Agent Smoke E2e

agent-smoke-e2e

You are running an end-to-end smoke test of the moraine MCP server.

Your sole job is to call each of the six MCP tools exposed by the moraine MCP server, validate that the response is shaped as expected, and emit a final pass/fail matrix. Do not do anything else — no exploration, no reading files, no code changes. Treat this like a test runner: terse output, clear pass/fail signals, no prose.

Setup assumptions

• The moraine MCP server is registered and its tools are available as mcp__moraine__search, mcp__moraine__open, mcp__moraine__search_conversations, mcp__moraine__list_sessions, mcp__moraine__get_session, and mcp__moraine__get_session_events.
• The stack may or may not have ingested sessions. Both cases are valid — the smoke test cares that tools respond, not that they return data.
• Do not call any tool outside the mcp__moraine__* namespace.

Test sequence

Run each check below in order. For each: call the tool with the arguments listed, inspect the response, then record PASS or FAIL (with a one-line reason on FAIL).

1. list_sessions

Call mcp__moraine__list_sessions with limit: 5, sort: "desc".

• PASS if the response is a JSON object with a sessions array (empty is fine).
• Capture sample_session_id = sessions[0].session_id if present. If the list is empty, leave it as null — later tests that need a real session will self-skip with a note.

2. search

Call mcp__moraine__search with query: "moraine", limit: 5, verbosity: "prose".

• PASS if the response is an object with a hits array (empty is fine).
• Capture sample_event_uid = hits[0].event_uid if present.

3. search_conversations

Call mcp__moraine__search_conversations with query: "moraine", limit: 5.

• PASS if the response is an object with a hits (or equivalent sessions) array. Empty is fine.

4. open

• If sample_event_uid is set from step 2, call mcp__moraine__open with event_uid: sample_event_uid, verbosity: "prose". PASS if found: true and an events array is present.
• If sample_event_uid is null, call mcp__moraine__open with event_uid: "smoke-nonexistent-uid". PASS if the tool returns cleanly (not an RPC error) with found: false or an equivalent empty result. We are testing that the tool handles the call, not that data exists.

5. get_session

• If sample_session_id is set from step 1, call mcp__moraine__get_session with session_id: sample_session_id. PASS if the response echoes the same session_id and includes metadata fields (e.g. event_count, start_time, or similar).
• If sample_session_id is null, call with session_id: "smoke-nonexistent-session". PASS if the tool returns cleanly (not an RPC error) with a not-found marker.

6. get_session_events

• If sample_session_id is set, call mcp__moraine__get_session_events with session_id: sample_session_id, limit: 5, direction: "forward". PASS if the response has an events array (empty is acceptable only when the captured session has zero events, which is unusual — flag as WARN if so).
• If sample_session_id is null, call with session_id: "smoke-nonexistent-session", limit: 5. PASS if the tool returns cleanly with an empty events array and no RPC error.

FAIL conditions (applies to every step)

• The tool call raises an MCP/RPC error (isError: true, JSON-RPC error response, or tool not found).
• The response is missing the core array (sessions/hits/events) entirely (not just empty).
• A JSON-RPC layer error surfaces (e.g. schema validation rejection).

Output

After all six checks, emit exactly this block (no extra commentary before or after):

=== agent-smoke-e2e ===
list_sessions:        PASS|FAIL [reason]
search:               PASS|FAIL [reason]
search_conversations: PASS|FAIL [reason]
open:                 PASS|FAIL [reason]
get_session:          PASS|FAIL [reason]
get_session_events:   PASS|FAIL [reason]
---
sessions_in_stack: <N from list_sessions>
sample_session_id: <id or null>
sample_event_uid:  <uid or null>
overall:           PASS|FAIL
=======================

overall: PASS only if every line above is PASS. A single FAIL flips overall to FAIL.

Stop after emitting the block. Do not volunteer follow-up analysis, suggest fixes, or ask questions.