Glow-Fix (3 Ursachen behoben):
- notifyGlow: schickt jetzt an targetTabId statt an den aktiven Tab des
Nutzers. Vorher: Agent arbeitet in Tab B, Glow geht an Tab A (den der
Nutzer gerade sieht) — Tab B leuchtet nie.
- executeForBroker: notifyGlow(true) beim ersten Tool-Call, nicht erst
bei 'round'. Vorher: targetTabId war bei 'round' noch null, currentTabId()
gab undefined zurück, Glow wurde nie gesendet.
- notifyTargetTab: Glow auf neuen Tab übertragen wenn Agent Tab wechselt.
focusWorkingTab (neue Config-Option):
- ExtensionConfig.focusWorkingTab: boolean (Default false)
- notifyTargetTab: bringt Tab in Vordergrund wenn focusWorkingTab=true
- Checkbox 'Tab live mitverfolgen' in Einstellungen > Anzeige & Sicherheit
- false = Agent arbeitet im Hintergrund (bisheriges Verhalten)
- true = Tab wird bei jedem Wechsel sichtbar
- MERCEDES_CONTEXT komplett neu: explizite Regel 'Social Intranet ist
immer der erste Anlaufpunkt' ganz oben, vor allem anderen.
- Vorgehen bei unbekannten Abkürzungen (AFM, CBS, ESS, ...): erst
Intranet-Suche, dann lesen, dann erklären — nicht googeln, nicht raten.
- Struktur: klare Abschnitte statt Fließtext-Wurst.
- handleBrokerMessage: 'round' schaltet notifyGlow(true), 'done'/'error'
schalten notifyGlow(false). Vorher wurde Glow nur im direct-Modus
(agent.ts) gesteuert — im Broker-Modus blieb er immer aus.
- ExtUserMessage: url?-Feld ergänzt.
- user_message an Broker: aktuelle Tab-URL wird mitgeschickt, damit
der Broker kontextspezifische Prompts einbauen kann.
- browser_computer left_click: afterState (url, urlChanged, readyState,
activeTag, activeLabel, dialogText) direkt im Ergebnis — kein sleep,
kein Polling, kein Blockieren.
- browser_computer screenshot: page (url, title, readyState, activeTag)
im Ergebnis — ein einziger evaluateInMainWorld-Call.
- browser_click (semantisch): snapshotAfterClick() nach Klick — gleiche
Felder, gleiche Logik, kein extra Warten.
- browser_navigate: gibt loading:true + hint zurück statt zu blockieren.
Die KI entscheidet selbst ob sie wartet.
- Keine adaptive Pace, kein waitForTabLoad, kein Polling — das war der
Grund warum alles eingefroren ist.
The settings were 18 rows on one level: the broker URL, the pace in
milliseconds and the model all shouted equally loud, although one is set
once and the other is changed daily.
- Model picker moved into the header. It is the switch you touch every day
and it sat at the end of the settings list. Source depends on driveMode:
the verified gateway models in direct mode, the broker's list otherwise;
hidden when there is nothing to choose.
- Settings grouped into five collapsible sections (Verbindung & Antrieb,
Verhalten, Anzeige & Sicherheit, Modell & Zugang, Memory). Only "Modell &
Zugang" starts open — that is what first-time setup needs.
- Existing rows were wrapped in place, not moved, so every element id and
every binding stays intact.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Nexus gateway (measured 2026-08-07, not assumed):
- Gemini speaks Google-GenAI (/v1beta/models/{id}:streamGenerateContent, header
api-key), NOT the Azure-OpenAI path — that returned 404 "no Route matched" and
was the cause of the reported failures. New agent/gemini_bridge.py translates
Bedrock Converse <-> Gemini in both directions.
- Only gemini-3.6-flash is subscribed; 2.5-flash/2.5-pro/3.1-flash-lite give 403,
every other name 404. Catalog corrected.
- Four Gemini rules, each previously an HTTP 400, now covered by tests:
thought signatures are mandatory, they belong to the TURN (not the individual
call), functionResponse turns must be homogeneous, arrays need `items`.
- Prompt caching is NOT available: cachePoint is accepted and ignored.
System prompt:
- Was an f-string; a code sample containing braces broke build_system_prompt at
request time (CLI and web both 500, import stayed green). Now a plain template
with __TOKEN__ placeholders. Regression guards in tests/test_system_prompt.py.
CLI:
- `agent resume` now prints the stored transcript. The history was always loaded
into the model context, only the terminal stayed empty.
Memory (new, all three surfaces):
- agent/memory.py stores notes about the user in one local file, written
atomically; memory.cnull.net remains an optional mirror that can never fail a
write. Tools memory_save/search/forget, injected into the prompt with a budget.
HTTP surface /api/memory for the extension.
Browser extension (agent/extension, first commit of the source):
- driveMode 'direct' talks to Nexus without the Python broker: Claude via
Bedrock converse, GPT via Azure-OpenAI, Gemini via Google-GenAI.
- browser_type no longer guesses the focused element — that wrote whole mails
into Outlook's subject line. Read-back now reports where the text actually
landed, so a mis-target is visible instead of silent.
- aria-labelledby is resolved across all ids (it is a list); contenteditable is
interactive and marked editable. Without this, subject and message body look
identical to the model.
- Hard block against sending mail, independent of riskMode.
- Runs survive the panel: events are buffered and replayed by sequence number.
- Image input (paste, file, drag&drop), on-page glow/spotlight, memory tools.
Cost: fixed tokens per round 11434 -> 6540 (-43%) by trimming tool schemas,
dropping gateway docs from the browser prompt and sending site knowledge only
where it applies.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
core
