Pine/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases 1 Wiki Activity

cleanup(install): drop the master.path record

Browse Source 
Nothing read it. The launcher, dsh-upgrade, and the installer's own re-run
all ignored the file, and the diagnostic it was meant to feed was never
built, so it was write-only state.

Git already owns the fact it recorded: a staging worktree's .git file names
the repository path, and `git worktree list` in that clone enumerates every
worktree depending on it. An installer-written copy only adds state that can
go stale while nothing validates it.

The containment caveat it documented is real and stays in the script header
and the Agent Note, now pointing at git's own records.
This commit is contained in:
Turtle
2026-07-31 21:17:28 +08:00
parent 00c5f2abd1
commit f3a1ff41b7
4 changed files with 12 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.agents/notes/implemented/process/2026-07-31-installer-adopts-existing-checkout.md
+4 -4
View File
@@ -16,11 +16,11 @@ In-repo mode still never clones and never modifies the working tree, but it now
The container owns staging worktrees and `current`; the repository is *discovered*, not owned. `git rev-parse --git-common-dir` resolves the shared git directory behind the checkout — for a linked worktree that is the real clone rather than the worktree itself — and its parent is the repository that serves as the upgrade base. A staging worktree branched from the checkout's `HEAD` is then created under `$DSH_SOURCE`, and `current` points at it. A clone anywhere on disk therefore converges on the same layout as a `curl` install, and the two paths share one worktree/exclude/lock/link sequence: they differ only in whether the repository was discovered by `git clone` or by `git rev-parse`.
`$DSH_SOURCE/master.path` records the resolved repository, and only when that repository lives outside the container. A container holding its own master is self-contained and gets no file, so the file's presence is itself the signal that this container depends on an outside path: each staging worktree holds an absolute gitdir pointer into that clone, so deleting the clone breaks them.
The installer records nothing about where that repository lives. A container whose repository sits outside it is not self-contained — each staging worktree holds an absolute gitdir pointer into that clone, so deleting the clone breaks them — but git already owns that fact: the worktree's `.git` file names the path, and `git worktree list` in the clone enumerates every worktree depending on it.
Adoption branches from `HEAD`, so committed work is what runs and uncommitted changes stay in the checkout. This is not prompted or warned about: the installer builds the layout and gets out of the way. Setting `DSH_SOURCE` to a different directory remains the one documented way to opt back into cloning a separate tree.
Every path comparison runs on physical paths through a `resolve_dir` helper, and every compared value is resolved at assignment rather than at the comparison. macOS resolves `/var` through a symlink to `/private/var`, so comparing a git-reported path against an unresolved one misclassified an existing managed install as a foreign clone and would have built a second container beside the real one. The same defect recurred twice more during review — once where a curl install's `REPO_ROOT` stayed unresolved and wrote a spurious `master.path`, and once where `x=$(resolve_dir …) || x=$fallback` left an empty path because the assignment succeeds even when the substitution fails. `resolve_dir` therefore echoes a missing path back itself, and callers that need "does not exist" test the directory explicitly. `git rev-parse --path-format=absolute` would do the same job but requires git 2.31+.
Every path comparison runs on physical paths through a `resolve_dir` helper, and every compared value is resolved at assignment rather than at the comparison. macOS resolves `/var` through a symlink to `/private/var`, so comparing a git-reported path against an unresolved one misclassified an existing managed install as a foreign clone and would have built a second container beside the real one. The same defect recurred twice more during review — once where a curl install's `REPO_ROOT` stayed unresolved and so compared unequal against every resolved path, and once where `x=$(resolve_dir …) || x=$fallback` left an empty path because the assignment succeeds even when the substitution fails. `resolve_dir` therefore echoes a missing path back itself, and callers that need "does not exist" test the directory explicitly. `git rev-parse --path-format=absolute` would do the same job but requires git 2.31+.
Before `current` is repointed, the installer rejects a staging path that resolves to the repository itself, enforcing the upgrade contract that the launcher never resolves to the master clone.
@@ -42,10 +42,10 @@ One layout now serves every install, so an adopted clone is upgradable by `dsh-u
The cost is that a contributor can no longer point PATH at a checkout and have `dsh` follow that working tree as they switch branches: the launcher now resolves to a staging worktree pinned to the `HEAD` adopted at install time. Re-running the installer adopts the current `HEAD` again.
The cost is that a container adopting an outside clone is no longer self-contained: deleting that clone breaks its staging worktrees. This is inherent to reusing an existing clone rather than a property of this design — the rejected symlink hides it rather than fixing it — and `master.path` is the mitigation, not a repair.
A container adopting an outside clone is also no longer self-contained: deleting that clone breaks its staging worktrees. This is inherent to reusing an existing clone rather than a property of this design — the rejected symlink hides it rather than fixing it — and git's own worktree records are what diagnose it.
## Testing
`scripts/install.sh` has no automated test, and this change does not add one: the user directed that `install.spec.ts` be left out of scope. That is a known gap on a shipped user-facing path, and the `/var` resolution defect above is exactly the class of bug a test would have caught first. The standing [`FIXME(install-ts)`](../../../../scripts/install.sh) asking for this workflow to move into a tested TypeScript entrypoint is correspondingly more pressing.
Verification was manual, through a throwaway harness driving the real script with a stubbed `pnpm`: adopting a standalone clone; adopting from a linked worktree into its existing container; an explicit `DSH_SOURCE` still opting back into cloning; a dirty tree adopting silently with no prompt or warning while its uncommitted file stays behind; a non-git checkout failing with guidance; and a `curl`-style clone install asserting both the built layout and the absence of `master.path`, which is the regression that caught the unresolved-`REPO_ROOT` defect. The interactive path was exercised under tmux from a dirty checkout, confirming the run reaches the launcher with no adoption prompt and ends with `dsh` running from the new staging worktree while the original checkout keeps its branch and its uncommitted file.
Verification was manual, through a throwaway harness driving the real script with a stubbed `pnpm`: adopting a standalone clone; adopting from a linked worktree into its existing container; an explicit `DSH_SOURCE` still opting back into cloning; a dirty tree adopting silently with no prompt or warning while its uncommitted file stays behind; a non-git checkout failing with guidance; and a `curl`-style clone install asserting the built layout, which is the regression that caught the unresolved-`REPO_ROOT` defect. The interactive path was exercised under tmux from a dirty checkout, confirming the run reaches the launcher with no adoption prompt and ends with `dsh` running from the new staging worktree while the original checkout keeps its branch and its uncommitted file.