Local code intelligence for developers and AI agents. Connect code, docs and infrastructure in a persistent graph, with CBM Lens, MCP tools, agent integrations and change-aware test selection.
C aiderastclaude-codecode-analysiscode-intelligencecodexcursorcypherdeveloper-toolsgemini-cligraph-visualizationkilocodeknowledge-graphmcpmcp-servermodel-context-protocolopencodesqlitetree-sitterwindsurf
docs(ui): record the second hand test round K29 to K44 and its checks
Bernhard Jackiewicz8c48d58d54ed
diff --git a/docs/development/pr-2068-call-feedback-plan.md b/docs/development/pr-2068-call-feedback-plan.mdindex 8e1a9ef9b..9b63f1c68 100644--- a/docs/development/pr-2068-call-feedback-plan.md+++ b/docs/development/pr-2068-call-feedback-plan.md@@ -295,3 +295,19 @@ Browser checks on the final tree: graph-ui/tools/handtest-fixes-galaxy.mjs 29/29Known limits: the 0.5B model can still word things loosely next to the listed facts (its text is labelled, unknown names are flagged); Behavior resolves os.environ.setdefault in manage.py-tpl to a QueryDict method (index call resolution, not frontend); with the chat open the Galaxy toolbar wraps to two rows below about 1,440 px.++## Second hand test round (4 October 2026)++Bernhard tested the branch a second time and recorded 16 findings (K29 to K44) in the same local correction plan. A headless run with the real model covered what he could not run himself (configuration files in Explore, the Endpoints check box, Behavior on cbm) and confirmed what his screenshots showed. Two read only reviewers then went through every change and every chat text the browser had produced; their findings were fixed in a third round. Everything is test first and checked in a headless browser with screenshots.++Galaxy: scene labels stay below the panels, and the panels are opaque so nothing shows through; the node hover card stays above them. Path labels keep off the hierarchy band title and are left out where no free spot exists. Expand +1 explains at once on hover and focus which render limit a layer will likely pass and what the limit does, and while a layer loads it says that "−" cancels it.++Chat: a layer that stopped at the render limit is called partial, with the same wording as the Galaxy tooltip. Counts carry their units ("Incoming: 23 relationships from 12 symbols"), caller lists name the callers first and the other incoming relationships apart, and German answers are German throughout. The note about unknown names ignores case and literals but still flags mangled names, and a dropped model sentence says what it claimed. The model is told the language of the question. Short general questions about a selection, an open code file or marked code get the listed facts, the lines read from the code and at most a checked sentence; a prompt without a question gets example questions; configuration files (YAML, JSON, TOML, INI) are answered with an outline read from the file and a line about what the file is for. "Ask the model" now adds the model's reply below the listed answer instead of replacing it, every free model answer says that it comes from the local model, and returning to a file says "Back to" and sends that file's own earlier turns again.++Architecture: Back and Forward stand first in the subtab row with the same words as in Galaxy and stay in one place in every subtab; the Recent menus of both workspaces close on a press elsewhere, Escape, a choice and any change of place. The Refresh buttons of Architecture, ADR, the project picker and the file impact show that they run and then the time, "no changes" or the error. The Behavior start list has suggestions, then all operations in alphabetical order with look alikes named by class, a filter field and the current start apart from the matches. A project switch keeps the workspace in the address.++Not changed, because correct: "Next lines" is disabled at the end of a file; browser Back and Forward switch projects (the Galaxy Back works inside Galaxy); with the chat open "All graph" reads "All"; older wrong answers in the chat came from the history of the first round.++Checks on the final tree (234e0f23): handtest-fixes-galaxy.mjs 29/29, handtest-fixes-architecture.mjs 25/25, handtest-fixes-chat.mjs 19/19 with the real model, call-feedback-acceptance.mjs 19/19, handtest-fixes-k27.mjs 55/55, all headless. scripts/ci/test-ui.sh is green with 295 test files and 3,858 tests, the style gate, the promise scan, 203 of 203 acceptance checks and the build. The C suites pass with sanitizers: 8,005 passed, 0 failed, 7 skipped in 143 suites.++Known limits: the local 0.5B model stays weak, above all in German, so its sentence is often dropped and an answer shows the facts and the code lines only; its free answers are labelled but not better. General questions about an Architecture selection still go to the model, because those facts exist only in English. Behavior resolves os.environ.setdefault in the Django project template to a QueryDict method, which comes from call resolution in the index. With the chat open the Galaxy toolbar wraps to two rows below about 1,440 pixels.