Skip to content

Latest commit

 

History

History
119 lines (93 loc) · 6.37 KB

PaseoShortcomings.md

File metadata and controls

119 lines (93 loc) · 6.37 KB

Paseo shortcomings

Candidate upstream issues, checked on 2026-10-03 against the local Paseo checkout and the installed 0.11.0-beta.3 host. PlexusPi uses the public 0.12.0 plugin contracts. No Paseo source was modified.

The documentation audit reconfirmed the resume-template and child-outline source gaps in the local checkout. These are upstream limitations, not permission to patch Paseo. Current PlexusPi features and the self-development workflow are in the provider README.

Plugin providers cannot supply native resume commands

Repro: right-click a PlexusPi conversation and choose Copy Resume Command. The app reports "Resume command not available".

packages/app/src/utils/provider-command-templates.ts hardcodes built-in provider IDs. It has pi and omp, but not plexuspi, omp-plugin, or OMP's profile-specific plugin IDs. workspace-screen.tsx's handleCopyResumeCommand reads only runtimeInfo.sessionId or persistence.sessionId.

The generic plugin adapter uses its bridge ID or an opaque plugin:{...} persistence handle for that value, not the native Pi/OMP transcript identity. Adding a hardcoded plexuspi template alone would still give the wrong session argument. The public provider contract has no native resume-command field.

Paseo-OMP confirmation: the inspected plugin registers omp-plugin in paseo-omp/server/provider/registration.ts, and profile providers in paseo-omp/index.server.ts. Those IDs have no host resume template, so the same menu resolves to unavailable. This is a source-contract confirmation, not a separate interactive OMP smoke test. Built-in omp is a different provider.

Needed API: let a provider return a native resume command (or a native resume identity with a provider-declared template). Don't interpret opaque persistence data as a CLI session ID.

PlexusPi workaround: the agent command-center action PlexusPi: Copy Pi resume command (stop session first) resolves the owned transcript server-side and copies a quoted terminal command. Stop/detach the Paseo session before running it; Pi's terminal process must not write the same transcript concurrently. Unsaved/ephemeral sessions cannot supply a resume command. The command preserves the daemon's Pi agent directory, but does not copy per-agent environment overrides or credentials. Supply any needed environment in the terminal yourself. The command uses POSIX shell quoting.

Provider child views trigger top-level prompt-index requests

Observed: opening the two explore children of root agent e648c66a-45e1-42a1-960a-b6099616d873 logged errors at 19:39:36 and 19:42:31 UTC. Both were agent.timeline.list_prompts.request failures with Agent not found: provider:<root>:superpi-child%3A<run> (observed under the earlier Superpi child id; the current prefix is plexuspi-child:).

provider-subagent-panel.tsx renders an AgentStreamView using a synthetic provider: stream ID. agent-stream/chat-outline/use-chat-outline.ts requests its prompt index. The daemon's handleAgentTimelineListPromptsRequest in packages/server/src/server/session.ts calls ensureAgentLoaded unconditionally, but provider children are not top-level stored agents.

The client catches the rejection without displaying an error. That explains why the child UI looked fine: execution and timeline rendering aren't broken; the optional chat outline cannot load. It is a real invalid request, not merely a misleading log level.

Needed fix: skip top-level outline requests for provider child streams, or support a provider-child prompt index. Don't register fake top-level agents or hide all errors to work around it.

Rewind cannot replace a plugin provider's mirrored history

Pi's active branch rewinds successfully, but abandoned rows can remain visible in Paseo. The generic adapter's streamHistory() stays append-only after session.revert; host forced hydration replays those same abandoned rows. The provider event contract has no history replacement/reset operation.

Needed API: an explicit supported history replacement after revert. See the executable verification and tests/rewind-host-contract.test.ts for the installed-host repro.

Plugin commands have no out-of-band dispatch hook

The host's PluginAgentSession does not implement tryHandleOutOfBand. Provider slash commands use the ordinary prompt path, where the host may interrupt the current turn before the provider sees them.

PlexusPi uses session.configure for live model/tier/context controls, so those controls can use configuration rather than chat commands. The slash-command alternatives should be used while idle; they may interrupt a running turn before reaching PlexusPi. Manual /compact is idle-only. This does not make arbitrary extension commands safe to invoke during generation.

Needed API: a provider-declared non-interrupting command/control path.

Generic permission cards cannot preserve Pi editor prefill

Pi's editor dialog supplies initial text, but the generic host question-card contract does not carry that prefill into its input. PlexusPi's Pi dialogs settings screen preserves it through a plugin RPC and native text input.

Needed API: initial values and multiline editor semantics on question inputs.

Do not file these as Paseo defects

  • Missing PlexusPi model labels and built-in slash-command advertisement were plugin bugs, fixed here. Pi's get_commands omits terminal built-ins; the plugin must advertise and dispatch the RPC-supported ones explicitly.
  • Long Context's old fixed target was PlexusPi policy, not a host defect. It is now replaced by Plexus policy metadata.
  • MCP translation, arbitrary Pi TUI widgets, and Windows descendant cleanup remain PlexusPi implementation limits unless a specific missing host contract is demonstrated. Direct steering now uses Pi's native RPC and the public prompt.steer capability; it does not need a host workaround.
  • Reload worked in the user's test. The earlier ERR_IPC_CHANNEL_CLOSED during worker shutdown is recorded, but not enough by itself to claim a persistent reload failure.
  • Plugin reload deliberately tears down owned Pi processes. Ending the PlexusPi development conversation by reloading its own provider is expected cleanup, not a host bug. Test on the isolated daemon and reload the active host from a separate terminal after saving the handoff.