adopt-project.py: --source <org/repo> reads a directory of that name under the current one, a trailing / clones into source-readonly/ itself, and a recorded clone that is gone or emptied is refused for the wrong reason #652

Lane: openreposhape-1 (openRepoShape-1)

plan --source <org/repo> should not read a directory of that name under the current one as the source. A trailing / should not make the clone land in source-readonly/ itself. A recorded clone that is gone or emptied should be refused for what it is. The plan should record whether plan made the clone (cloned:), and a clone no plan records should be removed.

From lane 1's triage of 2026-10-15, D3, approved by Brett Heap at 2026-10-16T11:21:17Z ("Approve all (Recommended)"). One pull request covers the lines below.

Parent: #457. Milestone: v1.1.0.

Lands after: #594, and D2's issue (#650, the --work-dir refusal and help text), which edits the same functions: _work_root, Plan.open_source, the head of cmd_check and the finally in main, as #594 leaves them. It is authored on D2's branch, or rebased once D2 lands.

Effort: M. Platform: all.

Lane openRepoShape-1 claims it at filing; see its CLAIMED comment.

Why this is P2.

• L1-24, probed on 396c0ff. With ./acme/Atlas present, plan --source acme/Atlas planned that directory, cloned nothing and exited 0: Source.open asks Path(spec).exists() first. So a person working in a ~/src/<org>/<repo> tree adopts a local copy they never named. Read again on 7ba64f0: Source.open still asks candidate.exists() first.
• L1-43, probed on 396c0ff. --source acme/Atlas/ cloned into <work>/source-readonly itself, and recorded acme/Atlas/. Read again on 7ba64f0: Source.open still names the clone by spec.split("/")[-1], which is empty here.
• L1-22 and L1-26, on adopt-project.py and tests/conftest.py: a run without --work-dir removes the empty work directory it made, and a recorded clone that is gone is cloned again, so no empty openreposhape-adopt-* is left and no source-unclonable follows (#526) #594's head. A recorded clone with no repository: that is gone (L1-22) is refused source-unclonable with a gh auth status hint. One a temp cleaner emptied (L1-26) is refused git-failed. In both cases, passing --source again works.

The fix. The plan records whether plan made the clone (cloned:, L1-27's narrowing). That answers L1-22, L1-26 and L1-27 together. A clone no plan records is removed (L1-23, #526's own subject).

Lines

• L1-22 (P2): a vanished local_path with no recorded repository goes to gh repo clone on POSIX and is refused source-unclonable, with a gh auth status hint; on Windows it is refused source-unresolvable, naming a --source nobody passed. _where_the_clone_is_gone could say the path is gone. On be808a7, adopt-project.py and tests/conftest.py: a run without --work-dir removes the empty work directory it made, and a recorded clone that is gone is cloned again, so no empty openreposhape-adopt-* is left and no source-unclonable follows (#526) #594's head, read. The line measured it on 652b686 (POSIX) and read it from the code (Windows). (source)
• L1-24 (P2): Source.open tests Path(spec).exists() against the current directory, so org/repo is read as a local path when a directory of that name exists there, even for the re-clone whose NOTE says "is cloned again". On 396c0ff, probed: with ./acme/Atlas present, plan --source acme/Atlas planned that directory, cloned nothing and exited 0. (source)
• L1-26 (P2): a recorded clone that a cleaner emptied or half-cleaned, directory still there, is not treated as gone: check refuses git-failed rather than cloning again. macOS's temp cleaner can leave clones that way. On be808a7, adopt-project.py and tests/conftest.py: a run without --work-dir removes the empty work directory it made, and a recorded clone that is gone is cloned again, so no empty openreposhape-adopt-* is left and no source-unclonable follows (#526) #594's head, read: exists() at :4075. The line measured it on 652b686. (source)

Rides along (backlog; tick it here if the same change covers it)

• L1-23: a clone no plan records stays in its temp directory: the fresh clone check or execute makes, each --source <org/repo> run, and a refused plan --source. Each check after the recorded clone is gone clones again and keeps 31 files where main left an empty source-readonly/. On be808a7, adopt-project.py and tests/conftest.py: a run without --work-dir removes the empty work directory it made, and a recorded clone that is gone is cloned again, so no empty openreposhape-adopt-* is left and no source-unclonable follows (#526) #594's head, read. The line measured it on 652b686. (source)
• L1-27: source.repository is read from origin on any host, so a moved non-GitHub local clone is re-cloned from github.com under that name; plan-stale keeps it safe. Narrowing: re-clone only a local_path this tool made, or record cloned: in the plan. Not re-checked since the line measured it on 652b686. (source)
• L1-43: --source acme/Atlas/ with a trailing slash is compared as acme/Atlas/, and spec.split("/")[-1] is empty, so the clone target becomes source-readonly/ itself. On 396c0ff, probed: --source acme/Atlas/ cloned into <work>/source-readonly itself, and recorded acme/Atlas/. (source)

Each line above is a line of #457's checklist comment (issuecomment-6200110061). Tick it there too when this closes.