Repository navigation
Restore clean portability before MainFrame public-core promotion #2
Description
Activity
Portability RC1 now has a bounded qualified candidate and Draft PR #3.
Exact product candidate:
16a7859304b393c2656b30e8f4ba5bf5570b4f00(treea8a7774fb72347a1470733243c0fff06b3f45379).Corrected clean qualification: https://github.com/camerontjs-dot/MindGraph/actions/runs/34532611069
Artifact:
10174170779, SHA-256496d67dce83d6d66f5d665bbab591d4d0b76c0ad846148141b88584c57e6b17d.Observed PASS: clean Python 3.11 install, 319 maintained tests, explicit MiniLM acquisition, missing-state fail closed, fresh 7/7 synthetic ingest, intended
antinettop result with portable provenance, no serialized hostsource_root, negative-noise ordering, and stable identity/provenance after delete/rebuild.The PR remains Draft. Promotion is still blocked on reconciliation with the tracked
mainframe-live/mindgraphmirror, which contains local evolution absent from standalone main, plus a decision on the 5.9G default dependency closure and explicit model-acquisition contract. Retrieval authority semantics were not weakened.2026-10-03 current-main re-baseline
Live standalone
mainis still8df9ae7fccdb742558950ef9bdedb96cc74df6d0.The original code-level portability blockers recorded here are no longer present on current main:
src/mindgraph/query.pyimportsLiteral;pyproject.tomlconstrains MCP tomcp>=1.0.0,<2;- the reintroduced
tests/test_pruning.pyis absent; - package extras are split into base,
semantic,mcp,full, anddev; - normal semantic ingest/query is cache-only and model acquisition is an explicit
bootstrap-modelsetup operation.
That means the next decision is qualification of the current split-profile package, not replaying the old three-file repair.
I am opening a fresh RC2 qualification issue against this exact current main. It will keep lightweight core/MCP reproducibility separate from the heavy semantic/model-acquisition leg and will measure the latter instead of hiding its dependency cost inside a generic PASS.
No MainFrame integration/promotion follows from this re-baseline.
RC2 update — 2026-10-03
Current-main portability RC2 (#48 / #49) ended
SUPPORTED_LIGHTWEIGHT_PROFILE_SEMANTIC_COST_OPEN.Observed on exact current product base
8df9ae7fccdb742558950ef9bdedb96cc74df6d0:- Python 3.11 and 3.12 fresh
.[mcp,dev]environments passed 368 tests, wheel build/install, missing-state refusal, fresh lexical ingest/query and byte-identical delete/rebuild provenance; - serialized output retained no
source_root; - semantic Python 3.11 closure worked but measured 1.07 GiB before the 87.4 MiB MiniLM cache;
- Python 3.10 was unavailable locally;
- hosted Actions for the research workflow executed zero jobs, so hosted coverage is NOT_RUN.
This materially narrows #2. The old collection/package defects are not the current blocker. Remaining public-core questions are hosted/cross-platform reproduction and whether the optional semantic footprint is acceptable for the intended profile. No promotion or release follows automatically.
- Python 3.11 and 3.12 fresh
Context
MainFrame v0.3.0 evaluated the current standalone MindGraph source for a bounded public-core retrieval profile and deferred promotion after the clean-environment gate failed.
Frozen source under test:
4d292179e1478b999bc6b7c36f9de564c07bd1d9ed7803dcf9718005fd190a12f0537b69dcb8b349Primary qualification record:
1016321546597f76a487d05a62f6ba2bdae40ad7a3d101597e077c5e230847b850dc00d4640Observed blockers
A clean GitHub-hosted Ubuntu / Python 3.11 installation of
.[dev]succeeded, butpytest -qfailed during collection before the synthetic ingest/query qualification could run.Observed failures:
src/mindgraph/query.pyreferencesLiteralwithout importing it, producingNameErrorduring collection.mcp>=1.0.0resolved to MCP 2.2.0, while the current tests importcreate_connected_server_and_client_sessionfrommcp.shared.memory; that symbol is unavailable in the resolved version.tests/test_pruning.pyimportspruning_pass, which is unavailable in the clean checkout/package environment.sentence-transformers, Torch, and a large Linux CUDA stack. That may be valid for the full semantic engine but is substantially heavier than the smallest MainFrame public retrieval profile needs.These observations do not establish that MindGraph's retrieval architecture is unsound. They establish that this exact source does not currently reproduce cleanly enough to promote into MainFrame's public core.
Target
Create the smallest portable profile that can clear a clean-clone qualification without weakening the retrieval authority boundary.
At minimum, a successor gate should prove:
Scope guidance
For a future MainFrame promotion, prefer a slim closure first: core indexing/persistence/query/provenance, minimal CLI, synthetic fixtures, and meaningful CI. Treat semantic/vector retrieval as an optional extra if its dependency/model cost cannot be made proportionate. Daemon, shared MCP/proxy lifecycle, project federation, and repair tooling should not enter the MainFrame core merely because they exist in standalone MindGraph.
Do not change the historical MainFrame v0.2.0 implementation or reinterpret its release. A future promotion should use a newly qualified exact MindGraph candidate and then pass through MainFrame's private positive-manifest publication boundary.