Untitled
Lessons Learned - Consult selectively
Load only relevant topics; new entries: concrete experience -> actionable rule, no secrets/personal data. Do not duplicate binding rules here: Root, CoderTeam, GitHub Behavior, Token Optimization.
1. Collaboration and tools
| Experience (date) | Derived rule / reference |
|---|---|
| Repository/workspace copies from PR #4 drifted in license/Public notes (2026-10-25 to 2026-11-01). Extra synchronization used context. | Only root entry + central supplements; load explicitly, no product-repository copies. Inspect before changes, verify independently, write coherently. Storage: Root; cleanup: GitHub C. |
| Three private worktrees were clearer than one repo per branch; Public intentionally has independent history (2026-10-25). Fixed suffixes do not prevent target confusion (2026-11-01). | <ProjectName>-Intern / -Public; verify actual remote/branch/SHA/visibility per GitHub A1/B2. |
BatteryBuddy was the first use case, not a technology template (2026-10-25). Runtime choice there: .NET Framework 4.8/WPF, Core netstandard2.0, ADR 0001. | Check target systems, preinstalled runtime, deployment/support against current requirements; do not transfer framework/license/path/approval assumptions. Project start: CoderTeam 6. |
| Private branch protection/rulesets required a plan upgrade (2026-10-25). User is a second pair of eyes for manual testing. | Do not make a repository public for that reason; use PR/CI/process protection. test -> main requires approval for the specific state, preferably merged by the user. GitHub A8, CoderTeam 4. |
| dot.net download failed due to a manual firewall approval, then succeeded after approval (2026-10-25). | If actually blocked, pause and ask; do not work around it. GitHub B2; separate Ollama startup approval in Token Optimization 3. |
| Three translation subagents at the time had no write tools (2026-10-25). | This was a limit of that environment, not all agents: check capabilities/permissions before delegation; do small tasks directly and verify delegated file results/tests. |
PR #30 was merged before CI (2026-10-25). Build output with Select -Last 5 hid errors; tests/previews used old binaries (2026-10-27). | Actually wait for current CI; do not treat "no checks" as green. Capture build exit code and keep errors visible. GitHub A6/B5. |
$input as a PowerShell parameter caused empty model tasks (2026-10-28). | Do not use automatic variables as parameters; inspect stored request for complete payload before large runs. An invalid harness run is not model failure. |
| Fast/correct prose from local models still led to incorrect test/merge decisions (2026-10-28). | Verify explanations and tool actions separately; deterministic tests, coverage/acceptance/GitHub remain central. Model policy: Token Optimization 1; evidence below. |
Selective Git approval worked with git AND rtk git; user confirmed no prompts when creating/deleting local branches (2026-11-04). | Allow wrappers selectively, not broadly; a successful tool call alone does not prove that no user prompt appeared. Configuration/limits: GitHub B7. |
| Public had its own build parameters without symbols/private revisions; later user README changes were fast-forwarded separately (2026-11-04). | Test exact Public binaries, tagged rebuild/package, and downloaded assets; do not overwrite user documentation with an internal snapshot. Procedure: GitHub A7. |
2. Windows, WPF, and hardware (BatteryBuddy experience)
Not a general architecture requirement; consider for relevant integrations.
| Topic / experience (2026) | Action |
|---|---|
WPF startup (2026-10-25): AppMain.xaml requires ApplicationDefinition + Page Remove; slider events fire in InitializeComponent. | Guard initialization with _loading; check build items and, if project configuration changed, clean obj deliberately if needed. |
UI evidence (2026-10-25/28): old EXE showed the wrong language; single-instance logic can redirect a new start to an old process. PrintWindow distorts RTL. | Rebuild before tests, check running instance path/start time; ask user to close an old instance, do not stop unrelated processes. Check RTL in the real window (confirmed by user at the time). |
| Footer/hover (2026-10-28): setting a tooltip does not prove it is reachable. | Check a real WPF window using InputHitTest and visible popup content; empty status fields must not intercept hover. |
Icons (2026-10-27/28): System.Drawing.Icon fails on PNG-compressed 256 px ICO entries, while WPF IconBitmapDecoder can read them; decoders may choose the 16 px entry. | Check actual PixelWidth/PixelHeight, not merely a non-empty image source; make large previews from original PNG. Smoke-test OS integrations too; optional icons must not disable the main feature, and errors must remain visible. |
User settings (2026-10-25): Environment.GetFolderPath(ApplicationData) ignores an overridden $env:APPDATA. | Inject or deliberately preserve/restore the path; tests must not overwrite real settings. |
xUnit (2026-10-25): .GetAwaiter().GetResult() failed on xUnit1031 when warnings were errors. | New task tests should use async Task/await, not block. |
| HID++/radio (2026-10-25): sleeping devices lose the first packet (wake-up); they respond sporadically while idle. | Retry first request, cache last state ("stale," not "missing"); measure idle AND active states. |
| Hardware test with user (2026-10-25): multitasking made a single measurement ambiguous, and script output was invisible to the user. | Give a start signal, define phases (e.g. 30 s idle / activity / idle), provide brief live chat updates, and announce phase changes. |
Toasts without installer (2026-10-27): Framework 4.8 with Microsoft.Windows.SDK.Contracts (metadata, no extra DLL) and AUMID under HKCU\Software\Classes\AppUserModelId\<id> (DisplayName/IconUri) worked without shortcut/COM activator/admin. Clicks from the notification center arrive at a running app. | Consider for a suitable unpackaged app; use random strictly validated short codes instead of IDs as arguments, XML with XElement, and keep balloon fallback. |
Tray history (2026-10-28): NIM_DELETE removes the icon, not persisted visibility; random test GUIDs leave residue. | Simulate the shell in automation, keep real NativeWindows/callbacks testable. Real shell tests only in a deliberate test account/VM with approval; precisely identify/preserve/approve cleanup, never clear the entire Explorer cache. |
| Multiple tray icons (2026-10-27): custom NativeWindows/stable hashed GUIDs create independent entries; shared menu survives all icons. | Keep domain keys stable; replace image only after native replacement, release menu only after all icons. Test registration/switching/Explorer re-registration/removal with a simulated shell; real checks as above, plus 16/24/32 px and RTL. |
| Portable GUIDs/accessibility (2026-10-27): Windows binds unsigned GUIDs to binary path; accessibility toggle changes checked state without Click. | GUID from normalized EXE path + device/category key, stable across restart/battery state/tooltip/update at same location; another location has separate preferences. Use Checked/Unchecked for exclusive checkboxes and test the actual ToggleProvider, not only mouse clicks. |
| EasySMX radio receiver (2026-11-04): XInput reported WIRED without a valid battery; Bluetooth also provided no battery level. | Device Manager visibility does not prove a battery API exists. Distinguish reported interface from physical connection; "no reading" is not "no battery." Treat diagnostics as a reasoned hypothesis, not a blanket claim that all dongles are wired. |
| Keyboard navigation (2026-11-04): user suspected missing tab stops; arrow keys correctly reached other tabs/devices. | Explain and test Tab between composite areas, arrows within tab headers/lists, forward and backward; do not add dynamic single-tab stops solely based on that assumption. UIA/contrast tests are not screen-reader certification. |
WPF test host (2026-11-04): a new Application per test failed due to AppDomain limits; AppDomain isolation was disposing. | Use a shared STA Application fixture with Application.Run, dispatcher calls, visible startup errors, and orderly shutdown/thread join; do not register real tray/autostart integrations for isolated UI tests. |
Footer themes/status (2026-11-04): local implicit button style with the same BasedOn key fell back to native style; Transparent also differed. | Name the shared style explicitly; check all affected rendered buttons through light/dark/light. Keep transient success separate from persistent footer; stop timer when replaced/edited/closed and preserve focus when hiding. Do not automatically hide errors. |
Tray exit during unsaved prompt (2026-11-04): ShowDialog pumped a nested dispatcher in Closing; another Close threw an exception. | Explicitly guard close attempts, reuse existing decision, clear pending app exit on cancel/veto, and continue shutdown only after Closed handlers. Regression-test repeated exit calls in a real modal dispatcher loop; do not suppress exceptions or force discard. |
3. Completed model experiments (2026-10-28)
Historical evidence, not operating instructions or a request for new benchmarks. Active selection/routes/guards are exclusively in Token Optimization. The four private lab repositories/local fixtures at the time were disposable; do not assume they exist now or derive dependencies/cleanup from them. Research documents are in Git/PR history; raw data is outside the product repository.
| Measurement | Result and limitation |
|---|---|
| Coding, 41 cases each | E4B 37/41 -> 32/41 after repair; Qwen 39/41 -> 40/41. Review/repair needs do not justify a default outsourcing path. |
| App acceptance: 24 workflows / 48 calls, two models x normal/strict | Complete workflows: E4B 3/6 and 2/6; Qwen 1/6 and 4/6. No false green, but 9/12 clean cases rejected; stricter prompts were not reliably better. |
| GitHub: 96 tasks, 24 per variant | Exactly executable contracts: E4B 3/24 and 7/24, Qwen 5/24 and 7/24. Correct prose with incorrect action; despite "CI is running, do not merge," Qwen included a merge action. |
| Total app/GitHub run | 144 valid calls, 118,608 local tokens; invalid harness run excluded. E4B about 8-10x faster, still not a reliable general agent. |
| Translation, core set | E4B 9.93x faster; dedicated Windows runner memory about 4.95 GiB vs. Qwen 9.91 GiB with CPU offload. Conservative route 2.60x in a warm theoretical calculation, excluding load/switch/review; no general quality/future guarantee. |
| UI test scope | 361 green baseline tests, real SettingsWindows de/en/ar; footer layout measured at 38 languages x 3 widths. Not 114 visually approved windows, not autonomous screenshot analysis or 4K/hardware/Explorer certification. |
| Historical central token estimate | Product translations avoided net 10,000-30,000, language research consumed 150,000-300,000 and coding research 60,000-140,000: about 180,000-430,000 additional effort; subsequent pilot/app/GitHub attempts at the time added about 80,000-180,000. |
The last estimate has low confidence, without complete usage/cache/reasoning/billing data. Distinguish local API tokens, central estimates, and actual billing; include setup/reading/review/repairs/failed attempts in the total. Do not describe research as amortized or as net savings. Hundreds of deterministic test cases do not require hundreds of LLM calls; additional model loops showed no proven positive net relief.
Reconstruction: Token Optimization 5, BatteryBuddy commit 3a220e7 and PRs #157/#159. Historical measurements are not current resource/app state.