Skip to content

M2 report: Framework 0.2

M2 was planned in proposal 0010 and built in seven steps, each with its own proposal (0011 to 0017). This report checks the M2 exit criteria from spec section 15 and release gates 1 to 5 (section 14.2). Numbers are from the test suite and the pilots, counted with o200k_base.

What was built

Area Requirements Proposal
Shared store (memory, Redis, any py-key-value backend), idempotent writes, ctx.deps EXE-4, SEC-6 (shared single use), spec 9.2 0011
Safe writes: preview, approve, commit Section 10, gate 4 0012
Results by reference and background jobs RES-5 0013
Resources and prompts CAP-2 0014
Migration adapter and guide, ouncemcp audit, URL checks DX-7, AUD-1 to AUD-13 (HTML excepted), part of SEC-9 0015
Client broker CLI-1 to CLI-8 0016
ouncemcp.toml, health and readiness, metrics, description lints Spec 9.2, 9.4, CAP-5 0017
Policies over stdio (stdio_policy) Decision 3 of proposal 0010 0010

Exit criteria

Criterion Result Evidence
Pilots pass release gates 1 to 5 Pass Gate table
The migration guide works on one existing FastMCP server Pass: markitdown-mcp, wrapped unchanged (step 1), then migrated examples/markitdown, tests/test_pilot_markitdown.py
audit totals match check within 1% for the same server Pass: 488 and 488 on the inventory example, in-process and over stdio tests/test_audit.py

Release gates 1 to 5

Gate Result Evidence
1. Conformance Pass. The first CI run of the official suite (0.2.0-alpha.12, revision 2026-07-28) on PR #14 passed 36 checks; its 5 failures are exactly the expected ones (completions, and tools returning image, audio, embedded-resource or mixed content). The job is not yet required in CI .github/workflows/conformance.yml, run on commit fadf7eb
2. Isolation Pass Security catalog on the ticketing pilot (cross-tenant, guessed names); handles and job ids refused for another caller (report pilot); idempotency keyed per tenant; ctx.cache partitions
3. Writes Pass A create that times out after the upstream write produces one ticket, and the retry gets the stored result (the plain FastMCP version writes twice); concurrent duplicates run once; two replicas sharing a store write once
4. Approvals Pass Approved writes run once after the person sees the preview; declines change nothing; receipts cannot be replayed or moved to other arguments; clients that cannot ask get approval_unavailable; enforced on stdio
5. Budgets Pass Every example, the migration steps and a fresh ouncemcp init project pass ouncemcp check; the RES-8 fuzz and RES-9 property tests pass

Pilots

Pilot Key result
Ticketing (examples/ticketing) No duplicate write on retry after a timeout; approval with the exact ticket shown; admin-only delete; 555 definition tokens against 283 for the plain version
Long-running report (examples/report) jobs.get 168 tokens and a 50-row page 1,351, against 52,002 for all 2,000 rows; works across two replicas and a restart; a job lost in a restart reports interrupted
Migration of markitdown-mcp (examples/markitdown) Result for a 60-section document from 14,760 tokens to 388 plus pages; metadata, loopback and private URLs refused; file: access opt-in and confined

Measured overhead

In-process calls through the full stack take about 0.22 ms p50 on the session machine; tracing adds 0.01 to 0.15 ms. This is not a benchmark (M3 has those).

Known limits

  • Single-use receipts and idempotency hold across replicas only with a shared atomic store (RedisStore); KeyValueStore claims are atomic per process.
  • check_url resolves names once and does not follow redirects for you; full SEC-9 stays in P1.
  • Interrupted jobs must be started again; jobs are not mapped onto the MCP tasks extension.
  • Five tests need network access to tiktoken's download host (the gpt2 encoding and stdio subprocess tests) and fail in a sandbox without it.
  • Host verification: Claude Desktop and Cursor pass (owner, 12 November 2026; see the compatibility matrix). VS Code and the Claude mobile app are not yet verified, and host versions and the BUD-7 field study are not yet recorded.