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 / adapter-integrations

Agent adapter integrations

Agent adapters connect Paperclip agents to local command-line tools, remote agent services, and other execution harnesses. The adapter registry gives each harness one shared identity across execution, model choice, authentication, capability checks, and the agent setup UI. For the wider agent lifecycle, see agent work management, issue execution, execution environments, agent instructions, native runner execution, runner evaluation and conformance, case handoffs and continuity, and reusable workflows.

Adapter registry

The adapter registry is the server-side source for the harnesses that Paperclip can run. Each adapter type maps to one server module. The UI uses the same type to select its label, form, capabilities, and transcript behavior.

The registry has three important views:

ViewPurpose
Registered adaptersAll built-in and loaded external adapter types.
Active adapterThe implementation used for new requests.
Enabled adaptersRegistered types that remain available in agent selection menus.

Built-in adapters are registered with the server. External adapters load from the persistent plugin store at startup. A package that uses the type of a built-in adapter can replace that implementation. Paperclip keeps the built-in implementation as a fallback, so an administrator can pause the override or remove it without losing the original type.

The registry waits for external loading before it validates adapter types. This prevents a valid local or npm adapter from being rejected during the startup window. A disabled adapter is hidden from new agent menus but remains usable by existing agents.

This sequence shows how an adapter becomes available to a run and to the UI.

flowchart LR
    A[Built-in modules] --> R[Adapter registry]
    P[Plugin store records] --> L[External package loader]
    L --> R
    R --> S[Server execution and model services]
    R --> I[Adapter listing API]
    I --> U[Agent setup UI]
    R --> T[Enabled or disabled selection]

References: server/…/registry.ts• L506-649 server/…/registry.ts• L807-817 server/…/plugin-loader.ts• L270-292 server/…/adapters.ts• L261-317

Adapter packages

An adapter package is a deployable unit that describes one agent harness to Paperclip. Its server entry defines the adapter type, execution behavior, environment checks, and optional model, session, skill, authentication, or configuration features.

Paperclip ships bundled adapters for common local and gateway-based runtimes. External packages follow the same adapter contract but have their own release cycle. They can come from npm or from a local directory.

Package sourceRuntime meaningUpdate path
BundledShips with the Paperclip server release.Update Paperclip.
npm packageInstalled into Paperclip's managed adapter directory.Reinstall from npm.
Local directoryLoaded directly from the selected path.Reload after rebuilding the package.

Instance administrators can install, remove, enable, disable, reload, or reinstall external adapters. Cloud-managed instances keep adapter code platform-managed and do not allow runtime code installation. Reloading or reinstalling refreshes the server module and its related UI metadata.

Bundled local adapters also own the way their runtime is made available in a target environment. OpenCode uses its official installer and exposes the installed binary through a normal non-login shell path. The Hermes Gateway adapter connects to an already-running Hermes service rather than launching a local agent process. The old Hermes Gateway package remains a deprecated compatibility shim for the replacement Hermes Paperclip adapter package.

References: server/…/plugin-loader.ts• L149-203 server/…/adapter-plugin-store.ts• L23-35 server/…/adapter-plugin-store.ts• L133-163 server/…/adapters.ts• L320-453 server/…/adapters.ts• L528-761 packages/…/package.json• L15-39 packages/…/index.ts• L1-103 packages/…/index.ts• L5-187 packages/…/index.ts• L1-78 packages/…/index.ts• L161-185 packages/…/package.json• L1-79

Adapter execution lifecycle

The adapter execution lifecycle turns an agent run into a controlled invocation of a local or remote harness. Local adapters prepare the workspace, runtime assets, environment, prompt, and optional Paperclip bridge. They then convert structured output into a shared run result.

Claude, Codex, Gemini, and Kimi select ACP by default when the target and runtime support it. ACP means the Agent Client Protocol path. An explicit CLI choice remains available when the target or its security settings cannot support ACP. The adapter does not silently change engines when prerequisites fail. Claude and Codex require Node 24.11 or newer for their ACP lane; Kimi has a lower Node floor. Remote ACP needs a bidirectional process target, and SSH ACP needs a Paperclip-managed account. SSH host-auth runs stay on the CLI lane.

Codex uses the concrete gpt-6.1-sol default rather than the older bare gpt-5.6 alias. Older saved aliases are normalized before execution. Reasoning and fast-mode controls follow the selected model. Managed SSH ACP runs use the selected account scope for their remote home and can preserve refreshed subscription credentials. ACP classifies explicit usage exhaustion as provider quota, while selected-model capacity is a transient upstream failure. This keeps retry behavior separate from authentication and model configuration errors.

Remote runs stage only the runtime data that the harness needs. Claude stages a sanitized configuration seed and selected skills. Codex stages a curated account home and copies refreshed subscription credentials back under its identity and freshness rules. Gemini and Kimi stage selected skills without changing operator-owned shared skill state. Grok stages project instructions and skills as temporary remote project assets. OpenCode and Pi stage isolated configuration and runtime assets. Paperclip restores workspace changes at the result boundary and removes temporary assets without replacing the provider result with a cleanup error.

Each adapter maps provider events into the same result contract. OpenCode consumes its JSONL stream as one event flow, which lets Paperclip keep assistant text, usage, cost, tool errors, and provider errors together. Sessions resume only when their working directory and execution identity still match the current run. Gemini, Kimi, Grok, and OpenCode retry an unavailable saved session with a fresh session; Pi creates a new session file when its saved file or working directory no longer matches.

This flow shows how a run request becomes a durable result across local, remote, and gateway adapters.

flowchart LR
    R[Run request] --> P[Resolve adapter lane and target]
    P --> A[Prepare workspace, auth, skills, and prompt]
    A --> H[Invoke local process, SDK, or gateway]
    H --> O[Parse structured output or events]
    O --> S[Shared run result]
    R --> C[Execution control]
    C --> D[Stop and phase diagnostics]
    D --> S
    S --> T[Remote teardown and session state]

Paperclip keeps stop control separate from provider behavior. A stop that arrives before an adapter publishes its execution control holds registration until the stop write settles. A stop waits for execution and cleanup to settle. If it remains unsettled past the report bound, Paperclip records diagnostics without treating them as proof that the provider stopped.

References: packages/…/acp.ts• L98-120 packages/…/acp.ts• L223-365 packages/…/acp.ts• L614-637 packages/…/execute.ts• L413-430 packages/…/cli-capabilities.ts• L40-58 packages/…/index.ts• L11-19 packages/…/acp.ts• L88-163 packages/…/execute.remote.test.ts• L114-168 packages/…/execute.ts• L583-600 packages/…/acp.ts• L71-84 packages/…/execute.ts• L211-227 packages/…/acp.ts• L64-76 packages/…/execute.ts• L344-410 packages/…/execute.ts• L415-495 packages/…/remote-project-assets.ts• L45-60 packages/…/execute.ts• L392-485 packages/…/execute.ts• L426-528 server/…/adapter-execution-control.ts• L4-153

Declarative adapter availability

The declarative adapter registry lets an operator define the complete set of adapter types that an instance offers. PAPERCLIP_ADAPTERS accepts inline JSON. PAPERCLIP_ADAPTERS_FILE points to a JSON file. If neither is set, Paperclip keeps its built-in availability behavior.

An entry names an adapterType and can mark it enabled. In managed Kubernetes execution, the same entry can describe runtime and network settings. The declared list is complete: an omitted known adapter becomes disabled, and a declared type without a loaded implementation is an error.

References: packages/…/adapter-registry.ts• L1-30 packages/…/adapter-registry.ts• L3-16 server/…/adapter-registry-bootstrap.ts• L27-97

Authentication flows

Interactive adapter login lets a user prepare credentials in the same environment that will run the agent. An adapter declares a login capability only when it needs this flow. API-key-only adapters use the normal environment and secret binding path.

The capability describes how to start the login command, read a prompt, select the target transport, choose the UI panel, and limit the login time. It does not store a secret. The server validates this capability before it registers the adapter.

The login flow follows this path:

  1. The user tests the adapter in the environment that a real run would use.
  2. A confirmed missing-auth result selects the matching login panel only when the target supports in-place login.
  3. The server starts an owner-scoped login session.
  4. The user follows a URL or enters a one-time code when required.
  5. The UI polls the session until it reaches success or a terminal failure.
  6. The server checks and promotes the resulting credential through its protected path.

Claude ACP runs a real hello probe for local and remote targets when its configured provider path does not already decide authentication. A sandbox gets the stable adapter_auth_missing signal because it can offer in-place login. A probe timeout, failed command, or incomplete response remains a warning and does not open the login panel. ACP authentication failures use the same Claude-specific run error as the CLI lane, so the UI keeps one sign-in recovery path.

Claude also supports a setup-token login path. It binds the browser prompt to a validated authorization URL, accepts one browser code, and delivers the minted credential only after the credential sink confirms receipt. The parser and runner keep the URL, code, and token out of logs and results, and fail closed when the output does not match the expected success record.

Codex credentials can come from an API key, a host or company login, a target sandbox login, or a company-scoped cache entry for the same account identity. A failed sandbox probe is unknown rather than proof that credentials are absent. SSH host-auth runs use the login and configuration already present in the remote host's native Codex home instead of copying host credentials from Paperclip. Other remote runs stage only the needed Codex home and copy back refreshed subscription credentials at teardown. Promotion is company-scoped and keeps the newer credential for the same account. When several credential sources exist in a sandbox, the adapter reports which source wins instead of hiding the precedence choice.

Gemini accepts API keys or Google account auth. Remote managed homes select the API-key auth mode only when the credential reaches the target. Kimi accepts OAuth login or the paired KIMI_MODEL_NAME and KIMI_MODEL_API_KEY settings. Grok uses API-key or device-login credentials and stages only the credential data needed by a remote run. OpenCode treats a provider response that proves a 401 as an authentication warning. Other provider failures remain probe failures, so the setup UI does not offer sign-in based on an unrelated upstream error.

The login target follows the same environment resolution as a real run: the agent default wins, then the instance default, then the local default. This keeps the test and login result meaningful for the later run.

sequenceDiagram
    participant UI as Agent setup UI
    participant API as Paperclip server
    participant S as Target sandbox
    participant P as Provider login

    UI->>API: Test adapter in resolved environment
    API-->>UI: Missing-auth result and login capability
    UI->>API: Start login session
    API->>S: Run adapter login command
    S->>P: Begin device, browser, or setup-token login
    P-->>S: One-time prompt or credential result
    S-->>API: Prompt and readiness status
    API-->>UI: URL, code, or terminal status

References: packages/…/login-capability.ts• L1-11 packages/…/login-capability.ts• L63-97 server/…/adapter-login-route-spine.ts• L64-106 server/…/device-login-service.ts• L44-87 server/…/device-login-service.ts• L731-798 packages/…/acp.ts• L649-831 packages/…/setup-token-parse.ts• L395-529 packages/…/setup-token-runner.ts• L193-423 packages/…/acp.ts• L538-677 packages/…/execute.ts• L339-452 packages/…/execute.remote.test.ts• L276-302 packages/…/execute.ssh-host-auth.test.ts• L82-218 packages/…/acp.ts• L459-490 packages/…/acp.ts• L349-373 packages/…/index.ts• L72-93 packages/…/test.ts• L402-469

Capability metadata

Capability metadata tells Paperclip which shared features an adapter supports. The server projects safe capability fields through GET /api/adapters. The UI uses that response instead of guessing from an adapter name.

CapabilityProduct behavior
Instructions bundleShows and resolves the managed instructions file.
SkillsEnables skill listing or synchronization.
Local agent JWTAllows the heartbeat to create a local agent token.
Materialized runtime skillsWrites skill entries to the runtime when the harness scans disk.
Model profilesEnables named secondary model settings.
ACPShows that the adapter can use the Agent Client Protocol path.
LoginSelects the interactive login panel and sandbox transport.

External adapters opt into these features through their returned module. Unknown types use an all-false UI fallback until the adapter listing arrives. The UI then registers external types and synchronizes disabled state. A native-runner candidate is discovery metadata for a qualifying target and model, not a durable adapter setting.

References: packages/…/types.ts• L420-512 server/…/adapters.ts• L102-132 server/…/adapters.ts• L167-203 server/…/adapters.ts• L267-317 ui/…/adapters.ts• L18-57 ui/…/use-adapter-capabilities.ts• L15-56 ui/…/use-disabled-adapters.ts• L19-56

Model configuration

Model configuration connects an agent's selected model to the adapter that can run it. An adapter may provide a fixed list, discover models from its runtime, refresh a cached list, define model profiles, or detect a model from local configuration.

The server resolves model choices in this order:

  1. Use an operator-declared list from PAPERCLIP_ADAPTER_MODELS when one exists.
  2. Otherwise ask the adapter to discover models when it provides discovery.
  3. Otherwise use the adapter's bundled model list.

An explicit refresh uses the adapter refresh hook first. If it returns no models, Paperclip falls back to discovery and then to the bundled list. Model profiles are separate from the primary model. A cheap profile can hold a lower-cost model and adapter settings without changing the agent's main model.

Claude defaults to Fable 5.2 when no model is set. An explicit model wins over ANTHROPIC_MODEL, while Bedrock and Vertex can keep their provider-owned model choice. The available reasoning levels and the required Claude Code version depend on the selected model. Standard API discovery and Bedrock discovery use different model identifiers, so the picker must follow the active provider mode.

Codex keeps the selected model consistent across CLI, ACP, and native configuration. The current default is gpt-6.1-sol; old bare gpt-5.6 values normalize to the concrete gpt-5.6-sol identifier. The UI exposes reasoning and fast-mode controls only when the selected model supports them. The adapter also records model-specific CLI floors and distinguishes models retired for ChatGPT sign-in from models that remain usable with API keys. Gemini, Grok, and Kimi keep their selected model and engine settings aligned across their CLI and ACP paths. Kimi maps supported effort settings to its provider-specific thinking control on the CLI lane and removes unsupported effort controls from ACP.

OpenCode and Pi use provider-aware model identifiers. OpenCode validates a selected provider/model pair in the target environment. Pi materializes custom provider configuration before it checks model availability and stages that configuration for remote runs. Remote model checks use the target environment rather than the controller's catalog.

The agent form fetches models for the selected adapter and environment. It can refresh the list, fetch profile definitions, and ask adapters that support detection to report local provider settings. It also preserves adapter-specific settings through built-in fields or a server-provided schema.

References: server/…/registry.ts• L663-730 server/…/adapter-models-env.ts• L1-40 server/…/agents.ts• L2347-2386 packages/…/index.ts• L1-58 packages/…/index.test.ts• L5-79 packages/…/cli-capabilities.ts• L40-58 packages/…/models.ts• L121-181 packages/…/index.ts• L11-19 packages/…/index.ts• L44-188 packages/…/index.test.ts• L16-77 packages/…/codex-args.test.ts• L27-176 packages/…/acp.ts• L95-125 packages/…/acp.ts• L88-132 packages/…/execute.ts• L269-279 packages/…/index.ts• L45-83 packages/…/execute.ts• L95-173 packages/…/execute.ts• L339-413 ui/…/AgentConfigForm.tsx• L688-770 ui/…/AgentConfigForm.tsx• L1142-1259

Adapter-facing UI integration

The UI combines static display metadata with dynamic adapter data so a newly installed adapter can appear in agent creation without a Paperclip source change. Known types get curated labels, descriptions, and icons. Unknown external types get a readable label and a generic external-adapter display.

The agent form chooses the configuration path from the adapter type. Built-in adapters can provide dedicated fields. External adapters can return a declarative schema with text, number, toggle, select, or combobox fields. The server resolves dynamic options before returning the schema. The UI applies defaults and visibility rules while it edits the agent configuration.

Native runner configuration is a separate boundary. The Codex form keeps auto as an unpinned engine choice while preserving explicit CLI or ACP pins. Its runner profile removes settings that the runner does not implement, keeps an explicit search-off choice, requires full-auto Codex permissions, and rejects unsupported profiles or unsafe bounded limits. Temporary candidate metadata can admit an exact discovered harness and model at save time, but it is never stored as runner configuration.

Run output has a separate parser path. An external package may export a self-contained UI parser. The server checks and caches it. The browser loads it on demand in a worker. If the parser is missing, incompatible, or fails, the UI uses a generic parser so raw output remains visible. Reloading or reinstalling an external adapter invalidates its server schema cache and UI parser cache.

This diagram shows how external adapter metadata and output parsing reach the browser.

flowchart TD
    E[External adapter package] --> M[Server adapter listing]
    M --> F[Agent form and adapter menus]
    E --> C[Config schema endpoint]
    C --> F
    E --> U[UI parser source]
    U --> W[Sandboxed parser worker]
    W --> V[Run transcript]
    F --> A[Adapter test and login panel]

References: ui/…/adapter-display-registry.ts• L21-30 ui/…/adapter-display-registry.ts• L167-195 ui/…/AdapterMark.tsx• L1-59 ui/…/schema-config-fields.tsx• L202-264 ui/…/schema-config-fields.tsx• L322-513 server/…/adapters.ts• L763-825 ui/…/dynamic-loader.ts• L210-284 ui/…/AgentConfigForm.tsx• L445-455 packages/…/build-config.ts• L34-327

Change checklist

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

  • When an adapter type or capability changes, keep the server module, adapter listing, UI capability defaults, and UI display registration aligned.
  • When a login capability changes, update validation, target transport, safe projection, and the panel mode selected by the UI.
  • When an environment test changes, keep local and remote credential sources distinct and preserve the stable missing-auth result used by sandbox login controls.
  • When Claude ACP changes, keep engine prerequisites, hello-probe status, auth error mapping, model environment values, remote config seeding, and workspace restore aligned with the CLI lane.
  • When Claude setup-token handling changes, preserve strict URL and success-record parsing, one-time code submission, credential-sink acknowledgement, and secret removal from logs and results.
  • When Claude model metadata changes, keep the default model, provider override rules, reasoning levels, model discovery mode, and Claude Code version gates aligned between the picker and both execution lanes.
  • When Codex model metadata changes, keep aliases, model order, reasoning and fast-mode support, CLI version floors, ChatGPT retirement messages, and native-runner model checks aligned.
  • When Codex credential handling changes, preserve identity matching, company isolation, sandbox-auth precedence, effective-home selection, and teardown copy-back rules.
  • When Codex SSH host-auth handling changes, preserve remote-home discovery, profile selection, pre-install change detection, managed-skill ownership, and cleanup of temporary runtime files.
  • When Gemini, Kimi, Grok, OpenCode, or Pi runtime preparation changes, keep target-side assets, provider auth state, managed-home scoping, model validation, session identity, and cleanup aligned.
  • When OpenCode or Pi provider configuration changes, keep remote model checks and staged configuration aligned with the local validation path.
  • When native runner form mapping changes, keep exact candidate-model matching, save-time admission, draft selection, non-persistence of temporary candidate context, and provider-specific field filtering aligned.
  • When an external package entry point or UI parser export changes, keep package exports, loader checks, caches, and invalidation in sync.
  • When stop handling changes, preserve the registration barrier, settlement wait, bounded diagnostics, and the rule that diagnostics do not prove termination.

Important failure modes

TriggerConsequenceGuard
Declared adapter type has no loaded implementationAvailability reconciliation fails instead of offering a non-runnable adapter.Registry reconciliation checks each declared type against loaded server adapters.
Claude, Codex, Gemini, or Kimi ACP prerequisites are missingThe selected ACP run cannot start.Each adapter reports an explicit setup error and does not silently switch engines.
Remote ACP target has no bidirectional process channelACP cannot maintain the provider session.Remote ACP is admitted only for a process-session bridge or managed SSH account.
Claude ACP auth probe times out or does not completeThe UI can mistake an untested target for a logged-in target.Probe failure stays a warning and does not produce the login affordance.
Claude setup-token output contains a look-alike URL or tokenUntrusted terminal text could be treated as a credential.Parsers bind values to exact anchors and accepted shapes, and fail closed.
Claude setup-token credential sink is missing or rejectsThe login could report success without storing the credential.The runner reports success only after the mandatory sink resolves.
Claude selects a model that the installed Claude Code cannot supportACP or CLI execution fails before the provider starts.The adapter checks the model's minimum Claude Code version and probes the target command when needed.
Codex model is sent as the bare gpt-5.6 valueCodex lacks model metadata and falls back to generic context limits.The adapter normalizes the alias to gpt-5.6-sol and advertises concrete model IDs.
Codex model needs a newer CLI than the target providesChatGPT-backed Codex rejects the model even though credentials are valid.The adapter records model-specific CLI floors and reports the version gap.
Codex login promotes a different or older credentialA company can switch accounts or lose a newer rotating token.Promotion is company-scoped and compares account identity and freshness before writing.
Codex subscription refresh is not copied back from a remote homeThe single-use refresh token is spent and later host runs need re-authentication.Remote subscription teardown copies back only a newer same-identity credential.
Codex sandbox auth probe failsA real sandbox login can be treated as absent or a missing login can fail without a clear cause.Probe failure is unknown, not absent, and the adapter reports credential-source precedence.
Gemini remote auth selector is written without a target credentialHeadless Gemini rejects the selected auth mode.The selector is written only when the credential reaches the remote run.
Kimi ACP receives CLI-only effort controlsThe ACP backend rejects the session configuration.ACP removes unsupported effort fields and keeps thinking control on the CLI path.
Grok remote project assets are staged outside the approved project pathsCleanup or setup could touch unrelated remote files.The remote asset helper validates path roots, names, parents, and cleanup receipts.
OpenCode remote model probe cannot runA transient probe issue can incorrectly block useful work.The probe is best effort; the real invocation remains authoritative.
Pi custom provider configuration is not staged with the remote runA model works locally but is unavailable on the target.Pi materializes and transfers managed provider configuration before execution.
A non-qualified native runner candidate is saved without its current exact modelA temporary or unsupported harness can enter durable runner configuration.Save-time validation accepts the candidate only when current metadata matches the selected model.
Stop arrives while adapter control is not registeredReadiness can expose a stale running state.Pending stop ownership blocks registration until the terminal write settles.
UI parser is absent, incompatible, or failsThe transcript can lose its adapter-specific presentation.The browser falls back to a generic parser and keeps raw output visible.

Key files

FileWhy to read it
server/…/registry.tsDefines shared adapter lookup, registration, overrides, model resolution, and capability-aware access.
server/…/plugin-loader.tsShows how npm and local adapter packages are validated, loaded, and reloaded.
server/…/adapters.tsDefines adapter management, safe capability projection, candidate discovery, and metadata exposed to the UI.
server/…/adapter-execution-control.tsDefines stop ownership, readiness barriers, settlement waits, and diagnostic reporting.
packages/…/types.tsDefines the adapter module contract and optional capabilities.
packages/…/login-capability.tsDefines the non-secret interactive login contract and validation rules.
server/…/device-login-service.tsCoordinates owner-scoped login sessions, target leases, credential promotion, and cleanup.
packages/…/index.tsDefines Claude defaults, model metadata, reasoning levels, and configuration contract.
packages/…/acp.tsShows Claude ACP selection, target checks, authentication probes, billing identity, and remote runtime setup.
packages/…/cli-capabilities.tsDefines Claude model version floors and target command capability checks.
packages/…/models.tsShows provider-aware Claude model discovery, fallback, and caching.
packages/…/setup-token-parse.tsShows the fail-closed setup-token prompt and credential boundaries.
packages/…/setup-token-runner.tsShows one-time prompt, code, credential delivery, timeout, cancellation, and secret cleanup.
packages/…/index.tsDefines Codex defaults, model aliases, reasoning and fast-mode metadata, and configuration contract.
packages/…/acp.tsShows Codex ACP prerequisites, credential staging, copy-back, billing identity, and typed failure classification.
packages/…/execute.tsShows credential readiness, target selection, managed homes, runtime assets, and execution handoff.
packages/…/codex-home.test.tsDemonstrates managed-home isolation, identity-aware credential retention, and remote home staging guarantees.
packages/…/execute.remote.test.tsDemonstrates managed SSH account scope, remote home staging, conversation continuity, and cleanup.
packages/…/execute.ssh-host-auth.test.tsDemonstrates native SSH-host authentication with Paperclip-owned skills and runtime configuration.
packages/…/acp.tsShows Gemini ACP setup, credential signals, managed home, and remote skill delivery.
packages/…/acp.tsShows Kimi ACP setup, supported controls, credential checks, and engine limits.
packages/…/execute.tsShows Grok credential-home selection, project asset staging, remote execution, and session recovery.
packages/…/remote-project-assets.tsDefines safe remote installation and cleanup of Grok instructions and skills.
packages/…/index.tsDefines OpenCode installation, provider/model identifiers, and configuration contract.
packages/…/execute.tsShows isolated OpenCode configuration, target-side model checks, session handling, and cleanup.
packages/…/execute.tsShows Pi provider configuration, skill and extension staging, session files, and remote execution.
ui/…/AgentConfigForm.tsxShows how model, capability, login, environment test, and adapter configuration meet in agent setup.
ui/…/adapter-display-registry.tsDefines labels, descriptions, icons, and display behavior for bundled and external types.
ui/…/schema-config-fields.tsxShows the adapter-neutral form path for dynamic external schemas.
ui/…/dynamic-loader.tsShows the browser loading boundary for external transcript parsers.