Skip to content

Run 2026-05-04 — first baseline campaign

Verdict: ✅ HARNESS-PIPELINE GREEN (LLM cells pending API credentials) Captured against: ai-memory-mcp v0.6.4 release binary @ feat/v0.6.4 HEAD Repo: alphaonedev/ai-memory-discovery-gate

What this run covers

The 2026-05-04 baseline run is the harness-pipeline smoke test — it validates every step of the discovery-gate runner except the LLM call itself. Cells with llm=harness-pipeline-smoke substitute a deterministic MCP-stdio simulation for the model API call, then score the result against the same pass criteria a real LLM run would.

A green smoke run proves:

  • The v0.6.3.1 baseline fixture restores cleanly
  • The v0.6.4 binary opens it, migrates schema v19 → v20, populates audit_log, preserves all rows
  • MCP stdio loop drives the canonical T1 first-call sequence (initialize → tools/list → memory_capabilities)
  • The capabilities response contains the v0.6.4 families block with all 8 families
  • The runner's verdict scoring + per-cell markdown emission work

A green smoke run does NOT prove that real LLMs (Claude / GPT / Grok / Gemini) will make those calls. That's what the next run (with API keys) tests.

Cells

Cell Outcome Wall clock Evidence
harness-pipeline-openclaw-t1-awareness-core ✅ PASS see cell json · md

Key signals

  • All 8 families surfaced by memory_capabilities (target ≥6)
  • tools/list returns 6 tools under --profile core (5 core + always-on memory_capabilities) — matches the v0.6.4 contract
  • Schema migration v19 → v20 non-destructive on the fixture (17 memories preserved, audit_log table created)
  • loaded_flags correctly reports only core family loaded under the --profile core profile

Pending

The first LLM-driven baseline campaign requires:

  1. xAI API key (for Grok 4.3 — simplest to wire; matches v0.6.3.1 A2A campaign)
  2. Anthropic API key (for Claude Opus 4.7 + Claude Sonnet 4.6)
  3. OpenAI API key (for GPT-5)

Set these in .env and re-run scripts/smoke-t1-local.sh with the --llm override to fan out into real cells. Multi-LLM × multi-harness × all-tier matrix runs land at docs/runs/<future-date>/.

Reproducibility

cd ai-memory-discovery-gate
DISCOVERY_GATE_BINARY=../ai-memory-mcp/target/release/ai-memory \
    bash scripts/smoke-t1-local.sh

The smoke script is fully self-contained — needs only the v0.6.4 binary and the gate's own fixture. No Docker, no API keys, no daemon stand-up.

Cell verdict (full)

See cells/harness-pipeline-openclaw-t1-awareness-core.json for the structured verdict and cells/harness-pipeline-openclaw-t1-awareness-core.md for the human-readable per-cell page.