From the checkout, agents read its AGENTS.md and docs: Hermes concluded
from docs/IDX_NEWS.md that news "is not implemented" and failed the news
case it had passed before the doc existed. Users' agents never see our
repo, so run every agent from an empty directory under the work dir.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The dev shell's bash has no compgen (no programmable completion), so the
npx-install check failed every agent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
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
opencode 2 (@opencode/cli) moved MCP servers to `mcp.servers`, turns
Code Mode on by default, replaced `permission` with ordered
`permissions` rules, and dropped `run --pure` for `--standalone` (runs
otherwise go through a shared background service). Pick the config and
flag from `opencode --version`.
Its `execute` tool stays allowed: with it denied, opencode 2.0.30 never
offers the idx tools to the model in `opencode run`, though the server
connects before the first step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
opencode kept its built-in tools, so it could route around the server
under test: on published 0.3.1 it answered a quote from web search, and
on the live `changes` case it queried the ownership SQLite file with
bash and timed out. Deny every tool but `idx_*`, as Claude, pi, and
Hermes already are.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The `changes` case names the fixture releases (2026-01-30, 2026-02-27),
which the live snapshot does not have, so under --live/--npx it failed
agents that correctly declined to answer about missing releases and
passed ones that called the tool anyway. A case's `live` fields now
replace the mock ones; this one asks about the two most recent releases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
All agents shared one npm cache, so only the first ran a cold
`npx -y idx-mcp` install and the others reused its installed binary,
never exercising their own first-run install.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
- 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
Findings from a Codex (gpt-6.1-sol, xhigh) review of v0.2.4..HEAD, each
verified in the code before fixing:
- high: ticker symbols reached cache file paths unchecked, so a symbol
like `../../REPORT` read (and deleted, if unparseable) .json files
outside the cache, reachable through MCP get_quotes. resolve_symbol now
accepts only ticker characters, and cache keys are encoded as single
safe path components.
- high: sync decided before downloading and replaced the database
without rechecking; an import that landed meanwhile was erased. It now
rechecks the local state before installing, and idx-mcp runs
ownership_sync exclusively with ownership reads.
- replaces_sqlite_sha256 never downgrades history: the client requires
the rebuild to keep at least the local release count, and the
publisher lists a replaced build only then.
- cache writes are atomic (temp file + rename), so concurrent readers
never see, and delete, a partial entry.
- idx-mcp caps symbols (20, deduplicated) and limits, and runs at most 4
tool calls at once.
- each Yahoo client gets its own cookie jar, paired with its crumb.
- ZIP entries (XLSX, KSEI archives) are read with a 256 MiB bound;
non-ASCII archive dates and huge percentages are parse errors instead
of panics or overflow.
- agent-smoke --keep deletes every copied agent config, not only auth
files.
CLI goldens unchanged; the MCP tool golden changes only to state the
new caps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Require bash 4.4+ up front, and match answer facts with jq regexes
(Oniguruma, (?i) everywhere) instead of GNU-only grep -P.
- Unparseable agent output fails that run instead of aborting the
matrix; check_run now also reports unparseable tool calls, which
previously passed the tool check silently.
- Copy the Hermes model block with awk up to the next top-level key, so
a final or differently ordered block is handled.
- --keep keeps run output but deletes the copied agent credentials.
- docs: full path to the tool golden.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- tests/agents/agent-smoke.sh runs tests/agents/cases.json through
Claude Code, Codex, pi, Hermes Agent, and opencode headless against a
local idx-mcp build, each with an isolated config, and checks the
tools called, their key arguments, and (mock data) facts in the
answer. Opt-in; never part of the gate.
- crates/idx-mcp/tests/golden/tools.json pins tool names, descriptions,
and input schemas, so changes to what agents see are reviewed.
- nix develop .#agents adds Claude Code, Codex, and opencode from
nixpkgs after the host PATH.
First run: 30/30 (5 agents x 6 cases, mock data).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
