Repository navigation
Added
-
Nine more Local Read kinds. The local read surface now serves
overlays.list,
overlays.removed,checkpoints.list,session.context,session.handoff,
session.value,update.status,security.statusandtelemetry.statusas JSON documents
(GET /v1/local/...and the matching--jsonCLI commands, e.g.lean-ctx overlays list --json,lean-ctx session context --json --session <id>). A local client reads them
instead of parsingctx_control,ctx_checkpointorupdate --statustext. Every body is a
function of the stored data, with no generated timestamps or counters, and is released
through the tool run of its tool. Project kinds need a project root; overlay removal only
accepts files inside the project. The status documents carry versions, switches and counts
only: no paths, digests, secrets or installation id. -
Shared parse cache across worktrees (#2247), opt-in. With
[cache] derivations_enabled = trueorLEAN_CTX_DERIVATION_CACHE=1, index builds reuse BM25 chunks, the deep analysis
behind graph and call-graph edges, and graph-scan signatures by exact content hash plus
language, path selector and build identity. A fresh worktree then parses only the files it
changed. Only pure parse results are shared: no project index, root, absolute path, retrieval
result or permission decision. The store (derivations_dir/
LEAN_CTX_DERIVATION_CACHE_DIR, default<cache dir>/derivations) must be a private
directory of the current user. It refuses links and FIFOs, writes atomically, treats
corruption as a miss, and evicts the oldest entries beyondderivations_max_bytes
(default 512 MiB). Output after a parser panic is never stored. The setting is read once per
process, so a running daemon picks up a change after a restart. Requested by @dkiaulakis.
Fixed
-
A Windows Cockpit can no longer exhaust system memory (#2252). On Windows, the Engine now
startsleanctx-cockpit.exein a job that caps its committed memory at 4 GiB and ends it
when the Engine goes away. A Cockpit that reaches the cap fails its allocation and is
restarted after the usual backoff, instead of committing tens of GiB until Windows runs out
of virtual memory.LEAN_CTX_COCKPIT_MEMORY_MBsets another cap (at least 256;0turns it
off); if the job cannot be set up, the Cockpit runs without a cap and the log says so. When
a Cockpit exits, the log names its exit status, on Windows its peak memory against the cap,
and the last whole lines of its stderr (at most 4 KiB), with the session tokens and
credential-shaped text redacted. Before, stderr was discarded and a restart was logged only
as "cockpit exited". Reported by @vik-borisov. -
ctx_readtask mode no longer runs for hours on large files. The MMR line ranking rebuilt
both token sets for every candidate and selected pair, so one read on a large file could run
for hours and a long-lived MCP server kept 150–550 % CPU. Token sets are now built once per
line and each round costs linear time. Selection order, ties and scores are unchanged; a test
checks the result against the previous implementation.
Upgrade
lean-ctx update # recommended (auto-downloads + refreshes shell hooks)
cargo install lean-ctx # or
npm update -g lean-ctx-bin # or
brew upgrade lean-ctxNote: After upgrading via cargo/npm/brew, run
lean-ctx setupto refresh shell aliases.lean-ctx updatedoes this automatically.
Full Changelog: v3.12.5...v3.12.6