Files
PaperClipAI/maps/feature-map/workspaces.md
T
06bc8db75c fix(runner): install verified API artifacts in execution workspaces (#18494)
## Thinking Path

> - Paperclip manages AI agents and their work.
> - Native agents use authorized API tools to read saved results.
> - Copying a large response as model-generated text can lose bytes.
> - The execution workspace needs an exact copy of the saved artifact.
> - This change copies the authorized bytes without shell API
credentials.
> - The runtime verifies size and SHA-256 before it reports success.
> - A live evaluation checks capture origin, installation and delivered
bytes.

## Linked Issues or Issue Description

**What happened?**

A native agent lost one byte while copying a 24576-byte API text page.
Existing capture and paging controls passed. The runtime had no verified
saved-artifact-to-workspace copy tool. A separate historical screening
refusal does not establish capture corruption.

**Expected behavior**

An authorized native run can install the complete saved artifact. The
stored and installed bytes must match in size and SHA-256. A conflicting
file must remain unchanged. A lost publication receipt must not cause an
unsafe repeated write.

**Steps to reproduce**

Run
`server/src/services/native-runtime/artifact-transfer.integration.test.ts`
on baseline `c46b8f35add9255fc71be294d27f5a7865bcb47f`. Its capability
assertion fails because `download_artifact` is absent. Apply this PR and
run the same regression with the filesystem and API controls. It
verifies 388131 synthetic bytes through the authority and installation
boundary.

Refs #18429. Related #17393, #17086 and #17285 cover paging, transfer
foundations and remote output reads. This PR adds the inbound
installation contract.

## What Changed

- Add `download_artifact` for supported standard-mode native runs, with
a 10 MiB limit.
- Reuse current company, task, run, asset-content and responsible-user
access checks. Check authority again before publication.
- Bind installation to the current execution lease. Stage through
confined Linux descriptors and publish with an atomic create-only link.
Keep API credentials out of execution commands.
- Verify bytes at storage and execution boundaries. Reject unsafe paths,
symlinks, different existing files, truncation and corruption. Reconcile
an unknown publication through verification only; an identical explicit
retry remains safe.
- Add API documentation, access-map entries and content-free audit
records.
- Preserve only the pinned Daytona SDK numeric HTTP status when failed
creation cleanup is unconfirmed. Keep provider exception messages and
headers private.
- Keep only the finished runner binary and portable provider pack in
their builder layers. Remove Rust targets/downloads and the Node install
workspace/store before export. Report tree sizes.
- Add one explicit Codex/Daytona Product E2E cell. Its independent
oracle checks the server capture audit, redacted tool receipts, stored
source, delivered attachment and browser link after reload. Uploaded
substitutes and one-byte loss fail.

## Verification

- Merge update at `6766b46553cab5b9787ade87a002eccaf50e68fc`: merge
current master and preserve both explicit suites in the catalog. Focused
catalog, image and oracle checks: 122 pass. Filesystem and authority
integration: 20 pass. Maps, module boundaries and whitespace checks
pass. New-head CI and Greptile are running. Full local build and
typecheck remain unavailable because cargo is absent.

- Behavioral red/green: the absent-tool regression failed before
implementation and passed afterward. The initial missing-package attempt
is recorded as setup failure.
- Linux filesystem and real PostgreSQL/authority integration: 19
filesystem tests and 1 authority test pass. Existing API authority,
capture and paging controls: 70 pass, 2 skipped. Provider transport in
these integration controls is simulated.
- Qualification calibration: 118 focused tests pass. The real
persisted-event redaction case and a byte-identical uploaded substitute
each failed before their oracle repair. The unchanged controls then
passed. Harness typecheck passes.
- Full recursive typecheck and build pass locally on the integrated
production source at `086f7a2901`. Later commits change only
qualification tests and documentation.
- Structural map/module checks and whitespace checks pass.
- Full local `pnpm test:run` was attempted by the implementation worker
and stopped under its 8 GiB memory limit, with no full-suite verdict.
Current-head CI must supply the full test result.
- The complete local Product E2E support suite has 13 failures. The
unchanged pre-qualification baseline reproduces all 13: host PostgreSQL
semaphore capacity, stale timeout/profile/workflow inventory
expectations, and expired fixed-date allowance fixtures. This is not a
full local suite pass.
- Earlier CI at `086f7a2901` failed a browser test during cleanup
cancellation with a database lock timeout. It remains recorded. All
browser shards passed on both later fully checked heads.
- Two qualification campaigns were canceled during setup before provider
execution while correcting the oracle. They provide no live product
verdict.
- Previous-head CI at `28fa001457f59df1c300c016dc569049e72d0061`: 53
checks pass, 3 intentional selection skips; no failed or pending checks.
- Real image/export regression: campaign 41272351780 compiled the Linux
runner and passed root/unprivileged CLI and transport checks, then
failed BuildKit export with `no space left on device` (exit 102,
filesystem 99.99%). No provider cell ran. Removing the Rust build trees
alone was insufficient: campaign 41273604815 also failed during
provider-pack export with exit 102. Commit `5bbb549c54` also
consolidates dependency installation and pack construction so the
disposable workspace/store can be removed in the same layer. All 122
image/oracle/catalog checks pass. Image and remote-pack publication then
passed in campaign 41275196741 after cleanup changes plus a scoped
60-to-120 GiB expansion of only its delete-on-termination builder disk.
The VM terminated and the expanded volume was deleted. This does not
establish that cleanup alone makes the default 60 GiB builder
sufficient.
- The Daytona admission probe then failed before agent creation, with
unconfirmed provider cleanup. The queued continuity case was canceled
because it shares this prerequisite. No artifact-transfer or continuity
success is claimed.
- Status-diagnostic regression: the original generic wrapper failed the
HTTP-status assertion. After the repair, the full plugin suite passes
501 tests, with 7 skips; its standalone TypeScript check passes.
Controls reject invalid status values, avoid accessors, preserve the
original create status and exclude credential canaries.
- Previous-head Greptile: 5/5; its summary names
`28fa001457f59df1c300c016dc569049e72d0061`. All review threads are
resolved.
- Live artifact transfer and three-turn warm continuity: NOT VERIFIED.
The diagnostic case in
https://github.com/paperclipai/paperclip/actions/runs/41277598289
confirmed Daytona HTTP 401 before agent creation. The configured
`DAYTONA_API_KEY` in the existing `runner-e2e-paid` environment needs
verification or refresh. The queued continuity control was not run after
its shared admission prerequisite failed. No further provider attempt is
planned until the credential changes. Unconfirmed provider cleanup
remains recorded.

## Risks

- Source-provenance changes can repeat the provider dependency install.
The combined build layer trades install-cache granularity for lower
retained disk use.
- Linux descriptor semantics and Node are required on the execution
host. Unsupported providers and restricted runtime profiles fail closed.
- Storage reads buffer at most 10 MiB. A provider outage can leave a
reserved staging file. Partial bytes never become the destination.
- A lost publication receipt can leave a complete file with an unknown
result. Recovery verifies the existing file without installation.
Conflicts are never overwritten.
- Artifact contents remain untrusted. The copy does not bypass screening
or authorize downstream execution.
- Live qualification uses the governed Codex API profile. It does not
establish every model, subscription profile, platform or historical
failure cause.

## Model Used

OpenAI GPT-6 in Codex assisted with diagnosis, code, tests and
qualification. The desktop runtime does not expose a more specific model
deployment ID or context-window size. The model used reasoning, source
inspection, tools and code execution. The native implementation agent is
configured with OpenAI `gpt-6.1-sol`.

## 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
- [ ] All Paperclip CI gates are green
- [ ] 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: devinfoley <139239+devinfoley@users.noreply.github.com>
Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-11-02 08:27:09 -06:00

227 lines
16 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
---
navigation: {"section":"Projects and execution","sectionOrder":40,"order":30}
---
# Workspaces, services, and files
A project or task can use a managed workspace with inspectable files, Git state, services, and runtime controls. Operators can follow its lifecycle through provisioning, use, and closure.
Feature ID: `workspaces`.
Implementation: [workspace list](../../ui/src/pages/Workspaces.tsx), [execution workspace](../../ui/src/pages/ExecutionWorkspaceDetail.tsx), [project workspace](../../ui/src/pages/ProjectWorkspaceDetail.tsx).
Agent access: [companion](../agent-access-map/workspaces.md).
## Sub-features
- `binding`: `workspaces/binding` — identify the workspace actually bound to a task/run.
- `provisioning`: `workspaces/provisioning` — observe readiness and failures instead of inferring readiness from a created record.
- `services`: `workspaces/services` — start/stop supported services and inspect logs and exposed endpoints.
- `files-git`: `workspaces/files-git` — inspect workspace files and repository state used for the deliverable.
- `closure`: `workspaces/closure` — close/reopen or reconcile through permitted controls without losing work.
For a historical failed export whose exact sandbox was destroyed, the task recovery
notice offers **Continue from saved sources** to an authorized board operator.
Select surviving task attachments, record lost work and external action outcomes,
then rebuild through fresh-session admission. The old failed run remains failed.
Bounded bootstrap retries retain this decision and recheck their system wake,
predecessors, saved files and current execution gates before provider startup.
Later cancelled runs require their own confirmed stop and board file repair.
Their evidence is preserved with the reconstruction and rechecked before dispatch.
Stop retains unfinished native files; startup also repairs missing historical
retention without replaying the cancelled turn.
Daytona stop-only cleanup confirms retention and accepts an already archived source.
Cancelled native file recovery can record a verified no-action outcome. Its
confirmation stays optional and resets when the source changes.
Retained sandboxes keep their existing export recovery path. Automated coverage:
[reconstruction](../../server/src/services/native-runtime/native-workspace-reconstruction.test.ts).
Live staging acceptance remains separate from these regression tests.
## How to get to it (user POV)
### `task-workspace`
Use the task’s workspace card or Files panel and follow its workspace link.
Native artifact-copy verification: in an admitted standard-mode task, save a
synthetic API response and call `download_artifact` with its saved identity and a
new workspace-relative filename. Independently inspect the file's complete byte
count and SHA-256 through the admitted filesystem. Retry the identical request;
reject a conflicting destination, symlink, revoked authority and foreign artifact.
The transfer does not change the content's untrusted status. See
[API tools](../../doc/runner-api-tools.md#verified-artifact-installation) for support
and limits. Provider-free regression evidence is not live remote qualification.
### `workspace-management`
Open `/workspaces`, a project workspace, or `/execution-workspaces/:workspaceId` and its Services/Configuration/Runtime logs views.
## Driving it with Playwright and the Paperclip CLI
Preconditions: follow the [baseline](./README.md#before-driving-a-journey). Enable isolated-workspace UI where needed. Use an owned disposable repository and environment; record workspace ID and branch.
### `task-workspace`
Automated: [workspace card](../../ui/src/components/IssueWorkspaceCard.test.tsx) and [task binding](../../server/src/__tests__/issue-runtime-workspace-binding.test.ts) cover host/binding contracts.
Manual: Run a task that writes a small file, follow the bound workspace, and inspect the file and revision. Reload and verify links still refer to the same workspace, not another checkout or a stale run.
For experimental native remote storage, follow the
[deployment prerequisites and explicit local import](../../doc/DEVELOPING.md#experimental-native-remote-workspace-storage).
With durable Postgres, S3, and a stable key configured, a new disposable
Paperclip Runner sandbox task with a supported fresh source enrolls by default.
Verify an explicit `remoteWorkspaceStorageEnabled: false` disables new enrollment,
and an existing host workspace or configured local folder stays on its legacy
path until explicitly imported. An enrolled workspace must keep remote authority
when enrollment is disabled and fail admission if persistence prerequisites are
lost. Write ordinary/binary files and, for Git, staged, unstaged, and unpushed
work. Finish the run and inspect its saved version. Replace the webserver and
verify the next task turn uses the same remote folder without a controller
checkout. Then release the saved allocation and verify a new one restores the
files, Git state, and managed same-project repositories without the old sandbox
or original Git remote. Read-only cross-project referenced folders must fail
admission. Separate workspaces may run concurrently; shared concurrent writers
must be rejected.
The Files panel must label **Live workspace** or **Saved copy** with its save
time. Confirm browse, preview, and download agree. A missing allocation with
unsaved work must report unavailability rather than show an old save as current.
Close must retain dirty/unfinished work and reject custom cleanup/teardown
commands; a clean saved workspace can release compute while retaining files.
Manual runtime commands, jobs, repairs, and service starts must reject remote
workspaces before local filesystem access; agent task execution remains supported.
The archive, storage, migration, retention, and lifecycle regression suites
exercise these contracts separately. This recipe records required deployment
checks; it does not claim a live provider or webserver replacement passed.
For a repository-free isolated task, complete an explicitly fresh retry, save a
marker, then send ordinary new input. It must retain the bound attempt directory
and marker. A later explicitly fresh retry must use a separate clean sibling.
The [production preparation regression](../../server/src/services/heartbeat/workspace-preparation.test.ts)
uses PostgreSQL and real private directories; it also rejects changed bindings.
The [directory boundary suite](../../server/src/services/isolated-task-directory.test.ts)
checks namespace, missing-directory and symlink refusal. Neither suite proves a
hosted provider turn.
For slow source preparation, configure a short PostgreSQL idle-transaction
timeout in a disposable instance. Hold preparation beyond that interval and
verify it retains its source lock, rejects competing preparation, and leaves
ordinary reads available. The [source-lock regression](../../server/src/services/isolated-remote-workspace-source-lock.test.ts)
uses real PostgreSQL; it does not prove provider startup or deployed recovery.
Create an authorized task whose selected checkout is also used by private historical
tasks and many old workspace records. Start and reuse it without a historical-source
permission failure or remote-clone fallback. Verify task-bound file browsing works,
while private historical task content and operation logs remain denied through app
APIs. Root traversal and symlink escapes must remain rejected. The OS and execution
environment own isolation between agents; Paperclip does not classify existing bytes
by the task that wrote them. Existing `workspace_source_restricted` failures remain
inspectable and recoverable, but new preparation does not produce that failure from
filesystem history. Current trust, credential, binding, and publication checks still
apply. The preparation-history benchmark supplies local service measurements, not
provider deployment qualification.
For reconstructable Git expiration, finish a disposable task tree, run the
internal lifecycle sweep after its 24-hour grace, and inspect candidate/reclaimed
bytes. Advance its upstream first; use the ordinary task/workspace reopen control
and verify the original commit, file contents, and subsequent persisted work.
Current runs, previews, recovery, unique files and shared base clones must stay.
The [expiration tests](../../server/src/__tests__/git-workspace-expiration.test.ts)
exercise real local Git and database paths, interruption and concurrent reopening;
cloud credentials and the deployed user journey require separate live evidence.
For task workspace selection, use the authorized task workspace API to choose a
task directory, project source, or existing workspace. Verify that a change
made during a run stays pending until the next normal admission. Submit another
choice using the returned revision, then retry the superseded request; it must
fail without replacing the newer pending choice. A projectless
task must retain its files across resume and reassignment. Use the repository
preparation API to add an authorized repository beneath that workspace;
repeat the request and verify that it reuses the checkout and pinned revision.
See [task binding tests](../../server/src/__tests__/task-workspace-binding.test.ts)
and [repository access tests](../../server/src/__tests__/task-workspace-repositories-trust.test.ts).
For native unfinished work, yield after preparation is staged and verify that
normal admission prepares the repository and continues the same task without
a timer or user message. A staging-only task that finishes must not wake again.
The saved continuation must disappear when the repository is ready or belongs
to another task. Source regressions cover those eligibility checks; a live
provider journey is required to qualify the complete automatic resume.
An existing projectless folder alias of a sandbox-restricted project must be
rejected for a standard local task before its adapter executes. Repeat after
relocating the configured source; retained ownership must still enforce the
boundary. This is separate from ordinary authorized file preview/download.
For a registered worktree selected for a sandbox or SSH environment, configured
and built-in provision scripts must not run on the controller during fresh
creation, reuse, or rebuild, including host runtime provisioning. Repeat with
standard trust and a protected read-only source, then with source write grants;
permissions must not change execution placement. Verify isolated source copying
and selected-environment dispatch still work. Repeat an explicit local isolated
source choice: copying remains available but controller scripts must stay absent,
including after binding or metadata changes and closed-worktree reopening. An
implicit trusted local project default remains a positive provisioning control.
Repeat with a newer configured default overriding an older primary source, then
with a newer primary overriding an older nonprimary source. Both configured and
built-in hooks must run for the implicit local default. Task creation and controller
classification share configured-default, primary, creation-time and ID precedence.
Explicit choices and saved copy-only markers still suppress hooks; changing the
default must not rotate an existing binding or clear its marker.
SSH keeps its explicit remote provision-command path; these checks
do not claim a new sandbox provision transport.
Low-trust closed-worktree reopening, including a board-comment reopen of a projectless task, must also suppress both kinds of host provision script. Returned checkpoints must reject file, symlink or missing parents before host writes, even if a nested file is marked unchanged. The retained host file and recovery evidence must survive rejection. These are automated safety contracts, not new live runtime acceptance.
An unchanged returned snapshot skips the host scan after acquiring the writer lock
and checking authority. Concurrent host edits remain untouched; changed files,
modes, symlinks and deletions still take the conflict path. The
[unchanged restore regression](../../packages/adapter-utils/src/workspace-restore-unchanged.test.ts)
checks both hooks under the lock. This is source-level coverage, not a cloud latency result.
For a child created by an admitted agent with an explicit existing workspace, verify that a matching sandbox boundary permits creation before the child has an ID. A changed child policy, unrelated parent, or forged creator run must fail. See the [child workspace trust regression](../../server/src/__tests__/task-workspace-child-trust.test.ts). This automated contract does not add live acceptance evidence.
### `workspace-management`
Automated: [workspace detail](../../ui/src/pages/ExecutionWorkspaceDetail.test.tsx) and [runtime service](../../server/src/__tests__/workspace-runtime.test.ts) cover UI and runtime contracts separately.
Manual: Provision a disposable workspace, wait for readiness, start an available service, open its endpoint, and inspect a real response and logs. Stop it and verify it actually stops. Exercise close/reopen only after inspecting pending changes and active runs.
## Gotchas
- UI visibility, default isolation policy, the local environment row, a workspace
record, and a running service are distinct prerequisites. A fresh instance
still needs an owned cwd/repository and runtime commands; `/workspaces` lists
existing work. Reopening depends on lifecycle, not a generic detail action.
- A link or running badge does not prove the service is reachable from the user’s browser.
- Workspace IDs, runtime leases, and repository paths are not interchangeable.
- Remote workspace snapshots require external Postgres, S3, and the original
encryption key. Edits after the last save still depend on the sandbox; there
is no periodic workspace backup. The experiment does not remove the instance
volume or establish zero-outage deployment for other data paths.
- Terminal workspace cleanup wakes after committed task/workspace/PR changes and
restores cooldown deadlines at startup. Empty lanes stop scanning. Display Git
results can lag by thirty seconds. Compatible records share one advisory Git
inspection per bounded page, while each record keeps its own delivery check.
Destructive cleanup starts a separate Git scan even if an older scan is running.
The [worker tests](../../server/src/__tests__/terminal-workspace-work.test.ts)
cover quiet lanes and deadlines. The
[service tests](../../server/src/__tests__/execution-workspaces-service.test.ts)
cover real PostgreSQL signals and retained worktrees. These tests do not
qualify an external wake alarm or a live provider deployment.
- Organizational project, source project, task files, personal `AGENT_HOME`, and
provider home have separate meanings. Private source ownership remains attached
to copied files after source configuration changes or workspace closure.
- Use [recovery](./recovery.md) for divergence/export repair and [CLI operations](./cli-operations.md) for local worktree tooling.
### Storybook verification
#### `closure`
Capability ID: `workspaces/closure`.
Scope: ui
Storybook: Close-readiness checks, pending changes/active runs, blocked/pending/error close, retry close and eligible reopening.
Production: Confirm owned workspace shutdown, preserved files/Git state and permitted lifecycle transition/reopening; Browser Use session closure is unrelated.
Workspace detail and operation-history reads do not inspect Git or remote storage. Delivery is unknown until explicit close readiness or lifecycle inspection. Verify that these reads work during an inspection outage, while destructive close still refuses unverified data.
Sources: Source: [ExecutionWorkspaceDetail.tsx](../../ui/src/pages/ExecutionWorkspaceDetail.tsx), [ExecutionWorkspaceCloseDialog.tsx](../../ui/src/components/ExecutionWorkspaceCloseDialog.tsx), [execution-workspaces.ts](../../server/src/routes/execution-workspaces.ts). Tests: [execution-workspaces-routes.test.ts](../../server/src/__tests__/execution-workspaces-routes.test.ts), [execution-workspaces-service.test.ts](../../server/src/__tests__/execution-workspaces-service.test.ts), [execution-workspace-reopen.test.ts](../../server/src/__tests__/execution-workspace-reopen.test.ts)