コンテンツにスキップ

Wave 47 — Dim W (AX Layer 3) migration PR — STATE

  • Date: 2026-05-12 (Wave 47 Phase 2 永遠ループ tick#4)
  • Dim: W — AX Layer 3 (WebMCP + A2A + Observability) per feedback_ax_4_pillars
  • Branch: feat/jpcite_2026_05_12_wave47_dim_w_migration
  • Worktree: /tmp/jpcite-w47-dim-w-mig (lane claim: /tmp/jpcite-w47-dim-w-mig.lane)
  • Base: origin/main @ cd5b7bbfb
  • PR: filled at push time

Purpose

Storage substrate + 3-axis seed ETL for the Dim W "AX Layer 3" surface — the third (and final) pillar of the AX 4-pillars framework per feedback_ax_4_pillars. Layers 1+2 (Access / Context / Tools / Orchestration discovery + metadata) are already complete; Layer 3 (WebMCP transports + A2A handshakes + Observability metrics) was the only remaining pillar without a persistent substrate.

Three independent axes captured by three small tables:

  • am_webmcp_endpoint — registry of WebMCP transport endpoints advertised by jpcite (HTTP transport variant of MCP for browser / in-page agent embeds). One row per (path, transport) with capability tag binding it to an AX pillar.
  • am_a2a_handshake_log — append-only audit of agent-to-agent handshake attempts (capability negotiation). Drives the AX funnel "Trustability" rollup and debugs interop with partner agents.
  • am_observability_metric — append-only metric stream (Layer 3 observability piece). Pairs metric_name + value + recorded_at; downstream rollups read this for AX-Layer-3 dashboards.

LLM-0 by construction (per feedback_no_operator_llm_api): schema is config + audit metadata only. ZERO columns imply LLM inference (no summary_text, no ai_explanation). The handshake log records WHAT was negotiated, not WHY. All natural-language explanation is rendered customer-side by the customer's own agent.

Deliverables

  • scripts/migrations/283_ax_layer3.sql (~145 LOC including docstring): three tables + 5 indexes + 1 UNIQUE backstop + 2 helper views (v_webmcp_endpoint_active, v_observability_recent). All CHECK constraints enumerated (transport enum, capability_tag enum, path-must-be-rooted, handshake exclusive state CHECK, temporal ordering CHECK on succeeded_at >= initiated_at).
  • scripts/migrations/283_ax_layer3_rollback.sql (~25 LOC): drops every table / index / view created.
  • scripts/etl/seed_ax_layer3.py (~205 LOC): seeds 3 WebMCP endpoints + 2 A2A handshake templates + 8 observability metrics on a fresh autonomath.db. Idempotent under daily re-run (UNIQUE + date-of-day dedup); --force re-appends audit rows.
  • tests/test_dim_w_ax_layer3.py (~310 LOC): 20 tests covering apply / idempotency / rollback / 7 CHECK constraints / ETL dry-run / apply-then-idempotent / --force audit append / both helper views / boot manifest registration in jpcite + autonomath mirror / LLM-0 grep / brand hygiene grep / no LLM-coupled columns in schema.
  • Boot manifest entries appended to BOTH scripts/migrations/jpcite_boot_manifest.txt and scripts/migrations/autonomath_boot_manifest.txt with full rationale block (transport enum + capability tag + handshake exclusive state + LLM-0 by construction + ¥3/req posture).

3-axis seed contents

am_webmcp_endpoint (3 endpoints)

  • /v1/mcp/sse × sse × tools — default MCP 2025-06-18 browser transport.
  • /v1/mcp/sse/health × sse × access — transport liveness probe.
  • /v1/mcp/streamable_http × streamable_http × tools — Wave 16 A8 long-poll fallback.

am_a2a_handshake_log (2 templates, succeeded_at set)

  • claude-3.5-sonnet → jpcite-mcp negotiating tools/search_programs.
  • cursor-mcp-client → jpcite-mcp negotiating resources/list.

am_observability_metric (8 seed metrics)

  • 4 Layer-3-specific: webmcp.endpoints_active, a2a.handshakes_total, a2a.handshake_success_rate, observability.metrics_emitted.
  • 4 AX pillar surface counts: access.surfaces_active=12, context.surfaces_active=9, tools.surfaces_active=139, orchestration.surfaces_active=6.

Bug-free verify (4 axes)

  • pytest 20 / 20 GREEN — tests/test_dim_w_ax_layer3.py runs cleanly against the local venv (/Users/shigetoumeda/jpcite/.venv). Covers migration apply / idempotency / rollback / CHECK constraints / ETL dry-run + apply + idempotent re-run / --force audit append / both helper views / boot manifest dual presence / LLM-0 grep / brand grep / no LLM-coupled columns in schema.
  • SQLite syntax check — sqlite3 :memory: < 283_ax_layer3.sql parses cleanly (BEGIN/COMMIT wrap; PRAGMA foreign_keys=ON). Rollback also parses cleanly and drops every artefact.
  • 3-axis seed verified — fresh apply yields exactly 3 am_webmcp_endpoint rows + 2 am_a2a_handshake_log rows + 8 am_observability_metric rows; idempotent re-run is noop; --force re-append on audit tables grows handshake 2→4 and metric 8→16 while endpoint stays at 3 (UNIQUE backstop).
  • LLM-0 + brand hygiene — grep guard inside the test suite asserts ZERO import anthropic / import openai / google.generativeai / claude_agent_sdk in seed_ax_layer3.py and ZERO legacy brand strings (税務会計AI, zeimu-kaikei.ai, autonomath.ai); also asserts ZERO summary_text / ai_explanation / ai_summary columns in the migration schema (non-comment lines only).

Hard constraints honored

  • No existing AX 4 pillars overwrite — confirmed no am_webmcp* / am_a2a* / am_observability_metric table existed before this migration; Layer 1+2 substrate (mcp.json / openapi / resources / prompts / discovery) is untouched.
  • No main worktree mutations — all work performed inside /tmp/jpcite-w47-dim-w-mig; main worktree untouched.
  • No rm / mv — purely additive (4 new files + 2 manifest appends).
  • No legacy brand — only jpcite; zero 税務会計AI / zeimu-kaikei.ai strings in any new file.
  • No LLM API import — seed_ax_layer3.py imports only stdlib (argparse / json / logging / sqlite3 / sys / pathlib).

Coexistence with parallel in-flight Dim worktrees

  • Dim T (mig 280) — main HEAD, present in manifest above 283.
  • Dim U (mig 281, credit_wallet) — in flight in /tmp/jpcite-w47-dim-u-mig. Migration number 281 already claimed in that worktree's tree; we leave 282 open for the next Dim worktree (likely dim-v).
  • Dim G (in flight, separate task #255) — distinct dimension and distinct migration number. No collision with 283.

PR diff scope

  • scripts/migrations/283_ax_layer3.sql (new)
  • scripts/migrations/283_ax_layer3_rollback.sql (new)
  • scripts/etl/seed_ax_layer3.py (new)
  • tests/test_dim_w_ax_layer3.py (new)
  • scripts/migrations/jpcite_boot_manifest.txt (append-only)
  • scripts/migrations/autonomath_boot_manifest.txt (append-only)
  • docs/research/wave46/STATE_w47_dim_w_pr.md (this file)

Next ticks

  • Wire a daily cron after merge to read the runtime AX Layer 3 signals and append to am_observability_metric (replacing the 8-row seed with real telemetry).
  • Wire am_a2a_handshake_log insert into the MCP session handler so every real handshake gets audited (currently only the 2 template seed rows exist).