.expect_text[]?' <<< "$case_json")
|
|
fi
|
|
}
|
|
|
|
# --- run --------------------------------------------------------------------
|
|
|
|
mapfile -t case_ids < <(jq -r '.[].id' "$cases_file")
|
|
if [[ ${#only_cases[@]} -gt 0 ]]; then
|
|
for id in "${only_cases[@]}"; do
|
|
[[ " ${case_ids[*]} " == *" $id "* ]] || { echo "unknown case: $id" >&2; exit 2; }
|
|
done
|
|
case_ids=("${only_cases[@]}")
|
|
fi
|
|
|
|
results="$work/results.jsonl"
|
|
: > "$results"
|
|
echo "agent-smoke: agents=${agents[*]} cases=${#case_ids[@]} runs=$runs"
|
|
for agent in "${agents[@]}"; do
|
|
if [[ -n "$npx_spec" ]]; then
|
|
# A shared cache would let later agents reuse the first one's install.
|
|
# Agents may filter env, so pass it explicitly.
|
|
export npm_config_cache="$work/npm-cache/$agent"
|
|
server_json="$(jq -c --arg cache "$npm_config_cache" '.env.npm_config_cache = $cache' <<< "$server_json")"
|
|
fi
|
|
"setup_$agent"
|
|
for id in "${case_ids[@]}"; do
|
|
# A case's "live" fields replace the mock ones, e.g. release dates.
|
|
case_json="$(jq -c --arg id "$id" --argjson live "$live" '.[] | select(.id == $id)
|
|
| if $live == 1 then . + (.live // {}) else . end | del(.live)
|
|
| walk(if type == "string" then gsub("\\{latest_release\\}"; $latest)
|
|
| gsub("\\{previous_release\\}"; $previous) else . end)' \
|
|
--arg latest "$latest_release" --arg previous "$previous_release" "$cases_file")"
|
|
prompt="$(jq -r '.prompt' <<< "$case_json") Use the idx MCP tools."
|
|
for ((run = 1; run <= runs; run++)); do
|
|
out="$work/runs/$agent/$id.$run"
|
|
mkdir -p "$(dirname "$out")"
|
|
: > "$out.calls"; : > "$out.answer"
|
|
started=$SECONDS
|
|
if timeout "$timeout_secs" bash -c "set -eo pipefail; $(declare -f "run_$agent"); $(declare -p work server_json exec_prefix); run_$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 $out.raw)"
|
|
else
|
|
status=$?
|
|
# 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
|
|
jq -nc --arg agent "$agent" --arg case "$id" --argjson run "$run" --arg status "$status" \
|
|
--arg failures "$failures" --argjson secs "$((SECONDS - started))" \
|
|
--slurpfile calls "$out.calls" \
|
|
'{agent: $agent, case: $case, run: $run, status: $status, secs: $secs,
|
|
tools: [$calls[] | .name], failures: ($failures | split("\n") | map(select(. != "")))}' >> "$results"
|
|
printf ' %-9s %-13s run %d %s %3ss %s\n' "$agent" "$id" "$run" "$status" "$((SECONDS - started))" \
|
|
"$(jq -r '[.[] | .name] | join(",")' <(jq -s . "$out.calls"))"
|
|
[[ -z "$failures" ]] || printf ' - %s\n' "${failures//$'\n'/$'\n - '}"
|
|
done
|
|
done
|
|
if [[ -n "$npx_spec" ]]; then
|
|
# Proof that this agent's own npx installed the package binary, rather
|
|
# than the server coming from somewhere else on PATH.
|
|
status=pass failures=
|
|
if ! compgen -G "$npm_config_cache/_npx/*/node_modules/idx-mcp/bin/idx-mcp" > /dev/null; then
|
|
status=fail failures="no idx-mcp binary under $npm_config_cache/_npx"
|
|
fi
|
|
jq -nc --arg agent "$agent" --arg status "$status" --arg failures "$failures" \
|
|
'{agent: $agent, case: "npx-install", run: 1, status: $status, secs: 0, tools: [],
|
|
failures: ([$failures] | map(select(. != "")))}' >> "$results"
|
|
printf ' %-9s %-13s %s\n' "$agent" npx-install "$status"
|
|
[[ -z "$failures" ]] || printf ' - %s\n' "$failures"
|
|
fi
|
|
done
|
|
|
|
echo
|
|
echo "agent-smoke: pass rate per agent and case"
|
|
jq -rs 'group_by(.agent)[] | "\(.[0].agent): " + (group_by(.case) | map("\(.[0].case) \(map(select(.status=="pass")) | length)/\(length)") | join(" "))' "$results"
|
|
consistent_failures="$(jq -s '[group_by([.agent, .case])[] | select(all(.status == "fail"))] | length' "$results")"
|
|
if [[ "$consistent_failures" -gt 0 ]]; then
|
|
echo "agent-smoke: FAIL ($consistent_failures agent/case pairs failed every run)"
|
|
exit 1
|
|
fi
|
|
echo "agent-smoke: PASS"
|