Repository navigation
Remove release pins and the formal-contract tooling - #547
Merged
Merged
Conversation
arXMCP no longer pins a sibling repo's tagged formalization or
re-serves it. Removed:
- tools/formal_release_pin.py, the CLI that pinned a topic repo's
release into notebooks.db;
- the arxmcp://formal/{notebook} and /{key} resources, with the
coverage census that only they used, and the retrieved_formal
wrap tag;
- the store's formal_releases methods.
notebooks.db goes to v7, which drops the formal_releases table. The
store's rule against dropping tables exists to protect user-authored
notebook metadata; a pin was re-derived from a git tag, so the rule
now names user-authored state explicitly. A new migration test seeds
a v6 file holding a notebook and a pin, and checks that v7 drops only
the pin table.
Refs #546.
Removed the offline statement tooling that existed for the cross-repo formal contract: - tools/statement_resolve.py, the resolver that checked whether a registry quote still appeared in the corpus; - tools/statement_mint.py, which prefilled candidate registry entries; - tools/_statement_common.py, their shared plumbing; - their tests; - the `contract` extra (mfc from math-formal-contract-lean, pinned to a commit), and .github/workflows/contract.yml, which ran only these tests. uv.lock loses only the mfc package and the extra. Kept: documents.arxiv_version and its backfill. They were built for the resolver's version guard, but the column is half the documents key and is hashed into the corpus manifest. Their comments now give that reason instead of citing the resolver. Refs #546.
With the release pin, the formal resources and the statement CLIs gone, CLAUDE.md described tooling that no longer exists. Deleted, with nothing added or reworded: - §4.10, the binding cross-repo contract boundary, including its "what has landed" table; - the §4.1 sentences about .github/workflows/contract.yml; - the four removed tools from the directory tree. Two tests read that text and would now fail or skip forever. Both were tests of documentation wording, which CREDO.md forbids: - test_claude_md_table_matches_live_behavior keeps its useful half: _declaration_names is still checked against the eight snippet shapes it has got wrong before. It no longer parses the table out of CLAUDE.md. - test_the_contract_state_date_is_not_behind_claude_md is deleted, since the dated table it compared is gone. Its section's entry leaves the AGENTS.md section baseline, which may only shrink. Refs #546.
This was referenced Nov 15, 2026
chris-dare-dev
pushed a commit
that referenced
this pull request
Nov 15, 2026
#547 removed the contract extra, whose jsonschema entry sat next to the runtime one this branch adds. uv.lock keeps the runtime requirement and drops the contract one.
chris-dare-dev
pushed a commit
that referenced
this pull request
Nov 15, 2026
#547 removed the formal resources from server/mcp_resources.py, next to where this branch adds the chunk resource. Kept the chunk resource and dropped everything formal: the resource count becomes four, and the read_corpus_version import goes with the code that used it.
chris-dare-dev
added a commit
that referenced
this pull request
Nov 15, 2026
server/tools.py imports jsonschema to validate lean_verify's result against the schema it advertises, but no runtime dependency declared it. It installed only because mcp requires it. The contract extra, which #547 removes, was the one place it was ever named. Declare it explicitly, at the >=4.20 floor mcp already requires, as the project does for jinja2 and python-multipart. uv.lock gains only the jsonschema lines; `uv lock --check` passes. Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Closes #546. This removes arXMCP's half of the cross-repo formal contract (CLAUDE.md §4.10). Delivery moves to continuous integration against
mainrather than a pinned record of a tagged sibling release. The design and scope are in #546.Three commits, one concern each. In total 23 files change: about 5,000 lines deleted and 58 added.
9521a03tools/formal_release_pin.py; thearxmcp://formal/{notebook}and/{key}resources, including the coverage-census code only they used; theretrieved_formalwrap tag; and the store'sformal_releasesmethods.notebooks.dbgoes to v7, which dropsformal_releases.2843c9etools/statement_resolve.py,tools/statement_mint.py,tools/_statement_common.pyand their tests; thecontractextra (mfc) and.github/workflows/contract.yml, which ran only those tests.uv.lockloses onlymfcand the extra.aa1c9c6contract.yml, and the four tools in the directory tree, as a pure deletion. Also two tests that read that text (see below).Kept:
documents.arxiv_versionandtools/notebook_arxiv_version_backfill.py. They were built for the resolver's version guard, but the column is half thedocumentsprimary key and is hashed into the corpus manifest. Their comments now give that reason instead of citing the resolver.Verification
Run on Ubuntu (Linux) at
aa1c9c6, built onmainate2357bd. Thatmainis still atnotebooks.dbschema 6, so this branch's v7 is the only v7. The failure baseline below was measured onmainatd9ddaa8, before #539, #540 and #541 merged. No test those PRs added shows up as a failure here.TestV6ToV7Migrationseeds a real v6notebooks.dbholding a notebook and aformal_releasesrow, reopens it, and checksuser_version == 7, that the table is gone, and that the notebook row survived. The other migration tests intest_notebook_api.pypass.mcp_resources,NotebooksStoreor wheel packaging, every failure is one in the baseline (the BGE-M3 download), and none is new. The backfill, wheel-packaging and dev-toolchain tests pass.pytest: 199 failures and errors, against the baseline's 214, in this sandbox.test_formal_release_pin.pyfailures; the rest are the same model-download failures (fixed separately by test: run chunker tests on a stand-in tokenizer #544 and fix(server): warm BGE-M3 through patchable loaders #545).test_proof_chain.py::TestPerfGate::test_perf_gate_500msfailed once under full-suite load and passed 3 of 3 in isolation. This branch does not touch it.ruff check .is clean,uv lock --checkpasses, andpython3 .github/credo_prose.py origin/mainexits 0.make test.Notes
server/notebooks_store.py) protects user-authored notebook metadata. A pin was re-derived from a sibling repo's git tag, and CLAUDE.md recorded that none had been made on the workstation. The rule's wording now names user-authored state explicitly. Restoring an older v6 backup migrates it to v7 the same way.test_claude_md_table_matches_live_behaviorkeeps its useful half:_declaration_namesis still checked against the eight snippet shapes it has got wrong before. It no longer parses CLAUDE.md.test_the_contract_state_date_is_not_behind_claude_mdis deleted along with the dated table it compared. Its entry leaves the AGENTS.md section baseline, which may only shrink.bridgeland-stab-leanwas built against this contract (its ADR-0001/0002/0004/0006,cross-repoissues and committed resolver output). It needs its own issue, which is drafted but not yet filed.server/tools.pyimportsjsonschemaat runtime, butjsonschemawas only ever declared in the removedcontractextra. It still installs, transitively throughmcp. This change does not alter that; a separate change declares it, as httpx is an undeclared dependency of the test suite — 60 test modules import fastapi.testclient and get httpx only by accident, transitively through the mcp SDK #377 did forhttpx.