Files
PaperClipAI/doc/runner-launch-preparation.md
T
Devin FoleyandPaperclip b4cea5b5a4 fix(runner): show disk exhaustion before recovery errors (#22448)
Show the confirmed sandbox disk failure before a later lost-turn or restore error in saved run errors and both run detail views. Preserve the original diagnostics and leave scarce inodes available when preparing the existing runner relaunch reserve.

Verified focused regressions, scoped typechecks, and full CI on the reviewed revision, including Runner checks and all eight end-to-end shards. Greptile is 5/5 with all review findings resolved.

Co-Authored-By: Paperclip <noreply@paperclip.ing>
2026-11-20 17:08:31 -08:00

5.4 KiB

Runner launch preparation

Core selects the provider, model, instructions, tools, and task permissions. The execution environment supplies executable paths and filesystem locations.

The built-in TypeScript implementation constructs the provider descriptor in createRunnerProviderLaunch. The transport sends this descriptor with run.prepare. Provider backends keep their existing validation and session rules.

By default, the SDK prepares local files. It stages remote files through the existing execution driver.

Managed sandbox storage

Core holds a 256 MiB runner relaunch reserve only in managed sandboxes. It creates the reserve when there is at least 2 GiB of additional free space and 1,024 free inodes. It leaves scarce capacity available to an existing agent. A relaunch releases that reserve before it stages the replacement runner. This reserve is recovery headroom; it does not limit the agent's downloads.

When runner diagnostics confirm a full sandbox disk, the saved run error and run details show that cause ahead of a later lost-turn or restore error. Raw failure diagnostics remain in the run's diagnostic evidence. Older saved runs with sandbox_disk_full get the same display without a data migration. Download budgets and image policy belong to the deployment environment.

Environment-owned preparation

An execution driver can supply prepareLaunch. The callback receives RunnerProviderLaunchRequirements. It returns PreparedRunnerProviderLaunch.

Requirements contain the selected provider, model, policy, run identity, and runtime context. They do not contain task text, process callbacks, provider credentials, or controller executable choices. Resource paths identify source locations. They do not grant filesystem access on the execution host.

The environment resolves its resources and prepares the provider descriptor. It can use bindRunnerLaunchResources to replace registered context locations. It can use createRunnerProviderLaunch for the existing descriptor projection. The transport does not perform its built-in file preparation in this mode. The process owner supplies the Runner artifact and process lifecycle.

The result contains a configuration ID, a requirements digest, and a provider descriptor. The transport checks Core's selected profile and policy before registration and launch. Core still supplies authorized tools and completion criteria through PRP. A failed preparation does not enable a fallback launch.

Codex probe policy

Core can select settings.disableCodexProjectInstructions for a hello probe. Preparation must preserve it as project_doc_max_bytes=0 before app-server, after any environment-supplied overrides. Core validates this setting alongside disableCodexWebSearch. Ordinary tasks retain project instructions when the setting is absent or false. This does not change the selected filesystem policy.

Working directory

In this mode the environment chooses the provider working directory. Core does not supply one. The backend uses workingDirectoryAuthority: "runner_prepared_launch", and its controller environment stays empty.

  • Core authorizes the agent, task, projects, and workspace binding.
  • The environment resolves that assignment into mounts and the descriptor's cwd. For a projectless task it selects its own default directory.
  • The transport accepts only a normalized absolute cwd that is not a filesystem root. Runnerd validates the directory on its own filesystem.
  • The transport reports that cwd to the driver. The driver records it in session context, snapshots, and recovery, and ignores the caller's directory.
  • Recovery and run attachment must keep the same cwd. A runner snapshot or recovered thread that reports another directory fails recovery. Core does not substitute a controller-local or temporary directory.

remote_runner remains the mode in which Core assigns the remote directory through PAPERCLIP_WORKSPACE_CWD.

The transport saves the exact requirements and prepared configuration in its private launch-preparation.json file. The file holds at most two records. A new attachment keeps the active run's record. Runner commands contain the provider descriptor once. They do not contain these recovery records.

Recovery within the same run must return the same configuration. This rule also applies during an interrupted warm attachment. A new run uses its current requirements. For a new run, runtimeContext: null removes the assigned context. Only an omitted context can reuse the previous context. Provider session compatibility checks remain in their existing backends. Close waits for preparation within its deadline. A timeout retains an unsettled checkpoint and prevents reuse.

Plugin task protocol

The environment task API defines the plugin preparation exchange. The native plugin integration connects the callback to task execution. A Launcher implementation remains the execution owner's responsibility.

Managed Git helper interpreter

githubLauncherSource({ interpreterPath }) can select an absolute Node interpreter owned by the execution environment. Without this option, generated helpers retain #!/usr/bin/env node. The option changes only the shebang; it does not change PATH, command discovery, credential handling, or helper behavior. Interpreter paths must be absolute and contain no whitespace or NUL bytes.