Plompy 0.6.0
Integrations and recovery without losing the thread
This is a historical release. Its notes describe this version, not the current setup. See the latest stable release before starting a new installation.
This version predates Plompy Local. It covers the Discord engine; the separate local MCP gateway and starter kit arrive in 0.13.0.
GitHub notes and assets for this version may use the earlier Szpont name. Historical filenames and source links are preserved.
What changed#
Guided integrations#
- Choose Codex, Claude Code or OpenCode before configuring integrations, with explicit
--runnertargeting and private managed MCP defaults for the host adapters. OpenCode model providers, including Go, use the OpenCode adapter; configuring tools does not switch existing Discord threads. - Add an integration setup CLI for GitHub, Camoufox, OpenAI Docs, Context7 and operator-reviewed MCP servers. Keep account authorization separate from local configuration checks; Camoufox still requires a separately provisioned browser and display.
- Give GitHub CLI a dedicated private credential directory and Docker volume; keep parent GitHub tokens out of agent subprocesses.
- Load bounded, runner-specific integration playbooks on each task. Preserve operator instructions, unrelated JSON settings and conflicting MCP entries; reject unsafe managed files.
- Include the configurator and GitHub CLI in the Codex-only image, with the existing instance lock protecting setup. Exercise packaged setup offline without model calls, account actions or browser downloads.
Recovery without losing the thread#
- Let
!recoverselect the task blocking the current thread, or accept an explicit task ID and owner instruction. Reconcile saved state and external action receipts before continuing the same provider-bound session. - Allow a fresh explicit
!workbudget in a blocked thread to reconcile prior actions and continue without requiring!new-session. Replace queued follow-ups atomically while retaining history, image inputs and previously charged execution. - Preserve the newest accepted recovery budget when recovering an earlier ancestor. Use the durable acceptance sequence rather than mutable timestamps or elapsed totals to order recovery state.
- Include the actual subprocess deadline in workflow prompts and request a checkpoint before that deadline. This guidance does not guarantee timely model completion or renew an exhausted budget automatically.
Release boundary#
- Retain public SQLite schema 4; no new migration is introduced. Keep verified private backups and reconcile external effects before recovery or rollback.
- Keep Claude Code and OpenCode as experimental host adapters, with operator-provided model access and isolation. The Docker image remains Codex-only; live Discord, model, browser and account acceptance remain deployment checks.
- Align the browser package's public version constant and MCP initialization metadata with the coordinated release version.
Downloads and source#
This GitHub release has no separately uploaded download assets. Its source and published distribution details are linked from the release page.
Original GitHub release · Source at v0.6.0 · Changelog at this tag
Before you upgrade#
Back up private state and credentials, review the version's migration boundaries, and retain a compatible rollback snapshot. A release does not update an existing deployment automatically.