PaperClipAI/doc/runner-api-tools.md

# Runner API escape hatch

`search_api` and `call_api` extend the native runner when an available dedicated
operation cannot express the requested work. Use `set_task_monitor` to schedule or clear a one-shot task check; after confirming
the receipt, finish with `yielded` and `continuation.kind: "monitor"` to await that
task’s timer. The monitor workflow remains preferred; REST calls use the same route authorization as legacy agents.
Existing tools remain preferred;
agents do not have to search before using them. Only two tool definitions are
advertised. The API catalog is returned on demand, never injected into the
initial prompt.

## Default availability and operator controls

The escape hatch is enabled by default. No environment variable is required.
The native `hire_agent` tool shares this availability policy.
Set `PAPERCLIP_RUNNER_API_TOOLS_ENABLED=false` on the server to disable these
tools. An explicit `true` also enables them; other explicit values fail closed.

Operators can restrict availability with `PAPERCLIP_RUNNER_API_TOOLS_COMPANY_IDS`,
a comma-separated list of company UUIDs. An unset list allows every company;
an explicitly empty list allows none. IDs must match exactly. This restriction
also applies when the enabled flag is unset.

A server-owned binding can disable these tools for a baseline eval but cannot
override an operator restriction. The server checks the policy when advertising
tools, when accepting a call, and immediately before HTTP dispatch after
preparing any files. Existing dedicated tools remain available. Operators must
update the environment of each server process and restart it for deployment-level
changes; this environment switch is not a live settings API.

All company, run, work-mode, credential, and route authorization checks below still apply.

## Discovery and requests

```json
{"query":"create project","limit":5}
```

Search is deterministic lexical ranking over OpenAPI paths, summaries and the
old skill reference. It supports task/issue and other terminology, exact
`METHOD /api/path/{parameter}` lookup, and opaque query/catalog-bound pagination.
Results include resolved request schemas, response descriptions, authorization
metadata, work modes, examples where available, and relevant dedicated tools
with their supported parameters. `limit` defaults to five and is capped at ten.

```json
{"operationId":"PATCH /api/projects/{id}","pathParams":{"id":"PROJECT_UUID"},"body":{"description":"Updated project description"}}
```

The catalog determines method and path. `companyId` is filled from the active
binding. Scalars and arrays are accepted in `query`. `body` defaults to JSON;
`contentType` supports text and raw uploads. `files` accepts entries containing
exactly one authorized `artifactId` or task-workspace `path`, and an optional
multipart `field`. No arbitrary URL, headers, authentication, or remote file URL
can be supplied. Routes still validate payloads and enforce permissions.

Writes put every request field in `body`. A POST or PUT whose request schema
names required fields, or a PATCH with a required body, is refused before it is
sent when `body` (and `files`) are missing; the error names the received
arguments and the shape to resend. `query` and the read-only options never carry
request fields.

For lossless input to a local helper, call the separate `save_api_response`
tool with an authorized non-sensitive GET `operationId`, `pathParams` and
`query`. It has no request body and cannot run a write, so `call_api`'s write
path carries no read-only save option. This saves the complete response bytes as an immutable artifact
even below the inline threshold. A zero-byte successful response instead returns
`empty: true`, `byteSize: 0` and `data: null`, without an asset or storage
reservation. JSON empty results such as `[]` are nonempty bytes and export normally.
Artifact receipts contain `artifactId`, `byteSize` and `sha256`. For artifacts up
to 10 MiB, pass those exact values to `download_artifact` with a workspace-relative
destination. Larger text snapshots (up to the 1 GiB capture limit) remain readable
through `GET /api/assets/{assetId}/content` with `responseText`; follow
`nextOffsetBytes` until null. They cannot be installed with `download_artifact`;
narrow or paginate the source query for local helper input. Pass downloaded files
to the helper instead of recreating JSON in model output. Pagination metadata remains in the receipt; each
artifact contains one complete HTTP response, not all pages of a list. Continue
list pagination explicitly. The file remains untrusted data.

Sensitive endpoints are rejected before dispatch. Older saved `call_api` calls
that carry `saveResponse: true` remain valid for recovery and follow the same
rules; `saveResponse` cannot be combined with `responseText`, and is rejected
before dispatch for mutations. Denied or partial HTTP responses
are not exported. Failed response receipts include the full `byteSize` and a
`truncated` flag when their inline diagnostic body is incomplete. It uses the same capture limits, storage admission, actor access
and integrity validation as automatic large-response artifacts. The original
request runs once. Default inline behavior is unchanged when the option is absent.

Requests have a 16 KiB URL limit and 10 MiB request/upload limit. **New response
captures are limited to 1 GiB of decoded bytes.** This is separate from the 24 KiB
inline/page limit. The receiver rejects an oversized Content-Length before
reading and counts actual bytes before writing, including chunked or compressed
responses. Oversized responses return `api_response_too_large`; narrow the query
or use the endpoint's own pagination. A mutation may already have committed, so
inspect its state rather than retrying it to obtain a smaller response.

Connection setup and stalled response reads time out after 30 seconds. An active
capture has a 10-minute total download deadline. Responses above 24 KiB stream
into a private temporary file, then into a company-owned asset scoped to the
issuing agent and responsible user (plus authorized board administrators).
Ordinary task or run visibility does not grant access to another actor’s API
snapshot. Explicit deliverables use the existing attachment/work-product flow. Memory stays
bounded by the inline prefix and stream buffers. Binary responses also become
assets; text previews are limited to 2,000 bytes.

Large captures reserve 1 GiB against a **durable 4 GiB per-run capture budget**
before creating a file. A completed capture settles to its actual byte count;
failed/interrupted captures retain the full reservation to bound retry loops.
A process restart does not reset that budget. Small inline responses without explicit export and reads
of existing assets need no reservation. Concurrent large captures are limited
to **two per company and four per server process**, including the storage upload
and temporary-file cleanup. Budget exhaustion returns `api_response_capture_limit`;
concurrency exhaustion returns `api_response_capture_busy` without queuing more
large transfers. A **20 GiB company-wide quota** counts all stored `runner-api` snapshots plus
unattached reservations. Admission uses a company database lock, so runs and
server processes share the same quota. Existing snapshots from before this
change count too. Attaching an asset converts its reservation to actual stored
bytes; deleting the asset frees that capacity. Small binary snapshots also need
storage admission. Ordinary inline text/JSON and existing-asset pages do not.

Operators can set `PAPERCLIP_RUNNER_API_COMPANY_CAPTURE_MAX_BYTES` to a positive
safe integer of at least 1 GiB. Invalid values fall back to 20 GiB. No zero or
unlimited setting is accepted. This quota covers API snapshots, not all company
attachments. Storage capacity and backend limits still apply.

Handled pre-storage failures release the company reservation after temporary
file cleanup, but retain the run's charge against retry loops. After a metadata
transaction fails, a locking read must prove the asset was not committed before
the uploaded object is removed. Confirmed cleanup refunds the company quota. A crash, failed
cleanup, or ambiguous storage write keeps an unattached reservation. Operators
must reconcile possible orphan files/objects before deleting that reservation
from `runner_api_response_reservations`; no automatic expiry silently refunds
possibly occupied space. Linked rows are removed with their asset. Deleting a
run does not release its unattached company reservations.

Temporary files are removed on success or failure. Long captures revalidate the
active run at least every MiB or at the next chunk after one second, and again
before returning the snapshot. Stopping the run stops its download.

To inspect saved text without creating another artifact, call its authorized
content operation with `responseText`:

```json
{"operationId":"GET /api/assets/{assetId}/content","pathParams":{"assetId":"RETURNED_ARTIFACT_ID"},"responseText":{"offsetBytes":0,"limitBytes":8192}}
```

The result contains `data` as text (including JSON), plus `responseText` with
`offsetBytes`, `nextOffsetBytes`, and `totalBytes`. Continue at `nextOffsetBytes`
until it is null. Windows end at UTF-8 boundaries. The limit defaults to 24 KiB
and accepts 4–24,576 bytes. The server rejects binary content, invalid UTF-8,
and offsets inside a code point or beyond the response. This option only works
with GET; never repeat a mutation to retrieve another part of its response. A
new write call that carries `responseText` is refused before it is sent, like
`saveResponse`: it never carries request fields, so resend the write once
without it and with its fields in `body`. Older saved calls stay valid for recovery.
Read the saved artifact for a stable snapshot instead of paging a changing live
response. A text-window call against a live response above 24 KiB also returns
that snapshot's artifact reference; continue on its content operation. A saved
asset page never creates another asset. An unpaged asset read also fetches only
a bounded preview and returns the existing reference instead of copying the file. Offsets and total sizes use safe integer
byte counts, including values above 2 GiB. Existing assets larger than the 1 GiB
capture limit remain readable because each request transfers only a bounded range.
Request/upload limits and all route
authorization remain.
Saved asset pages use authenticated HTTP byte ranges. The storage provider reads
only the requested window, with at most two extra bytes for UTF-8/EOF handling.
The client validates `Content-Range`, the total size, and the received byte count;
it rejects unsupported or inconsistent ranges instead of downloading the whole
asset for every page. Other GET routes are fetched once in full to create the
snapshot, so use the returned asset for subsequent pages. S3 uses streamed
multipart uploads for large snapshots. Asset sizes are stored as PostgreSQL
`bigint`, preserving the existing numeric API shape.

### Verified artifact installation

Native standard-mode runs can use `download_artifact` to install an authorized
saved API artifact into the admitted execution workspace. Pass `artifactId`, a
workspace-relative `destination`, and the exact saved `byteSize` and `sha256`.
The parent directory must already exist. The transfer is bounded to 10 MiB;
larger snapshots remain readable through the existing paged content API.

The controller reuses asset content authorization (including company, capture
owner and responsible-user restrictions), checks the complete stored bytes, then
uses the lease-owned command runner for remote execution. The execution host
checks size and SHA-256 on a private staging file before atomic create-only
publication. Linux directory descriptors reject parent/destination symlinks and
traversal. Existing different files are never overwritten; an identical retry
verifies the existing file. A lost promotion acknowledgement is an unknown outcome:
the native recovery path rechecks current authority and verifies the destination
against the same size and hash without installing again. Missing or conflicting
destinations return tool errors. An explicit identical retry remains safe. Interrupted staging can
leave a reserved scratch file if cleanup cannot reach the provider, but never a
completed destination. File synchronization after the turn remains owned by the
normal workspace lifecycle.

Current support is Linux local execution and remote native command-runner
transports (including Daytona and SSH). Plugin-owned tasks without an installer,
Dot, Muse, Grok and nonstandard work modes do not advertise this capability.
The API-tool rollout policy can disable it. Every attempt rechecks live task/run
and content authority before storage reads and promotion, and writes content-free
`runner.artifact_install` audit records. No API credentials enter the execution
command. Stored content remains untrusted: installation does not promote source
trust, execute artifact instructions, or provide an alternative for content that
screening denied. Existing API paging, upload and deliverable tools keep their
current contracts.

The synthetic regression in
`server/src/services/native-runtime/artifact-transfer.integration.test.ts` verifies
388131 Unicode/escape-containing bytes and independent downstream hashing with
real PostgreSQL and filesystem operations. Filesystem fault tests exercise the
actual command with simulated provider transport; they do not establish a live
Daytona RPC or deployed recovery.


### Image generation for OpenAI API-key Codex agents

Codex shows its built-in `image_gen` tool only when it signs in with ChatGPT.
A Codex agent that uses an OpenAI API key does not get that tool. The runner
also keeps the provider key out of the agent's shell, so the imagegen skill's
`scripts/image_gen.py` fallback cannot run.

For these agents, the controller advertises `generate_image`. The tool is
advertised only when all of these conditions are true:

- The run is a native Codex standard-mode run.
- The run was admitted on an OpenAI `api_key` connection with no custom routing.
- The run has a workspace.

Dot, Muse, Grok, native reviews and plugin-owned tasks do not get the tool.

The server reads the key from that same connection on each call. It sends the
key only to `https://api.openai.com/v1/images/generations` (model
`gpt-image-2`). The key never enters the provider process, the execution
command or the tool result.

Every call does these checks again before any spend:

- The live run, task and budget authority.
- `issue:mutate` permission.
- The connection grant, audience and agent installation.

A budget hard-stop that pauses the agent refuses the call before OpenAI is
contacted.

The PNG is saved as a new, create-only file in the workspace root:

- Local hosts use symlink-confined opens.
- Remote targets use the same installer as `download_artifact`.

The result returns the relative path, the size and the SHA-256.

Each generated image records one `cost_events` row (`metered_api`,
`costStatus: "estimated"`) from a reviewed `gpt-image-2` rate card. The row is
recorded even when saving the image fails. Each call also writes
content-free `runner.image_generate` audit records.

A per-call receipt in the run result makes a retried or recovered call replay
the first result instead of paying again.

### Media files and upload limits

The 1 GiB limit covers new snapshots returned by `call_api`, such as large JSON
exports or binary API downloads. It does not raise attachment upload limits or
limit files an agent creates and edits inside its workspace. Saved asset downloads
stream from storage and support byte ranges, including video seeking.

`PAPERCLIP_ATTACHMENT_MAX_BYTES` separately defaults to 10 MiB for uploads and
native file handoffs. `call_api` uploads also have their own 10 MiB limit. Several
upload and handoff paths buffer complete files in memory; raising those defaults
to GiB sizes requires streaming ingestion and corresponding admission/budget
controls first. For a future video attachment workflow, 2 GiB per streamed file
is a reasonable default, with an operator override and storage quotas. Do not
claim that this response-paging change enables GiB attachment uploads.

Tool responses identify the HTTP route with `apiOperationId`. The native protocol
reserves `operationId` and `callId` for semantic tool-call identity; API metadata
must not masquerade as that envelope. Saved mutation receipts are normalized at
the tool boundary as well, without repeating their HTTP request.
All redirects are refused. Interrupted mutation responses have an
unknown outcome, requiring inspection before another mutation.
Mutation responses with HTTP 5xx, HTTP 408, redirects, or malformed JSON also
retain an unknown outcome. A server may have committed the write before it
failed to return a valid response.

## Authority and replay

The server revalidates the active native run, assigned task and actor, then
creates a server-held agent JWT bound to that company and run. Requests go
through the actual HTTP router with its authorization, validation and domain
audit behavior. Skill-test runs retain the same task-scoped `skill_test` JWT
as legacy REST runs, including after later edits to the task mode. An additional `runner.api_called` receipt attributes mutations
to the run even where older route audit events omit that field.
The run and work mode are checked again after asynchronous file preparation, so
a stopped run cannot dispatch an upload prepared under its earlier binding.

`call_api` is an authenticated REST transport, with the same endpoint and payload
permissions as a legacy agent using REST. It does not maintain a second denylist
for routines, runtime services, skill tests, recovery, invitations, secrets,
exports, lifecycle fields or approval routes. Existing HTTP routes enforce their
company/resource scope, ownership, grants, redaction, run requirements and board
or approval gates. A board-only route still denies an agent call. Dedicated tools
remain preferred for common workflows, including native completion and monitors.

Ask and pre-acceptance Plan retain their existing read restrictions and specific
mutation exceptions (task titles and conversational confirmation recording).
Protocol endpoints still require their existing clients. Active-run validation,
server-owned credentials, file containment, transfer bounds and replay handling
also remain in place. REST parity does not grant board identity or new permissions.

Mutation call IDs reserve a durable receipt in the run's existing `resultJson`
before dispatch. Replays return the recorded result. Reusing an ID with different
arguments is rejected. A crash after reservation leaves an unknown outcome and
never automatically resends the mutation. The limit is 512 mutation receipts per
run. No database migration is needed.

Workspace uploads use the existing workspace resource containment checks,
no-symlink file opens covering every path component, and bounded descriptor reads.
Local uploads require Linux or macOS; authorized artifacts work on other hosts.
Artifacts must belong to the bound company and pass the same actor access checks
as downloads, before and after reading their bytes. Private API response snapshots
cannot be used as public company logos or bundled into company exports. An owner
can explicitly publish a copy through the existing attachment upload path.
Request bodies and uploads go
through the route's validation and authorization; the fallback does not reject
fields that an agent can supply through REST. Route-authorized responses normally use the bounded inline/artifact capture and
replay mechanism. Credential-bearing responses (secret values, issued keys,
invites, webhook secrets (including restored triggers) and gateway credentials) stay inline, with a 10 MiB
response bound. They never spill into plaintext files or saved response assets.
Their durable mutation receipts encrypt the response with the existing local
secret provider. Replays decrypt it only inside the authorized run; no HTTP
request repeats. A working secret store is checked before credential dispatch.
Other API responses retain the existing 24 KiB inline and 1 GiB capture limits.

API request and response bodies are omitted from persisted native tool events,
checkpoints and completion results. Saved events retain operation identity, HTTP
status and a payload digest for replay conflict detection. The model still receives
the original authorized HTTP response; this redaction affects durable history.
Completion retries compare expected and persisted events through the same
idempotent redaction projection. Repeated projection preserves the original
payload digest. Provider recovery uses the same projection to compare saved
completion results with provider history. Changed private bodies still conflict.
Restart preparation preserves the run's admitted REST credential scope, even
when the task mode changes. See [the run-log guide](run-log-events.md) for stored fields.

The server-issued REST credential for the bound run is never returned to the model.
Streaming, WebSocket, MCP and authentication
handshakes are documented as protocol operations requiring their existing clients.

## Catalog maintenance

`runner-api-catalog.ts` builds from the server OpenAPI registry. Experimental
pipeline, Cases and smoke-lab routes now share their validators with discovery.
Seven Cases/pipeline route shapes are multiplexed by resource identity: the Cases
router intentionally forwards unknown resources to the pipeline router. Their
separate catalog entries explain which resource identifier is required. Registry
authorization descriptions are documentation; actual route checks are authoritative.

Regenerate old-skill enrichment after editing its API reference:

```sh
node scripts/generate-runner-api-reference.mjs
node scripts/generate-runner-api-reference.mjs --check
node scripts/generate-runner-experimental-api-metadata.mjs
node scripts/generate-runner-experimental-api-metadata.mjs --check
```

Mounted-route coverage tests include experimental routes. Three WebSocket mounts
are explicitly classified in the catalog. Shared protocol-action catalogs,
provider projections and generated compatibility checks include both tools.

## Verification and paid evals

The companion `paperclip-evals` worktree contains `evals/runner-api-tools`.
Its README documents explicit case/model selectors, the cumulative budget ledger,
fixture reset, progressive batches, and Evalbook generation. No command defaults
to running the entire paid suite. Capability, forced operation contracts and
paired common-operation regressions are reported separately.

Provider-free integration tests exercise real runnerd → PRP → authority → HTTP,
route validation and audit, stale bindings, Ask/Plan restrictions, identity
spoofing, file containment, uncertain mutation receipts and fixture isolation.

The Evalbook viewer uses the existing shared viewer and stylesheet on master.
The report retains actual persisted-state summaries for private local inspection;
public replay continues to withhold company-state details.

The ACPX sidecar includes the upstream terminal-usage accounting correction from
`origin/codex/evalbook-default-chat-sept6`. Its qualified Claude executable requires
Linux x64. The first macOS stage records a zero-cost ACPX admission failure. A later user-authorized
OpenCode/OpenRouter Sonnet profile reached a real HTTP read, but the attempt failed
on a missing harness completion contract and incomplete terminal accounting. The
harness contract is corrected. The missing fourth request was subsequently
recovered from the matching OpenRouter session and generation billing record;
the original failed attempt remains immutable. New attempts retain an append-only,
flushed event journal and bounded provider trace outside disposable runtime files.
Provider-free startup succeeds for OpenRouter Sonnet and DeepSeek. See
`doc/plans/2026-09-07-runner-api-production-readiness.md` for remaining release gates.