Back to changelog

v1.2.0 · Understand repository code · Bound Memory growth · Assess write evidence · Generate Experiences and import Skills in Codex

PowerContext 1.2.0 adds optional native repository understanding, Memory capacity controls and an evidence-based write gate, and introduces an experimental ZCode integration. In Codex, users can generate Experience and Skill candidates and import external Skills for review. It also adds paired continuation evaluation with Codex and Bub and automates release-version maintenance.

Highlights

  • Index mixed Python, JavaScript, TypeScript, and Go Git worktrees; locate definitions, inspect static relationships and possible affected tests, and read source evidence through CLI, HTTP, Python Client, Runtime, and MCP queries. Code understanding requires the code extra, is disabled by default, and needs no generation or embedding model. Local indexing has been verified on Linux; macOS and Windows local indexing remain unverified.
  • Refresh code evidence with local index and sync commands. Queries detect changed source content and reject stale evidence; optional prepare_context and Codex or Claude Code hook settings can include bounded code context. Embedded seekdb stores graph data in its database with local evidence files; other deployments use a separate SQLite cache. Changing the backend or database path requires a new index, and static relationships do not prove runtime calls or that tests can be skipped.
  • Inspect Memory capacity through Runtime, Client, HTTP, or MCP. Default per-Revision budgets are 5,000 active entries, 10,000 manifest entries, and 4 MiB of canonical content; writes that increase an exceeded dimension return memory_capacity_exceeded. Optional in-process compaction can remove eligible untagged inactive entries from the current manifest. Compaction is disabled by default and does not delete historical Revisions or cap total database size.
  • Optionally assess Memory candidates against their cited evidence before writing. Decision assistance and the Memory write gate are both disabled by default. When enabled, the gate can refuse unsupported or over-budget writes and surface reasons through explicit operations, flush results, and worker diagnostics; decision-backend failures fall back to the existing write path. Held extraction windows still advance their cursor and are not automatically retried. Enable and calibrate the gate separately from extraction.
  • Update Memory full-text and vector search projections incrementally while reusing unchanged embeddings, preserve canonical Memory error details through HTTP, and fix SQLite Topic Memory publication on builds without RETURNING support.
  • Use Codex tools to generate or propose Experience and Skill candidates and discover or import external Skills, with explicit review before approval. Bind a checkout to an existing Scope for later sessions, reuse saved credentials for native MCP, and distinguish basic connectivity from full tool coverage in doctor diagnostics. External Skill discovery reads configured roots on the Server host; importing or approving a Skill does not install or execute it.
  • Connect ZCode CLI and Windows desktop through the new community, experimental integration, including setup, doctor, prompt capture, prepared-context injection, MCP operations, and Handoff workflows. Live coverage includes CLI 0.16.9 and Windows desktop 3.14.3; other desktop versions and direct remote HTTPS ingress remain unverified.
  • Honor Hermes pre-compression checkpoint API v2 and skip automatic writes outside primary Agent contexts. Keep WorkBuddy recall focused on the submitted turn and bound Scope lookup time; allow DSH cold readiness probes to finish. Improve native service startup waits and avoid background throttling on macOS.
  • Support embedding providers that reject the dimensions request field while retaining configured vector dimensions. Allow bounded Qwen thinking settings for Topic Memory, and record model usage outside model deadlines so accounting failures do not fail the model operation.
  • Compare PowerContext off and on across independent continuation sessions with either Codex or Bub, recording host, model, and reasoning settings and grading the final recall step. Add LoCoMo-Plus evaluation and a LongMemEval-V2 smoke adapter. LongMemEval-V2 validation is limited to the smoke workflow; full-tier execution remains unvalidated.
  • Maintain release references with make version-bump VERSION=1.2.0 and verify consistency with make version-check. Before publishing, make release-check also checks for matching bilingual website release notes and installation links. Package versions remain derived from powercontext-v tags, and existing release notes retain their history.
  • Upgrade the Server, clients, and Agent plugins together, reopen Agent sessions, and back up existing databases before startup schema updates. Check existing Memory against the new capacity defaults before resuming growth. Code indexing, decision assistance, the write gate, and compaction require explicit opt-in. Older deployments with unfinished Artifact processing migration must complete it before starting the new configuration; Windows remains experimental.
uv tool install --force "powercontext[cli,server]==1.2.0"
View the full GitHub release