## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - Agents call connected apps, such as Google Workspace, through
Paperclip's tool gateway
> - Paperclip saves part of each tool result, approval previews, app
views and the full run transcripts that contain tool output
> - No user action deleted this data. Removing the connection left it in
place
> - Google's restricted-scope review requires that users can delete the
data an app obtained, and that the privacy policy says how
> - This pull request adds "Delete data read through this connection"
for any connection, on the connection's page and as an option when you
remove it
> - The benefit is that the picker text ("They stay there until you
delete them") is now true, and the privacy policy can link to a help
page
## Linked Issues or Issue Description
**Subsystem affected**
Connections (tool gateway audit tables, run logs, run events) and the
Connectors UI.
**Problem or motivation**
Paperclip stores content from a connection's tool calls in several
places and gives users no way to delete it:
- `tool_invocations.arguments_summary/result_summary` and
`tool_call_events.arguments_summary/request_summary/result_summary` keep
the first ~4,000 characters of each redacted value
(`summarizeToolValue`, `tool-access-policy.ts` `redactArguments`). No
retention limit.
- `tool_action_requests` keeps `canonical_arguments_summary`,
`preview_markdown` and `signed_arguments`. The linked approval card
(`issue_thread_interactions`) keeps copies in `payload.detailsMarkdown`,
`payload.toolAction.previewMarkdown`,
`payload.toolAction.argumentsSummaryJson` and
`result.toolAction.resultSummary`.
- `tool_mcp_app_views` keeps `tool_input` and `tool_result` for MCP app
views.
- Run transcripts hold full tool output: the run log (`local_file` with
an optional S3 mirror, or `s3_segments`/`s3_archive` object journals),
`heartbeat_run_events.message/payload`, and
`heartbeat_runs.stdout_excerpt/stderr_excerpt`. Neither run-log store
had a delete path.
- Connection removal (`removal.ts`) archives the connection and revokes
credentials. It deletes no content.
**Proposed solution**
A durable, resumable, company-scoped deletion request per connection.
See "Design" below.
**Alternatives considered**
- Delete synchronously in the request. Rejected: an interrupted request
leaves partial state that needs a person to retry it.
- Delete only the run events that mention the connection. Rejected: tool
output also reaches the model's later messages, so the safe unit is the
whole run transcript.
- Delete the audit rows. Rejected: counts, outcomes and hashes feed the
activity history and usage counts.
## Design
**Rule.** The job never drops a reference to content it has not
confirmed deleted. It is marked complete, and writes the
`tool_connection.content_deleted` activity entry, only after every step
succeeded. A step that cannot confirm its work throws. The request then
stays pending, the UI says "Paperclip keeps trying until it finishes",
and the worker retries it.
**Store inventory.** Every place that stores tool arguments, results or
previews. I found them from `summarizeToolValue`, `redactArguments`,
`writeAudit`, and every column or JSON field named like summary,
preview, arguments, result, toolInput or toolResult.
| Store | Content | Cleared by this PR | Proof |
| --- | --- | --- | --- |
| `tool_invocations.arguments_summary`, `result_summary`, `progress` |
Redacted request/result preview, provider progress | Yes: `summary`
becomes `""` with `contentDeletedAt`; `progress` null; size and sha256
kept | content-deletion test 1 |
| `tool_call_events.arguments_summary`, `request_summary`,
`result_summary` | Same previews | Yes, same | test 1 |
| `tool_access_audit_events.details.argumentsSummary` (`writeAudit`) |
Argument preview | Yes, same; decision, ids, size, hash kept | test 1 |
| `tool_action_requests.canonical_arguments_summary`,
`preview_markdown`, `signed_arguments` | Approval preview and arguments
| Yes, for settled requests | test 1, card retry test |
| Approval card in `issue_thread_interactions`:
`payload.detailsMarkdown`, `payload.toolAction.previewMarkdown`,
`payload.toolAction.argumentsSummaryJson`,
`result.toolAction.resultSummary` | Approval preview, arguments,
executed result | Yes. Cards are found from every settled request of a
batch, in the same transaction | test 1, card retry test |
| `heartbeat_runs.context_snapshot` and `agent_wakeup_requests.payload`
(`toolAction(s)[].resultSummary`,
`paperclipAgentMessage.untrustedToolResults[].resultSummary`) | Approved
action result delivered to the agent | Yes, for settled runs and settled
wakes | test 1 |
| `tool_mcp_app_views.tool_input`, `tool_result`, `model_context` | App
view input, result, shared context | Yes: rows deleted | test 1 |
| Run log: object journal, archive journal, original `local_file` log
and its legacy S3 mirror | Full transcript | Yes, every copy, for each
settled run that called the connection or received an approved result |
run-log-store tests, test 1 |
| `heartbeat_run_events.message`, `payload` | Transcript events | Yes,
except structural events (lifecycle, terminal, presentation, usage,
compaction, session/turn start, `native.*`) | test 1 |
| `heartbeat_runs.stdout_excerpt`, `stderr_excerpt`, and every string
`summary`, `result`, `message`, `text`, `output`, `content` or `body` at
any depth of `result_json` (including `nativeResult`, `acceptedResult`,
`semanticResult` and their `continuation`) | Output excerpts and final
answer | Yes; schema and control fields kept | test 1, canary |
| `approvals.payload` (formal board approval) | Tool name, risk and
hashes only | Not needed: no content | — |
| `tool_mcp_tasks` | Request hash and sealed provider task id; no result
| Not needed: no content | — |
| `tool_mcp_elicitations.questions` and their question cards | Questions
the provider asked a person | No. These are provider prompts, not tool
results, and the answered question stays in the task. Follow-up if
wanted | — |
| `tool_mcp_skills`, `tool_mcp_skill_files`, `tool_catalog_entries` |
Skills and tool descriptions the provider publishes | No. These are
provider definitions, not data read through tool calls | — |
| Transcripts of later runs that read an app view's `model_context` |
Shared app context | No. Nothing records which runs read it. The help
page says so. Follow-up | — |
| `issue_comments.body`, issue documents (including continuation
summaries), files, chat messages | Task content written by agents or
Paperclip | No, by design. The user removes it by editing or deleting it
or the task. The confirmation says in bold that anything agents wrote
into tasks is not removed, and so does the help page. The canary
excludes `issue_comments.body` by name for this reason | canary asserts
it is kept |
| Raw provider traces | Full provider traffic | No. Admin opt-in,
encrypted, expire within 24 hours | — |
| Feedback exports already sent | Run logs the user chose to share | No.
Already outside Paperclip | — |
| Content of runs of a deleted task or company | Same stores | No.
Follow-up | — |
**Durable request (schema change).** I added one table,
`tool_connection_content_deletions`, and migration `0490`. I followed
`doc/DATABASE.md`: edited the schema, exported it from
`packages/db/src/schema/index.ts`, ran `pnpm db:generate`, and ran the
migration drift and compatibility tests. The classifier marks it
`online_compatible` (a new table and its indexes only). No existing
table can hold a durable per-connection request with a cutoff time,
cursor, counts and retry state. Without it, an interrupted deletion
could resume only when a person pressed the button again.
**Resume after restart.** The worker uses the shared
`registerDeadlineWork` lifecycle on a new
`tool-connection-content-deletion` queue. The request insert signals the
queue in the same transaction. At server start the worker runs once,
`next()` finds any pending row, and processing continues from the saved
phase and cursor. Nobody presses anything. `next()` returns a deadline
only while a pending row exists, so an idle instance sends no recurring
query. A failure records `attempts` and `last_error` and retries after
30 seconds.
**One run selection.** `runsWithStoredContentSql` selects the runs whose
transcripts are deleted: runs that made a tool call or tool-call event
through the connection, and runs whose context lists one of the
connection's approved action requests, limited to runs that still hold a
log, an excerpt, result text or a content event. The dialog count, its
in-progress warning and the job's `runs` phase all use this one
function.
**Phases.** `invocations`, `events`, `auditEvents`, `appViews`,
`deliveredResults` (50 action requests per page; matching runs and wakes
are locked with `FOR UPDATE` before their documents are rewritten),
`runs` (the shared selection, 50 run ids per page, then one run at a
time), `done`.
**Canary test.** `tool-connection-content-deletion-canary.test.ts` puts
a unique marker in a tool call's arguments and result, an approved
action with a preview, a result delivered to a later run and its wake, a
native run that quotes the marker in its answer and events, a run log
migrated from a local file with an S3 mirror to an object journal, an
MCP App view and the audit tables. It checks that the marker is first
present in the database, the object store, the legacy mirror and the log
files. After the deletion it scans every `text`, `varchar`, `json` and
`jsonb` column of every table found through `information_schema` (no
hand-written table list), every object in both stores and every file
under the log directory. Any hit fails with the table and column or the
key. It runs three times: a complete deletion, a deletion interrupted
twice by storage failures and resumed, and two deletions for different
connections on one task at the same time. The only exclusion is
`issue_comments.body`, named with its reason.
**Batching and time budget.** Tool rows are read in pages of 200 with
keyset pagination on the existing `(company_id, created_at)` indexes,
bounded by the request time. Each page's writes and the saved progress
commit in one transaction. Counts are therefore exact, and a crash
repeats at most one page, idempotently. Runs advance one at a time from
the shared selection. Each 1,000-event page is its own step. Log removal
is its own step and runs before its progress transaction, so object
storage calls never run inside a database transaction. The 10-second
budget is checked after every step. Nothing changes on the agent boot or
tool-call path.
**In-flight work.** A run counts as settled only when its status is
terminal, `finished_at` is set, cost accounting is done, and no native
finalization is open. Calls, approvals, events, app views, delivered
results and transcripts tied to other runs or wakes are skipped and
counted. The confirmation says "N runs are still working and will be
skipped. Delete again after they finish."
**Permissions.** Board users with configure access (connection manager
or creator) can delete, like `DELETE /tool-connections/:id`. Agents,
viewers and other companies are refused. The activity entries
`tool_connection.content_deletion_requested` and
`tool_connection.content_deleted` hold counts only.
**UI reads.** The dialog loads counts once when it opens. While a
deletion runs, it polls `GET
/tool-connections/:id/stored-content/deletion`, a single indexed row
read, every 2 seconds. Polling stops when the job completes or the
dialog closes. The Delete button is not gated on counts.
## What Changed
- `packages/db`: new `tool_connection_content_deletions` table and
migration `0490` (online compatible: new table and its indexes only).
- `packages/shared`: `ToolConnectionStoredContent`,
`ToolConnectionContentDeletion` and counts types; optional
`contentDeletedAt` on `ToolRedactedValueSummary`.
- `server`: `RunLogStore.remove`, which deletes every copy of a run's
log (`composeRunLogStore`); `content-deletion.ts` service; deadline
worker wired into startup and pre-drain; `GET
/tool-connections/:id/stored-content`, `POST` and `GET
/tool-connections/:id/stored-content/deletion`, and `DELETE
/tool-connections/:id?deleteStoredContent=true`; OpenAPI entries.
- `ui`: "Delete data" row on the connection's Permissions tab with a
confirmation that shows counts, skipped runs, and "This cannot be
undone"; "Also delete data read through this connection" checkbox in the
Connectors removal dialog.
- Docs: user help
`docs/guides/board-operator/deleting-connection-data.md` (for the
privacy policy link) and an API note in
`doc/connections/CONNECTOR-PLAYBOOK.md`.
- `tests/runner-e2e/source-sdk-public-inputs.json` refreshed for the
shared type change.
`server/src/routes/tool-access.ts` (3,010 lines) and
`ui/src/pages/apps/Browse.tsx` (1,722 lines) grow by about 40 and 20
lines. The routes reuse that file's connection access checks; the dialog
reuses the existing removal confirmation. The new logic lives in its own
files.
## Verification
- `server/src/__tests__/tool-connection-content-deletion.test.ts`
(embedded PostgreSQL): content is gone from every "Yes" store in the
inventory; rows, hashes and sizes stay; another connection and another
company are untouched; a running run is skipped and is deleted after it
settles; a storage failure leaves the request pending with `retrying`,
and the next pass completes it; an approval-card update that fails once
is cleared on the retry; an unrecognised log backend keeps the pointer,
keeps the job pending and writes no completion entry; a zero time budget
yields after every step and still finishes; a repeat request returns the
pending one; agents and outsiders are refused; removal with
`deleteStoredContent=true` queues the deletion.
- `run-log-store.test.ts` and `object-run-log-store.test.ts`: `remove`
deletes local and mirrored logs and object journals, removes a migrated
run's original local file and mirror with its journal, refuses an object
log when object storage is not configured, and is idempotent.
- `ui/.../StoredContentDeletion.test.tsx`: confirmation text, counts,
skipped-run warning, and the request.
- Red: these tests fail on `master` because `remove`, the service and
the routes do not exist. Green on this branch.
- Ran: the tests above, `openapi-routes`,
`server-startup-feedback-export`, `tool-connection-removal`,
`packages/db` migration drift and compatibility tests, all
`ui/src/pages/apps` tests (950), `pnpm check:token-gates`, server and UI
`tsc --noEmit`. `pnpm verify` exits 2 (global CI input changed), so
required CI supplies broad coverage.
## Risks
- Deleting is permanent by design. The confirmation says so.
- A `local_file` log written on another replica's local disk stays on
that pod's disk until the pod is replaced; its S3 mirror is deleted. New
logs on S3-configured instances use object journals, which are deleted.
- On versioned buckets, earlier object versions remain until the
bucket's lifecycle policy removes them (the existing journal delete path
does not remove versions).
- Not in this PR (follow-ups): the "No" rows of the inventory (except
agent-written task content, which is out of scope by design), mainly the
same deletion when a task or a company is deleted, and later runs that
read an app view's shared context.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, via Claude Code with
tool use and code execution.
## 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
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Apps, Connections, and Integrations
Audience: internal engineers and product contributors working on integrations.
Start here when adding a provider: Connection authoring runbook. It is the canonical agent tutorial from provider research and protocol classification through manifest generation, branding, secrets, deterministic tests, real-account proof, and PR submission.
Chat and email channels: conversation-channel contract, question inventory and quality bar, and per-provider question coverage index.
Runtime authentication: AI Connections.
MCP protocol versions (2026-07-28 stateless MCP and older handshake servers): MCP-PROTOCOL-VERSIONS.md.
MCP extensions: MCP Apps (interactive views in the task chat), MCP Skills (skills that connected apps publish, reviewed by the board and delivered to agents), MCP Tasks (long provider calls that survive timeouts, restarts, and deploys), MCP elicitation (an app's question to the person, answered in the task chat), and Machine sign-in (OAuth client credentials for organization connections where no person signs in).
Machine sign-in (OAuth client credentials for organization connections where no person signs in).
Long-term memory: Experimental memory connectors.
Provider notes: Google Workspace, Gmail, Asana, Sentry, PostHog, Braintrust, Neon, Superagent, Gauge, Ditto, Circleback, 1Password, Unikraft, AgentMail, Telem.AI, iMessage Photon, Enterpret, Supabase, and Vercel. Optional credential custody: Vercel Connect.
Post-read action: classify a new integration request, pick the right Paperclip layer to change, and avoid creating a parallel connection framework.
Decision Record
Board decisions from PAP-13211 make the Apps v2 substrate on the PAP-10341 branch canonical:
- D1: Apps v2 is the substrate. The active model is
tool_applications,tool_connections, catalog entries, profiles, policy rules, action requests, gateway sessions, audit events, and runtime slots. Connections v1 is retired as an implementation path. - D2: one credential authority per connection, brokered projections. The
default is the Paperclip instance vault: durable third-party credentials live
in
company_secretsas secret refs. A reviewed remote MCP method may instead opt in to Vercel Connect, in which case Vercel is the durable credential authority and Paperclip stores only the connector reference and redacted grant metadata. A connection must never mix those two sources. Adapter config, plugin config, harness credential files, and run environments may receive only brokered or projected credentials. - D3: the vocabulary and three-door IA are product law. The default product doors are Apps, Connections, and Review. Protocol and operator-depth concepts live behind Developer or Advanced surfaces.
- D4: unification lands on PAP-10341. Pages, CircleBack-style harness MCP OAuth, provider gallery work, and plugin-provided integrations converge on this branch instead of spawning new integration substrates.
- D5: inbound stays thin. External clients that call Paperclip use scoped Paperclip tokens and existing profiles/rules. They do not get a separate permission model.
Canonical Object Model
Use connection as the unifying noun. A connection is four things:
- A stored credential reference.
- A capability catalog.
- A governance layer.
- An audit trail.
Everything else is an axis on that object:
| Axis | Values | It answers |
|---|---|---|
| Direction | outbound, inbound | Who is the client? |
| Transport | MCP, native REST/OpenAPI, OAuth app install, webhook | How do bytes move? |
| Auth mode | OAuth, API key/PAT, app installation, none | What does the secret represent? |
| Credential owner | company, user, run | Whose identity acts? |
| Packaging | catalog entry, plugin, skill | How does it ship? |
MCP is a transport, not a product category. "Install the Discord app", "connect Google Drive", and "add an MCP endpoint" all produce governed connections with different transport/auth values.
Layer Stack
When you are unsure where a change belongs, place it on the narrowest layer that solves the problem:
| Layer | Owns | Examples |
|---|---|---|
| Surface | user-facing Apps, Connections, Review, Developer/Advanced screens | gallery cards, setup wizard, review queue |
| Governance | profiles, bindings, allow/ask-first/block rules, quarantine, audit | read-only profile, ask-first write policy |
| Capability | action catalogs, schemas, risk classes, changed-tool review | search_issues, create_comment, schema hash |
| Credential | company_secrets, OAuth broker, credential resolver, token broker |
Slack bot token ref, Google OAuth refresh token ref |
| Identity | actor attribution and token exchange | board user, agent run, first-party service identity |
| Transport | how the external system is reached | remote HTTP MCP, local stdio, REST/OpenAPI, webhook |
The agent should not hold a durable provider credential. It should hold a Paperclip run/session token; the server or broker resolves the connection, checks governance, invokes the provider, and writes audit.
Identity vs. connections
Signing a user in and connecting a resource are different planes with different owners, different token profiles, and different homes. Do not merge them. This section is the public, connections-side statement of the identity model so connector implementers inherit the rule without depending on private identity-service documentation or re-deriving it.
| Plane | Question | Lives where | Token profile |
|---|---|---|---|
| P1. Sign-in methods | Who are you? | paperclip-id (id.paperclip.ing → Account) |
Minimal-scope provider tokens (openid email profile), used once to authenticate, encrypted at rest, never exported |
| P2. Connections (Apps) | What may your agents touch? | Paperclip App instances (tool_connections), acquired via the connect broker for hosted + self-hosted |
Rich-scope, long-lived resource tokens in the instance's encrypted vault; per-agent grants; risk-tier policy defaults |
| P3. Login with Paperclip | Who may authenticate against us? | paperclip-id OIDC provider + DB-backed client registry |
Our ES256 ID/access tokens issued by us to registered RPs (instances, the broker, future third parties) |
Everything in doc/connections/ — the First-30 matrix,
the connection authoring runbook, and the connect-broker work —
lives on plane P2. It never acquires, stores, or brokers a P1 sign-in token.
The standing rule (D7)
Adopted as a standing rule (decision D7) with the identity-model plan. State it verbatim in any P2 design so the app-store work cannot drift into merging the planes:
Sign-in tokens are never reused as resource tokens; id.paperclip.ing never stores resource tokens; no connections hub on the ID service.
P2 tokens flow broker → instance vault as pass-through only; the id.paperclip.ing Account page therefore must not grow a "Connections" hub. The reasons to hold the planes apart (from the plan §3):
- Scope discipline. Sign-in wants the narrowest grant; connections want deliberately broad ones. One button that does both is how you grant repo access just to log in.
- Blast radius. id.paperclip.ing holding every customer's Vercel/Slack/GitHub resource tokens would make it the single juiciest target in the fleet; the broker is intentionally pass-through.
- Self-hosted symmetry. Instances own their vaults, so self-hosters don't depend on our uptime to use their own connections.
- Legibility. Sign-in and connections answer different user questions, and every product we benchmarked (Vercel, Railway, GitHub, Google) keeps them on separate pages with separate names.
The explicit Vercel Connect exception does not change D7 or merge P1 and P2.
The operator chooses Vercel as the P2 credential authority for an individual
connection. id.paperclip.ing is not involved, and neither sign-in tokens nor
provider tokens pass through it. The deployment's Vercel access token or
workload OIDC identity is bootstrap authority for that external vault, not a
provider resource credential.
Naming alignment
Use the surface-correct name for each plane; they intentionally differ:
| Surface | Plane | Name to use |
|---|---|---|
| Paperclip App instances | P2 | "Connections" |
| id.paperclip.ing Account | P1 | "Ways to sign in" |
| id.paperclip.ing admin | P3 | "OIDC clients" (until the app store productizes it) |
Packaging Rule
Default to a catalog entry when an integration can be described as metadata: manifest, auth config, action catalog, resource filters, and policy defaults.
Use a plugin only when the integration needs product code such as custom UI pages, its own tables, workers, migrations, routines, or specialized ingestion. A plugin may bundle catalog entries, but it must not bypass the connection, profile, policy, credential, and audit model.
Use a skill for agent instructions. Skills may use connections; they must not own durable tokens.
Canonical Docs
- Glossary defines product and internal terms.
- Identity vs. connections is the public statement of the P1/P2/P3 boundary and the D7 standing rule for connections work.
- Security threat model harvests the keeper from PAP-2359 and maps it onto Apps v2.
- First-30 matrix harvests the keeper from PAP-2432 and is the source matrix for connector playbook work.
- Connecting any remote MCP server is the baseline: how an operator connects a standards-compliant remote MCP endpoint with no Paperclip code change, and how sign-in resolves a client.
- Connection authoring runbook is the one end-to-end, agent-executable guide for adding a vendor as a catalog entry on Apps v2: research, connection-type selection, OAuth/API-key/generated-URL setup, encrypted credential handling, branding, implementation, browser and live-provider testing, verification, and PR submission.
- Vercel Connect operator guide documents the optional external credential source, deployment flags, runtime resolution, recovery, and smoke requirements.
- MCP access governance remains the operator runbook for the current gateway, profile, policy, approval, runtime, and audit APIs.
Migration Notes
Connections v1 contributed useful policy, UX, and rollout thinking, but its
implementation branch is no longer the target. When you see old tickets or code
using connections, connection_grants, or a provider-directory mental model,
translate the intent into Apps v2:
| Connections v1 intent | Apps v2 home |
|---|---|
| Provider directory | Apps gallery / tool_applications |
| Configured provider instance | Connection / tool_connections |
| Grant allowlist | Profiles, profile bindings, policies |
| Resource filters | Policy/profile conditions plus provider config |
| Tool broker | Tool gateway and runtime supervisor |
| Connection UX tail | Apps, Connections, Review, Developer/Advanced IA |
Do not add new work to the retired v1 branch. If an old ticket still describes a valid product gap, retarget it to an active Apps v2 issue or close it as superseded with a link to the replacement.
For Slack chat bot creation, installation, recovery, and live qualification, see Automatic Slack app setup.