Files
PaperClipAI/doc/AGENT-ARTIFACTS.md
T
Devin FoleyandPaperclip 6a43890326 Fix image attachments in remote native workspaces (#20939)
## Thinking Path

> - Paperclip manages AI agents and their work.
> - Native runs must receive the files that users attach to their
messages.
> - Remote wake preparation returned file metadata without readable
bytes.
> - The historical chat reader also rejected every remote target.
> - Command-backed sandboxes and SSH targets already provide a bound
command transport.
> - This pull request stages verified files through that transport and
clears them after the turn.
> - Agents can then open attached images with their normal image reader.

## Linked Issues or Issue Description

Refs #19740. This change covers SSH and command-backed sandbox
execution, including Daytona. Plugin Runner tasks without a command
transport still need a separate versioned file-transfer contract. This
PR does not close that broader issue. The open attachment-recovery PR
#20778 concerns provider-session attachment journals, not input files.

**What happened?**

A user uploaded an image and sent it to an agent. The remote native run
received its filename and size, but no readable file. Wake preparation
reported `remote_workspace_staging_unsupported`. The historical read
fallback also refused remote execution.

**Expected behavior**

Copy the exact admitted attachment into the bound remote workspace. Let
the provider read the image. Preserve company, task, source-comment,
size, and hash checks. Clear temporary file contents when the turn ends.

**Steps to reproduce**

1. Select a native agent on a command-backed sandbox or Linux SSH
environment.
2. Attach a PNG to its current task or in-app chat message.
3. Ask the agent to inspect it.
4. Check that wake preparation supplies a readable workspace-relative
image path.

## What Changed

- Add a remote staging owner with bounded commands, stdin transfer, hash
verification, opaque paths, and safe image extensions.
- Confine Linux filesystem operations through directory descriptors.
Refuse symlinks and replacement files during cleanup.
- Bind current wake files and authorized historical external-chat reads
to the selected lease transport.
- Join active writes during cancellation. Retain failed cleanup work
until scope close.
- Tell providers to open image files with an image-capable reader.
- Handle stdin stream errors in the shared SSH transport. Failed uploads
return failure instead of raising an unhandled EPIPE or reporting
success.
- Add database, transport, cancellation, integrity, cleanup,
executor-wiring, and Linux filesystem regressions. Document support and
limits.

## Verification

- Prepared an isolated source checkout with `node
scripts/agent-preflight.mjs --skip-rust-debug`.
- Focused suites: `pnpm exec vitest run
server/src/services/native-runtime/remote-attachment-staging.test.ts
server/src/services/native-runtime/native-runner-file-handoff.test.ts
server/src/services/native-runtime/chat-attachment-read.test.ts` — 74
passed, 4 Linux-only tests skipped.
- Executor checks: `pnpm exec vitest run
server/src/services/native-runtime/native-session-executor.test.ts -t
'attachment staging to the remote|stages from the authenticated|verifies
frozen staged admission'` — 7 passed.
- Server typecheck: `pnpm --filter @paperclipai/server typecheck`,
followed by the server layered check after the final changes.
- `git diff --check` passed. The local secret, PII, and
private-reference diff scan found no matches.
- `pnpm verify` returned exit 2 because the documentation change
requires broad CI. No full local suite is claimed.
- The four Linux filesystem tests were skipped locally because this
macOS host has no available Docker daemon. Linux CI executed all 17
remote-staging tests successfully, including those four confinement
tests: [server shard
10](https://github.com/paperclipai/paperclip/actions/runs/43102035096/job/130796646266).
- SSH regression: the real pipe test reproduced two unhandled EPIPE
errors before the fix, plus false success when the peer exited with code
zero. The fixed SSH tests and existing argv tests pass (12 tests). The
remote-staging transport suite also passes (13 tests; 4 Linux tests
require CI).
- `pnpm --filter @paperclipai/adapter-utils typecheck` passed after the
SSH change.
- Final revision `1c5a9288b8d5f267ddeac84638ff6075dbbbf8ec`: all
applicable CI jobs passed, including typecheck, build, general and
serialized tests, Runner checks, canary dry run, and all eight browser
shards. [CI
run](https://github.com/paperclipai/paperclip/actions/runs/43102035096).
Linux CI also passed all 6 native SSH tests: [server shard
5](https://github.com/paperclipai/paperclip/actions/runs/43102035096/job/130796646349).
- Greptile reviewed that final revision at 5/5 with no remaining
findings. The review thread is resolved. GitHub reports the PR mergeable
with a clean merge state.
- No live model turn or deployed release is claimed.

## Risks

- Remote staging requires Linux, Node, and a command transport bound to
the selected lease. Plugin Runner tasks without that transport retain
their unavailable result.
- Each file uses bounded independent commands. Slow transports can delay
preparation or cleanup. File bytes stay out of command arguments and
provider-facing error messages.
- An abrupt controller or remote-host failure can bypass cleanup.
Ephemeral sandbox disposal remains the backstop. Retained files are
never blindly overwritten or deleted.
- Previous-release provider sessions and checkpoints retain their format
and identity. New turns receive newly staged input files. Existing
restart-admission checks still apply. No migration, runner-image change,
or account-credential transfer is required.

## Model Used

OpenAI `gpt-6.1-sol` through the Codex harness in T3 Code, with xhigh
reasoning, repository inspection, code execution, and tests. No
subagents were used. The context-window size is not exposed by this
runtime.

## 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>
2026-11-11 14:12:09 -08:00

304 lines
14 KiB
Markdown

# Agent Artifact Upload Workflow
Generated files that a board user or reviewer should inspect as deliverables
must be attached to the Paperclip issue before the agent chooses a final
disposition. A local workspace path is not enough, because cloud users and
reviewers often cannot access the agent's disk.
## Read attached files in native runs
Native runs receive workspace-relative paths for the exact files admitted in
their current wake. On Linux SSH and command-backed sandbox targets, Paperclip
copies verified bytes through the bound lease transport before the provider
turn starts. The agent needs no Paperclip API key to open those paths. Image
inputs must be opened with the provider's image-capable file reader.
The same transport supports authorized historical external-chat reads. Their
existing conversation and source-comment checks still apply. External-chat
history tools do not grant access to unrelated tasks or in-app conversations.
Targets without a command transport still report attachment access as unavailable.
Temporary slots are isolated per staging operation. Cleanup clears the verified
file after the turn and refuses symlinks or replacement files. Abrupt controller
or remote-host failure can interrupt cleanup; sandbox disposal remains the
backstop for ephemeral environments. Retained provider sessions and checkpoints
keep their existing format and identity across this change.
## Upload files from a task
People can open the task's Artifacts tab and use the upload icon, or drop files
anywhere in the panel, including its header and existing artifact cards. Uploads
save directly to the task. They do not send a message, assign an agent, or
release a pause hold. Files from people and agents appear together in Artifacts.
Markdown imports create task documents. Repeated filenames get a new document
key instead of replacing an existing revision.
Hidden system keys and artifact-review prefixes are reserved; imports with
those filenames receive a visible upload document key. Other supported files use
the existing attachment API and its type and size limits. Failed uploads show
an inline Retry action while successful files remain available.
## Native runner
When `register_deliverable` is available, use it for files in the bound local or
remote workspace. Supply a workspace-relative `contentRef`, basename `filename`,
`contentType`, exact `byteSize` and SHA-256, `title`, and a stable `idempotencyKey`.
The tool verifies the file, stores an attachment and artifact work product, and
binds it to the response. Generic API tools and a legacy API key are unnecessary.
`register_deliverable` stores the file on its current task. A completion report
can also cite a file on a direct child task that is done, because the child's
own completion verified that file. A parent that delegated a file to a child
links the child's file and does not register a copy. The child's file does not
count as delivery of a file that the parent's own request names.
Named JSON file outputs and explicit publication requests require the same
verified delivery evidence as other file outputs. A saved summary document does
not replace a requested downloadable JSON file. Reading JSON inputs or discussing
their contents inline does not by itself require a file attachment.
JSON output intent is recognized from a direct filename or an explicit output
destination. Ambiguous descriptions of input files joined by `to` remain text
requests; use `as`, `into`, or an explicit output file when an artifact is required.
A `Deliverable:` or `Deliverables` label also states the output. A file named
on that line, or in the list directly below it (blank lines between items are
allowed), is a requested file. This includes relative paths such as
`dist/index.html` and web page files such as `index.html` and `styles.css`.
Items that describe a repository change, pull request, commit, or branch do
not count, nor do files named as inputs (`a comparison of input.csv and
baseline.csv`), files the item forbids (`do not create report.pdf`), files to
be edited in place (`update styles.css`) unless the item also asks to publish
them, files to be written inline in the reply (`index.html and styles.css inline
in the reply`), or items that say no files are created. An inline or no-upload
instruction about another output in the same item does not cancel a named file
(`index.html and a short summary in the reply`).
The completion gate permits a narrow sentence-bound exception for one requested
file when the following sentence explicitly qualifies it, for example:
`Write memory.txt. This is personal memory, not a task deliverable.` or
`Write internal-proof.txt. This is an internal verification file, not a deliverable.`
Each qualifier applies to one file in the immediately preceding request sentence.
Separate file requests can each qualify when each has its own following qualifier.
Multiple files requested in one sentence, unqualified report requests, and explicit
attachment/download requests still need verified accessible publication evidence. Saying only “for internal verification”
does not establish this exception. Expected denied native writes in a negative
test are separately recognized as denied attempts, not requested deliverables.
Wait for the receipt. It includes `attachmentId`, `contentPath`, and
`downloadPath`, along with the existing command, revision, entity references,
and disposition. Reuse the original key after an ambiguous result. A receipt
confirms storage and response binding in Paperclip; it does not confirm delivery
to an external chat provider. If registration fails, use the returned error to
resolve the failure or explain the limitation; do not describe a workspace path
as an uploaded file.
With workspace-relative file access, create or copy task deliverables into the
directory returned by `get_workspace.cwd` before registration. `AGENT_HOME`
holds personal instructions and memory. If registration reports a missing
workspace file, copy the file to that task directory and retry with a relative
`contentRef`.
## Legacy adapters
Use Bash to run the helper bundled with the Paperclip skill from the repo root; installed skill files may not retain executable permissions:
```sh
bash skills/paperclip/scripts/paperclip-upload-artifact.sh path/to/output.webm \
--title "Walkthrough render" \
--summary "Rendered walkthrough for review"
```
The helper uses the authenticated Paperclip API from the current heartbeat
environment:
- `PAPERCLIP_API_URL`
- `PAPERCLIP_API_KEY`
- `PAPERCLIP_COMPANY_ID`
- `PAPERCLIP_TASK_ID`
- `PAPERCLIP_RUN_ID`
It uploads the file to
`POST /api/companies/{companyId}/issues/{issueId}/attachments` and creates an
artifact work product on `POST /api/issues/{issueId}/work-products` by default.
The command prints issue-safe markdown links for the final task comment.
## Task artifact presentation
Existing and newly arriving agent attachments, work products, and documents add
the task's Artifacts tab without selecting it, opening the side panel or mobile
drawer, or changing the current document/file link. If the pane is closed, the
tab is available when the user opens it. This uses stored object IDs, so it works
with either runner. Uploading a file and registering its work product counts as
one arrival. Revisions and repeated query refreshes preserve dismissed tabs and
the user's selection. Plans retain their existing Plan-tab behavior;
unregistered user input attachments remain in the conversation.
## Uploaded Artifacts vs Workspace Files
Use uploaded artifacts for deliverables: videos, PDFs, screenshots, archives,
reports, rendered HTML, or any file the board should inspect without needing the
agent's checkout. Attachment-backed artifact work products set `type` to
`artifact` and `provider` to `paperclip`, with metadata canonicalized from the
uploaded `attachmentId`.
Use `workspace_file` metadata only for important files that intentionally remain
in a project or execution workspace, such as source files, committed markdown
plans, or generated files whose meaning depends on the checkout. Workspace-only
references are useful signposts, but they are not durable uploads.
Expected work product metadata shape:
```json
{
"resourceRef": {
"kind": "workspace_file",
"issueId": "<issue-id>",
"workspaceKind": "execution_workspace",
"workspaceId": "<execution-workspace-id>",
"relativePath": "doc/plans/example.md",
"line": 1,
"column": 1,
"displayPath": "doc/plans/example.md:1:1"
}
}
```
`workspaceKind` is `execution_workspace` or `project_workspace`. `line` and
`column` are optional. `relativePath` must be relative to that workspace root;
do not store host-local absolute paths as workspace references.
Workspace file links resolve only inside registered Paperclip workspaces. The
default target is the current issue's execution workspace first, then its
project workspace. A link may target another same-company project workspace only
when it carries both that `projectId` and `workspaceId`. Paperclip does not
resolve arbitrary machine-wide filesystem paths, absolute host paths, home
paths, or relative paths that escape the selected workspace.
## Completion Pattern
When a task produces a user-inspectable deliverable file:
1. Generate and verify the file locally.
2. Upload it with `skills/paperclip/scripts/paperclip-upload-artifact.sh`.
3. Keep the artifact work product unless the file is incidental; pass
`--no-work-product` only for supporting files that should not be promoted.
4. Link the printed attachment URL in the final issue comment.
5. Then set the final issue status.
For a response that is explicitly intended for an external chat conversation,
also pass each intended file with `paperclipai issue comment --attachment-id
<id>`. Paperclip binds only those exact uploaded files to that comment; other
task attachments remain internal.
Final comments should name and link the uploaded artifact or work product, not
just the local filesystem path. For workspace-only files, include the work
product title and recorded relative path. Local paths can be included as
diagnostic context, but they cannot be the only access path. Browse/search is a
fallback for recovering workspace files when the issue link or chip is not
available, not the preferred way to deliver files to users.
## Video Examples
Upload an `.mp4` render:
```sh
bash skills/paperclip/scripts/paperclip-upload-artifact.sh dist/demo.mp4 \
--title "Demo video render" \
--summary "MP4 render for board review"
```
Upload a `.webm` render:
```sh
bash skills/paperclip/scripts/paperclip-upload-artifact.sh out/walkthrough.webm \
--title "Walkthrough video" \
--summary "WebM walkthrough render"
```
The helper detects `.mp4`, `.webm`, and `.mov` content types. If a renderer uses
an unusual extension, pass the MIME type explicitly:
```sh
bash skills/paperclip/scripts/paperclip-upload-artifact.sh render.bin \
--title "Demo video render" \
--content-type video/mp4
```
## Direct API Pattern
If the helper is unavailable, use the same API shape:
```sh
curl -sS -X POST \
"$PAPERCLIP_API_URL/api/companies/$PAPERCLIP_COMPANY_ID/issues/$PAPERCLIP_TASK_ID/attachments" \
-H "Authorization: Bearer $PAPERCLIP_API_KEY" \
-H "X-Paperclip-Run-Id: $PAPERCLIP_RUN_ID" \
-F 'file=@"dist/demo.mp4";type=video/mp4'
```
Then create a work product when the uploaded file is the deliverable:
```sh
curl -sS -X POST \
"$PAPERCLIP_API_URL/api/issues/$PAPERCLIP_TASK_ID/work-products" \
-H "Authorization: Bearer $PAPERCLIP_API_KEY" \
-H "X-Paperclip-Run-Id: $PAPERCLIP_RUN_ID" \
-H "Content-Type: application/json" \
--data-binary @artifact-work-product.json
```
Use `type: "artifact"`, `provider: "paperclip"`, and metadata containing the
uploaded `attachmentId`. The server canonicalizes `contentType`, `byteSize`,
`contentPath`, `openPath`, `downloadPath`, and `originalFilename`.
The optional `executionWorkspaceId` on work-product create and update requests
must identify an execution workspace in the same company. A project workspace
ID is a different identifier and cannot be used here. Omit the field when no
execution workspace is available, or send `null` to clear an existing link.
Invalid references return `422` without changing the work product or the current
primary product.
## Verification
The file-delivery integration suite runs the real helper through queue and
HTTP/2 gateways against a disposable API, database, and storage. It also tests
native registration with generic API tools disabled, duplicate retries, Unicode
filenames, company isolation, and downloads after deleting the workspace.
```sh
pnpm exec vitest run server/src/__tests__/file-delivery-bridges.test.ts
```
To run the same suite on disposable Daytona sandboxes, install the standalone
Daytona plugin's dependencies and set `DAYTONA_API_KEY` in the test process:
```sh
PAPERCLIP_FILE_DELIVERY_DAYTONA=1 pnpm exec vitest run server/src/__tests__/file-delivery-bridges.test.ts
```
The live fixture deletes each sandbox before checking that its attachments
remain downloadable from Paperclip. It does not run unless explicitly enabled.
## Text attachment previews
In the task chat layout, select a text attachment in Artifacts, a work-product
card, or a chat attachment chip to open a named right-side tab. Reopening the
same attachment focuses its existing tab. Tabs can be switched and closed.
On mobile, the same viewer opens in the task details drawer.
Markdown work products keep their expandable review document, annotations,
revision indicator, and document links. **Open in tab** is a separate action.
Work-product cards keep **Download** as a direct original-file download.
Text file cards provide **Open in tab** beside their existing actions. CSV cards
keep this action before and after loading their data preview.
Markdown attachments offer **Rendered** and **Raw** views. Other supported text
files display literal text. Image references and diagram source remain inert;
opening a preview does not load attachment-selected media URLs. The viewer
provides a download action. Preview reads
are limited to 512 KiB; oversized, unsupported, or unavailable files show an
explicit fallback instead of attempting an unbounded render. A failed read can
be retried. Workspace files continue to use the existing workspace file viewer.
Storybook: **Tasks / Text file tabs** covers opening from Artifacts, Markdown,
plain text, empty files, oversized files, missing attachments, and a narrow panel.