version 0.5.2

PATCH

version 0.5.1

Changed
No library code changes in this release.
Modernized the local manual E2E smoke test (test_mcp.py, not part of the pytest suite) to cover the full 0.5.0 tool set — it was stale and only exercised tools present before 0.5.0. Now verifies ctx_resolve, ctx_bug, ctx_doc, how_to_ctx, and the ctx_export/ctx_import object-form round-trip (re-confirming the 0.5.0 import fix) against a live server.
Notes
Purely a maintenance/version bump. If you're already on 0.5.0, there's nothing functional to gain from upgrading — safe to skip this release for production installs.
99 tests green (pytest suite, unaffected by this change).
Upgrade (optional):

pip install -U one-ctx

version 0.5.0

Fixed
ctx_import no longer fails on the natural usage pattern. Passing the object returned by ctx_export straight back into ctx_import used to crash (TypeError: ... not dict) because it only accepted a hand-stringified JSON string. This was the most common reported import failure — assistants naturally pass the export result as-is. ctx_import now accepts the object, a JSON string, or a list of project objects.
Added
ctx_doc — verbatim per-project documents. Unlike the WHAT/DONE/NOW/MAP buckets (auto-merged, summarized, sometimes lossy), docs are stored and returned exactly as written. Use kind="plan" for an implementation plan, kind="instructions" for project rules, kind="context" for a full detailed brief — or any free-form kind. Wired through ctx_export/ctx_import, ctx_reset, and the CLI (ctx doc ).
ctx_get(view="detailed") — one call returns the buckets plus the full verbatim update history, notes, and docs. Use it when you need the complete picture (e.g. loading an implementation plan), not just the compact summary. Bounded by CTX_MAX_DETAIL_CHARS (default 20000).
Project instructions doc surfaced automatically — if set via ctx_doc(project, "instructions", ...), it's returned at the top of every ctx_get so any connected assistant sees project rules first.
how_to_ctx() — a built-in usage guide the model can call anytime it's unsure how to use these tools. New users: send Call how_to_ctx() and follow that guide for this and every session. as your first prompt and the assistant teaches itself the workflow.
MCP server-level instructions — clients that support the MCP instructions capability now get a short "how to use one-context" pointer automatically at connect time, no manual prompt required.
Notes
No breaking changes to existing tools or the DB schema (new project_docs table only, added via the normal migration path).
99 tests green (Ubuntu + Windows CI).
No new dependencies, no new network surface, no API keys required — the local, zero-config default path is unchanged.
Upgrade:

pip install -U one-ctx
Then restart your MCP client.

version 0.4.0

One-context now protects against cross-project mixing, survives concurrent writes from multiple AI tools, ships a real test suite + CI, and can back up and restore your context. Fully backward-compatible — existing ~/.ctx/ctx.db databases upgrade automatically.

Context isolation

• ctx_resolve(repo_path) — reverse-lookup from a workspace folder (or any subfolder, case-insensitive) to its linked project. The folder is the source of truth, not a remembered name.
• Folder-ownership guards — ctx_update / ctx_note / ctx_map now refuse to auto-create a different project for a folder already linked elsewhere, preventing typo-split projects (e.g. pocketlab vs pocket-lab) from forking your context.
• ctx_list now returns each project's linked repo_path.

Reliability

• Atomic concurrent writes — the full read-merge-write runs inside a single BEGIN IMMEDIATE transaction with a busy-timeout, so two tools (Claude, Cline, Codex…) updating the same project can no longer silently clobber each other. Verified against 40 concurrent writers with zero lost updates.
• 83-test pytest suite + GitHub Actions CI on Ubuntu & Windows × Python 3.10 / 3.12.

Security

• ctx serve now binds 127.0.0.1 by default (was 0.0.0.0). HTTP mode is no longer exposed to your whole network unless you opt in.
• Optional bearer-token auth via CTX_AUTH_TOKEN for network exposure. With auth on, /health hides the tool list and CORS is restricted to localhost origins. stdio mode (the default) has no network surface.

Durability

• ctx_export / ctx_import MCP tools and ctx export / ctx import CLI commands — JSON backups (idempotent merge or replace) plus human-readable Markdown.
• Bucket size caps (CTX_MAX_BUCKET_CHARS, CTX_MAX_MAP_CHARS) trim oldest entries on write; ctx_get view="brief" returns just the recent slice of DONE for context-tight sessions.

Install

pip install -U one-ctx
# or, in any MCP client:
uvx --from one-ctx one-context stdio
   
Full changelog: https://github.com/m4vic/one-context-mcp/compare/v0.3.0...v0.4.0


---

That `compare` link at the bottom auto-renders a full commit diff on GitHub — nice touch for a release.

Once you've published it, tell me and I'll finalize the 0.5.0 privacy-hardening plan (private-project flag, scoped search, last-writer visibility) and we start building. The plan's already drafted and ready — I just need to add the "0.4.0 release is done" checkpoint and re-confirm scope with you before writing code.

version 0.3.0

Adds ctx_bug bug tracking, fixes a concurrent-write race in ctx_update, adds Gemini as an optional cloud merge provider, and fixes an invalid Anthropic model id default."

one-context-mcp v0.2.1

one_ctx-0.2.1.tar.gz