v1.5.0 · Source and Artifact plugins · Gateway alpha · Historical database migrations
PowerContext 1.5.0 adds managed Source and Artifact plugins, content-free preparation receipts, and a metadata-only Atomic Memory manifest. The Rust PowerContext Gateway debuts in alpha with memory injection and inference admission checks. Migration tooling adds normalization for unmanaged historical Server storage on SQLite, seekDB, and OceanBase, subject to the release upgrade restrictions below. Desktop and local Ops improve connection and Scope selection, identity recovery, credential storage, and managed-runtime installation and repair.
Highlights
- Extend Source ingestion with trusted local plugins for synchronous writes and ongoing collection. Hosts discover capabilities and manage collection targets through the public MCP; separate collect Runtimes retain checkpoints and explicit lifecycle state. Plugin loading requires deployment administration, and local collect process management currently requires Linux.
- Generate custom Artifact families with managed plugins, versioned schemas, exact Source lineage, and declared text or vector search. Administrators load, enable, disable, reload, or unload bindings; initial registrations are disabled. Saved results remain readable when generation is disabled or plugin code is removed. Included incident-report and ticket-summary examples demonstrate the workflow.
- Request include_receipt during context preparation to inspect selected exact revisions, retrieval and fallback modes, omissions, configuration, and stage timings without returning query or evidence bodies. Receipts are bounded, ephemeral diagnostics; unavailable or oversized receipts leave the original prepared context usable. Normal host injection keeps the default response shape.
- Browse a bounded, paginated Atomic Memory manifest before reading selected exact revisions. New writes retain a short description, with a text preview for explicit writes and descriptions generated within existing extraction calls. Historical descriptions can remain absent; reading the manifest never generates summaries. Extraction instructions also preserve subject, scope, and applicability when facts concern different entities.
- Opt in to final inference request and complete-response checks at owned provider transports. Enforcing mode refuses forbidden content or incomplete inspection; disabled remains the default. Public context admission and short-lived exact Atomic Memory reuse proofs let external executors check their final payloads and current permissions. This protection does not provide general-purpose redaction or a provider cache.
- Introduce the optional Rust PowerContext Gateway in alpha, with OpenAI-compatible Chat Completions and Responses endpoints, request checks before sending, and complete-response checks before release. Opt-in memory injection for Chat Completions and capture of turns that pass release checks support later recall; both are disabled by default. Build and run the Gateway separately, either standalone or alongside Core over a private Unix domain socket. The alpha buffers complete responses for inspection and does not provide real-time token streaming to clients. Its APIs and behavior may change.
- Configure reversible input encoding separately for Atomic Memory extraction/reconciliation and Experience generation/incubation. Shadow mode sends original inputs; experimental active mode selects an exactly restorable, smaller representation when eligible. Encoding is off by default and does not establish equivalent model understanding, privacy redaction, or lower total task cost.
- Guide Desktop connection activation into exact Scope selection and gate each action on discovered operation compatibility. Revalidate Server and Principal identity around private reads and writes; identity changes clear active context, while uncertain writes retain a read-only recovery draft. Add bounded session diagnostics and Linux Secret Service credential storage, with an explicit session-only option when persistent storage is unavailable.
- Install and repair fixed managed local Server runtimes through independent Ops preflight and apply operations, with a constrained native Desktop bridge. Repair preserves configuration and data and leaves the service stopped. Existing managed SQLite startup checks use a private main-file and WAL snapshot through the exact target bundle. Resource candidates remain unqualified; a public production catalog and clean-machine native distribution acceptance are separate requirements.
- Add experimental Cindy scoped recall and automatic memory hooks, with a separate native MCP connection for the full tool catalog. Tighten pre-compaction lifecycle handling, bound recall queries in Pi, OpenCode, DSH, and Bub, run Codex SessionStart on forked sessions, and support Claude Code hooks on the macOS system Python 3.9.
- Bound all query-embedding waits within one context preparation budget and reduce repeated reads in Candidate review, Experience incubation, Handoff Reports, and archive compatibility checks. Add durable recovery intents for external ownership writes, safe processing failure classifications, and Code publication journals that keep newly orphaned generations reclaimable. Identical rendered code evidence no longer changes with query time.
- Deploy API and background roles on Kubernetes using the new Helm chart with an external OceanBase database, pinned image, existing Secrets, and explicit database initialization. Linux and macOS Servers can also use a private Unix domain socket for same-user proxies. Socket transport retains authentication and requires explicit external-scheme configuration for HTTPS Dashboard access.
- Default new seekDB and OceanBase selections to the powercontext database while preserving explicit database names. Legacy seekDB data in test requires selecting that database explicitly; no data is moved automatically. Provision only confirmed missing databases during authorized initialization, preserve read-only maintenance, and improve seekDB restart recovery and OceanBase full-text schema reflection.
- Add a bilingual website blog with author profiles and evaluation tooling for recovery from failed MCP calls, capture ownership, and encoded-input quality. Evaluation fixtures and published measurements do not certify every Agent host, model, or native platform.
- Upgrade Server, Python SDK, CLI, and integrations together: older clients reject the new context capability field even when protection is disabled. Prepare requests no longer accept explain, and responses no longer include top-level reason; use include_receipt and receipt.retrieval.searches[].fallback_reason, checking that the optional receipt is present. Update custom Atomic Memory prompts for the v2 extraction/reconciliation definitions.
- Add recognition and normalization of unmanaged historical Server storage through server db-migrate on SQLite, embedded seekDB, and OceanBase. Inspect the actual structure and data state before joining the existing migration chain, preserving Scope identities and compatible extensions. Reuse legacy Memory-to-Atomic and Processing conversions, with read-only plans, existing backup and verification requirements, SQLite rollback, and recoverable native DDL operations. Legacy Memory conversion with vector search requires the configured embedding provider during apply; existing Atomic embeddings remain preserved. Historical normalization does not expand the installed-release acceptance matrix or certify every legacy layout.
- Existing managed databases require explicit migration for owner recovery, Atomic Memory descriptions, Source plugin registrations, and Artifact plugin storage through v1_5_0_001. Normal publication requires all twelve installed-1.3.0 upgrade combinations across SQLite, embedded seekDB, and OceanBase, including both seekDB Code configurations, against the exact target artifact. No intermediate release is added automatically; the registered 1.4.0 bridge remains outside that acceptance matrix. The policy remains upgrade-required pending acceptance and review of the support declaration.
- Upgrades from 1.3.0 and 1.4.0 require passing full Server acceptance on SQLite, seekDB and OceanBase. These migration paths are not yet verified; do not upgrade existing databases.
uv tool install --force "powercontext[cli,server]==1.5.0"View the full GitHub release