Introducing a new way to review your PRs.

Read the announcement

paperclipai / paperclip

The open-source app everyone uses to manage agents at work

View on GitHub

30 documents · Updated Nov 20, 2026

docs / native-runner-execution

Native runner execution

Native runner execution gives Paperclip one provider-neutral way to run an agent and keep control of the work around it. Paperclip owns run authority, policy, tools, results, durable evidence, and recovery. A runner owns the provider session and reports its work through the Paperclip Runner Protocol (PRP).

Parent: Agent work management

Related pages: Issue execution, Execution environments, Agent instructions, Adapter integrations, Runner evaluation and conformance, Case handoffs and continuity, Reusable workflows

Execution boundary

Control-plane authority

The control plane is the source of truth for a run. It admits the provider choice, binds the run to its task and environment, grants tool access, records events, accepts the result, and decides the task status. The runner reports provider facts and proposes a result, but provider text or process exit cannot complete the task by itself.

This boundary keeps provider work separate from workspace finalization, usage accounting, and task status. Paperclip can therefore retry a lost delivery or workspace step without starting another provider turn.

References: server/…/paperclip-control-plane-port.ts• L185-211 server/…/paperclip-control-plane-port.ts• L541-664 server/…/native-session-executor.ts• L770-816

Execution topology

The same run model supports local processes, remote environments, hosted providers, and short-lived external bridges. The browser reads Paperclip's durable projections. It does not connect to a provider session.

Execution shapePaperclip responsibilityProvider responsibility
Local runnerOwn the process, workspace binding, and loopback route.Run the provider session.
Remote runnerSelect a trusted route and retain environment ownership.Run in the remote environment and report through PRP.
Hosted providerBind the managed service, limits, and session identity.Run the model loop in the hosted service.
External bridgeIssue a short-lived service binding with no workspace access.Perform bounded semantic operations.

Transport selection is part of admission. Paperclip selects the route before remote assets or provider endpoints are prepared, and the route cannot change silently after dispatch. Resource preparation belongs to connection registration. Activation happens only after the control plane has registered the route and its ownership.

This diagram shows the main path from an assigned task to a settled task decision.

flowchart LR
    A[Task and run] --> B[Admitted authority]
    B <-->|Authenticated PRP| C[Runner]
    C --> D[Native session backend]
    D --> E[Provider or bridge]
    E --> D
    D --> F[Durable events and result proposal]
    F --> B
    B --> G[Workspace finalization and task status]

References: packages/…/runner-execution-driver.ts• L5-10 packages/…/runner-execution-driver.ts• L41-62 packages/…/runnerd-codex-transport.ts• L1110-1145

Paperclip Runner Protocol

Protocol contract

PRP is the language-neutral contract between a controller and a runner. Commands request bounded work or lifecycle changes. Events report ordered provider-neutral facts. A structured result and terminal record describe the proposed disposition of the run.

PRP currently supports versions 1 through 3. Version 2 adds durable session goals. Version 3 adds external-provider operations and runner attachment downloads. A peer must negotiate a supported version and reject unsupported required fields or meanings.

DirectionRecordPurpose
Control plane to runnerCommandStart, steer, stop, recover, snapshot, or close work.
Runner to control planeEventReport session, turn, item, tool, request, usage, result, and terminal facts.
Runner to control planeResult proposalOffer a structured work disposition with evidence.
Control plane to runnerResult or denialAccept, reject, or classify the requested action.

References: packages/…/replay-contract.ts• L34-117 packages/…/README.md• L3-29 packages/…/durable-prp-control-plane.ts• L52-130

Run identity

Every command and event belongs to a company, task, run, normalized session, turn, item, runner instance, environment lease, and provider identity. The normalized session identifies Paperclip continuity. The provider session identity is separate so recovery can detect a provider that opened the wrong conversation.

Provider, model, permission mode, workspace, runtime profile, and tool contract are part of admitted behavior when they affect execution. A replacement controller or provider cannot change them without a governed transition.

References: packages/…/replay-contract.ts• L83-101 packages/…/native-session-backend.ts• L54-83 packages/…/durable-prp-control-plane.ts• L885-893

Native execution input

Native execution input is the versioned launch contract from Paperclip to the runner. It carries task scope, workspace or service access, session policy, runtime context, interaction responses, provider identity, credential references, and the completion contract.

The contract also defines a model envelope. Only task content, selected context, constraints, workspace view, and completion rules enter provider-visible model input. Provider settings, credentials, release identity, and recovery data stay execution-only.

References: packages/…/native-execution.ts• L13-29 packages/…/native-execution.ts• L174-213 packages/…/native-execution.ts• L290-322

Completion contract

The completion contract states the task objective and the criteria that a result must address. A provider uses the paperclip_finish or paperclip_block semantic operation to report a disposition. Paperclip checks that report against the current contract and durable evidence before it accepts a result.

The report separates the user-facing answer from control-plane wait data. A completed report can be done or needs_review. A yielded report must name a real continuation with a stable idempotency key, or omit continuation when the control plane has an authorized saved wait. A blocked report must identify its reason, owner, scope, and unblock action. Provider-facing schemas accept known legacy aliases, but the bridge normalizes them before strict PRP validation. The first accepted report for a turn is final for that turn.

Questions and permission prompts are runtime requests tied to the active turn. If the provider is lost before an answer is safe to replay, Paperclip records a wait or expiry instead of sending the answer to another provider session.

References: packages/…/completion-result.ts• L1-32 packages/…/completion-result.ts• L181-228 packages/…/completion-result.ts• L230-301 packages/…/completion-result.ts• L401-464

Session backends and provider drivers

Session backend contract

A native session backend is the common boundary for opening, continuing, steering, interrupting, snapshotting, recovering, and closing a provider session. It exposes provider-neutral events, runtime requests, structured results, usage, and capability claims so the control plane does not depend on provider mechanics.

Unsupported controls are explicit. A provider that cannot safely resume, steer, report usage, or preserve a runtime request reports that capability as unavailable rather than receiving a best-effort promise.

Cancellation first revokes publication and mutation authority. The returned cleanup then stops provider work. This prevents late provider callbacks from changing a cancelled run. Slow or failed cleanup remains an ownership gate for the affected task until it settles or needs operator recovery.

References: packages/…/native-session-backend.ts• L139-149 packages/…/native-session-backend.ts• L223-330 packages/…/native-session-runtime.ts• L765-780 packages/…/native-session-runtime.ts• L1102-1210

Provider families

The runner selects a provider implementation from the admitted execution input. The factory exposes only provider runtimes included in the release slice. Persisted input for a future provider does not make that provider executable.

Provider familySession modelMain continuity rule
CodexApp-server session through the native or runner-owned transport.Match the saved conversation and turn history.
OpenCodePinned local HTTP/SSE server facade.Match the saved OpenCode session and workspace.
ACPX profilesQualified sidecar session for a selected agent profile.Match the profile, model, policy, workspace, and provider lifetime.
Muse and external botsBroker-backed runner session.Use the admitted broker and Rust Runner authority.
Managed providersManaged service or harness session.Match service identity, limits, memory, and cost policy.
Cursor CloudRemote agent observed by the local driver.Match the saved remote agent, run, thread, turn, and checkpoint identity.

The provider family changes the available controls, not the control-plane boundary. Capabilities such as goals, steering, follow-up turns, usage, artifacts, and recovery are advertised per integration.

References: packages/…/native-backend-factory.ts• L48-135 packages/…/harness-driver-backend.ts• L31-54 packages/…/codex-app-server-driver-impl.ts• L248-299

Provider admission

Admission proves that the selected provider matches the stored run contract before model work starts. It checks provider identity, model, execution shape, workspace or service binding, permissions, runtime context, session identity, and tool route.

ACPX adds qualification for the installed command, release, model, profile, platform, credentials, and provider lifetime. Its adapter also rejects an unsafe platform or a missing provider-lifetime fence before it creates a provider session. OpenCode requires a qualified pinned runtime and model shape. A failed admission does not erase ownership. Late resources remain under a cleanup owner until shutdown is proved.

References: packages/…/acpx-runtime-adapter.ts• L201-239 packages/…/acpx-runtime-adapter.ts• L769-830 packages/…/installation-integrity.ts• L40-50

Provider isolation

Each provider receives a controlled home, environment, credential boundary, workspace boundary, and runtime context. Paperclip treats authentication and metadata checks as proof of access to the selected provider, not as proof that inference will succeed.

Provider errors map to stable classes such as authentication, permission, rate limit, quota, invalid request, and model unavailable. These classes guide status and recovery without making provider message text part of the contract.

Provider observability is an opt-in view of PRP events. It exports bounded, redacted turn, tool, and usage spans when configured. It does not become a second source of run state or result authority.

References: packages/…/installation-integrity.ts• L153-220 packages/…/permission-policy.ts• L37-66 packages/…/company-observability.ts• L72-127

Durable control-plane transport

Transport lease

The control plane issues a one-use bootstrap ticket for one admitted run. The runner exchanges it for a short-lived connection lease. Authentication binds the lease to the run, runner artifact, environment lease, protocol version, expiry, and revocation epoch.

Lease renewal extends the same connection. It does not replace the runner, provider, session, turn, or authority. A controller restart can detach a live runner, but a replacement controller must prove the exact owner and binding before it resumes delivery.

Runner artifacts are part of this proof. A relaunched runner can replace the parked artifact only when its version is a valid forward-compatible release and its digest is recorded. An older, malformed, or incompatible artifact remains refused.

References: packages/…/durable-recovery.md• L34-61 packages/…/durable-prp-control-plane.ts• L147-171 packages/…/runner-version.ts• L1-40 packages/…/runnerd-codex-transport.ts• L239-253

Secure frame channel

The runner can dial a secure control-plane endpoint or connect to an authorized provider-ingress route. Plaintext WebSocket use is limited to loopback. Secure connections use mutual HMAC proof and encrypted, direction-specific frames with session-bound counters.

Invalid identity, expired lease, replayed frame, out-of-order counter, oversized frame, or wrong route fails closed. The channel authenticates the connection. The run contract still authorizes each command and tool effect.

References: packages/…/durable-recovery.md• L527-579 packages/…/durable-prp-control-plane.ts• L711-720 server/…/native-runner-execution-driver.ts• L50-95

Durable commands and events

The control plane journals commands before delivery. The runner journals a command outcome before it sends that outcome. A retry with the same command identity and fingerprint replays the saved outcome instead of repeating the effect.

The runner writes events to an outbox before sending them. The control plane commits or deduplicates them and returns a cumulative source cursor. The runner removes events only after a valid acknowledgement. Streaming deltas may be batched, but every non-streaming event, checkpoint, and result waits behind those deltas. Priority-zero lifecycle, semantic-tool, and terminal facts retain storage space when lower-priority output reaches backpressure.

Semantic tool input is also durable before dispatch. If the controller changes owner or cannot persist the journal, the effect does not start. A saved input without a result remains pending for controlled recovery rather than being run twice.

References: packages/…/control-plane-port.ts• L25-45 packages/…/control-plane-port.ts• L89-127 packages/…/durable-prp-control-plane.ts• L1023-1084 packages/…/native-session-runtime.ts• L1285-1323

Checkpoints and warm attachment

A checkpoint is a bounded recovery view of control state, identity, commands, cursors, provider evidence, and session state. It is not a second transcript. External persistence can hold this view outside the runner host, while a local marker becomes a fail-closed state until the committed checkpoint is restored.

Checkpoints retain the runner version and digest that produced the state. They store changed command and recovery records rather than copying the full stream, while keeping the identity and terminal anchors needed to rebuild control state.

Warm attachment moves a retained runner from one run authority to another without creating a second provider session. The old owner must leave an acknowledged event cut. The new owner proves the transition receipt, lease, checkpoint, runner identity, and target authority before it sends new work.

References: packages/…/durable-prp-control-plane.ts• L358-503 packages/…/durable-prp-control-plane.ts• L896-1008 packages/…/runnerd-codex-transport.ts• L474-536 packages/…/codex-harness-session.ts• L117-181

Event replay and recovery

Replay rules

Replay validates event schema, source order, identity binding, duplicate behavior, semantic-call pairing, and terminal meaning before it rebuilds state. A source event ID is applied at most once. A source gap remains visible as a gap. Replay does not invent a missing fact.

Each semantic call needs one input and exactly one result or reconciled outcome. Conflicting payloads under one event identity or conflicting results for one call are integrity failures that stop automatic recovery.

References: packages/…/replay-contract.ts• L251-439 packages/…/replay-contract.ts• L442-524 packages/…/durable-prp-control-plane.ts• L1023-1084

Provider recovery

Provider recovery restores runner state and the event cursor before it asks a backend to restore the provider session. Same-session recovery requires the saved provider identity, workspace or service binding, active turn, admitted configuration, and tool contract to agree.

If continuity cannot be proved, policy either opens a recorded replacement session or stops the run. A replacement receives bounded history and the current completion contract. It does not inherit an old turn as completion authority.

Runner process recovery follows the same rule. It may relaunch a replacement after an authenticated loss, but it stops after a bounded crash loop or an unverified termination. A full sandbox disk is a stable launch failure and does not cause repeated retries during the recovery grace.

ACPX uses a prompt-dispatch journal to distinguish a turn that was received by the sidecar from a turn whose first provider prompt was dispatched. A complete received record without dispatched proves that the prompt did not reach the provider. Missing or torn evidence remains unknown and does not authorize a retry.

Codex recovery also reconciles provider history. It can adopt the exact accepted turn or a provider-side terminal that arrived while the controller was detached. It refuses a different session, workspace, provider identity, active turn, or ambiguous history.

References: packages/…/native-session-backend.ts• L85-137 packages/…/codex-app-server-driver-impl.ts• L407-820 packages/…/codex-harness-session.ts• L763-903 packages/…/sandbox-disk-full.ts• L1-11 packages/…/turn-dispatch-journal.ts• L8-35

Controller recovery

Controller recovery combines stored session, checkpoint, owner, process, and provider evidence. It can wait for a retained owner, restore an exact session, attach to a verified remote runner, or start normal admission when no provider work exists.

EvidenceRecovery result
A valid owner still holds the lease.Wait for that owner to release it.
Session, checkpoint, provider identity, and cursor match.Restore the existing session.
A verified remote runner remains bound.Attach to the retained execution.
A newer runner is forward-compatible and its digest is accepted.Record the new artifact and continue recovery.
No provider event, tool call, or accepted result exists.Admit a new attempt.
A live process lacks exact authority evidence.Wait or block. Do not adopt or kill it by guesswork.

Warm reattachment never launches a second runner while the verified retained process is alive. The execution driver separates adoption authorization from launch preparation and requires a durable configuration binding for the exact execution.

References: server/…/native-controller-recovery.ts• L79-123 server/…/native-controller-recovery.ts• L181-245 packages/…/runner-execution-driver.ts• L16-40 packages/…/runner-execution-driver.ts• L90-101

Result settlement

The runner reports a structured result proposal and a terminal state. The control plane validates both against the completion contract, current run state, durable evidence, pending interaction, governance, and workspace finalization before it stores one accepted result.

An accepted result starts finalization. Workspace export, log projection, usage settlement, and task status remain separate recovery steps. A finalization retry reuses the accepted result and never opens another provider session.

References: server/…/paperclip-control-plane-port.ts• L541-664 packages/…/completion-result.ts• L230-301 packages/…/native-session-runtime.ts• L748-763 server/…/native-run-finalizer.ts• L218-299

Semantic tool dispatch

Provider-neutral tool catalog

A semantic tool is a stable Paperclip operation with a name, input and output schema, effect class, allowed work modes, placement, and required claims. Provider bindings convert that operation to the provider's tool shape.

The catalog covers task context and progress, completion, human input, delegation, approvals, workspaces, skills, documents, cases, workflows, deliverables, chat, connectors, responsibilities, routines, and bounded API access. Restricted profiles receive smaller catalogs. Discovery helps a provider find tools. It never grants authorization.

Optional tools are grouped into namespaces such as active_task, documents, delegation, governance, workspace, skills, responsibilities, and routines. This lets an agent ask for a small, relevant set instead of receiving every optional operation at session start.

References: packages/…/semantic-action-types.ts• L1-20 packages/…/semantic-action-types.ts• L68-123 packages/…/semantic-action-catalog.ts• L133-151 packages/…/semantic-action-catalog.ts• L499-540 packages/…/paperclip-discovery.ts• L13-59

Run-scoped discovery

Discovery projects only operations that are both bound to the run and allowed by its current context. It filters by namespace and query, returns a bounded result, and marks truncation when more matches exist. A later invocation checks authority again, so a discovery response is not a permission token.

The catalog now includes company responsibility, recurring routine, and skill-detail actions as optional standard-mode work. Their descriptions define the product rule at the same boundary as their schemas. For example, routine changes are revisions, responsibility changes can become proposals, and skill detail changes remain separate from SKILL.md content.

References: packages/…/paperclip-discovery.ts• L67-124 packages/…/paperclip-discovery.ts• L133-186 packages/…/responsibilities.ts• L1-45 packages/…/update-routine.ts• L1-51 packages/…/update-skill-details.ts• L1-31

Run-scoped dispatch

The tool authority is the application boundary for every provider call. It checks the current run, task state, work mode, claims, tool advertisement, input schema, size limits, and resource scope before it invokes a handler.

The server reloads authority at effect time. This prevents stale discovery from keeping mutation power after reassignment, a terminal transition, a mode change, or a session replacement. The runner-side bridge adds session authentication, schema validation, request limits, cancellation, and duplicate-call protection. Built-in completion calls use the same authenticated roundtrip as other semantic tools, so the provider does not see an accepted result before Paperclip validates it.

References: server/…/paperclip-runner-tool-authority.ts• L618-731 packages/…/dispatcher.ts• L106-179 packages/…/dispatcher.ts• L465-523 packages/…/permission-policy.ts• L19-66

Mutation receipts

Mutation dispatch uses a stable call identity, input digest, and durable receipt. Paperclip records the attempt before an effect and stores the sanitized outcome after it. An identical retry returns the saved outcome. A conflicting or unknown receipt does not run the effect a second time.

Read operations can be replayed when their scope is still valid. Receipt-backed writes can be recovered when their effect and receipt share one transaction, such as revisioned document writes. Filesystem or external effects without a proven result remain unknown and require reconciliation rather than an automatic retry.

References: packages/…/types.ts• L370-438 packages/…/types.ts• L440-466 server/…/paperclip-runner-tool-authority.ts• L450-610 packages/…/native-session-backend.ts• L158-171

Conformance contracts

Cross-language protocol parity

Conformance means that the JSON schemas, generated validators, TypeScript runner, Rust runner, server persistence, and fixtures agree on versions, identities, event order, duplicates, gaps, semantic results, and terminal meaning. The checked-in schemas are the language-neutral source of truth.

Generated protocol declarations, semantic catalogs, capability inventories, manifests, runner artifact bindings, and replay goldens are compatibility artifacts. A source change is incomplete when one of those artifacts no longer describes the same contract.

The capability inventory is generated from normative capability specifications and tracks evaluation groups and legacy aliases. It gives conformance work one shared inventory rather than separate provider-specific lists.

References: packages/…/README.md• L3-23 packages/…/README.md• L113-133 packages/…/capability-contract.ts• L1-46 packages/…/semantic-action-catalog.ts• L681-703 packages/…/index.ts• L1-18

Control-plane port

The control-plane port is the narrow boundary used by a standalone runner. It opens a run, checkpoints session state, appends ordered events, replays events, and accepts one structured result without importing Paperclip server or database code.

The port also states the durability rules for cancellation and streaming. A cancelled append, checkpoint, or finalization must settle without a new commit. A port that pipelines streaming events must still make every later event and result wait for the earlier deltas.

The production port adds persisted run binding, validation, redaction, deduplication, replay cursors, result fingerprints, semantic-input recovery, and durable task finalization signals. Live publication is an observation path, not the source of truth.

References: packages/…/control-plane-port.ts• L25-45 packages/…/control-plane-port.ts• L47-80 packages/…/control-plane-port.ts• L89-127 server/…/paperclip-control-plane-port.ts• L87-138

Driver qualification

A provider driver is qualified when it preserves the session contract across admission, start, input, interruption, recovery, result settlement, usage, and close. Its capability claims describe what Paperclip can safely preserve, not every feature of the upstream provider.

ACPX qualification also covers the selected native distribution, provider lifetime, session-update projection, provider-started work, background activity, and tool identity. OpenCode qualification covers its pinned runtime and the provider-neutral event and request boundary. Evaluation uses deterministic protocol fixtures and controlled live runs. Provider traces, usage observations, and process evidence remain bounded diagnostic evidence. They do not replace the durable PRP result or authorize a task effect.

References: packages/…/native-session-backend.ts• L30-36 packages/…/native-session-backend.ts• L239-347 packages/…/installation-integrity.ts• L33-50 packages/…/company-observability.ts• L157-225

Change checklist

When these files or sections are changed, remember to consider these:

  • PRP schemas, generated validators, protocol versions, or event families: update TypeScript, Rust, server persistence, replay rules, manifests, and fixtures together.
  • Native execution input or provider identity: update model projection, admission, provider descriptors, recovery, and compatibility checks together.
  • Native backend selection: keep the release slice, provider schema versions, broker requirements, and explicit unsupported-provider failures aligned.
  • Provider profiles or remote runner artifacts: preserve release identity, environment leases, model checks, credential fencing, and reattachment rules.
  • Runner version or artifact digest handling: update ticket admission, checkpoint headers, relaunch acceptance, and warm recovery as one contract.
  • Session capabilities, goals, requests, or usage: keep unsupported behavior explicit and preserve request identity, turn ownership, expiry, and replacement-session rules.
  • Transport leases or warm attachment: preserve route selection, handshake binding, expiry, revocation, event cursors, transition receipts, and ownership fences.
  • Checkpoints or external persistence: keep run binding, owner checks, sparse cursor rules, redaction, atomic replacement, and fail-closed markers aligned.
  • Semantic tools or assigned MCP tools: update the shared catalog, provider projections, schemas, work-mode rules, fresh authorization, limits, and recovery receipts.
  • Completion result schemas: keep canonical, provider-facing, bridge-facing, descriptions, continuation rules, and first-accepted-result behavior consistent.
  • Result settlement or workspace finalization: preserve accepted-result identity, finalization ownership, retry scope, usage provenance, log cursors, and retained resources.
  • Responsibility, routine, or skill-detail actions: keep the canonical action description, namespace, allowed mode, live binding, and catalog entry aligned.

Important failure modes

TriggerConsequenceGuard
Provider, model, workspace, session, or tool contract differs from admission.Launch or recovery stops before provider work.Versioned launch input and run-bound admission.
More than one provider state authority exists after restart.Recovery refuses to choose a provider.One-provider selection and conflict detection.
Remote route is not eligible before staging.Remote preparation and endpoint acquisition do not start.Transport selection before artifact preparation.
Relaunched runner is older, malformed, or has an unaccepted digest.The parked runner remains authoritative and recovery stops.Semver resume rule plus artifact binding in tickets and checkpoints.
Command or event acknowledgement is lost.The record is replayed without a second logical effect.Durable journals, stable identities, cumulative cursors, and deduplication.
Checkpoint persistence fails or loses its owner.The last committed recovery view stays authoritative; adoption is blocked.Atomic owner-checked checkpoints and fail-closed markers.
Prompt-dispatch evidence is missing, torn, or says dispatched.Recovery cannot safely resend the prompt.Provider shutdown proof and fail-closed unknown state.
Runner replacement staging reports sandbox_disk_full.Recovery fails immediately instead of retrying the same impossible launch.Stable disk-full failure code and retry suppression.
Tool discovery is stale or the run changes owner, mode, or status.The call is rejected without an application effect.Current authority reload, claims, bounds, and schema checks.
A mutation result is unknown or conflicts with its receipt.Paperclip does not repeat a possibly applied effect.Stable call identity, input digest, durable receipt, and reconciliation.
Provider output ends without an accepted structured result.The run remains failed, waiting, interrupted, or recoverable.Completion contract validation and control-plane settlement.
Workspace finalization loses its lease after result acceptance.The result remains durable while workspace or status projection waits for repair.Separate finalization owner and bounded retry.
A provider permission request is not bound to an admitted Paperclip action.The request is rejected or delegated for approval.Exact catalog lookup and fail-closed provider permission policy.
A responsibility, routine, or skill-detail action is exposed without its current mode or policy.A company change is denied or reaches the wrong authority boundary.Shared descriptor, live binding, and invocation-time authorization.

Key files

FileWhy to read it
packages/…/README.mdProtocol compatibility, session goals, fixtures, manifests, and attachment commands.
packages/…/replay-contract.tsPRP versions, event identity, semantic-call pairing, and validation.
packages/…/native-execution.tsLaunch input, provider selection, model projection, and completion contract.
packages/…/completion-result.tsCompletion, blocker, verification, attention, and continuation contracts.
packages/…/native-session-backend.tsShared session, recovery, cancellation, capability, and cleanup boundary.
packages/…/harness-driver-backend.tsTranslation from provider drivers to the normalized native session boundary.
packages/…/durable-prp-control-plane.tsDurable commands, events, leases, checkpoints, semantic-input recovery, and warm transitions.
packages/…/control-plane-port.tsStandalone runner durability, streaming, replay, and result boundary.
packages/…/runner-version.tsCompatibility rule for relaunching a runner from saved state.
packages/…/sandbox-disk-full.tsStable failure code that stops impossible runner relaunch retries.
packages/…/runner-execution-driver.tsEnvironment preparation, route registration, adoption, and process ownership.
packages/…/runnerd-codex-transport.tsRunner launch, provider transport, turn identity, process recovery, and warm reuse.
packages/…/native-session-runtime.tsSession admission, event consumption, recovery continuation, cleanup, and finalization coordination.
packages/…/installation-integrity.tsQualified provider distribution, executable identity, and provider lifetime proof.
packages/…/acpx-runtime-adapter.tsQualified ACPX runtime, provider events, requests, provider cycles, and retained cleanup.
packages/…/permission-policy.tsFail-closed provider permission decisions and Paperclip action allowlist.
packages/…/codex-app-server-driver-impl.tsCodex session admission, resume checks, turn reconciliation, and provider identity.
packages/…/codex-harness-session.tsCodex turns, runtime requests, goals, reconciliation, snapshots, and close.
packages/…/company-observability.tsRedacted, bounded provider observability projected from PRP events.
packages/…/semantic-action-catalog.tsCanonical Paperclip semantic operations and their provider-neutral schemas.
packages/…/paperclip-discovery.tsRun-bound optional tool projection and namespace search.
packages/…/dispatcher.tsTest control-plane discovery, validation, authorization, dispatch, and denial results.
packages/…/types.tsSemantic authorization, receipt, binding, and result contracts.
packages/…/responsibilities.tsResponsibility discovery, filing, proposals, and company changes.
packages/…/update-routine.tsRecurring work and schedule revision contract.
packages/…/update-skill-details.tsCompany skill catalog metadata update contract.
packages/…/capability-contract.tsGenerated capability inventory and evaluation contract metadata.