NeuralMind v4.8.2 — the low-severity bug-hunt fixes

v4.8.1 fixed the high- and medium-severity bugs found by testing v4.7.0 end to end. This release fixes the low-severity ones, plus a few closely related problems found while fixing them. None of them lose data. They are wrong answers, misleading messages, and a test suite that dirtied the checkout.

The ones you’re most likely to notice:

• A mistyped project path is an error. It used to create a new, empty project at that path and answer from it.
• The daemon and neuralmind serve answer malformed requests with a 4xx. They used to drop the connection, return a 500, or raise inside the handler.
• Ingesting a Markdown file that build already indexed stores it once. The same text used to come back twice in query context.
• feedback and doctor say what actually turns query memory on.

What changed

Project paths and decisions

• A project path that doesn’t exist is an error. query, search, wakeup, stats, the decisions subcommands that open a store (record, query, amend, audit, export, restore, invalidate) and every MCP tool used to create <path>/.neuralmind/ for a mistyped path and run against an empty index or decision store. The CLI now exits 2 with project path does not exist: <path>. MCP tools return code: "project_not_found", with a hint to pass an absolute path when the path was relative, and the daemon answers 404. Nothing is created. build in an existing directory works as before.
• decisions restore and decisions invalidate report failures. An unknown id made restore print a traceback and invalidate print success. A database error was logged and then reported as success. Both now print Decision not found: <id> or Could not … decision <id>: … and exit 1. The MCP invalidate tool returns not_found or storage_error.
• decisions record --confidence must be between 0 and 1. 7 was accepted and stored as 1.0, and so was nan. The CLI and the MCP record tool now reject the value.
• search --n must be at least 1. --n 0 and --n -3 returned one result. The MCP tools now enforce the numeric bounds their schemas declare.
• stats . names the project. It printed Project: with nothing after it, in the CLI and in the MCP neuralmind_stats tool.

Memory and learning

• feedback and doctor say how to turn query memory on. With memory off, feedback good|bad said “run a query first”, and the next query wasn’t recorded either. doctor advised NEURALMIND_MEMORY=1, which changes nothing: the variable is on by default, and logging also needs a yes in ~/.neuralmind/memory_consent.json. Both now name what’s off and what turns it on.
• validate stops warning about stale synapses on a fresh index. Every edge it flagged pointed at a community_<id> node or a compliance-control key, which NeuralMind writes on purpose. Synapses to deleted nodes, or to communities a rebuild dropped, are still reported.
• The synapse memory export reports what each association earned. .neuralmind/SYNAPSE_MEMORY.md added a pair’s weight and activation count across namespaces, so a weight could read 2.00, and it tagged edges “(long-term)” by count alone. A pair now shows its strongest namespace, and “long-term” means protected from decay: at least five activations, weight at or above 0.20, and not ephemeral. Expect fewer long-term tags.
• The SynapseDynamics API does what its docstrings say. Resource limiting can trigger (ten potentiations per node until replenished), retrieval-induced forgetting weakens competitor edges that exist, and lateral inhibition keeps the strongest results instead of sometimes returning none. Nothing in the CLI, hooks or MCP tools calls this API.

Servers, demo and ingest

• The daemon and neuralmind serve answer malformed requests with a 4xx. A non-numeric Content-Length dropped the connection with no response. A non-object JSON body or a non-integer n was a 500. A non-ASCII token raised inside the handler. They now get a 400 with a JSON error naming the problem: 413 for a body over 1 MiB, 401 for the token. The daemon reads flags such as force strictly; "false" used to count as true. Both servers close a request that stalls for 30 seconds, and the graph view’s session-cookie check is constant-time.
• neuralmind demo stops reporting stale files. It copied the bundled project with the bundle’s timestamps, so when the bundled graph was older than its sources, every run printed “N files changed since the graph was built”.
• Ingesting a Markdown file that’s already in the project stores it once. The built-in graph already holds a Markdown file’s text, heading by heading. neuralmind ingest docs/guide.md and the neuralmind_ingest_document MCP tool stored a second copy under the absolute path, so the text came back twice in query context. Both now skip such a file and list it under already_indexed. A file edited since the last build is skipped too, but listed under needs_build: the graph still holds the old text, and neuralmind build indexes the edit. Any other file inside the project is stored under its project-relative path.

For contributors

• pytest leaves the checkout clean. The graph tests built the committed fixture projects in place, and each build rewrote their extraction caches. They now build throwaway copies, and the cache files are no longer tracked.

What the agent actually sees post-install

Environment variables

None added or changed.

Upgrade notes

• A project path that doesn’t exist now fails. A script that relied on a command creating <path>/.neuralmind/ should create the directory, or run neuralmind build, first.
• Daemon clients: flags must be booleans (or "true", "false", 1, 0), request bodies are capped at 1 MiB, and a missing project is a 404.
• SYNAPSE_MEMORY.md shows fewer long-term tags. It now counts only associations that are actually protected from decay.

Related

• CLI reference: search · decisions · exit codes
• v4.8.1 release notes · v4.8.0 release notes