[Deicyde]

Premise

• User: projects that once gave articles an article_id with autoform migrate article-ids, and maintainers of the code behind it.
• Problem: 0.7.0 deprecated autoform migrate article-ids for removal in 0.8.0 (Read literature-links annotations as an article's Lean link #601). It still prints a warning[deprecated]: line on every run, and its plan, its frontmatter writer and its autoform-article-id-plan/v1 schema still ship.
• Replaces: autoform links set BLUEPRINT ARTICLE gives an article its article_id, and autoform migrate links blueprint --write gives one to each annotation it writes. Nothing replaces the --check gate or the --json plan, as 0.7.0's notes said.
• Deletes: the command, plan_article_ids, write_article_ids, ArticleIdWriteError, the entry and plan classes, IDENTITY_PLAN_SCHEMA and the frontmatter writer _with_article_id.

Line delta: +73/-552 in total.

• code: +3/-240 (article_identity.py +1/-180, __main__.py +1/-59, links_migration.py +1/-1);
• docs: +8 (CHANGELOG.md);
• tests: +62/-312 (tests/test_article_identity.py +47/-284, tests/test_annotations.py -24, tests/test_links_migration.py +15/-4).

What changes

• CLI. autoform migrate article-ids fails as an invalid choice, exit 2 from argparse. autoform migrate keeps links, which now dispatches directly.
• Python. article_identity.py keeps only the three helpers links set and migrate links import: _proposed_id, the ID a new article is proposed; _read_regular, which reads a regular file and refuses anything else; and _replace_if_unchanged, which replaces a file unless it changed since it was read. They stay where they are: moving _proposed_id into links_migration.py would make links_set.py import it, and links_migration.py already imports links_set.py. The graph still reads article_id until 0.9.0, unchanged.
• Generated files. No template, generated workflow or skills/setup file runs migrate article-ids, on main or in any release, so neither copy of autoform_audit.py changes. Their one article_id line validates the assumption contract, which the 0.9.0 removal of article IDs owns.
• Tests.
  • The plan, write and command tests go: tests/test_article_identity.py's 17 plan, write and CLI tests, and tests/test_annotations.py's two tests of --write beside blueprint/links/.
  • The two graph tests of article_id stay.
  • Where a write test covered a helper that migrate links relies on, and no other test did, the helper now has a direct test: refusing a symlink, keeping the file mode with and without os.fchmod, and refusing a file changed before or during the replacement. A new test checks that the proposed ID is deterministic and well formed, as the deleted plan test did, and that it depends on both the path and the source hash.
  • The write tests of CRLF and CR frontmatter, the byte-order-mark refusal and the hash check against the plan tested only the removed writer. So did the write through an unreadable link, which tests/test_graph.py covers for loading.
  • tests/test_links_migration.py's two tests that took the expected ID from plan_article_ids take it from _proposed_id. A new CLI test checks that migrate article-ids exits 2 as an invalid choice.
• CHANGELOG. Under Unreleased, an Upgrading from 0.7.0 entry and a Removed entry, the latter pairing with 0.7.0's Deprecated entry for Read literature-links annotations as an article's Lean link #601. They were added in a second commit once the PR had its number:

### Upgrading from 0.7.0

- **Article-ID gate.** Nothing replaces `autoform migrate article-ids
  --check`: a CI step you added that runs it exits 2 until you delete it. No
  generated workflow runs it (#796).

### Removed

- **`autoform migrate article-ids`**, deprecated in 0.7.0, is gone, with its
  `--check` gate, its `--write` and its `--json` plan.
  `autoform links set BLUEPRINT ARTICLE` gives an article its `article_id`,
  and `autoform migrate links blueprint --write` gives one to each annotation
  it writes (#796).

Checks

• Full CI on scratch at 2b19874, this PR's code commit on main at 8c618dc: every job passed. That covers test on 3.11 and 3.13 (7,171 and 7,170 passed, ruff included), the real-Lean, skeleton and REPL suites, WebKit, Windows and the Node read-back cards.
  https://github.com/facebookresearch/autoform-bot/actions/runs/43190503980
• Mutants on scratch, run after their tests passed unmutated (6 tests), each failing its tests. Six in one run: the read following a symlink, the replacement without its check, the mode lost where os.fchmod is missing, the proposed ID ignoring the source hash or the path, and migrate article-ids restored.
  https://github.com/facebookresearch/autoform-bot/actions/runs/43190587234
• ruff check with the pinned 0.15.17 on the changed Python files: clean.
• git merge-tree, with main at 8c618dc and the open PRs' heads as of 2026-11-13 01:57Z: this commit adds no conflict with any of them. Prove2me - Add skill files and modifications to the autoform-bot CLI to allow for easy integration with prove2.me #351 and fix(site): check links and cards in the HTML5 element tree #621 already conflict with main, in files this PR doesn't touch, and the others merge cleanly.
• Tests for new behavior, docs for changed APIs, a CHANGELOG line.
• make lint and make test, through CI on scratch.

Posted by Claude Code, at Jack's request.

[Deicyde]

Review at 17271f3: no blocker. I checked it against path identity's next step and against migrate links.

• Nothing the next path identity step needs is removed. That step switches the writers and re-keys articles in migrate links. It uses neither plan_article_ids and write_article_ids nor the autoform-article-id-plan/v1 JSON, all of which name articles by article_id. _proposed_id, _read_regular and _replace_if_unchanged stay, since links_migration.py and links_set.py import them. Their tests stay too: determinism, the symlink refusal, the file mode, the changed-file refusal and an edit made during the flush.
• migrate links is unchanged. links_migration.py changes one docstring line. migrate now dispatches straight to _migrate_links, as _migrate did for links. That is its only subcommand, and argparse still requires it.
• The removed command fails plainly. autoform migrate article-ids BLUEPRINT --check exits 2 with argparse's invalid-choice error, as test_cli_refuses_migrate_article_ids checks.
• CHANGELOG. The Removed entry pairs with 0.7.0's Deprecated entry for autoform migrate article-ids (Read literature-links annotations as an article's Lean link #601) and names the same replacement, links set, plus migrate links --write for annotations. The Upgrading entry says a CI step that still runs --check exits 2, which matches argparse's exit for an invalid choice. Both go under the existing Unreleased headings, so no heading is duplicated.

Notes, none blocking:

1. test_each_fact_reads_back_the_same_from_the_annotation and test_the_bundled_example_migrates_with_its_graph_unchanged now take the expected ID from node.article_id or _proposed_id(node.id, node.source_sha256), the same expression links_migration.py uses where it proposes one. plan_article_ids computed the same value, so the checks are unchanged.
2. plan_article_ids was the only place that refused an authored article_id equal to the one proposed for another article, and its test goes with it. migrate links never had that check, and this PR leaves that as it was.

Checks: this PR's CI on 17271f3, every job passed (https://github.com/facebookresearch/autoform-bot/actions/runs/43363927528). The 6 mutants of the code commit were each caught (https://github.com/facebookresearch/autoform-bot/actions/runs/43190587234); the CHANGELOG commit changes no code.

Posted by Claude Code, at Jack's request.