verify

name	verify
description	Build, launch, and drive prism's real surfaces (CLI, MCP stdio server, downstream MCP bridge, benchmarks) to verify changes end-to-end.

Verifying prism changes

Build

go build -o ./prism ./cmd/prism

Runtime environment gotchas

• Model runtime config comes from ~/.prism/config.env (points at a homelab SGLang server, sglang.barton.local, serving Qwen2.5-Coder-14B). Explicit env vars override it: PRISM_MODEL_RUNTIME_ENGINE=ollama PRISM_MODEL_RUNTIME_BASE_URL=http://127.0.0.1:11434 PRISM_MODEL_RUNTIME_MODEL=qwen3.5:9b
• Local Ollama is usually up (./prism config doctor); qwen3.5:9b supports native tool calling. The SGLang endpoint returns tool calls as fenced JSON text (no --tool-call-parser), so use Ollama to exercise the tool loop.
• Always pass --state-dir <tmpdir> to avoid mutating ~/.prism.
• Agent latency budgets (~45s) can trip on Ollama cold model loads (per-agent num_ctx differences force reloads). A status: timeout on the first run of a new agent is usually cold load — retry once warm.
• macOS: no timeout command; use the Bash tool timeout instead.

Drive the surfaces

CLI basics: ./prism agent list, ./prism agent show github-cli, ./prism config doctor.

Benchmarks (fixtures, mock server, runner, parse): ./prism benchmark run --mock --mock-delay-ms 1 --json (determinism: hash 3 runs).

MCP stdio server — pace the input or responses never flush:

# session.jsonl: initialize, notifications/initialized, then tools/list / tools/call
(while IFS= read -r l; do printf '%s\n' "$l"; sleep 0.4; done < session.jsonl; sleep 3) \
  | ./prism --state-dir "$STATE" mcp serve 2>/dev/null > out.jsonl

Downstream MCP bridge — register prism itself as a real downstream server:

./prism --state-dir "$STATE" mcp server add-command prism-self "$PWD/prism" mcp serve
./prism --state-dir "$STATE" mcp server tools prism-self

Note: add-command stops flag parsing at the first positional arg — put --timeout-ms/--max-bytes BEFORE <name> <command> or they become args to the command. Edit $STATE/mcp-servers.yaml directly when in doubt.

Unresponsive-downstream simulation: /usr/bin/tail -f /dev/null (stays alive, never answers) → should fail with context deadline exceeded at timeout_ms.

Live tool loop end-to-end (real model, real downstream):

echo "Use call_mcp_tool to call 'list_agents' on server 'prism-self'..." \
  | env PRISM_MODEL_RUNTIME_ENGINE=ollama PRISM_MODEL_RUNTIME_BASE_URL=http://127.0.0.1:11434 \
        PRISM_MODEL_RUNTIME_MODEL=qwen3.5:9b \
    ./prism --state-dir "$STATE" run linear --skills linear-issue-management --stdin

Check artifacts[] for mcp_tool_call entries to confirm the model actually made the call rather than answering from the evidence block.