CLI, MCP And TUI Issues

Known problems in the eulix command line, the terminal screens, the MCP server and the model client.

Scope: eulix-cli (internal/cli, internal/tui, internal/mcp, internal/llm, internal/cache, internal/utils) Last verified against: ad744e9 (2026-11-03)

These entries were found by reading the current code; I did not run the screens or the server. Where an entry says probably, I am inferring the behaviour from the code. Read Overview for the legend.


CLI-1: Each chat answer is probably saved to history twice

Severity: Medium · Status: Open · Verification: Read in code

Symptom. In eulix history, each chat question likely appears two times.

Cause. eulix chat gives the same history manager to the query router and to the chat screen. The router’s QueryEngine saves the reply when it has a manager, and the chat screen saves it again when the reply arrives.

Workaround. Delete duplicates by ID with eulix history delete <id>.

Fix direction. Let only one component save. The screen is the better choice, because it already holds the split reasoning and answer.


CLI-2: The history viewer deletes the wrong entry, does not refresh and has no way back

Severity: Medium · Status: Open · Verification: Read in code

Three problems in the history screen (internal/tui/history-tui.go):

  1. d in the list can delete the wrong entry. The deletion uses m.selected, which is set only when you press Enter to open an entry. In the list it is still 0 (the newest entry) or the last entry you opened, not the highlighted one.
  2. After a deletion the list does not refresh. The delete command works on a copy of the model and returns nothing. The database entry is removed, but the screen still shows it.
  3. There is no way back to the chat. /history replaces the chat model entirely, and q in the list quits the whole program.

Workaround. Open an entry with Enter and delete it from the detail view, where m.selected is correct. Quit and reopen to see the result. Use eulix history delete <id> for precision.

Smaller issues in the same screen: answers are wrapped as plain text rather than rendered as Markdown, and long titles are cut by bytes, which can split a multi-byte character. See History Viewer. The tui package has no tests at all.


CLI-3: [checksum] settings are ignored

Severity: Low · Status: Open · Verification: Read in code

checksum.change_threshold and checksum.force_reanalyze_threshold are read, validated and clamped, but no command uses them. eulix chat uses fixed limits of 10% (ask first) and 30% (refuse).


CLI-4: eulix query prints a prompt and never calls the model

Severity: Low · Status: By design · Verification: Read in code

For factual questions (locations, usage, call graphs, entry points, dependencies, metrics, todos) eulix query prints the answer. For everything else it prints the retrieved context and reasoning instructions, ready to paste into a model. It never contacts a model, though an older README said it did. Use eulix chat for model answers.


CLI-5: Model requests cannot be cancelled, are not streamed, and have three smaller defects

Severity: Medium · Status: Open · Verification: Read in code

All in internal/llm/client.go:

  • No cancellation. Requests are built with http.NewRequest, without a context, and the only limit is a five-minute HTTP timeout. A slow local model generating max_tokens = 8192 can hit that limit, and the chat screen cannot abort a request.
  • No streaming. The client waits for the whole reply, so the chat shows nothing until the model has finished.
  • Only the first Anthropic content block is returned (resp.Content[0].Text). That is fine for plain text. If extended thinking or tool use is ever enabled, the first block may not be text and the answer would come back empty.
  • A debug-log failure blocks the answer. With DebugConfig on, if .eulix/debug/llmdebug.log cannot be written (read-only checkout, full disk), LlmResponse returns that error instead of asking the model.
  • Typos in key names are silent. A misspelled key in eulix.toml (for example modle under [llm]) is ignored without a warning. Reporting unknown keys would need the TOML decoder’s metadata and a policy for old, removed keys.

CLI-6: History entries with equal timestamps list in arbitrary order

Severity: Low · Status: Open · Verification: Read in code

ListAll sorts by CreatedAt alone with an unstable sort. On a coarse clock (Windows), two entries can share a timestamp and swap places between calls. Sorting by ID, which only grows, fixes it.


CLI-7: Developer-facing loose ends

Severity: Low · Status: Open · Verification: Read in code

  • DebugLogger data races. Log and Flush read the closed flag without the lock, and Lines and Write do not lock. Harmless today, but go test -race will flag any test that logs while the logger is closing.
  • Lint config is never loaded. The file is eulix-cli/golangci.yml; golangci-lint looks for .golangci.yml, and the file uses the version 1 format that golangci-lint 2.x rejects. make lint runs only the default linters.
  • A misleading comment. chat.go says the checksum “is already saved by checksum.Run()”. Run only compares; the parser writes the baseline during eulix analyze. The code is right; the comment is not.
  • Dead repair tools. internal/fixers (GLaDOS, Aspirine) is not imported by anything, the glados and aspirine commands are stubs, and the code assumes a binary layout that no longer matches the quantized files. See Fixers.
  • Leftovers in the chat screen. StateTyping is never set, and the welcome text and title bar are fixed strings (the title says “local-first” even with a cloud provider).

CLI-8: “Call graph of” questions can look up the wrong word

Severity: Medium · Status: Open · Verification: Reproduced (2026-11-10)

Symptom. eulix query "call graph of main" answers 'of' not found in call graph. The same happens for any name with no capital letter or underscore, and for C++ names such as Vault::unlock. Through MCP, “call graph of Vault::unlock” looks up Key instead. Names such as DecodeJSONFile and payment_total work.

Cause. When the classifier finds no symbol, extractEntityName in internal/query/router_helpers.go prefers a word that looks like code, then falls back to the first word not on its stop list. “of” is not on that list. isLikelySymbol rejects names containing :, so under MCP the capitalised Key from the appended Key symbols: line wins (see MCP-3).

Workaround. Ask “who calls main” or “who calls Vault::unlock”. Both find the symbol.

Fix direction. Add “of” to the stop words, accept :: in isLikelySymbol, and leave the appended Key symbols: line out when picking the symbol.


MCP-1: A repository’s own .eulix/ and eulix.toml are trusted

Severity: High · Status: Open · Verification: Carried over

Symptom. The MCP server binds any directory that contains .eulix/kb.json. A cloned repository can therefore ship its own knowledge base and eulix.toml and be queried without you ever running eulix analyze. With an editor that sends workspace roots, the binding is automatic.

Impact. The known abuses are closed (see Fixed Issues: code execution through a model name, reading arbitrary files through the index, and leaking ignored files). The trust itself remains: a hostile repository still chooses the configuration, including which model is downloaded (EMB-2), the language-model provider address, and the contents of the knowledge base.

Workaround. Only open repositories you trust in an editor that uses Eulix; or pass --repo explicitly and decline automatic roots.

Fix direction. Bind only when the checksum file shows the knowledge base was built on this machine, or ask the user once per repository.


MCP-2: A panic in a handler takes the MCP server down

Severity: Low · Status: Open · Verification: Read in code

The SDK does not recover panics, and the package has no recover. Fuzzing the query parsers found no panic, but one deeper in retrieval would kill the editor’s server. A recover that turns the panic into a tool error would keep it alive.


MCP-3: MCP text-search fallback gaps

Severity: Low · Status: Open · Verification: Read in code

  • It does not remove identical content across files, so repositories with backup trees waste budget. Use .euignore or lexical_exclude.
  • retrievalConfig.etiya.lexical_exclude is read only when the engine is etiya.
  • The client “roots” feature the server uses to find a repository is deprecated in newer protocol versions, and clients on those versions never send the message that triggers it. For them, use --repo, set_repo_path or repo_path.
  • Multi-hop retrieval depends on an implicit contract: anchors are handed to the router by appending a Key symbols: line to the question. Changing the router’s prompt format could silently weaken it, and the line already sends one call graph question to the wrong symbol (CLI-8).

Severity: Low · Status: Limitation · Verification: Reproduced (2026-11-10, with the eulix binary)

Symptom. A symlinked source file is left out of text search, and its source does not appear in answers. This applies even when the link points to another file inside the repository.

Cause. This is a deliberate choice, and it applies to MCP and eulix query alike. Following links safely would mean checking every target against the hidden-file and .euignore rules, on both Unix and Windows. Refusing links needs none of that. The parser’s walker does not enable link following either, so these files are normally not indexed.

A .eulix folder that is a symlink, or that contains one, is refused outright, and eulix analyze, eulix chat, eulix query and the MCP server stop with an error naming the link.

Workaround. Keep a real copy of the file in the repository, and remove any link from .eulix. Opening a repository through a symlinked path still works, because binding resolves it to the real folder first.

Not tested. Windows junctions have not been tested. The symlink tests skip when the system does not allow creating symlinks, so check for skips in a Windows run.


Severity: Low · Status: Partly fixed · Verification: Read in code

The serious routes are closed: see Fixed Issues. The .eulix check runs when a router or context builder is built, so these narrower cases remain:

  • A link added after that check. Metrics, entry points and dependencies are read again for each question. Until the next build, a link added by a pull can pull in a JSON file from outside the repository. Debug logs are not affected, because each write refuses such links itself.
  • A symlinked .eulix in eulix chat. The history database opens before the check, so an empty history.db can be created in the link’s target before the command stops.
  • A replaced folder outside MCP. Only the MCP server checks that the bound folder is still the same one. A long-running eulix chat does not.
  • Files outside .eulix. eulix.toml and .euignore are read through links. Their contents act as settings and rules and are not shown in answers. This belongs with MCP-1.

Not tested. None of this has run on native Windows, including whether junctions are refused.


CLI-9: On Windows, a bound repository folder can’t be moved or renamed

Severity: Medium · Status: Open · Verification: Reproduced (Windows only)

Symptom. While eulix mcp (or eulix chat) has a project loaded, renaming or moving the project folder or its .eulix fails with “Access is denied”.

Cause. The query router always opens .eulix/debug/context_debug.log and susbstemDebug.log and keeps them open, even with debugging off. Windows refuses to rename a folder that holds an open file.

Workaround. Wait for the MCP idle release, or stop the server first.

Fix direction. Open the debug files only when debugging is on, or open them for each write. Test: TestWinIssue_BoundRepoFolderCannotBeMoved (go test -tags winissues ./internal/mcp). Details: Windows-issues.md.


CLI-10: On Windows, a project opened through a directory junction doesn’t work

Severity: Medium · Status: Open · Verification: Reproduced (Windows only)

Symptom. Binding a repository through a junction (mklink /J) succeeds, but every MCP query is refused with “changed after it was bound”, and lexical search finds nothing. Running eulix analyze through a junction gives an empty checksum scan.

Cause. Since Go 1.23 a junction is not a symlink: filepath.EvalSymlinks leaves it unresolved, os.Stat (at bind) and os.Lstat (at query) disagree, and filepath.WalkDir does not enter a junction root.

Workaround. Open the project by its real path.

Fix direction. Resolve the root to its final path (for example GetFinalPathNameByHandle) before pinning and walking it. Tests: TestWinIssue_RepoBoundThroughJunctionIsUnusable, TestWinIssue_ScanThroughJunctionRootFindsNothing (-tags winissues).


CLI-11: “Where are X run” is answered with where X is defined

Severity: Medium · Status: Open · Verification: Reproduced with eulix query (2026-11-16, 07907ad, on Flask 3.1.3); MCP not checked

Symptom. eulix query "Where are before_request and after_request hooks run during a request?" prints only Function 'before_request' defined at: src/flask/sansio/scaffold.py:460. The question asks where the hooks run (preprocess_request and process_response, reached from full_dispatch_request in app.py), and after_request is dropped.

Cause. Not traced. The router appears to treat a question that starts with “Where” and contains a symbol as a location question, and answers it from the index without retrieval (the factual-answer path in CLI-4).

Workaround. Ask without “where”: “How does Flask run before_request and after_request hooks during a request?” retrieves app.py:904-951 (full_dispatch_request and finalize_request) along with the hook registration code.

Fix direction. Classify a “where” question as a location lookup only when it asks where something is defined or declared, not where it is run, called or used.


MCP-6: Project memory adds unrelated “Key symbols” to later queries

Severity: Medium · Status: Open · Verification: Reproduced (2026-11-16, 07907ad, on Flask 3.1.3)

Symptom. What a query returns depends on the queries asked before it in the same server session. On Flask, after one query returned the Flask class summary, the next query naming full_dispatch_request was sent as … Key symbols: full_dispatch_request, root_path, host_matching, static_folder, static_host, static_url_path, subdomain_matching, import_name, url_map. The eight extra names are constructor parameters from the class docstring and unrelated to the question, and most of the reply was unrelated tests.

Cause. projectMemory.Observe (internal/mcp/strategy_memory.go) splits a trusted result on blank lines and treats every symbol in a block as related to every other, keeping up to maxRelated (8) per symbol. A class summary is one block that lists all of the class’s methods and documents every parameter, so each of its methods becomes related to the first eight names in the docstring. expand then adds those names as anchors whenever one of the methods is asked about. The result that taught the memory had cleared the 0.55 quality bar even though it was a poor answer (RET-13).

Workaround. Switch the active repository away and back (set_repo_path) or restart the MCP server; both clear the memory.

Fix direction. Don’t learn co-occurrence from class summaries or other large blocks, or cap the symbols taken per block, and rank remembered neighbours below the query’s own anchors.