Docs / Updates

Changelog

What changed in LLMGRAPH, newest first. Each entry links to the page it changed.

0.4.5 (2026-10-27)

Changed

• The stop check says when part of its list is not confirmed yet. If some broken callers are not yet labelled certain when your agent stops, the message now says "and N more are not confirmed yet" and points to check_my_changes for the full list, so your agent does not take a partial list for the whole one.
• **telemetry show** now prints the switch-off command as npx -y llmgraph-mcp telemetry off, the form that works with npx.

0.4.4 (2026-10-26)

Your agent can no longer finish a turn while a caller it broke is still broken, and the check after an edit can list the tests that call your change.

New

• A stop check in the Claude Code plugin, version 0.3.0. When your agent ends a turn after editing code, the plugin asks LLMGRAPH for callers that your edits break. It counts only lines marked certain, never weaker ones. If any are found, the agent is shown them once and keeps working: "Fix them, or tell the user why each is fine." It stays silent when the engine is not running, has not caught up, or is an older version. Switch it off with hook off, the setting stop = "off" under [hook], the variable LLMGRAPH_STOP_HOOK=off, or by removing the plugin. See The LLMGRAPH plugin.
• Tests that call the changed code. check_my_changes can list the test files that call the code you edited, each labelled like any other caller. It is off by default: your agent asks for it with tests: true, or you turn it on with tests = "on" under [check]. See check_my_changes.

Changed

• The plugin's skill now says when a plain text search is the better tool (unique names, renames that must also change comments and strings) and when LLMGRAPH is (names that also appear in unrelated code, indirect effects, callers missed after an edit).

0.4.3 (2026-10-24)

Usage counts and crash reports can now be switched off, and calls through an alias are never silently missing.

New

• Usage counts and crash reports can be switched off. They stay on by default; npx -y llmgraph-mcp telemetry off stops them and telemetry on starts them again. The licence check stays on: the product needs it. A report already being sent when you switch off may still arrive. See what leaves your machine.

Changed

• Calls through an alias are flagged. When a function or method is stored under another name (for example const objectType = ZodObject.create, exported as object), who_calls and check_my_changes say so: "passed as a value in 1 place (objectType, exported as object): calls through it are not listed", and calls that may go through the alias are shown as maybe_more, never stronger. See who_calls.
• A bare name also names the methods called that. Asking about render now says "matches 9 definitions and 20 methods with that name" and lists the methods, instead of leaving them out. Ask Class.method for one of them.

0.4.2 (2026-10-22)

LLMGRAPH now keeps one engine per computer instead of one per agent session, and keeps it inside a processor and memory budget.

New

• One shared engine. One llmgraph-engine.exe serves every agent session on the computer. An agent session starts it when it needs one, never at boot. It leaves after 20 idle minutes (setting engine.idle_minutes; 0 gives each session its own engine, as in 0.4.1). ps shows it; stop ends it, and stop --repo <folder> closes one project. See How LLMGRAPH is put together.
• A governor. The engine uses no more than half the processor cores, held by Windows for the engine and every helper it starts, and no more memory than the smaller of 2048 MB and 35 MB for every 100 MB of the computer's memory. Questions come first; background work waits for them. See The governor.
• A search hint in the Claude Code plugin, version 0.2.0. When your agent searches for a function or class name, the plugin offers LLMGRAPH's answer as extra context first. It never blocks or changes the search, gives counts only when the files are confirmed unchanged, and says nothing for file-name searches. Switch it off with hook off, the setting grep = "off" under [hook], or by removing the plugin. See The LLMGRAPH plugin.

Changed

• Callers are confirmed faster after you ask. After you ask about a function, LLMGRAPH confirms its callers with the type checker in the background; asking again a moment later shows the stronger labels as the checker confirms them. See who_calls.
• The check after a save is faster. Once LLMGRAPH has caught up with a save (a second or two on the projects we tested), check_my_changes answers in well under a second. See check_my_changes.
• Crash reports say how the engine ended. A report now names one of six classes: access_violation, stack_overflow, out_of_memory, killed, other_code or unknown.
• An index saved by a newer version is refused. An older LLMGRAPH answers N17 and never overwrites it. See Messages you may see.
• **Files in languages LLMGRAPH does not read are reported as cannot_see,** not counted as nothing. See Reading an answer.

Fixed

• After the engine was ended without closing its index (for example killed), the next start could answer "still reading this repository" to every question. It now ends in a real answer or a precise message.
• TypeScript projects with a missing project reference. A tsconfig.json that points to a project file that is not there used to stop the TypeScript checker. LLMGRAPH now uses the project files that are there.

Known limits

• A TypeScript project file that has files of its own and also points to a missing project file is left out whole.