fix: preserve shared Git sources during online workspace migration (#19688)

## Thinking Path

> - Paperclip preserves the work produced by its agents.
> - Remote workspace migration must preserve legacy Git history and
local edits.
> - Linked checkouts can share administration and objects with other
tasks.
> - A task lock alone cannot exclude writers of a shared or borrowed
store.
> - The online source owner must derive and retain the full physical
ownership closure.
> - This change reuses task, project, lifecycle and finalizer locks for
that closure.
> - Idle legacy Git work can then migrate while unrelated remote turns
continue.

## Linked Issues or Issue Description

Refs: #19642.

**What happened?**

Online migration keeps legacy Git sources blocked because several task
checkouts can share a project base or a private Git cache. A read-only
source observation found 160 pending Git records, 120 distinct paths and
nine common stores. Its largest project cohort contained 126 records and
110 task owners. An ordinary-directory fence cannot preserve these
sources by itself.

**Expected behavior**

Derive the physical source owners from stable Git metadata. Hold their
existing source and lifecycle keys through publication. Retain all
history and recovery obligations. Reject foreign aliases, unknown stores
and changed sources. Let unrelated object-backed work continue.

**Steps to reproduce**

1. Create a managed project checkout and linked task worktrees.
2. Keep a settled prepared recovery source on a sibling task.
3. Keep unrelated remote work running.
4. Request an online import of an idle leaf checkout.
5. The existing owner keeps the shared source blocked.

**Paperclip version or commit**

This branch builds on merged #19642 (`7ac99718a4`) and fresh master
`fb3cf1a072`. It adds the Git closure. It does not add a writers-stopped
shortcut.

**Deployment mode**

Hosted native Runner with remote workspace storage. Local qualification
uses owned synthetic Git repositories, disposable PostgreSQL databases
and the actual archive and storage owners.

## What Changed

- Read bounded stable Git pointers, common-directory metadata and local
alternates. Derive the complete physical source ownership cohort.
Preserve all object stores.
- Retain common and borrowed store owners in preparation, lifecycle,
runtime, expiration and actual reopen paths. Keep company boundaries and
the existing 128-key bound.
- Hold exact sibling native recovery proofs under their existing
physical finalizer keys. Recheck descriptors, allocations, recovery
actions and source identities before capture and publication. Preserve
original states and warnings.
- Require leaf-first nested imports. Reuse the existing immutable child
download, restore and source comparison before parent exclusion.
- Bind the self-contained archive owner to the held Git namespaces. Keep
its three walks, history normalization, fsck, bytes and digest format.
Reject bare, detached, network, unknown or changed Git storage.
- Check both foreign path fields. Keep foreign metadata private.
Document the proof, bounds, source observation and remaining eligibility
limits.

## Verification

- Final revision `2934e41c923b93e16508b978c849b747c86be645`: all checks
are terminal, with 56 successful and four skipped, zero failures and
zero pending. The [required CI
aggregate](https://github.com/paperclipai/paperclip/actions/runs/42326464408/job/128266279210)
passed. Build, types, all server/Runner/serialized suites, end-to-end
aggregate and Snyk passed. [Fresh exact-head Greptile
review](https://github.com/paperclipai/paperclip/pull/19688#issuecomment-6699760104)
is 5/5 with no open threads. GitHub reports MERGEABLE and CLEAN.

- Review then found two correctness gaps: an ordinary managed directory
inside an unrelated host repository was blocked, and a Git parent could
miss a completed nested operation after its child snapshot became
missing or deleted. Three actual-owner regressions reproduced both gaps
in 9.70 seconds. The corrected owner passes 12 focused cases in 17.08
seconds. It imports and restores only the ordinary root bytes, preserves
the outer repository, rejects both missing/deleted child receipts before
a parent journal, and retains valid/changed child coverage and
private-cache contention. Its wrapper exited 0 and cleanup completed.
The first bootstrap attempt failed before tests ran because the host
exhausted its 32 shared-memory segment limit; no shared services were
stopped and no pass is claimed for that attempt. The final qualified
source/test type graph, module boundaries and diff checks pass.
- Receipt scans keep canonical path resolution. A lexical SQL prefix can
omit children reached through symlinked ancestors. The observed full
metadata closure completed in 1.17 seconds; no measured bottleneck
justifies weakening that alias proof. Writer admission still rejects
uncertain Git ownership before mutation. Expiration returns its typed
skip, and reopen returns its structured failure. Runtime and lifecycle
do not bypass an unknown source. No broad error-contract change was
added.

- Full CI at `720104ea975e` exposed five existing boundary cases across
three server shards. Four exact cases reproduced locally as failures in
25.98 seconds. Reopen now preserves its structured failure when source
inspection rejects the path, without reconstructing it. Runtime
recognizes only the exact regular ignore-only `.git/info/exclude`
fragment; migration still rejects that fragment. The original
arbitrary-path and file-URL import/delete-host/resume fixtures keep both
supported input forms and use the genuine offline setup path after their
sole writer finishes. The production mode gate is unchanged. A new
real-owner control proves the offline flag still rejects a running
legacy database writer.
- The final focused run passed all ten cases across four suites in 22.75
seconds, including all five CI failures, invalid-pointer admission,
fragment rejection, offline writer exclusion and actual nested reopen
lock ordering. Its wrapper exited 0 and cleanup completed. The known
invalid-pointer case remains fail-closed; the separate unexpected-I/O
diagnostic retains safe error metadata without command text or
credentials. Existing Git stderr redaction coverage remains intact. The
final metadata reader suite passed 20 cases in 2.68 seconds. The
qualified server/test source type graph, module boundaries and diff
checks pass. No production assertions or timeout defaults were weakened.
Fresh CI and review are required for the repair revision.

- A later review found that the ordinary early return still used only
one foreign path field. The actual ordinary providerRef-only regression
failed as incorrectly eligible. One source-path owner now enumerates
both local locators for company/foreign ordinary and Git scans, layout
checks, early admission, lifecycle/reopen/runtime and
expiration/move/preparation retention. Capture and display keep their
original preferred path. The actual owner canary passed 49 cases across
two suites in 109.15 seconds. It covers
cwd-only/ref-only/different/symlink locators, same-company mismatches,
foreign owners, lifecycle ordering, Git/recovery imports and
expiration/reopen. Three selected-locator fallback cases also failed
before their guard; the final eight-case control passed in 31.05
seconds, including the configured alias and valid import/private-child
controls. All wrappers exited 0 and completed cleanup. The
metadata/locator suite passed 19 cases. The final source/test type
graph, module boundaries and diff checks pass.

- Review found two P1 gaps. Both configured-directory symlink cases
failed before canonicalization. The actual parent import lost an
already-imported child's unique private cache and failed its archive
fence instead of retaining that owner. The corrected real-owner canary
passed four cases in 20.78 seconds: configured symlink runtime
provisioning, private-cache contention before the parent journal,
successful parent publication after release, and unchanged/changed child
coverage. The metadata suite passed 15 cases in 4.14 seconds. Both
wrappers exited 0 and cleanup completed.
- A fresh source read-only observation found 160 pending Git records and
nine common stores. An additional exact receipt query found 1,243
completed migration receipts and 557 nested source paths under pending
Git roots. All 126 records in the largest project were present in the
captured Git metadata set. For all 110 present candidate workspace
records, the derived closure fits the unchanged 128-key cap. The worst
case needs 110 issue keys and one project key, with 17 keys of headroom.
Its 173 imported children introduce no physical issue owner outside
those 110 logical task owners. This checks the observed key bound only.
It does not claim migration eligibility or a coherent snapshot.
- The initial stacked CI did not run four Runner test lanes: trusted
master passed new --package-lane/--native-server options to the old base
lane script. The branch is now rebased onto actual merged master and
retargeted before synchronization. It includes the canonical current
lane parser without an unrelated Runner patch. Fresh new-head CI is
required.

- The actual importer/source owner run passed 36 cases across three
suites in 187.75 seconds. It covers sibling prepared recovery, unrelated
running remote work, common-cache contention, changed
coordinator/allocation/path state during upload, exact nested coverage,
dirty unpushed work, and existing expiration/reopen owners. Wrapper exit
0 and cleanup completed.
- A separate previously hanging lifecycle-worker restore passed in 17.33
seconds with the unchanged 15-second test-body limit. An earlier run
timed out in three expiration/reopen cases and blocked cleanup; its
wrapper later exited 1. These failures remain recorded. They are not
counted as passes.
- Two boundary controls first failed: a foreign providerRef-only alias
was incorrectly eligible, and actual reopen did not admit its distinct
cache owner at the initial source boundary. The narrow guards now
inspect both foreign locators and derive physical Git keys before the
raw lifecycle transaction. The foreign alias and existing real nested
reopen contention passed. The private-cache control first retained
contradictory recorded source provenance and stayed fail-closed after
release. The corrected present legacy-worktree fixture passed in 12.61
seconds. These failed attempts remain recorded; they are not passes.
- The metadata reader and serialized/local archive controls use actual
Git repositories. They compare archive bytes and digests, remove every
original source, restore staged and unstaged work, and run fsck. The
metadata reader passed 13 cases in 2.85 seconds. The local/serialized
archive controls passed nine cases in 9.27 seconds. All processes exited
0. An additional actual importer control changed a relative commondir
pointer to the absolute name of the same store during upload.
Publication was rejected, the uploaded journal remained available, and
original files stayed intact (one case, 5.86 seconds; wrapper exit 0 and
cleanup complete).
- The final rebased full local server/test type graph passes with
db/shared/Runner source aliases because reused dependency declarations
predate new master exports. The first rebased attempt reported those
stale declaration fields/tables, and is not counted as a pass. Module
boundaries and diff checks pass. Standard package builds, types and
broad regression checks remain required in CI.
- No hosted import, admission hold, cancellation, provider mutation,
source filesystem write or shared-cache Git command was performed for
this change. The source observation is read-only and is not a locked
eligibility proof.

## Risks

- A missing writer owner could make physical source exclusion
incomplete. Existing mutation owners retain the same
company/task/project keys, and uncertain metadata fails closed.
- A project can contain a large cohort. Limits remain finite. A bound
failure keeps the source blocked.
- Bare, detached, unowned or network Git stores remain blocked online.
Missing archived sources retain their existing authenticated
reconstruction path.
- Exact recovery proofs preserve authority and obligations. They do not
infer provider termination or mark a prepared descriptor finalized.
- This change does not claim all observed Git records are immediately
eligible. Existing preparation, provider, recovery, runtime and
source-change blockers remain.

## Model Used

OpenAI GPT-6 through Codex, with reasoning, tool use and code execution.
The host does not expose a more specific backend model ID or
context-window size.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>
This commit is contained in:
DottaandPaperclip authored and GitHub committed 2026-11-06 20:08:19 -06:00
1 parent a894b11a2b
commit d5c4f7dd02
20 files changed
+1151 -134

No files matched your search

@@ -0,0 +1,90 @@
# Online Git source preservation
The online importer must exclude every writer of a selected checkout, its Git
administration, and all common or borrowed object stores. A checkout's task key
alone does not cover a shared project base or a private cache owned by another
task. Unrelated remote work must continue.
The source owner reads bounded, stable Git pointer, `commondir`, and alternates
metadata. It maps each physical directory to the existing company/task source
key or managed-project source key. It derives a transitive company cohort from
those owners and declared project sources. It rejects foreign physical aliases,
unknown stores, bare repositories, symlink metadata, network alternates, and
uncertain or excessive metadata. It preserves every object store; HEAD alone is
not a source closure. The bounds are 1,000 company records, 256 cohort records,
128 source keys, 32 borrowed stores, and four alternate levels. Foreign minimal
paths use the existing bounded scan of 256 records per page, at most 40 pages.
Configured directory aliases resolve through the existing canonical path owner
before Git inspection. Git pointer files and administration/object-store links
retain their strict checks. A configured symlink does not make a direct symlink
migration source admissible.
The runtime retains the existing plain-task-directory treatment of an exact
regular `.git/info/exclude` fragment. It is not an operational Git repository.
Git pointers and object stores retain their topology checks, and migration
rejects the fragment instead of treating it as an ordinary snapshot. A blocked
reopen returns its existing structured failure without attempting reconstruction.
Arbitrary self-hosted Git paths and file URLs retain the genuine offline
maintenance path; online migration requires a proven physical source owner.
Physical exclusion enumerates both `cwd` and `providerRef` through one source
path owner. Company/foreign alias checks, early admission, layout checks and
filesystem mutation owners use that set. A managed or Git source in either
locator requires a proven scope before the legacy fallback. Capture and display
keep their existing preferred-path selection. Missing archived sources retain
their existing reconstruction path.
Preparation, lifecycle, runtime-service, expiration, and reopen owners retain
the same physical keys before filesystem mutation or lifecycle exclusion. The
importer holds all cohort task/project and lifecycle keys through publication.
It also holds the existing per-run physical finalizer keys for each exact native
recovery proof. It re-inspects the complete source proof before capture and
publication. Changed descriptors, allocations, owners, aliases, or Git metadata
reject publication. Prepared sources and unresolved recovery warnings retain
their original state and evidence. They are not marked finalized or cleared.
Present nested workspaces must import before their parent. The existing owning
nested-import path downloads and restores each immutable child snapshot and
compares it with a fresh source capture before excluding that child. Missing or
changed coverage remains blocked. The archive retains its three source walks,
history normalization, Git/fsck checks, manifest ordering, and byte format. Its
self-contained remote function checks that every Git store belongs to the held
physical source roots. It preserves the original source files.
Ordinary capture does not read an enclosing host repository outside its root.
Its canonical ordinary proof remains valid there. Any `.git` marker at the
selected root still requires Git admission. A completed nested operation stays
in the parent layout guard unless the derived child proof covers that exact
operation and physical path. Missing or revoked snapshot coverage cannot turn
the leftover child into an ordinary parent directory.
Receipt path scans retain canonical resolution for aliases. A lexical SQL prefix
could hide a child reached through a symlinked ancestor. The finite receipt bound
remains in force; query optimization must preserve that alias coverage.
Imported child paths come from the same exact completed migration/snapshot
selector used by nested source verification. Their physical Git owners remain
in the parent fence after their database rows become remote. The selector bounds
company migration receipts to 10,240 and binds each child path, operation,
snapshot, digest and Git metadata proof. Current remote child execution does not
reacquire authority over those historical host files.
A source metadata observation at `2026-11-06T23:15:08.979Z` examined 160 pending
Git records and 120 unique paths. It found 92 Git pointers, 49 Git directories,
19 absent archived roots, and nine common/object stores. No alternate or uncertain
metadata was observed. Project cohort sizes were 126, 27, 4, 1, 1, and 1; the
largest contained 110 distinct task owners. These observations bound useful
cohorts. They are not a coherent snapshot, acquired lock proof, or claim that all
160 are eligible. Missing archived sources keep their existing authenticated
reconstruction path. Remaining provider, recovery, preparation, runtime, and
source-change blockers remain authoritative.
Focused verification uses the real importer and local storage owner. It covers
prepared sibling recovery, unrelated running remote work, private cache ownership,
source changes during upload, leaf-first exact child coverage, and existing Git
expiration/reopen consumers. Archive tests exercise both local calls and the
serialized remote function, compare archive bytes and digests, remove every
original source, restore uncommitted/staged work, and run fsck. No hosted import,
hold, cancellation, provider call, or shared-cache Git command is part of these
local tests.