refactor: remove the remaining version control calls - #23
Merged
Merged
Conversation
Worktree.release, .finish, and .merge_branch shelled out to git to tear down and merge a code worktree Plastic itself never creates. Arm.disarm now only releases the delivery lock and reports the git worktree remove instruction (Arm.worktree_removal_instruction) for a workspace the agent already provisioned; arm/disarm/repair drop the runner: parameter they no longer need. Worktree.ShellRunner and the git-shelling helpers still called by the node-graph runner (current_branch, remove_worktree, prune, git_repo?) stay for now; part B of intent 390 removes them with the runner modules that use them.
…tead of running them ExecWorktree and end-intent shelled out to git to check the code worktree for dirty or unmerged state and to finish/merge/remove it. Plastic runs no version control command: exec-worktree now prints the commit, merge, and worktree-removal steps for the agent to run by hand, and a delivered close no longer refuses on an unmerged branch or a dirty worktree (the exit 5 and exit 9 refusals are retired). Disarm still names the worktree-removal instruction when a worktree was provisioned.
… without git ScaffoldIntent.resolve_repo_dir now reads projects.yml only, and detect_base_branch returns "main" unless a project's project.yml names a different flow base -- neither shells out to git any more. diffstat is replaced by diffstat_instruction, which builds the git diff --stat command as a string for a caller to print or run itself; BackfillIntent and VerifyIntent's diffstat check both print it instead of running a git_runner/runner seam.
NodeInput's ledger block ran an injected git log after a reclaim to list landed commits. Plastic runs no version control command: the block, its git_runner seam, and the byte-truncation helper built for it are gone, and full_ledger_text no longer takes files:, repo_dir:, or git_runner:.
…of running it MaintenanceGit branched, committed, and merged a scoped store change through git. Plastic runs no version control command: maintenance-run now applies each tool's change directly and prints the git add/commit instruction naming the touched paths. MaintenanceGit and its test are deleted.
WorktreeSweep and sweep-store-worktrees existed to clean up orphaned store worktrees from a retired provisioning path; they shelled out to git to detect and remove them. That path is long gone, so this drops the module, its CLI, its test, and the three stale core_files manifest entries left over from earlier deletions in this same effort.
InstallerCore#git_init_if_absent (deleted alongside the worktree-sweep cleanup) made ~/.plastic its own git repository so update.rb could later commit re-synced core files. Plastic runs no version control command: install.rb drops the call site, update.rb's comment now describes commit_core_files as landing only when the home already carries its own repository, and the install/new-intent tests that built a real .git or asserted one exists after a fresh install no longer do either.
…ce paths architecture.md, internals.md, locks-and-worktrees.md, and completion-and-done.md described Worktree.release/finish, a git-shelled diffstat and em-dash diff, a git-committing maintenance-run, a git-initialized plastic_home, and an end-intent that checks for a merged, clean code worktree. All of that is gone (intent 390: Plastic runs no version control command); these chapters now describe the printed instructions that replaced each of those checks.
test/exec_worktree_test.rb still called ExecWorktree.run with the retired finisher:/status_checker:/runner: seams and asserted on the dirty-worktree and unmerged-branch refusals those seams drove. It never got fixed alongside the run() rewrite in the earlier commit, so verify-change caught 14 argument errors. Rewritten to exercise the current run(): the printed commit/merge/worktree-remove report, the nothing-provisioned and no-worktree-recorded paths, the lock/session and usage-failure exits, and a guard that the module never shells out.
…g them Absorb, rewind, sweep, watch, dispatch and the worktree module no longer shell out to git for merge, reset, worktree add/remove, or freshness by commit. Each prints the instruction and asserts on plain files and directories instead. The scope check is dropped: NodeReturn's closed schema carries no file list to measure a diff against. The reaper can no longer tell a merged branch from an unmerged one, so a terminal node's worktree is always named for removal now, recorded as a lost capability.
setup_real_repo becomes setup_repo: a plain directory tree registered in projects.yml, never a git repository - RunnerDispatch only prints its worktree instruction and RunnerAbsorb only prints its merge instruction now, so neither needs one. The two real-subprocess rows (concurrency cap, CLI arm) pre-create each dispatched node's own worktree directory by hand, standing in for the git worktree add a human would run from the printed instruction, so node-run's -C target still resolves to the right node.
node_run_cli_test.rb: rename setup_real_repo to setup_repo (a plain directory, never a real git repo); the stub codex commit mode drops its real git add/commit and reports a fixed fake sha. runner_cli_test.rb: drop the git! helper and the two merge-in-progress tests (the abort-on-a-stuck-merge step is gone); the stale-node-worktree reap test asserts the printed removal instruction against plain directories instead of running a real merge and removal; the CHANGELOG test checks the new intent 390 Unreleased line for five checks instead of the historical six. CHANGELOG.md: note under Unreleased that the gate drops its scope check and now runs five, and that git-touching steps print their instruction instead of running it.
Replaces the real throwaway git repo (init, worktree add, the git helper) with a plain directory tree - NodeWorktree.provision never shells out to git any more, it only ever checks Dir.exist?, so a directory at the exact repo/worktree path is the whole fixture these rows need. The two worktree-rollback rows now note that a failed input build or a refused running write has nothing of its own to roll back on the worktree side, since provisioning creates nothing itself; the node-worktree-naming row and the preexisting-worktree row pre-create the node's worktree directory, standing in for an agent having already run the printed git worktree add instruction.
…unning it Owner ruling 2026-09-24: Plastic runs no version control command. store_commit_instruction replaces store_commit - it names the exact git add/commit, scoped to the completing intent's own directory plus INDEX.md (D17: never git add -A), for the closer to run; the dry-run preview and the real close both print it, and git_toplevel/relative_to drop out with the git calls they only served. worktree_changed? drops its git status/rev-list checks (Plastic can no longer read them itself) and fails open: a code worktree that exists reads as changed, exactly as an unreadable one used to. Tests: end_intent_test.rb's five real-git-repo fixtures (the backfill, store-commit, no-commit, and unrelated-file rows, plus the checklist row) become plain directories asserting on the printed instruction and the files it names. end_intent_future_close_test.rb drops its real git repository entirely - the store is a plain directory and the close's own commit is checked against printed stdout, never git log.
…running it Owner ruling 2026-09-24: Plastic runs no version control command. commit_core_files now prints the git add/commit for ~/.plastic's core files instead of shelling out through an injected runner; the test asserts against the printed instruction instead of captured argv.
Plastic itself runs no version control command (owner ruling 2026-09-24), so git is no longer a runtime dependency this gate needs to probe for - only of the printed instructions a person or agent runs afterward.
NodeWorktree and Worktree compute every path and branch deterministically from projects.yml and the intent worktree's own path, with no git call in that path (owner ruling 2026-09-24). The scratch fixture now builds a plain directory tree instead of a real git repository, and the git-probe skip row it existed to prove is gone along with the probe itself.
Plastic runs no version control command. The delivered screen's shipped version now comes only from outcome.md's own record, falling back to naming the merge sha (D11) instead of a `git tag --contains` guess, and prints the command a person can run themselves when a merge sha is known but no version was recorded. The branch segment falls back to the literal "main" once a real project repository is known, instead of reading the repository's actual current or remote-default branch.
Plastic runs no version control command, so restore-intent-v1 no longer shells out to `git show`/`git rev-parse` to read the v1 snapshot. --at is now a free-form label kept for the report and revisions.md entry; a new --snapshot-dir names a directory the caller has already populated with the v1 content, one file per path, at the same relative path it has under --plastic-home. With no --snapshot-dir, or a missing file, the tool prints the exact `git -C <home> show <at>:<path>` command to run and aborts with no write. The test fixture now builds the v1 snapshot as a plain directory copy instead of a real git repository.
architecture.md and internals.md still described end-intent checking a merge with `git merge-base --is-ancestor`, disarm running `git worktree remove`, and node_worktree.rb shelling out through a `ShellRunner` - all retired by intent 390. Update both docs to match: Plastic prints the worktree, merge, and dirty-check instructions instead of running them, and the retired exit codes (5 and 9) name what used to live there.
Minitest/MultipleAssertions flagged the store-commit-instruction test at 5 assertions against a limit of 3.
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
1 participant
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
Plastic now runs no version control command anywhere, not only in session commit and auto take. Each remaining place writes its record and prints the git instruction on its
next:line for the person or agent to run. Code whose step no longer matters is deleted.Why
The owner ruled on 2026-09-24 that every git call in the script library goes, after pull request 21 removed the first two. Tests that built real repositories were slow and could not be scored by the mutation tester.
How
scripts/end-intentand the exec worktree step print the store commit, merge and removal steps. The dirty and merged checks are gone.projects.yml. The diff commands are printed.restore-intent-v1takes a snapshot directory and prints the commands that fill it.Tests