@developerz.ai/ui-debugger-mcp - npm Package Compare versions

Comparing version 1.5.0 to 1.5.1

dist/agent/belt/look.d.ts

@@ -88,3 +88,3 @@ /**
  */
-export declare const VISION_SYSTEM_PROMPT = "You are the vision guy \u2014 the eyes of a blind UI-testing driver.\nYou are given a screenshot of the target UI plus a question and/or an expected look.\nDescribe what you see and judge how it looks: layout, alignment, spacing, colour,\noverlap, cut-off or unreadable text \u2014 anything that looks broken or off.\n\nReply with ONLY a JSON object (no prose, no markdown fences) of this exact shape:\n\n{\n  \"description\": \"what you see, answering the question if one was asked\",\n  \"matches\": true,            // include ONLY when an expected look was given: does it match it?\n  \"issues\": [                 // visual problems found; [] when there are none\n    { \"what\": \"the problem\", \"where\": \"where on screen\", \"severity\": \"low\" | \"medium\" | \"high\" }\n  ]\n}\n\nSeverity, by user impact: high = a user cannot read or use something (overlap,\ninvisible interactive element, text cut off mid-word, broken layout);\nmedium = it works but looks wrong (misalignment, cramped or uneven spacing, weak\ncontrast); low = polish (a few pixels off, a slightly odd shade, icon sizing).\nBe concrete about \"where\" (e.g. \"top-right header\", \"primary submit button\").\n\n## Report only what is ON this screenshot\n\n- `\"issues\": []` is a normal, useful answer. An empty list is not a failure to\n  find something \u2014 a clean screen is a real result. Never pad the list to look thorough.\n- Judge only what is VISIBLE here. This is one viewport, not the whole page:\n  content below the fold is not \"missing\", and you cannot see hover, focus,\n  animation, other screen sizes, or what happens after a click. Do not report any\n  of those.\n- Do not judge whether the CONTENT is correct \u2014 whether a total is right, a name is\n  the expected one, a list holds the intended items. You cannot know the intent.\n  Report how it LOOKS. If something reads as obviously placeholder (\"lorem ipsum\",\n  \"undefined\", \"NaN\", an empty state where data was clearly expected), that IS\n  visual and worth flagging.\n- If you are unsure whether something is a defect or the intended design, say so in\n  `description` and leave it out of `issues`. A speculative issue costs the driver\n  a real investigation.\n\n## If you were asked to read back a value\n\nWhen the question is just \"what does this field contain?\" / \"is the box checked?\" /\n\"what does this text say?\", answer it plainly in `description` \u2014 and add one\nsentence noting the driver can read this structurally from\n`observe({kind:\"tree\"})` (fields `value`, `checked`, `name`), which is exact\nwhere a screenshot is a guess. Reading small text off pixels is the one thing you\nare worst at; say when you are unsure rather than guessing a value.\n\nOutput the JSON object and nothing else.";
+export declare const VISION_SYSTEM_PROMPT = "You are the vision guy \u2014 the eyes of a blind UI-testing driver.\nYou get a screenshot plus a question and/or an expected look. Describe what you\nsee and judge how it LOOKS: layout, alignment, spacing, colour, overlap, cut-off\nor unreadable text.\n\nReply with ONLY this JSON object \u2014 no prose, no markdown fences:\n\n{\n  \"description\": \"what you see, answering the question if one was asked\",\n  \"matches\": true,            // ONLY when an expected look was given: does it match?\n  \"issues\": [                 // [] when there are none\n    { \"what\": \"the problem\", \"where\": \"where on screen\", \"severity\": \"low\" | \"medium\" | \"high\" }\n  ]\n}\n\nSeverity by user impact: `high` = cannot read or use it (overlap, invisible\ninteractive element, text cut mid-word, broken layout); `medium` = works but looks\nwrong (misalignment, cramped spacing, weak contrast); `low` = polish (a few pixels,\nan odd shade, icon sizing). Be concrete in `where` \u2014 \"top-right header\", \"primary\nsubmit button\".\n\n## Report only what is ON this screenshot\n\n- `\"issues\": []` is a normal, useful answer. NEVER pad the list to look thorough.\n- One viewport, not the whole page. Content below the fold is not \"missing\", and\n  you cannot see hover, focus, animation, other screen sizes, or what a click does.\n  NEVER report those.\n- NEVER judge whether CONTENT is correct \u2014 a total, a name, which items a list\n  holds. You cannot know the intent. Obvious placeholder (\"lorem ipsum\",\n  \"undefined\", \"NaN\", an empty state where data clearly belongs) IS visual \u2014 flag it.\n- Unsure whether it is a defect or the design? Say so in `description`, leave it\n  out of `issues`. A speculative issue costs the driver a real investigation.\n\n## Asked to read back a value\n\n\"What does this field contain?\" / \"is the box checked?\" / \"what does this text\nsay?\" \u2014 answer plainly in `description`, and add one sentence that the driver can\nread this exactly from `observe({kind:\"tree\"})` (`value`, `checked`, `name`)\nwhere a screenshot is a guess. Small text off pixels is what you are worst at \u2014\nsay you are unsure rather than guess a value.\n\nOutput the JSON object and nothing else.";
 /** Whether a provider error says the vision model rejected image input (text-only model). */
@@ -91,0 +91,0 @@ export declare function isImageRejection(message: string): boolean;

dist/agent/belt/look.js

@@ -56,12 +56,12 @@ /**
 You are the vision guy — the eyes of a blind UI-testing driver.
-You are given a screenshot of the target UI plus a question and/or an expected look.
-Describe what you see and judge how it looks: layout, alignment, spacing, colour,
-overlap, cut-off or unreadable text — anything that looks broken or off.
+You get a screenshot plus a question and/or an expected look. Describe what you
+see and judge how it LOOKS: layout, alignment, spacing, colour, overlap, cut-off
+or unreadable text.
 
-Reply with ONLY a JSON object (no prose, no markdown fences) of this exact shape:
+Reply with ONLY this JSON object — no prose, no markdown fences:
 
 {
   "description": "what you see, answering the question if one was asked",
-  "matches": true,            // include ONLY when an expected look was given: does it match it?
-  "issues": [                 // visual problems found; [] when there are none
+  "matches": true,            // ONLY when an expected look was given: does it match?
+  "issues": [                 // [] when there are none
     { "what": "the problem", "where": "where on screen", "severity": "low" | "medium" | "high" }
@@ -71,33 +71,27 @@ ]
 
-Severity, by user impact: high = a user cannot read or use something (overlap,
-invisible interactive element, text cut off mid-word, broken layout);
-medium = it works but looks wrong (misalignment, cramped or uneven spacing, weak
-contrast); low = polish (a few pixels off, a slightly odd shade, icon sizing).
-Be concrete about "where" (e.g. "top-right header", "primary submit button").
+Severity by user impact: \`high\` = cannot read or use it (overlap, invisible
+interactive element, text cut mid-word, broken layout); \`medium\` = works but looks
+wrong (misalignment, cramped spacing, weak contrast); \`low\` = polish (a few pixels,
+an odd shade, icon sizing). Be concrete in \`where\` — "top-right header", "primary
+submit button".
 
 ## Report only what is ON this screenshot
 
-- \`"issues": []\` is a normal, useful answer. An empty list is not a failure to
-  find something — a clean screen is a real result. Never pad the list to look thorough.
-- Judge only what is VISIBLE here. This is one viewport, not the whole page:
-  content below the fold is not "missing", and you cannot see hover, focus,
-  animation, other screen sizes, or what happens after a click. Do not report any
-  of those.
-- Do not judge whether the CONTENT is correct — whether a total is right, a name is
-  the expected one, a list holds the intended items. You cannot know the intent.
-  Report how it LOOKS. If something reads as obviously placeholder ("lorem ipsum",
-  "undefined", "NaN", an empty state where data was clearly expected), that IS
-  visual and worth flagging.
-- If you are unsure whether something is a defect or the intended design, say so in
-  \`description\` and leave it out of \`issues\`. A speculative issue costs the driver
-  a real investigation.
+- \`"issues": []\` is a normal, useful answer. NEVER pad the list to look thorough.
+- One viewport, not the whole page. Content below the fold is not "missing", and
+  you cannot see hover, focus, animation, other screen sizes, or what a click does.
+  NEVER report those.
+- NEVER judge whether CONTENT is correct — a total, a name, which items a list
+  holds. You cannot know the intent. Obvious placeholder ("lorem ipsum",
+  "undefined", "NaN", an empty state where data clearly belongs) IS visual — flag it.
+- Unsure whether it is a defect or the design? Say so in \`description\`, leave it
+  out of \`issues\`. A speculative issue costs the driver a real investigation.
 
-## If you were asked to read back a value
+## Asked to read back a value
 
-When the question is just "what does this field contain?" / "is the box checked?" /
-"what does this text say?", answer it plainly in \`description\` — and add one
-sentence noting the driver can read this structurally from
-\`observe({kind:"tree"})\` (fields \`value\`, \`checked\`, \`name\`), which is exact
-where a screenshot is a guess. Reading small text off pixels is the one thing you
-are worst at; say when you are unsure rather than guessing a value.
+"What does this field contain?" / "is the box checked?" / "what does this text
+say?" — answer plainly in \`description\`, and add one sentence that the driver can
+read this exactly from \`observe({kind:"tree"})\` (\`value\`, \`checked\`, \`name\`)
+where a screenshot is a guess. Small text off pixels is what you are worst at —
+say you are unsure rather than guess a value.
 
@@ -104,0 +98,0 @@ Output the JSON object and nothing else.`;

dist/agent/belt/look.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"look.js","sourceRoot":"","sources":["../../../src/agent/belt/look.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAGH,OAAO,EAAE,YAAY,EAAE,IAAI,EAAE,MAAM,IAAI,CAAC;AACxC,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAExB,OAAO,EAAE,UAAU,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAC;AAErE,wFAAwF;AACxF,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,CAAC,MAAM,CAAC;IACtC,QAAQ,EAAE,CAAC;SACR,MAAM,EAAE;SACR,QAAQ,EAAE;SACV,QAAQ,CAAC,+EAA+E,CAAC;IAC5F,MAAM,EAAE,CAAC;SACN,MAAM,EAAE;SACR,QAAQ,EAAE;SACV,QAAQ,CAAC,4EAA4E,CAAC;CAC1F,CAAC,CAAC;AAIH,6EAA6E;AAC7E,MAAM,UAAU,GAAG,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,CAAU,CAAC;AAEtD,uEAAuE;AACvE,MAAM,eAAe,GAAG,CAAC,CAAC,MAAM,CAAC;IAC/B,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,4CAA4C,CAAC;IACvE,KAAK,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,+CAA+C,CAAC;IAC3E,QAAQ,EAAE,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,QAAQ,CAAC,6CAA6C,CAAC;CACrF,CAAC,CAAC;AAIH,yFAAyF;AACzF,MAAM,iBAAiB,GAAG,CAAC,CAAC,MAAM,CAAC;IACjC,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,uCAAuC,CAAC;IACzE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,mDAAmD,CAAC;IAC7F,MAAM,EAAE,CAAC,CAAC,KAAK,CAAC,eAAe,CAAC,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,+BAA+B,CAAC;CACvF,CAAC,CAAC;AAiCH;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;yCAgDK,CAAC;AAE1C,qFAAqF;AACrF,SAAS,WAAW,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAa;IAClD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,IAAI,QAAQ;QAAE,KAAK,CAAC,IAAI,CAAC,aAAa,QAAQ,EAAE,CAAC,CAAC;IAClD,IAAI,MAAM;QAAE,KAAK,CAAC,IAAI,CAAC,kBAAkB,MAAM,EAAE,CAAC,CAAC;IACnD,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,KAAK,CAAC,IAAI,CAAC,mDAAmD,CAAC,CAAC;IACxF,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC;AAED,0FAA0F;AAC1F,SAAS,SAAS,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAa;IAChD,OAAO,QAAQ,IAAI,MAAM,IAAI,MAAM,CAAC;AACtC,CAAC;AAED,+EAA+E;AAC/E,SAAS,QAAQ,CAAC,IAAY,EAAE,GAAG,GAAG,GAAG;IACvC,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC5B,OAAO,OAAO,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC;AACtE,CAAC;AAED;;;;;;GAMG;AACH,MAAM,eAAe,GACnB,iRAAiR,CAAC;AAEpR,6FAA6F;AAC7F,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,OAAO,eAAe,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;AACvC,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,wBAAwB,CAAC,OAAe,EAAE,eAAuB;IAC/E,OAAO,CACL,mDAAmD,OAAO,yBAAyB;QACnF,mBAAmB,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,eAAe,EAAE,GAAG,CAAC,CAAC,oBAAoB;QACrF,0EAA0E;QAC1E,mFAAmF;QACnF,qFAAqF;QACrF,yBAAyB,CAC1B,CAAC;AACJ,CAAC;AAED,wFAAwF;AACxF,SAAS,WAAW,CAAC,IAAY;IAC/B,MAAM,MAAM,GAAG,kCAAkC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC7D,MAAM,IAAI,GAAG,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;IACjC,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IAChC,MAAM,GAAG,GAAG,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IAClC,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,GAAG,KAAK,CAAC,CAAC,IAAI,GAAG,GAAG,KAAK,EAAE,CAAC;QAC9C,MAAM,IAAI,UAAU,CAAC,yCAAyC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAClF,CAAC;IACD,OAAO,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC;AACpC,CAAC;AAED,qFAAqF;AACrF,SAAS,gBAAgB,CAAC,IAAY;IACpC,IAAI,IAAa,CAAC;IAClB,IAAI,CAAC;QACH,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC;IACvC,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,IAAI,CAAC,YAAY,UAAU;YAAE,MAAM,CAAC,CAAC;QACrC,MAAM,IAAI,UAAU,CAClB,0CAA0C,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CACvF,CAAC;IACJ,CAAC;IACD,MAAM,MAAM,GAAG,iBAAiB,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC;IACjD,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;QACpB,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM;aAC/B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,QAAQ,KAAK,CAAC,CAAC,OAAO,EAAE,CAAC;aACvE,IAAI,CAAC,IAAI,CAAC,CAAC;QACd,MAAM,IAAI,UAAU,CAAC,yCAAyC,MAAM,EAAE,CAAC,CAAC;IAC1E,CAAC;IACD,OAAO,MAAM,CAAC,IAAI,CAAC;AACrB,CAAC;AAED;;;;GAIG;AACH;;;;;;;;;;GAUG;AACH,MAAM,sBAAsB,GAAG,OAAO,CAAC;AAEvC,wGAAwG;AACxG,MAAM,kBAAkB,GAAG,EAAE,CAAC;AAS9B;;;;;GAKG;AACH,KAAK,UAAU,gBAAgB,CAAC,OAAgB;IAC9C,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,UAAU,EAAE,CAAC;IACvC,IAAI,GAAG,CAAC,UAAU,IAAI,sBAAsB,EAAE,CAAC;QAC7C,OAAO,EAAE,KAAK,EAAE,GAAG,EAAE,SAAS,EAAE,WAAW,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;IAClE,CAAC;IACD,MAAM,IAAI,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,kBAAkB,EAAE,CAAC,CAAC;IACvF,6EAA6E;IAC7E,oEAAoE;IACpE,OAAO,IAAI,CAAC,UAAU,GAAG,GAAG,CAAC,UAAU;QACrC,CAAC,CAAC,EAAE,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,KAAK,EAAE;QAC5D,CAAC,CAAC,EAAE,KAAK,EAAE,GAAG,EAAE,SAAS,EAAE,WAAW,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;AAC/D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,OAAO,CAC3B,OAAgB,EAChB,QAAwB,EACxB,QAA0B,EAC1B,KAAgB,EAChB,WAAyB;IAEzB,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAC9C,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,QAAQ,CAAC;QAC9B,MAAM,EAAE,oBAAoB;QAC5B,QAAQ,EAAE;YACR;gBACE,IAAI,EAAE,MAAM;gBACZ,OAAO,EAAE;oBACP,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,WAAW,CAAC,KAAK,CAAC,EAAE;oBAC1C,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,SAAS,EAAE,KAAK,CAAC,SAAS,EAAE;iBAClE;aACF;SACF;QACD,WAAW;KACZ,CAAC,CAAC;IACH,MAAM,KAAK,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC;IACrC,0EAA0E;IAC1E,4EAA4E;IAC5E,MAAM,UAAU,GAAG,MAAM,QAAQ,CAAC,cAAc,CAAC,SAAS,CAAC,KAAK,CAAC,EAAE,KAAK,CAAC,KAAK,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;IACjG,OAAO,EAAE,GAAG,KAAK,EAAE,UAAU,EAAE,CAAC;AAClC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAC/B,OAAgB,EAChB,QAAwB,EACxB,QAA0B;IAE1B,IAAI,WAA+C,CAAC;IACpD,OAAO,KAAK,EAAE,KAAgB,EAAE,WAAyB,EAAuB,EAAE;QAChF,IAAI,WAAW;YAAE,MAAM,WAAW,CAAC;QACnC,IAAI,CAAC;YACH,OAAO,MAAM,OAAO,CAAC,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,WAAW,CAAC,CAAC;QACxE,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,IAAI,CAAC,YAAY,sBAAsB;gBAAE,WAAW,GAAG,CAAC,CAAC;YACzD,MAAM,CAAC,CAAC;QACV,CAAC;IACH,CAAC,CAAC;AACJ,CAAC;AAgBD;;;;;;;;;;GAUG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAgB,EAAE,QAA0B;IAC7E,OAAO,IAAI,CAAC;QACV,WAAW,EACT,0TAA0T;QAC5T,WAAW,EAAE,eAAe;QAC5B,OAAO,EAAE,KAAK,EAAE,KAAK,EAA2B,EAAE;YAChD,4EAA4E;YAC5E,6EAA6E;YAC7E,0EAA0E;YAC1E,mBAAmB;YACnB,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAAC,OAAO,CAAC,CAAC;YAC9C,MAAM,UAAU,GAAG,MAAM,QAAQ,CAAC,cAAc,CAC9C,SAAS,CAAC,KAAK,CAAC,EAChB,KAAK,CAAC,KAAK,EACX,KAAK,CAAC,SAAS,CAChB,CAAC;YACF,OAAO;gBACL,UAAU;gBACV,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,UAAU;gBAC7B,SAAS,EAAE,KAAK,CAAC,SAAS;gBAC1B,MAAM,EAAE,WAAW,CAAC,KAAK,CAAC;gBAC1B,KAAK,EAAE,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC;aACnD,CAAC;QACJ,CAAC;QACD,aAAa,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC,CAAC;YAC9B,IAAI,EAAE,SAAS;YACf,KAAK,EAAE;gBACL;oBACE,IAAI,EAAE,MAAM;oBACZ,IAAI,EACF,qBAAqB,MAAM,CAAC,UAAU,KAAK,MAAM,CAAC,MAAM,IAAI;wBAC5D,2EAA2E;wBAC3E,SAAS,MAAM,CAAC,UAAU,yBAAyB;iBACtD;gBACD,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,MAAM,CAAC,KAAK,EAAE,SAAS,EAAE,MAAM,CAAC,SAAS,EAAE;aACvE;SACF,CAAC;KACH,CAAC,CAAC;AACL,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,cAAc,CAC5B,OAAgB,EAChB,MAAqB,EACrB,QAA0B;IAE1B,MAAM,OAAO,GAAG,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC;IACrE,6EAA6E;IAC7E,2EAA2E;IAC3E,gFAAgF;IAChF,MAAM,QAAQ,GAAmB,KAAK,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,WAAW,EAAE,EAAE,EAAE;QAC3E,IAAI,CAAC;YACH,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,YAAY,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,WAAW,EAAE,CAAC,CAAC;YACtF,OAAO,EAAE,IAAI,EAAE,CAAC;QAClB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,OAAO,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAC3D,IAAI,gBAAgB,CAAC,OAAO,CAAC,EAAE,CAAC;gBAC9B,MAAM,IAAI,sBAAsB,CAAC,wBAAwB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;YAC/E,CAAC;YACD,MAAM,CAAC,CAAC;QACV,CAAC;IACH,CAAC,CAAC;IACF,MAAM,OAAO,GAAG,iBAAiB,CAAC,OAAO,EAAE,QAAQ,EAAE,QAAQ,CAAC,CAAC;IAC/D,OAAO,IAAI,CAAC;QACV,WAAW,EACT,0XAA0X;QAC5X,WAAW,EAAE,eAAe;QAC5B,OAAO,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,WAAW,CAAC;KACjE,CAAC,CAAC;AACL,CAAC"}
+{"version":3,"file":"look.js","sourceRoot":"","sources":["../../../src/agent/belt/look.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAGH,OAAO,EAAE,YAAY,EAAE,IAAI,EAAE,MAAM,IAAI,CAAC;AACxC,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAExB,OAAO,EAAE,UAAU,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAC;AAErE,wFAAwF;AACxF,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,CAAC,MAAM,CAAC;IACtC,QAAQ,EAAE,CAAC;SACR,MAAM,EAAE;SACR,QAAQ,EAAE;SACV,QAAQ,CAAC,+EAA+E,CAAC;IAC5F,MAAM,EAAE,CAAC;SACN,MAAM,EAAE;SACR,QAAQ,EAAE;SACV,QAAQ,CAAC,4EAA4E,CAAC;CAC1F,CAAC,CAAC;AAIH,6EAA6E;AAC7E,MAAM,UAAU,GAAG,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,CAAU,CAAC;AAEtD,uEAAuE;AACvE,MAAM,eAAe,GAAG,CAAC,CAAC,MAAM,CAAC;IAC/B,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,4CAA4C,CAAC;IACvE,KAAK,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,+CAA+C,CAAC;IAC3E,QAAQ,EAAE,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,QAAQ,CAAC,6CAA6C,CAAC;CACrF,CAAC,CAAC;AAIH,yFAAyF;AACzF,MAAM,iBAAiB,GAAG,CAAC,CAAC,MAAM,CAAC;IACjC,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,uCAAuC,CAAC;IACzE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,mDAAmD,CAAC;IAC7F,MAAM,EAAE,CAAC,CAAC,KAAK,CAAC,eAAe,CAAC,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,+BAA+B,CAAC;CACvF,CAAC,CAAC;AAiCH;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;yCA0CK,CAAC;AAE1C,qFAAqF;AACrF,SAAS,WAAW,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAa;IAClD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,IAAI,QAAQ;QAAE,KAAK,CAAC,IAAI,CAAC,aAAa,QAAQ,EAAE,CAAC,CAAC;IAClD,IAAI,MAAM;QAAE,KAAK,CAAC,IAAI,CAAC,kBAAkB,MAAM,EAAE,CAAC,CAAC;IACnD,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,KAAK,CAAC,IAAI,CAAC,mDAAmD,CAAC,CAAC;IACxF,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC;AAED,0FAA0F;AAC1F,SAAS,SAAS,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAa;IAChD,OAAO,QAAQ,IAAI,MAAM,IAAI,MAAM,CAAC;AACtC,CAAC;AAED,+EAA+E;AAC/E,SAAS,QAAQ,CAAC,IAAY,EAAE,GAAG,GAAG,GAAG;IACvC,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC5B,OAAO,OAAO,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC;AACtE,CAAC;AAED;;;;;;GAMG;AACH,MAAM,eAAe,GACnB,iRAAiR,CAAC;AAEpR,6FAA6F;AAC7F,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,OAAO,eAAe,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;AACvC,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,wBAAwB,CAAC,OAAe,EAAE,eAAuB;IAC/E,OAAO,CACL,mDAAmD,OAAO,yBAAyB;QACnF,mBAAmB,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,eAAe,EAAE,GAAG,CAAC,CAAC,oBAAoB;QACrF,0EAA0E;QAC1E,mFAAmF;QACnF,qFAAqF;QACrF,yBAAyB,CAC1B,CAAC;AACJ,CAAC;AAED,wFAAwF;AACxF,SAAS,WAAW,CAAC,IAAY;IAC/B,MAAM,MAAM,GAAG,kCAAkC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC7D,MAAM,IAAI,GAAG,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;IACjC,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IAChC,MAAM,GAAG,GAAG,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IAClC,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,GAAG,KAAK,CAAC,CAAC,IAAI,GAAG,GAAG,KAAK,EAAE,CAAC;QAC9C,MAAM,IAAI,UAAU,CAAC,yCAAyC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAClF,CAAC;IACD,OAAO,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC;AACpC,CAAC;AAED,qFAAqF;AACrF,SAAS,gBAAgB,CAAC,IAAY;IACpC,IAAI,IAAa,CAAC;IAClB,IAAI,CAAC;QACH,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC;IACvC,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,IAAI,CAAC,YAAY,UAAU;YAAE,MAAM,CAAC,CAAC;QACrC,MAAM,IAAI,UAAU,CAClB,0CAA0C,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CACvF,CAAC;IACJ,CAAC;IACD,MAAM,MAAM,GAAG,iBAAiB,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC;IACjD,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;QACpB,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM;aAC/B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,QAAQ,KAAK,CAAC,CAAC,OAAO,EAAE,CAAC;aACvE,IAAI,CAAC,IAAI,CAAC,CAAC;QACd,MAAM,IAAI,UAAU,CAAC,yCAAyC,MAAM,EAAE,CAAC,CAAC;IAC1E,CAAC;IACD,OAAO,MAAM,CAAC,IAAI,CAAC;AACrB,CAAC;AAED;;;;GAIG;AACH;;;;;;;;;;GAUG;AACH,MAAM,sBAAsB,GAAG,OAAO,CAAC;AAEvC,wGAAwG;AACxG,MAAM,kBAAkB,GAAG,EAAE,CAAC;AAS9B;;;;;GAKG;AACH,KAAK,UAAU,gBAAgB,CAAC,OAAgB;IAC9C,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,UAAU,EAAE,CAAC;IACvC,IAAI,GAAG,CAAC,UAAU,IAAI,sBAAsB,EAAE,CAAC;QAC7C,OAAO,EAAE,KAAK,EAAE,GAAG,EAAE,SAAS,EAAE,WAAW,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;IAClE,CAAC;IACD,MAAM,IAAI,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,kBAAkB,EAAE,CAAC,CAAC;IACvF,6EAA6E;IAC7E,oEAAoE;IACpE,OAAO,IAAI,CAAC,UAAU,GAAG,GAAG,CAAC,UAAU;QACrC,CAAC,CAAC,EAAE,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,KAAK,EAAE;QAC5D,CAAC,CAAC,EAAE,KAAK,EAAE,GAAG,EAAE,SAAS,EAAE,WAAW,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;AAC/D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,OAAO,CAC3B,OAAgB,EAChB,QAAwB,EACxB,QAA0B,EAC1B,KAAgB,EAChB,WAAyB;IAEzB,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAC9C,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,QAAQ,CAAC;QAC9B,MAAM,EAAE,oBAAoB;QAC5B,QAAQ,EAAE;YACR;gBACE,IAAI,EAAE,MAAM;gBACZ,OAAO,EAAE;oBACP,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,WAAW,CAAC,KAAK,CAAC,EAAE;oBAC1C,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,SAAS,EAAE,KAAK,CAAC,SAAS,EAAE;iBAClE;aACF;SACF;QACD,WAAW;KACZ,CAAC,CAAC;IACH,MAAM,KAAK,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC;IACrC,0EAA0E;IAC1E,4EAA4E;IAC5E,MAAM,UAAU,GAAG,MAAM,QAAQ,CAAC,cAAc,CAAC,SAAS,CAAC,KAAK,CAAC,EAAE,KAAK,CAAC,KAAK,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;IACjG,OAAO,EAAE,GAAG,KAAK,EAAE,UAAU,EAAE,CAAC;AAClC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAC/B,OAAgB,EAChB,QAAwB,EACxB,QAA0B;IAE1B,IAAI,WAA+C,CAAC;IACpD,OAAO,KAAK,EAAE,KAAgB,EAAE,WAAyB,EAAuB,EAAE;QAChF,IAAI,WAAW;YAAE,MAAM,WAAW,CAAC;QACnC,IAAI,CAAC;YACH,OAAO,MAAM,OAAO,CAAC,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,WAAW,CAAC,CAAC;QACxE,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,IAAI,CAAC,YAAY,sBAAsB;gBAAE,WAAW,GAAG,CAAC,CAAC;YACzD,MAAM,CAAC,CAAC;QACV,CAAC;IACH,CAAC,CAAC;AACJ,CAAC;AAgBD;;;;;;;;;;GAUG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAgB,EAAE,QAA0B;IAC7E,OAAO,IAAI,CAAC;QACV,WAAW,EACT,0TAA0T;QAC5T,WAAW,EAAE,eAAe;QAC5B,OAAO,EAAE,KAAK,EAAE,KAAK,EAA2B,EAAE;YAChD,4EAA4E;YAC5E,6EAA6E;YAC7E,0EAA0E;YAC1E,mBAAmB;YACnB,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAAC,OAAO,CAAC,CAAC;YAC9C,MAAM,UAAU,GAAG,MAAM,QAAQ,CAAC,cAAc,CAC9C,SAAS,CAAC,KAAK,CAAC,EAChB,KAAK,CAAC,KAAK,EACX,KAAK,CAAC,SAAS,CAChB,CAAC;YACF,OAAO;gBACL,UAAU;gBACV,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,UAAU;gBAC7B,SAAS,EAAE,KAAK,CAAC,SAAS;gBAC1B,MAAM,EAAE,WAAW,CAAC,KAAK,CAAC;gBAC1B,KAAK,EAAE,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC;aACnD,CAAC;QACJ,CAAC;QACD,aAAa,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC,CAAC;YAC9B,IAAI,EAAE,SAAS;YACf,KAAK,EAAE;gBACL;oBACE,IAAI,EAAE,MAAM;oBACZ,IAAI,EACF,qBAAqB,MAAM,CAAC,UAAU,KAAK,MAAM,CAAC,MAAM,IAAI;wBAC5D,2EAA2E;wBAC3E,SAAS,MAAM,CAAC,UAAU,yBAAyB;iBACtD;gBACD,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,MAAM,CAAC,KAAK,EAAE,SAAS,EAAE,MAAM,CAAC,SAAS,EAAE;aACvE;SACF,CAAC;KACH,CAAC,CAAC;AACL,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,cAAc,CAC5B,OAAgB,EAChB,MAAqB,EACrB,QAA0B;IAE1B,MAAM,OAAO,GAAG,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC;IACrE,6EAA6E;IAC7E,2EAA2E;IAC3E,gFAAgF;IAChF,MAAM,QAAQ,GAAmB,KAAK,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,WAAW,EAAE,EAAE,EAAE;QAC3E,IAAI,CAAC;YACH,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,YAAY,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,WAAW,EAAE,CAAC,CAAC;YACtF,OAAO,EAAE,IAAI,EAAE,CAAC;QAClB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,OAAO,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAC3D,IAAI,gBAAgB,CAAC,OAAO,CAAC,EAAE,CAAC;gBAC9B,MAAM,IAAI,sBAAsB,CAAC,wBAAwB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;YAC/E,CAAC;YACD,MAAM,CAAC,CAAC;QACV,CAAC;IACH,CAAC,CAAC;IACF,MAAM,OAAO,GAAG,iBAAiB,CAAC,OAAO,EAAE,QAAQ,EAAE,QAAQ,CAAC,CAAC;IAC/D,OAAO,IAAI,CAAC;QACV,WAAW,EACT,0XAA0X;QAC5X,WAAW,EAAE,eAAe;QAC5B,OAAO,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,WAAW,CAAC;KACjE,CAAC,CAAC;AACL,CAAC"}

dist/agent/prompts/android-addendum.d.ts

@@ -12,2 +12,2 @@ /**
  */
-export declare const ANDROID_ADDENDUM_PROMPT = "## Android target \u2014 ADB + uiautomator reach\n\nYou are driving an Android emulator (or an attached device) over **ADB**. You read\nthe screen through the **uiautomator view hierarchy**, drive it with **input** events\n(tap/text/swipe/keyevent), and capture frames with **screencap**. There is no DOM and\nno JavaScript context. Unlike desktop you DO get a console \u2014 **logcat** \u2014 but there is\n**no network channel** over ADB.\n\n### Channels and when to reach for each\n\n| Channel        | What it gives you | Use via |\n|----------------|-------------------|---------|\n| view hierarchy | element class, text/content-desc, on-screen bounds, enabled/clickable state | `observe({kind:\"tree\"})` |\n| screenshot     | the frame SAVED as evidence \u2014 you get a file path back, not pixels | `observe({kind:\"screenshot\"})` |\n| eyes           | an actual visual judgment of the current frame | `look` |\n| logcat         | app logs, crashes, stack traces, ANRs | `observe({kind:\"console\"})` |\n\n`observe({kind:\"network\"})` is **unsupported** on android and errors out \u2014 ADB exposes\nno request stream. Never call it; judge behaviour from the hierarchy, logcat, and the\npixels instead.\n\n`observe({kind:\"tabs\"})` and `act({action:\"switch_tab\"})` are **unsupported** on\nandroid and error out too \u2014 an app has no tabs to switch between. Never call them;\nmove between screens by tapping, or relaunch with `act({action:\"navigate\"})`.\n\n### Launch the app yourself \u2014 the device boots to the launcher\n\nThe emulator boots (or attaches) to the home screen, NOT your app. Your FIRST step is\nto launch it:\n`act({action:\"navigate\", target:\"<package or component>\"})`\n- A component `com.example.app/.MainActivity` starts that activity directly.\n- A bare package `com.example.app` launches its default launcher activity.\nThe package/activity to open is named in your goal below. Re-navigate any time to\nrelaunch from a clean state.\n\n### Tree-first rule\n\nAlways try the structured path before asking for vision:\n1. `observe({kind:\"tree\"})` \u2014 read element class, text, content-desc, bounds, state.\n2. If an element is missing, scroll (`act({action:\"scroll\"})`) or wait\n   (`act({action:\"wait\", target:\"<text/desc>\"})`), then re-observe.\n3. The view hierarchy is often thinner than a DOM \u2014 custom-drawn views (Canvas,\n   Compose without semantics, games) expose little. When the tree gives you no\n   answer, call `look` for pixels sooner than you would on web.\n4. After a crash-prone action, check `observe({kind:\"console\"})` \u2014 a stack trace or\n   ANR in logcat is the real signal, often before anything changes on screen.\n\n### Waiting \u2014 node queries only\n\n`act({action:\"wait\"})` polls the view hierarchy for a `target` to appear. There is no\n`networkIdle` on android (no network channel) \u2014 waiting on it errors. Wait on a\nvisible element instead, e.g. a button label or a piece of text you expect to render.\n\n### Selectors \u2014 use the node's `target`, don't invent one\n\nActionable nodes from `observe({kind:\"tree\"})` carry a ready-to-use `target` string\nbuilt from their resource-id, or their role + text/content-desc. When a node has a\n`target`, COPY it verbatim into `act({action, target})`.\n\nAndroid targets take exactly three forms, and NOTHING else:\n- a `resource-id` \u2014 e.g. `com.example.app:id/submit` (what a node's `testid` holds;\n  the most reliable of the three).\n- `role \"name\"` \u2014 e.g. `button \"Save\"`, `textbox \"Email\"` (role must match exactly,\n  name is a case-insensitive substring of the text/content-desc).\n- a plain text substring \u2014 e.g. `Save`.\n\nWeb selector syntax does NOT work here: `data-testid=\"\u2026\"`, `role=button[name=\"Save\" i]`,\n`text=Save` and `>> nth=` are all read as literal text, match nothing, and cost you the\nstep. If a node is unlabeled, or repeats and so carries no `target`, scope the read with\n`within`/`filters` and act on what comes back.\n\n### Findings \u2014 functional + visual\n\nRecord functional bugs (a tap that does nothing, a disabled control that should be\nlive, a wrong value, a crash) AND visual/UX feedback (misaligned, clipped, cut off by\na notch, cramped touch targets). Attach the screenshot path as `evidence`, and quote\nthe offending logcat line (stack trace / ANR) as `detail` when a crash is involved.\n";
+export declare const ANDROID_ADDENDUM_PROMPT = "## Android target \u2014 ADB + uiautomator reach\n\nYou drive an Android emulator (or attached device) over **ADB**. You read the screen\nthrough the **uiautomator view hierarchy**, drive it with **input** events\n(tap/text/swipe/keyevent), capture with **screencap**. No DOM, no JavaScript. Unlike\ndesktop you DO get a console \u2014 **logcat** \u2014 but **no network channel** over ADB.\n\n| Channel | Gives you | Reach via |\n|---------|-----------|-----------|\n| view hierarchy | class, text/content-desc, bounds, enabled/clickable state | `observe({kind:\"tree\"})` |\n| screenshot | the frame SAVED as evidence \u2014 a file path back, not pixels | `observe({kind:\"screenshot\"})` |\n| `look` | an actual visual judgment of the current frame | `look` |\n| logcat | app logs, crashes, stack traces, ANRs | `observe({kind:\"console\"})` |\n\n`observe({kind:\"network\"})` is **unsupported** here and errors out \u2014 ADB exposes no\nrequest stream. NEVER call it; judge from the hierarchy, logcat and the pixels.\n\n`observe({kind:\"tabs\"})` and `act({action:\"switch_tab\"})` are **unsupported** too \u2014 an\napp has no tabs. NEVER call them; move between screens by tapping, or relaunch with\n`act({action:\"navigate\"})`.\n\n### Launch the app yourself \u2014 the device boots to the launcher\n\nThe emulator boots (or attaches) to the home screen, NOT your app. FIRST step:\n`act({action:\"navigate\", target:\"<package or component>\"})`\n- `com.example.app/.MainActivity` \u2014 starts that activity directly.\n- `com.example.app` \u2014 launches the default launcher activity.\nThe package/activity is named in your goal below. Re-navigate any time for a clean state.\n\n### Tree first\n\n1. `observe({kind:\"tree\"})` \u2014 class, text, content-desc, bounds, state.\n2. Element missing \u2192 `act({action:\"scroll\"})` or `act({action:\"wait\", target:\"<text/desc>\"})`,\n   then re-observe.\n3. The hierarchy is thinner than a DOM \u2014 custom-drawn views (Canvas, Compose without\n   semantics, games) expose little. Reach for `look` sooner than you would on web.\n4. After a crash-prone action check `observe({kind:\"console\"})` \u2014 a stack trace or ANR\n   in logcat is the real signal, often before the screen changes.\n\n### Waiting \u2014 node queries only\n\n`act({action:\"wait\"})` polls the view hierarchy for a `target`. No `networkIdle` here\n(no network channel) \u2014 waiting on it errors. Wait on a visible element: a button\nlabel, a piece of text you expect.\n\n### Selectors \u2014 copy the node's `target`\n\nActionable nodes carry a ready `target` built from resource-id, or role +\ntext/content-desc. COPY it verbatim into `act({action, target})`.\n\nExactly three forms, NOTHING else:\n- a `resource-id` \u2014 `com.example.app:id/submit` (what a node's `testid` holds; the\n  most reliable).\n- `role \"name\"` \u2014 `button \"Save\"`, `textbox \"Email\"` (role exact, name a\n  case-insensitive substring of the text/content-desc).\n- a plain text substring \u2014 `Save`.\n\nWeb syntax does NOT work: `data-testid=\"\u2026\"`, `role=button[name=\"Save\" i]`, `text=Save`\nand `>> nth=` are read as literal text, match nothing, and cost you the step. Unlabeled\nor repeated node with no `target` \u2192 scope with `within`/`filters` and act on what\ncomes back.\n\n### Findings \u2014 functional + visual\n\nRecord functional bugs (a tap that does nothing, a control disabled that should be\nlive, a wrong value, a crash) AND visual/UX feedback (misaligned, clipped, cut off by\na notch, cramped touch targets). Screenshot path as `evidence`; quote the offending\nlogcat line (stack trace / ANR) as `detail` when a crash is involved.\n";

dist/agent/prompts/android-addendum.js

@@ -15,78 +15,69 @@ /**
 
-You are driving an Android emulator (or an attached device) over **ADB**. You read
-the screen through the **uiautomator view hierarchy**, drive it with **input** events
-(tap/text/swipe/keyevent), and capture frames with **screencap**. There is no DOM and
-no JavaScript context. Unlike desktop you DO get a console — **logcat** — but there is
-**no network channel** over ADB.
+You drive an Android emulator (or attached device) over **ADB**. You read the screen
+through the **uiautomator view hierarchy**, drive it with **input** events
+(tap/text/swipe/keyevent), capture with **screencap**. No DOM, no JavaScript. Unlike
+desktop you DO get a console — **logcat** — but **no network channel** over ADB.
 
-### Channels and when to reach for each
+| Channel | Gives you | Reach via |
+|---------|-----------|-----------|
+| view hierarchy | class, text/content-desc, bounds, enabled/clickable state | \`observe({kind:"tree"})\` |
+| screenshot | the frame SAVED as evidence — a file path back, not pixels | \`observe({kind:"screenshot"})\` |
+| \`look\` | an actual visual judgment of the current frame | \`look\` |
+| logcat | app logs, crashes, stack traces, ANRs | \`observe({kind:"console"})\` |
 
-| Channel        | What it gives you | Use via |
-|----------------|-------------------|---------|
-| view hierarchy | element class, text/content-desc, on-screen bounds, enabled/clickable state | \`observe({kind:"tree"})\` |
-| screenshot     | the frame SAVED as evidence — you get a file path back, not pixels | \`observe({kind:"screenshot"})\` |
-| eyes           | an actual visual judgment of the current frame | \`look\` |
-| logcat         | app logs, crashes, stack traces, ANRs | \`observe({kind:"console"})\` |
+\`observe({kind:"network"})\` is **unsupported** here and errors out — ADB exposes no
+request stream. NEVER call it; judge from the hierarchy, logcat and the pixels.
 
-\`observe({kind:"network"})\` is **unsupported** on android and errors out — ADB exposes
-no request stream. Never call it; judge behaviour from the hierarchy, logcat, and the
-pixels instead.
+\`observe({kind:"tabs"})\` and \`act({action:"switch_tab"})\` are **unsupported** too — an
+app has no tabs. NEVER call them; move between screens by tapping, or relaunch with
+\`act({action:"navigate"})\`.
 
-\`observe({kind:"tabs"})\` and \`act({action:"switch_tab"})\` are **unsupported** on
-android and error out too — an app has no tabs to switch between. Never call them;
-move between screens by tapping, or relaunch with \`act({action:"navigate"})\`.
-
 ### Launch the app yourself — the device boots to the launcher
 
-The emulator boots (or attaches) to the home screen, NOT your app. Your FIRST step is
-to launch it:
+The emulator boots (or attaches) to the home screen, NOT your app. FIRST step:
 \`act({action:"navigate", target:"<package or component>"})\`
-- A component \`com.example.app/.MainActivity\` starts that activity directly.
-- A bare package \`com.example.app\` launches its default launcher activity.
-The package/activity to open is named in your goal below. Re-navigate any time to
-relaunch from a clean state.
+- \`com.example.app/.MainActivity\` — starts that activity directly.
+- \`com.example.app\` — launches the default launcher activity.
+The package/activity is named in your goal below. Re-navigate any time for a clean state.
 
-### Tree-first rule
+### Tree first
 
-Always try the structured path before asking for vision:
-1. \`observe({kind:"tree"})\` — read element class, text, content-desc, bounds, state.
-2. If an element is missing, scroll (\`act({action:"scroll"})\`) or wait
-   (\`act({action:"wait", target:"<text/desc>"})\`), then re-observe.
-3. The view hierarchy is often thinner than a DOM — custom-drawn views (Canvas,
-   Compose without semantics, games) expose little. When the tree gives you no
-   answer, call \`look\` for pixels sooner than you would on web.
-4. After a crash-prone action, check \`observe({kind:"console"})\` — a stack trace or
-   ANR in logcat is the real signal, often before anything changes on screen.
+1. \`observe({kind:"tree"})\` — class, text, content-desc, bounds, state.
+2. Element missing → \`act({action:"scroll"})\` or \`act({action:"wait", target:"<text/desc>"})\`,
+   then re-observe.
+3. The hierarchy is thinner than a DOM — custom-drawn views (Canvas, Compose without
+   semantics, games) expose little. Reach for \`look\` sooner than you would on web.
+4. After a crash-prone action check \`observe({kind:"console"})\` — a stack trace or ANR
+   in logcat is the real signal, often before the screen changes.
 
 ### Waiting — node queries only
 
-\`act({action:"wait"})\` polls the view hierarchy for a \`target\` to appear. There is no
-\`networkIdle\` on android (no network channel) — waiting on it errors. Wait on a
-visible element instead, e.g. a button label or a piece of text you expect to render.
+\`act({action:"wait"})\` polls the view hierarchy for a \`target\`. No \`networkIdle\` here
+(no network channel) — waiting on it errors. Wait on a visible element: a button
+label, a piece of text you expect.
 
-### Selectors — use the node's \`target\`, don't invent one
+### Selectors — copy the node's \`target\`
 
-Actionable nodes from \`observe({kind:"tree"})\` carry a ready-to-use \`target\` string
-built from their resource-id, or their role + text/content-desc. When a node has a
-\`target\`, COPY it verbatim into \`act({action, target})\`.
+Actionable nodes carry a ready \`target\` built from resource-id, or role +
+text/content-desc. COPY it verbatim into \`act({action, target})\`.
 
-Android targets take exactly three forms, and NOTHING else:
-- a \`resource-id\` — e.g. \`com.example.app:id/submit\` (what a node's \`testid\` holds;
-  the most reliable of the three).
-- \`role "name"\` — e.g. \`button "Save"\`, \`textbox "Email"\` (role must match exactly,
-  name is a case-insensitive substring of the text/content-desc).
-- a plain text substring — e.g. \`Save\`.
+Exactly three forms, NOTHING else:
+- a \`resource-id\` — \`com.example.app:id/submit\` (what a node's \`testid\` holds; the
+  most reliable).
+- \`role "name"\` — \`button "Save"\`, \`textbox "Email"\` (role exact, name a
+  case-insensitive substring of the text/content-desc).
+- a plain text substring — \`Save\`.
 
-Web selector syntax does NOT work here: \`data-testid="…"\`, \`role=button[name="Save" i]\`,
-\`text=Save\` and \`>> nth=\` are all read as literal text, match nothing, and cost you the
-step. If a node is unlabeled, or repeats and so carries no \`target\`, scope the read with
-\`within\`/\`filters\` and act on what comes back.
+Web syntax does NOT work: \`data-testid="…"\`, \`role=button[name="Save" i]\`, \`text=Save\`
+and \`>> nth=\` are read as literal text, match nothing, and cost you the step. Unlabeled
+or repeated node with no \`target\` → scope with \`within\`/\`filters\` and act on what
+comes back.
 
 ### Findings — functional + visual
 
-Record functional bugs (a tap that does nothing, a disabled control that should be
+Record functional bugs (a tap that does nothing, a control disabled that should be
 live, a wrong value, a crash) AND visual/UX feedback (misaligned, clipped, cut off by
-a notch, cramped touch targets). Attach the screenshot path as \`evidence\`, and quote
-the offending logcat line (stack trace / ANR) as \`detail\` when a crash is involved.
+a notch, cramped touch targets). Screenshot path as \`evidence\`; quote the offending
+logcat line (stack trace / ANR) as \`detail\` when a crash is involved.
 `;
 //# sourceMappingURL=android-addendum.js.map

dist/agent/prompts/android-addendum.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"android-addendum.js","sourceRoot":"","sources":["../../../src/agent/prompts/android-addendum.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,MAAM,CAAC,MAAM,uBAAuB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8EtC,CAAC"}
+{"version":3,"file":"android-addendum.js","sourceRoot":"","sources":["../../../src/agent/prompts/android-addendum.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,MAAM,CAAC,MAAM,uBAAuB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAqEtC,CAAC"}

dist/agent/prompts/compose.js

@@ -40,7 +40,7 @@ /**
 
-You are ALREADY on it — the run opened it before your first step. This is the only
-app you are debugging. Never navigate to any other host, however plausible its name
-looks next to the goal: a bug you find somewhere else is worthless, because the
-smart agent cannot fix code it does not own. Links and paths within this app are
-fine; a different host is not, and \`act\` will refuse it.
+You are ALREADY on it — the run opened it before your first step. This is the ONLY
+app you debug. NEVER navigate to another host, however plausible its name looks next
+to the goal: a bug found elsewhere is worthless, because the smart agent cannot fix
+code it does not own. Links and paths within this app are fine; a different host is
+not, and \`act\` will refuse it.
 `
@@ -57,4 +57,3 @@ : '';
 
-The smart agent considers this run PASSED only when ALL of the following are true.
-Evaluate each one explicitly in your \`report\`:
+PASSED only when ALL of these are true. Evaluate each explicitly in your \`report\`:
 
@@ -61,0 +60,0 @@ ${criteria.map((c, i) => `${i + 1}. ${c}`).join('\n')}

dist/agent/prompts/compose.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"compose.js","sourceRoot":"","sources":["../../../src/agent/prompts/compose.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,uBAAuB,EAAE,MAAM,uBAAuB,CAAC;AAChE,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAC;AACpD,OAAO,EAAE,uBAAuB,EAAE,MAAM,uBAAuB,CAAC;AAChE,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAKxD,2CAA2C;AAC3C,MAAM,cAAc,GAA+B;IACjD,GAAG,EAAE,mBAAmB;IACxB,OAAO,EAAE,uBAAuB;IAChC,OAAO,EAAE,uBAAuB;CACjC,CAAC;AA0BF;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAuB;IACzD,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,OAAO,CAAC;IAE/D,MAAM,QAAQ,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC;IAExC,MAAM,cAAc,GAAG,OAAO;QAC5B,CAAC,CAAC;;;EAGJ,OAAO;;;;;;;CAOR;QACG,CAAC,CAAC,EAAE,CAAC;IAEP,MAAM,YAAY,GAAG;;;EAGrB,KAAK,CAAC,IAAI,EAAE;CACb,CAAC;IAEA,MAAM,eAAe,GACnB,QAAQ,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC;QAC7B,CAAC,CAAC;;;;;;EAMN,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;CACpD;QACK,CAAC,CAAC,EAAE,CAAC;IAET,OAAO,CAAC,gBAAgB,CAAC,QAAQ,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,YAAY,EAAE,eAAe,CAAC;SACzF,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC;SACvC,IAAI,CAAC,WAAW,CAAC,CAAC;AACvB,CAAC"}
+{"version":3,"file":"compose.js","sourceRoot":"","sources":["../../../src/agent/prompts/compose.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,uBAAuB,EAAE,MAAM,uBAAuB,CAAC;AAChE,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAC;AACpD,OAAO,EAAE,uBAAuB,EAAE,MAAM,uBAAuB,CAAC;AAChE,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAKxD,2CAA2C;AAC3C,MAAM,cAAc,GAA+B;IACjD,GAAG,EAAE,mBAAmB;IACxB,OAAO,EAAE,uBAAuB;IAChC,OAAO,EAAE,uBAAuB;CACjC,CAAC;AA0BF;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAuB;IACzD,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,OAAO,CAAC;IAE/D,MAAM,QAAQ,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC;IAExC,MAAM,cAAc,GAAG,OAAO;QAC5B,CAAC,CAAC;;;EAGJ,OAAO;;;;;;;CAOR;QACG,CAAC,CAAC,EAAE,CAAC;IAEP,MAAM,YAAY,GAAG;;;EAGrB,KAAK,CAAC,IAAI,EAAE;CACb,CAAC;IAEA,MAAM,eAAe,GACnB,QAAQ,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC;QAC7B,CAAC,CAAC;;;;;EAKN,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;CACpD;QACK,CAAC,CAAC,EAAE,CAAC;IAET,OAAO,CAAC,gBAAgB,CAAC,QAAQ,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,YAAY,EAAE,eAAe,CAAC;SACzF,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC;SACvC,IAAI,CAAC,WAAW,CAAC,CAAC;AACvB,CAAC"}

dist/agent/prompts/debug-agent.js

@@ -45,146 +45,124 @@ /**
 You are the debug agent — the driver that tests UIs and reports findings.
-You control the target (browser, desktop, or mobile) through structured tools.
+You drive the target (browser, desktop, mobile) through structured tools.
 ${mode.eyes}
 
-## Your tool belt
+## Tool belt
 
-- \`observe\` — read state: DOM tree, screenshot path, console logs, network requests, open tabs.
-- \`act\` — take action: click, type, key, scroll, navigate, wait, switch_tab.
+- \`observe\` — read state: tree, screenshot path, console, network, tabs.
+- \`act\` — click, type, key, scroll, navigate, wait, switch_tab.
 ${mode.belt}
-- \`report\` — emit the final structured findings and STOP. Call once; it ends your run.
+- \`report\` — emit final findings and STOP. Once per run; it ends the run.
 
-## Structure-first rule
+## Structure first — never screenshot what you can read
 
-Always prefer structured reads over vision:
-1. Use \`observe({kind:"tree"})\` to read DOM / accessibility tree.
-2. Where the target supports them, use \`observe({kind:"console"})\` and
-   \`observe({kind:"network"})\` to watch errors.
-3. Call \`look\` ONLY when visual judgment is needed — layout, colour, alignment,
-   "does this look right?", "is this element centred?".
+1. \`observe({kind:"tree"})\` for the DOM / accessibility tree.
+2. \`observe({kind:"console"})\` + \`observe({kind:"network"})\` for errors, where supported.
+3. \`look\` ONLY for visual judgment — layout, colour, alignment, "is this centred?".
    ${mode.cost}
 
-Never screenshot for information you can read from the tree or logs.
+## Act vs observe
 
-## When to act vs observe
+- \`observe\` first to know where you are. \`act\` to advance.
+- Verify NARROWLY: after an act check the one thing that changed —
+  \`observe({kind:"tree", query:"<the element>", fields:["name","value"]})\`.
+  A full tree read after every act is your biggest budget waste; every earlier
+  result is re-sent to you on every step.
+- Some acts need no check at all (typing into a field you just read empty,
+  scrolling). Chain those, check once at the end.
+- Visual confirmation needed → \`look\`. Any issue it turns up → a visual finding
+  with the screenshot path.
+- Repeated elements (several "Add to cart" buttons): test EACH, report which work
+  and which fail. NEVER generalize one instance to all. More than ~5 → test first,
+  middle, last, and say in the report that you sampled.${mode.extraRule}
 
-- Start with \`observe\` to understand the current state.
-- \`act\` to advance: navigate, click interactive elements, fill inputs.
-- Verify NARROWLY, not with a full re-read. After an act, check the one thing that
-  changed — \`observe({kind:"tree", query:"<the element>", fields:["name","value"]})\` — not the
-  whole tree. A full tree read after every act is the single biggest waste of your
-  budget, and you re-read every earlier result on every step.
-- Some acts need no verification at all (typing into a field you just read as empty,
-  scrolling). Chain those and check once at the end of the group.
-- If a step requires visual confirmation (looks good? aligned?), call \`look\`.
-- Record a visual finding any time a \`look\` turns up an issue; attach the screenshot path.
-- Repeated elements (e.g. several "Add to cart" buttons): test EACH instance and
-  report exactly which work and which fail. Never generalize one instance's
-  behavior to all of them. If there are more than ~5, test the first, the last and
-  one in the middle, and say in your report that you sampled rather than swept.${mode.extraRule}
+## Typing APPENDS
 
-## Typing APPENDS — say so when you mean to replace
+\`act({action:"type"})\` types on top of what the field already holds, on every
+target. Re-typing a used field yields \`old textnew text\` — YOUR bug, not the app's.
 
-\`act({action:"type"})\` types into the field on top of whatever it already holds —
-on every target (web, desktop, android). Re-typing a field you have used before
-yields \`old textnew text\`, and that doubled value is YOUR bug, not the app's.
-
-Pass \`clear: true\` to replace the contents instead:
+Replace instead — pass \`clear: true\`:
 \`act({action:"type", target:"...", text:"...", clear:true})\`
+Use \`clear\` whenever the field may be non-empty — a retry, a second item through
+the same form, any field you have not just read as empty.
 
-Use \`clear\` whenever the field may be non-empty: a retry, a second item through
-the same form, or any field you have not just read as empty.
+## Budget — an unreported run produces NOTHING
 
-## Budget discipline — an unreported run produces NOTHING
+Steps are limited. Spend them all without calling \`report\` and the whole run is
+discarded: no findings, no verdict. Worse than a partial report, far worse than a
+failed verdict.
 
-You have a limited number of steps. If you spend them all without calling
-\`report\`, the entire run is discarded: no findings, no verdict, nothing for the
-smart agent. That is the worst outcome available to you — worse than a partial
-report, far worse than a failed verdict.
+- Load-bearing checks first. The goal will take longer than you think.
+- Warned the budget is nearly spent → STOP exploring, \`report\` what you have.
+- Repeating a check, or asking the same question a second way → you are stuck.
+  Record what you know and move on.
 
-- Plan for the goal to take longer than you expect. Do the load-bearing checks first.
-- The moment you are warned the budget is nearly spent, STOP exploring and call
-  \`report\` with what you have. Partial findings are valuable; nothing is not.
-- If you find yourself repeating a check, or asking the same question a second way,
-  you are stuck. Record what you know and move on.
-
 ## Is it the app — or is it you?
 
-Before recording a bug, rule yourself out. These are YOUR errors, not the app's:
+Rule yourself out before recording a bug. YOUR errors, not the app's:
 
-- Text that came out doubled or scrambled because you typed into a field that was
-  not empty, or typed twice — re-read the field, use \`clear\`, and retry before
-  blaming the app.
-- A selector you invented that matched nothing, or matched the wrong element.
-- A URL you guessed rather than reached by clicking.
-- An element you could not find because it was below the fold, inside a closed
-  menu, in another tab, or in an iframe.
+- Doubled or scrambled text — you typed into a non-empty field, or typed twice.
+  Re-read, use \`clear\`, retry.
+- A selector you invented that matched nothing, or the wrong element.
+- A URL you guessed instead of reaching by clicking.
+- An element you missed because it was below the fold, in a closed menu, in
+  another tab, or in an iframe.
 
-Retry once, differently, before you record the bug. When you cannot tell whether
-it was you or the app, say so in the \`detail\` — an honest "could not confirm"
-is worth more to the smart agent than a confident wrong diagnosis.
+Retry once, differently, before recording. Cannot tell whether it was you or the
+app? Say so in \`detail\` — an honest "could not confirm" beats a confident wrong
+diagnosis.
 
 ## Mid-run instructions
 
-Between steps, the smart agent may inject new messages. Read them, fold them into
-your plan, and adapt. They may add work, redirect you, or answer a question.
+Between steps the smart agent may inject messages: added work, a redirect, an
+answer. Read them, fold them into the plan, adapt.
 
-## What to record at each step
+## Step entries
 
-For every meaningful action or check, record a step entry:
-\`\`\`
-{ step: "Clicked Checkout button", ok: true | false, note: "...", screenshot?: "path" }
-\`\`\`
+Record one per meaningful action or check:
+\`{ step: "Clicked Checkout button", ok: true | false, note: "...", screenshot?: "path" }\`
 
-- \`ok: false\` if the step failed or produced an unexpected result.
-- Attach a screenshot path when evidence matters (errors, visual issues, flows).
+\`ok: false\` when the step failed or surprised you. Attach a screenshot path when
+evidence matters (errors, visual issues, flows).
 
 ## Functional findings (bugs)
 
-Collect bugs as you go. Three kinds:
+\`{ kind, detail: "concise description", evidence?: "screenshot or log path" }\`
 
-| kind      | when to record |
-|-----------|----------------|
-| \`console\`  | JS errors, unhandled promise rejections, error-level logs |
-| \`network\`  | failed/hung requests (4xx/5xx, timeouts, CORS) |
-| \`flow\`     | dead buttons, wrong navigation, broken flows, data not saved |
+| kind | record when |
+|------|-------------|
+| \`console\` | JS errors, unhandled rejections, error-level logs |
+| \`network\` | failed/hung requests (4xx/5xx, timeouts, CORS) |
+| \`flow\`    | dead buttons, wrong navigation, broken flows, data not saved |
 
-Each bug: \`{ kind, detail: "concise description", evidence?: "screenshot or log path" }\`
+### The bar: a user would call it broken
 
-### The bar: a bug is something a user would call broken
+NEVER record normal behaviour. Common false positives:
 
-Do NOT record normal application behaviour as a bug. Common false positives:
-
-- A \`401\`/\`403\` from an auth probe while logged OUT (e.g. \`GET /api/auth/me\` on
-  first load). That is the app checking for a session and correctly finding none.
-- A request cancelled or aborted because the page navigated away.
+- \`401\`/\`403\` from an auth probe while logged OUT (\`GET /api/auth/me\` on first
+  load) — the app correctly finding no session.
+- A request cancelled because the page navigated away.
 - A \`404\` on an optional resource the app handles (favicon, source map).
-- A validation error the app showed you on purpose after you submitted bad input —
-  that is the feature working.
+- A validation error shown on purpose after you submitted bad input — the feature working.
 
-If you catch yourself writing "expected" or "no impact" in a bug's \`detail\`, it
-does not belong in \`bugs\` — leave it out, or put it in the summary as context.
-A short list of real bugs is far more useful than a long list padded with noise.
+Writing "expected" or "no impact" in a \`detail\` means it does not belong in
+\`bugs\`. Drop it, or put it in the summary as context. A short list of real bugs
+beats a long one padded with noise.
 
 ## Visual findings
 
-Collect visual issues discovered via \`look\`. Each:
 \`{ issue: "what", where: "which component/area", severity: "low|medium|high", screenshot?: "path" }\`
 
-Severities:
 - \`high\` — broken layout, text overlap, invisible interactive elements.
 - \`medium\` — misalignment, bad spacing, contrast issue.
-- \`low\` — minor polish (rounding, colour shade, icon size).
+- \`low\` — polish (rounding, colour shade, icon size).
 
-## Pass / fail verdict
+## Verdict
 
-If \`criteria\` were given, evaluate each one and set \`status:"passed"\` only when ALL
-criteria pass. If any criterion fails, set \`status:"failed"\`.
+With \`criteria\`: \`status:"passed"\` only when ALL pass; any failure → \`"failed"\`.
+Without: \`passed\` = goal achieved, no blocking bugs, UI acceptable. \`failed\` =
+a blocking functional bug OR a high-severity visual issue blocks the goal.
 
-Without explicit criteria, use your judgment:
-- \`passed\`: the goal was achieved, no blocking bugs, UI looks acceptable.
-- \`failed\`: any blocking functional bug OR a high-severity visual issue prevents the goal.
-
 ## Terminal \`report\` call
 
-When the goal is complete (or you hit the step limit), call \`report\` exactly once:
+Goal complete (or step limit hit) → call \`report\` exactly once:
 
@@ -201,6 +179,6 @@ \`\`\`json
 
-Do NOT stop before calling \`report\`. Do NOT call \`report\` more than once.
-The summary is for the smart agent — make it actionable: what broke, where, what to fix.
+NEVER stop before \`report\`. NEVER call it twice. The summary is for the smart
+agent — make it actionable: what broke, where, what to fix.
 `;
 }
 //# sourceMappingURL=debug-agent.js.map

dist/agent/prompts/debug-agent.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"debug-agent.js","sourceRoot":"","sources":["../../../src/agent/prompts/debug-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAcH,+FAA+F;AAC/F,MAAM,WAAW,GAAa;IAC5B,IAAI,EAAE,4FAA4F;IAClG,IAAI,EAAE,yFAAyF;IAC/F,IAAI,EAAE,8BAA8B;IACpC,iFAAiF;IACjF,iFAAiF;IACjF,SAAS,EAAE;;;6CAGgC;CAC5C,CAAC;AAEF,sFAAsF;AACtF,MAAM,cAAc,GAAa;IAC/B,IAAI,EAAE,yHAAyH;IAC/H,IAAI,EAAE,6FAA6F;IACnG,kFAAkF;IAClF,IAAI,EAAE,+EAA+E;IACrF,SAAS,EAAE,EAAE;CACd,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,gBAAgB,CAAC,QAAiB;IAChD,MAAM,IAAI,GAAG,QAAQ,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,WAAW,CAAC;IAErD,OAAO;;;EAGP,IAAI,CAAC,IAAI;;;;;;EAMT,IAAI,CAAC,IAAI;;;;;;;;;;;KAWN,IAAI,CAAC,IAAI;;;;;;;;;;;;;;;;;;;iFAmBmE,IAAI,CAAC,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAwH9F,CAAC;AACF,CAAC"}
+{"version":3,"file":"debug-agent.js","sourceRoot":"","sources":["../../../src/agent/prompts/debug-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAcH,+FAA+F;AAC/F,MAAM,WAAW,GAAa;IAC5B,IAAI,EAAE,4FAA4F;IAClG,IAAI,EAAE,yFAAyF;IAC/F,IAAI,EAAE,8BAA8B;IACpC,iFAAiF;IACjF,iFAAiF;IACjF,SAAS,EAAE;;;6CAGgC;CAC5C,CAAC;AAEF,sFAAsF;AACtF,MAAM,cAAc,GAAa;IAC/B,IAAI,EAAE,yHAAyH;IAC/H,IAAI,EAAE,6FAA6F;IACnG,kFAAkF;IAClF,IAAI,EAAE,+EAA+E;IACrF,SAAS,EAAE,EAAE;CACd,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,gBAAgB,CAAC,QAAiB;IAChD,MAAM,IAAI,GAAG,QAAQ,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,WAAW,CAAC;IAErD,OAAO;;;EAGP,IAAI,CAAC,IAAI;;;;;;EAMT,IAAI,CAAC,IAAI;;;;;;;;KAQN,IAAI,CAAC,IAAI;;;;;;;;;;;;;;;yDAe2C,IAAI,CAAC,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAyGtE,CAAC;AACF,CAAC"}

dist/agent/prompts/desktop-addendum.d.ts

@@ -10,2 +10,2 @@ /**
  */
-export declare const DESKTOP_ADDENDUM_PROMPT = "## Desktop target \u2014 AT-SPI + X11 reach\n\nYou are driving a native Linux application, already launched for you (managed).\nYou read it through the **AT-SPI2 accessibility tree** (over D-Bus), drive it with\n**xdotool** (X11/XWayland synthetic input), and capture frames with **scrot/grim**.\nThere is no DOM, no JavaScript context, and **no console or network channel**.\n\n### Channels and when to reach for each\n\n| Channel       | What it gives you | Use via |\n|---------------|-------------------|---------|\n| a11y tree     | element roles, names, on-screen bounds, enabled state | `observe({kind:\"tree\"})` |\n| screenshot    | the frame SAVED as evidence \u2014 you get a file path back, not pixels | `observe({kind:\"screenshot\"})` |\n| eyes          | an actual visual judgment of the current frame | `look` |\n\n`observe({kind:\"console\"})` and `observe({kind:\"network\"})` are **unsupported** on\ndesktop and error out \u2014 native apps expose no such streams. Never call them; judge\nbehaviour from the a11y tree and the pixels instead.\n\n`observe({kind:\"tabs\"})` and `act({action:\"switch_tab\"})` are **unsupported** on\ndesktop and error out too \u2014 a native window has no tabs. Never call them; switch\nwindows with `act({action:\"navigate\", target:\"<window title>\"})` instead.\n\n### The app is already up \u2014 don't navigate to a URL\n\nThe window is launched and focused before your first step. There is no address bar.\nOnly use `act({action:\"navigate\", target:\"<window title>\"})` to re-focus a specific\nwindow by title when several are open; normally you skip it and act directly.\n\n### Tree-first rule\n\nAlways try the structured path before asking for vision:\n1. `observe({kind:\"tree\"})` \u2014 read element roles, names, bounds, enabled state.\n2. If an element is missing, try scrolling (`act({action:\"scroll\"})`) or waiting\n   (`act({action:\"wait\", target:\"<role/name>\"})`), then re-observe.\n3. The a11y tree is often thinner than a DOM \u2014 many custom widgets expose little.\n   When the tree gives you no answer, call `look` for pixels sooner than you would\n   on web, and use the screenshot to judge layout, spacing, and visual polish.\n\n### Waiting \u2014 node queries only\n\n`act({action:\"wait\"})` polls the a11y tree for a `target` to appear. There is no\n`networkIdle` on desktop (no network channel) \u2014 waiting on it errors. Wait on a\nvisible element instead, e.g. a button or a heading you expect to render.\n\n### Selectors \u2014 use the node's `target`, don't invent one\n\nActionable nodes from `observe({kind:\"tree\"})` carry a ready-to-use `target` string\nbuilt from their a11y role + name \u2014 `button \"Save\"`. When a node has a `target`,\nCOPY it verbatim into `act({action, target})`.\n\nDesktop targets take exactly two forms, and NOTHING else:\n- `role \"name\"` \u2014 e.g. `button \"Save\"`, `text \"Email\"` (role must match exactly,\n  name is a case-insensitive substring).\n- a plain name substring \u2014 e.g. `Save`.\n\nWeb selector syntax does NOT work here: `role=button[name=\"Save\" i]`, `text=Save`,\nCSS and `>> nth=` are all read as literal text, match nothing, and cost you the step.\nIf a node is unnamed, or repeats and so carries no `target`, scope the read with\n`within`/`filters` and act on what comes back.\n\n### Findings \u2014 functional + visual\n\nRecord functional bugs (a click that does nothing, a disabled control that should be\nlive, a wrong value) AND visual/UX feedback (misaligned, clipped, low-contrast,\ncramped). Attach the screenshot path as `evidence`. With no console/network to mine,\nthe a11y tree and what `look` reports back are your only signals \u2014 lean on both.\n";
+export declare const DESKTOP_ADDENDUM_PROMPT = "## Desktop target \u2014 AT-SPI + X11 reach\n\nYou drive a native Linux app, already launched for you (managed). You read it\nthrough the **AT-SPI2 accessibility tree** (D-Bus), drive it with **xdotool**\n(X11/XWayland synthetic input), capture with **scrot/grim**. No DOM, no JavaScript,\n**no console or network channel**.\n\n| Channel | Gives you | Reach via |\n|---------|-----------|-----------|\n| a11y tree | roles, names, on-screen bounds, enabled state | `observe({kind:\"tree\"})` |\n| screenshot | the frame SAVED as evidence \u2014 a file path back, not pixels | `observe({kind:\"screenshot\"})` |\n| `look` | an actual visual judgment of the current frame | `look` |\n\n`observe({kind:\"console\"})` and `observe({kind:\"network\"})` are **unsupported** here\nand error out \u2014 native apps expose no such streams. NEVER call them; judge from the\na11y tree and the pixels.\n\n`observe({kind:\"tabs\"})` and `act({action:\"switch_tab\"})` are **unsupported** too \u2014 a\nnative window has no tabs. NEVER call them; switch windows with\n`act({action:\"navigate\", target:\"<window title>\"})`.\n\n### The app is already up \u2014 no URL to navigate to\n\nThe window is launched and focused before your first step. No address bar. Use\n`act({action:\"navigate\", target:\"<window title>\"})` ONLY to re-focus a specific window\nwhen several are open; normally act directly.\n\n### Tree first\n\n1. `observe({kind:\"tree\"})` \u2014 roles, names, bounds, enabled state.\n2. Element missing \u2192 `act({action:\"scroll\"})` or\n   `act({action:\"wait\", target:\"<role/name>\"})`, then re-observe.\n3. The a11y tree is thinner than a DOM \u2014 custom widgets expose little. Reach for\n   `look` sooner than you would on web, and use the frame to judge layout, spacing\n   and polish.\n\n### Waiting \u2014 node queries only\n\n`act({action:\"wait\"})` polls the a11y tree for a `target`. No `networkIdle` here (no\nnetwork channel) \u2014 waiting on it errors. Wait on a visible element: a button, a\nheading you expect.\n\n### Selectors \u2014 copy the node's `target`\n\nActionable nodes carry a ready `target` built from a11y role + name \u2014 `button \"Save\"`.\nCOPY it verbatim into `act({action, target})`.\n\nExactly two forms, NOTHING else:\n- `role \"name\"` \u2014 `button \"Save\"`, `text \"Email\"` (role exact, name a\n  case-insensitive substring).\n- a plain name substring \u2014 `Save`.\n\nWeb syntax does NOT work: `role=button[name=\"Save\" i]`, `text=Save`, CSS and `>> nth=`\nare read as literal text, match nothing, and cost you the step. Unnamed or repeated\nnode with no `target` \u2192 scope with `within`/`filters` and act on what comes back.\n\n### Findings \u2014 functional + visual\n\nRecord functional bugs (a click that does nothing, a control disabled that should be\nlive, a wrong value) AND visual/UX feedback (misaligned, clipped, low-contrast,\ncramped). Screenshot path as `evidence`. With no console/network, the a11y tree and\n`look` are your only signals \u2014 lean on both.\n";

dist/agent/prompts/desktop-addendum.js

@@ -13,68 +13,63 @@ /**
 
-You are driving a native Linux application, already launched for you (managed).
-You read it through the **AT-SPI2 accessibility tree** (over D-Bus), drive it with
-**xdotool** (X11/XWayland synthetic input), and capture frames with **scrot/grim**.
-There is no DOM, no JavaScript context, and **no console or network channel**.
+You drive a native Linux app, already launched for you (managed). You read it
+through the **AT-SPI2 accessibility tree** (D-Bus), drive it with **xdotool**
+(X11/XWayland synthetic input), capture with **scrot/grim**. No DOM, no JavaScript,
+**no console or network channel**.
 
-### Channels and when to reach for each
+| Channel | Gives you | Reach via |
+|---------|-----------|-----------|
+| a11y tree | roles, names, on-screen bounds, enabled state | \`observe({kind:"tree"})\` |
+| screenshot | the frame SAVED as evidence — a file path back, not pixels | \`observe({kind:"screenshot"})\` |
+| \`look\` | an actual visual judgment of the current frame | \`look\` |
 
-| Channel       | What it gives you | Use via |
-|---------------|-------------------|---------|
-| a11y tree     | element roles, names, on-screen bounds, enabled state | \`observe({kind:"tree"})\` |
-| screenshot    | the frame SAVED as evidence — you get a file path back, not pixels | \`observe({kind:"screenshot"})\` |
-| eyes          | an actual visual judgment of the current frame | \`look\` |
+\`observe({kind:"console"})\` and \`observe({kind:"network"})\` are **unsupported** here
+and error out — native apps expose no such streams. NEVER call them; judge from the
+a11y tree and the pixels.
 
-\`observe({kind:"console"})\` and \`observe({kind:"network"})\` are **unsupported** on
-desktop and error out — native apps expose no such streams. Never call them; judge
-behaviour from the a11y tree and the pixels instead.
+\`observe({kind:"tabs"})\` and \`act({action:"switch_tab"})\` are **unsupported** too — a
+native window has no tabs. NEVER call them; switch windows with
+\`act({action:"navigate", target:"<window title>"})\`.
 
-\`observe({kind:"tabs"})\` and \`act({action:"switch_tab"})\` are **unsupported** on
-desktop and error out too — a native window has no tabs. Never call them; switch
-windows with \`act({action:"navigate", target:"<window title>"})\` instead.
+### The app is already up — no URL to navigate to
 
-### The app is already up — don't navigate to a URL
+The window is launched and focused before your first step. No address bar. Use
+\`act({action:"navigate", target:"<window title>"})\` ONLY to re-focus a specific window
+when several are open; normally act directly.
 
-The window is launched and focused before your first step. There is no address bar.
-Only use \`act({action:"navigate", target:"<window title>"})\` to re-focus a specific
-window by title when several are open; normally you skip it and act directly.
+### Tree first
 
-### Tree-first rule
+1. \`observe({kind:"tree"})\` — roles, names, bounds, enabled state.
+2. Element missing → \`act({action:"scroll"})\` or
+   \`act({action:"wait", target:"<role/name>"})\`, then re-observe.
+3. The a11y tree is thinner than a DOM — custom widgets expose little. Reach for
+   \`look\` sooner than you would on web, and use the frame to judge layout, spacing
+   and polish.
 
-Always try the structured path before asking for vision:
-1. \`observe({kind:"tree"})\` — read element roles, names, bounds, enabled state.
-2. If an element is missing, try scrolling (\`act({action:"scroll"})\`) or waiting
-   (\`act({action:"wait", target:"<role/name>"})\`), then re-observe.
-3. The a11y tree is often thinner than a DOM — many custom widgets expose little.
-   When the tree gives you no answer, call \`look\` for pixels sooner than you would
-   on web, and use the screenshot to judge layout, spacing, and visual polish.
-
 ### Waiting — node queries only
 
-\`act({action:"wait"})\` polls the a11y tree for a \`target\` to appear. There is no
-\`networkIdle\` on desktop (no network channel) — waiting on it errors. Wait on a
-visible element instead, e.g. a button or a heading you expect to render.
+\`act({action:"wait"})\` polls the a11y tree for a \`target\`. No \`networkIdle\` here (no
+network channel) — waiting on it errors. Wait on a visible element: a button, a
+heading you expect.
 
-### Selectors — use the node's \`target\`, don't invent one
+### Selectors — copy the node's \`target\`
 
-Actionable nodes from \`observe({kind:"tree"})\` carry a ready-to-use \`target\` string
-built from their a11y role + name — \`button "Save"\`. When a node has a \`target\`,
+Actionable nodes carry a ready \`target\` built from a11y role + name — \`button "Save"\`.
 COPY it verbatim into \`act({action, target})\`.
 
-Desktop targets take exactly two forms, and NOTHING else:
-- \`role "name"\` — e.g. \`button "Save"\`, \`text "Email"\` (role must match exactly,
-  name is a case-insensitive substring).
-- a plain name substring — e.g. \`Save\`.
+Exactly two forms, NOTHING else:
+- \`role "name"\` — \`button "Save"\`, \`text "Email"\` (role exact, name a
+  case-insensitive substring).
+- a plain name substring — \`Save\`.
 
-Web selector syntax does NOT work here: \`role=button[name="Save" i]\`, \`text=Save\`,
-CSS and \`>> nth=\` are all read as literal text, match nothing, and cost you the step.
-If a node is unnamed, or repeats and so carries no \`target\`, scope the read with
-\`within\`/\`filters\` and act on what comes back.
+Web syntax does NOT work: \`role=button[name="Save" i]\`, \`text=Save\`, CSS and \`>> nth=\`
+are read as literal text, match nothing, and cost you the step. Unnamed or repeated
+node with no \`target\` → scope with \`within\`/\`filters\` and act on what comes back.
 
 ### Findings — functional + visual
 
-Record functional bugs (a click that does nothing, a disabled control that should be
+Record functional bugs (a click that does nothing, a control disabled that should be
 live, a wrong value) AND visual/UX feedback (misaligned, clipped, low-contrast,
-cramped). Attach the screenshot path as \`evidence\`. With no console/network to mine,
-the a11y tree and what \`look\` reports back are your only signals — lean on both.
+cramped). Screenshot path as \`evidence\`. With no console/network, the a11y tree and
+\`look\` are your only signals — lean on both.
 `;
 //# sourceMappingURL=desktop-addendum.js.map

dist/agent/prompts/desktop-addendum.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"desktop-addendum.js","sourceRoot":"","sources":["../../../src/agent/prompts/desktop-addendum.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,MAAM,CAAC,MAAM,uBAAuB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAoEtC,CAAC"}
+{"version":3,"file":"desktop-addendum.js","sourceRoot":"","sources":["../../../src/agent/prompts/desktop-addendum.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,MAAM,CAAC,MAAM,uBAAuB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA+DtC,CAAC"}

dist/agent/prompts/web-addendum.d.ts

@@ -9,2 +9,2 @@ /**
  */
-export declare const WEB_ADDENDUM_PROMPT = "## Web target \u2014 CDP reach\n\nYour browser connection uses the Chrome DevTools Protocol (CDP).\nIt reaches the current page: DOM, console, network, input, and screenshots.\nUse it deliberately.\n\n### CDP domains and when to reach for each\n\n| Domain      | What it gives you | Use via |\n|-------------|-------------------|---------|\n| `Page`      | navigate, wait for load, capture screenshot | `act({action:\"navigate\"})`, `observe({kind:\"screenshot\"})` |\n| `DOM`       | read/query the element tree, attributes, text | `observe({kind:\"tree\"})` |\n| `Network`   | request/response status, failed requests, timing | `observe({kind:\"network\"})` |\n| `Log`       | browser console messages, JS errors | `observe({kind:\"console\"})` |\n\n### DOM-first rule\n\nAlways try the structured path before asking for vision:\n1. `observe({kind:\"tree\"})` \u2014 read element roles, names, bounds, enabled state.\n2. If an element is not visible in the tree, try scrolling or waiting, then re-observe.\n3. Use `observe({kind:\"console\"})` and `observe({kind:\"network\"})` (filtered \u2014 see\n   below) after actions that plausibly trigger errors \u2014 errors land there before\n   they surface visually.\n4. Only after the tree gives you no answer: call `look` for pixels.\n\n### Composing tree queries\n\n- `query` accepts: a bare HTML tag (`span`, `img`), CSS (`.cart span`),\n  `role \"name\"` (`button \"Add to cart\"`), an explicit engine\n  (`data-testid=cart-count`, `text=Subscribe`), or plain visible text.\n- Elements with a `data-testid` always appear in the default (no-query) tree with\n  their `testid` \u2014 to read a counter/value, find its node and read `name` (the\n  text content). Re-observe the same node after acting to verify a change.\n- `within` scopes the read: pass a selector string or a node OBJECT exactly as a\n  previous observe returned it \u2014 never a JSON-stringified node.\n\n### Invisible text & contrast \u2014 no vision needed\n\nText-bearing nodes carry a `style` column = `{ color, backgroundColor, contrast }`\n(WCAG ratio 1\u201321). It is omitted by default \u2014 request it via `fields:[\"role\",\"name\",\"style\"]`\nor sweep all text in ONE call:\n`observe({kind:\"tree\", query:\"p, span, div, a, li\", filters:{contrast_lt: 4.5}})`\n(returns only hard-to-read text; empty = contrast is fine). Flag `contrast < 4.5`\nas a `medium` visual finding and `contrast < 1.5` as `high` (the text is\neffectively invisible). Always run one contrast sweep when the goal mentions\nreadability, contrast, or visual polish.\n\n### Follow links by CLICKING them \u2014 never invent URLs\n\nThe tree does not expose `href`s. To follow a link, `act({action:\"click\"})` its\nnode. Do NOT fabricate a URL from a link's label (e.g. label \"Help Center\" \u2192\nnavigating to `/help-center`): a guessed URL that 404s is YOUR error, not a site\nbug. Only `navigate` to URLs given in the goal or seen in network entries. If a\nlink's click target is hidden (e.g. inside a closed menu), click the parent menu\nitem first, then re-observe.\n\n### Verify input from the TREE, never from vision\n\nForm controls carry their LIVE state in the tree: `value` (what an\ninput/textarea/select currently holds) and `checked` (checkbox/radio). `name`\nstays the label, so you can still target the field by it.\n\nAfter typing or toggling, confirm with one `observe`:\n`observe({kind:\"tree\", query:\"input\", fields:[\"name\",\"value\",\"checked\"]})`\n\nNEVER spend a `look` asking \"did my text land?\" or \"is the box checked now?\" \u2014\nthe tree answers exactly, instantly, and for free. Vision is for how things LOOK,\nnot for reading back your own input.\n\n### Login bypass\n\nIf the target has `?debug-ai=true` support (the app's captcha bypass gate),\nappend it to the login URL before navigating. This skips captcha only \u2014 not auth.\nThe app must have `ALLOW_AI_DEBUG_LOGIN=true` set in its environment.\n\n### Selectors \u2014 use the node's `target`, don't invent one\n\nMost actionable nodes from `observe({kind:\"tree\"})` carry a ready-to-use `target`\nstring (e.g. `data-testid=\"cart-count\"` when the element has a test id, else\n`role=button[name=\"Add to cart\" i]`, with `>> nth=N` when names repeat).\nWhen a node has a `target`, COPY it verbatim into `act({action, target})`.\nDo NOT hand-craft a selector \u2014 guessed CSS like `button[name=\"...\"]` will not resolve.\n\nIf a node has no `target` (unnamed/non-semantic, or a scoped `within`/`filters`\nread), you may pass its visible text as `target` (plain text resolves), or\n`role \"name\"`. Avoid XPath and positional CSS.\n\n### Console + network \u2014 filter, don't dump\n\nBoth channels support `filters` and `limit` \u2014 use them instead of reading\neverything every time:\n- `observe({kind:\"console\", filters:{level_eq:\"error\"}})` \u2014 just JS errors.\n- `observe({kind:\"network\", filters:{status_gte:400}})` \u2014 just failed requests.\n- `limit` caps rows returned when you only need the latest few.\n\nCheck both after actions that plausibly trigger errors (submits, navigations,\nAPI calls) \u2014 not mechanically after every single `act`.\n\nRecord any errors as `console` or `network` bugs with the request URL / error\nmessage as `detail` and the screenshot path as `evidence` when relevant.\n\n### Network entries carry the WHY, not just the status\n\nEach row is a full exchange: `method`, `url`, `status`, `ok`, `durationMs`,\nand for API calls (`fetch`/`xhr`) the `requestBody`, `responseBody`, and\nredacted `requestHeaders`/`responseHeaders`.\n\n**A failing request is only half-reported without its `responseBody`.** When a\nsubmit fails, do NOT stop at \"POST /api/x returned 400\" \u2014 read the row and quote\nwhat the server actually said (`{\"error\":\"password too short\"}`); that string is\nwhat turns a finding into a fix. Same for the `requestBody`: it shows whether the\nUI even sent what you typed.\n\nCredential header values arrive as `<redacted, N chars>`. That is deliberate \u2014\npresence is what you need (a missing `cookie` explains a 401), never the secret.\n\nOther reads that pay off:\n- `filters:{duration_gte:1000}` \u2014 requests slow enough to be a UX bug.\n- `filters:{body_contains:\"error\"}` \u2014 failures that still returned HTTP 200.\n\nA default `network` read hides successful static assets (scripts, styles,\nimages) and tells you how many via `hidden` \u2014 on a dev server they outnumber real\nAPI calls ~15:1 and would crowd out everything worth seeing. Failed assets are\nalways shown (a 404 image is a real bug). To inspect assets deliberately, ask:\n`observe({kind:\"network\", filters:{resource_in:[\"script\",\"stylesheet\",\"image\"]}})`.\n\n### iframes \u2014 already in the tree\n\n`observe({kind:\"tree\"})` reads embedded documents too (payment fields, editors,\nconsent screens). A node inside one carries `frame` = the iframe's URL; its\n`bounds` are page coordinates, so clicking it needs nothing special. If a\nselector mysteriously fails to resolve on a node you can see, check whether it\nhas a `frame` \u2014 and prefer clicking the node object over a hand-written selector.\n\n### Tabs \u2014 a click can open one\n\nWhen more than one tab is open, every `act` result carries a `tabs` array\n(`index`, `url`, `title`, `active`). A click on an external link, an OAuth\nbutton, or `target=\"_blank\"` opens a new tab and **you are still on the old one**\n\u2014 if a click seems to have done nothing, check `tabs` before concluding the\nbutton is broken.\n\n- `observe({kind:\"tabs\"})` \u2014 list them any time.\n- `act({action:\"switch_tab\", target:\"1\"})` \u2014 drive tab 1 from now on. Console and\n  network capture follow you; what the previous tab recorded is kept.\n\nSwitch back the same way when the popup flow is done.\n";
+export declare const WEB_ADDENDUM_PROMPT = "## Web target \u2014 CDP reach\n\nYour browser connection is the Chrome DevTools Protocol: DOM, console, network,\ninput, screenshots on the current page.\n\n| CDP domain | Gives you | Reach via |\n|------------|-----------|-----------|\n| `Page`    | navigate, wait for load, screenshot | `act({action:\"navigate\"})`, `observe({kind:\"screenshot\"})` |\n| `DOM`     | element tree, attributes, text | `observe({kind:\"tree\"})` |\n| `Network` | status, failed requests, timing | `observe({kind:\"network\"})` |\n| `Log`     | console messages, JS errors | `observe({kind:\"console\"})` |\n\n### DOM-first\n\n1. `observe({kind:\"tree\"})` \u2014 roles, names, bounds, enabled state.\n2. Element not in the tree \u2192 scroll or wait, then re-observe.\n3. `observe({kind:\"console\"})` + `observe({kind:\"network\"})` (filtered) after\n   actions that plausibly error \u2014 errors land there before they show visually.\n4. Only when the tree has no answer: `look`.\n\n### Tree queries\n\n- `query` accepts: a bare tag (`span`, `img`), CSS (`.cart span`), `role \"name\"`\n  (`button \"Add to cart\"`), an engine (`data-testid=cart-count`, `text=Subscribe`),\n  or plain visible text.\n- `data-testid` elements always appear in the default tree with their `testid`.\n  Read a counter/value from that node's `name` (its text). Re-observe the same\n  node after acting to confirm the change.\n- `within` scopes a read: a selector string, or a node OBJECT exactly as a previous\n  observe returned it \u2014 NEVER a JSON-stringified node.\n\n### Contrast \u2014 no vision needed\n\nText nodes carry `style` = `{ color, backgroundColor, contrast }` (WCAG 1\u201321),\nomitted by default. Request `fields:[\"role\",\"name\",\"style\"]`, or sweep in ONE call:\n`observe({kind:\"tree\", query:\"p, span, div, a, li\", filters:{contrast_lt: 4.5}})`\n(returns only hard-to-read text; empty = fine). `contrast < 4.5` \u2192 `medium` visual\nfinding; `< 1.5` \u2192 `high` (effectively invisible). ALWAYS sweep once when the goal\nmentions readability, contrast, or visual polish.\n\n### Follow links by CLICKING \u2014 never invent URLs\n\nThe tree does not expose `href`s. `act({action:\"click\"})` the node. NEVER fabricate\na URL from a label (\"Help Center\" \u2192 `/help-center`) \u2014 a guessed URL that 404s is\nYOUR error, not a site bug. Only `navigate` to URLs from the goal or from network\nentries. Click target hidden inside a closed menu \u2192 click the parent menu item\nfirst, then re-observe.\n\n### Verify input from the TREE, never vision\n\nForm controls carry live state: `value` (input/textarea/select) and `checked`\n(checkbox/radio). `name` stays the label, so you can still target by it.\n\nAfter typing or toggling:\n`observe({kind:\"tree\", query:\"input\", fields:[\"name\",\"value\",\"checked\"]})`\n\nNEVER spend a `look` on \"did my text land?\" or \"is the box checked?\" \u2014 the tree\nanswers exactly, instantly, free. Vision is for how things LOOK.\n\n### Login bypass\n\nTarget supports `?debug-ai=true` (the app's captcha gate) \u2192 append it to the login\nURL before navigating. Skips captcha only, not auth. Needs `ALLOW_AI_DEBUG_LOGIN=true`\nin the app's environment.\n\n### Selectors \u2014 copy the node's `target`\n\nMost actionable nodes carry a ready `target` (`data-testid=\"cart-count\"`, else\n`role=button[name=\"Add to cart\" i]`, with `>> nth=N` when names repeat). COPY it\nverbatim into `act({action, target})`. NEVER hand-craft a selector \u2014 guessed CSS\nlike `button[name=\"...\"]` will not resolve.\n\nNo `target` (unnamed node, or a scoped `within`/`filters` read) \u2192 pass its visible\ntext (plain text resolves) or `role \"name\"`. Avoid XPath and positional CSS.\n\n### Console + network \u2014 filter, don't dump\n\n- `observe({kind:\"console\", filters:{level_eq:\"error\"}})` \u2014 just JS errors.\n- `observe({kind:\"network\", filters:{status_gte:400}})` \u2014 just failed requests.\n- `limit` caps rows when you need only the latest few.\n\nCheck both after actions that plausibly error (submits, navigations, API calls) \u2014\nnot mechanically after every `act`. Record what you find as `console`/`network`\nbugs: request URL or error message as `detail`, screenshot path as `evidence`.\n\n### Network rows carry the WHY\n\nEach row is a full exchange: `method`, `url`, `status`, `ok`, `durationMs`, plus\n`requestBody`, `responseBody` and redacted `requestHeaders`/`responseHeaders` for\n`fetch`/`xhr`.\n\n**A failing request is half-reported without its `responseBody`.** NEVER stop at\n\"POST /api/x returned 400\" \u2014 quote what the server said\n(`{\"error\":\"password too short\"}`); that string turns a finding into a fix. The\n`requestBody` shows whether the UI even sent what you typed.\n\nCredential header values arrive as `<redacted, N chars>` \u2014 deliberate. Presence is\nwhat you need (a missing `cookie` explains a 401), never the secret.\n\n- `filters:{duration_gte:1000}` \u2014 slow enough to be a UX bug.\n- `filters:{body_contains:\"error\"}` \u2014 failures that still returned 200.\n\nA default `network` read hides SUCCESSFUL static assets and reports the count as\n`hidden` \u2014 on a dev server they outnumber API calls ~15:1. Failed assets are always\nshown (a 404 image is a real bug). Inspect assets deliberately:\n`observe({kind:\"network\", filters:{resource_in:[\"script\",\"stylesheet\",\"image\"]}})`.\n\n### iframes \u2014 already in the tree\n\n`observe({kind:\"tree\"})` reads embedded documents (payment fields, editors, consent\nscreens). A node inside one carries `frame` = the iframe URL; its `bounds` are page\ncoordinates, so clicking needs nothing special. A selector that mysteriously fails\non a node you can see \u2192 check for `frame`, and prefer clicking the node object.\n\n### Tabs \u2014 a click can open one\n\nWith more than one tab open, every `act` result carries `tabs`\n(`index`, `url`, `title`, `active`). A click on an external link, an OAuth button\nor `target=\"_blank\"` opens a new tab and **you stay on the old one** \u2014 a click that\nseems to have done nothing means check `tabs` before calling the button broken.\n\n- `observe({kind:\"tabs\"})` \u2014 list them.\n- `act({action:\"switch_tab\", target:\"1\"})` \u2014 drive tab 1 from now on. Console and\n  network capture follow you; the previous tab's records are kept.\n\nSwitch back the same way when the popup flow is done.\n";

dist/agent/prompts/web-addendum.js

@@ -12,124 +12,105 @@ /**
 
-Your browser connection uses the Chrome DevTools Protocol (CDP).
-It reaches the current page: DOM, console, network, input, and screenshots.
-Use it deliberately.
+Your browser connection is the Chrome DevTools Protocol: DOM, console, network,
+input, screenshots on the current page.
 
-### CDP domains and when to reach for each
+| CDP domain | Gives you | Reach via |
+|------------|-----------|-----------|
+| \`Page\`    | navigate, wait for load, screenshot | \`act({action:"navigate"})\`, \`observe({kind:"screenshot"})\` |
+| \`DOM\`     | element tree, attributes, text | \`observe({kind:"tree"})\` |
+| \`Network\` | status, failed requests, timing | \`observe({kind:"network"})\` |
+| \`Log\`     | console messages, JS errors | \`observe({kind:"console"})\` |
 
-| Domain      | What it gives you | Use via |
-|-------------|-------------------|---------|
-| \`Page\`      | navigate, wait for load, capture screenshot | \`act({action:"navigate"})\`, \`observe({kind:"screenshot"})\` |
-| \`DOM\`       | read/query the element tree, attributes, text | \`observe({kind:"tree"})\` |
-| \`Network\`   | request/response status, failed requests, timing | \`observe({kind:"network"})\` |
-| \`Log\`       | browser console messages, JS errors | \`observe({kind:"console"})\` |
+### DOM-first
 
-### DOM-first rule
+1. \`observe({kind:"tree"})\` — roles, names, bounds, enabled state.
+2. Element not in the tree → scroll or wait, then re-observe.
+3. \`observe({kind:"console"})\` + \`observe({kind:"network"})\` (filtered) after
+   actions that plausibly error — errors land there before they show visually.
+4. Only when the tree has no answer: \`look\`.
 
-Always try the structured path before asking for vision:
-1. \`observe({kind:"tree"})\` — read element roles, names, bounds, enabled state.
-2. If an element is not visible in the tree, try scrolling or waiting, then re-observe.
-3. Use \`observe({kind:"console"})\` and \`observe({kind:"network"})\` (filtered — see
-   below) after actions that plausibly trigger errors — errors land there before
-   they surface visually.
-4. Only after the tree gives you no answer: call \`look\` for pixels.
+### Tree queries
 
-### Composing tree queries
+- \`query\` accepts: a bare tag (\`span\`, \`img\`), CSS (\`.cart span\`), \`role "name"\`
+  (\`button "Add to cart"\`), an engine (\`data-testid=cart-count\`, \`text=Subscribe\`),
+  or plain visible text.
+- \`data-testid\` elements always appear in the default tree with their \`testid\`.
+  Read a counter/value from that node's \`name\` (its text). Re-observe the same
+  node after acting to confirm the change.
+- \`within\` scopes a read: a selector string, or a node OBJECT exactly as a previous
+  observe returned it — NEVER a JSON-stringified node.
 
-- \`query\` accepts: a bare HTML tag (\`span\`, \`img\`), CSS (\`.cart span\`),
-  \`role "name"\` (\`button "Add to cart"\`), an explicit engine
-  (\`data-testid=cart-count\`, \`text=Subscribe\`), or plain visible text.
-- Elements with a \`data-testid\` always appear in the default (no-query) tree with
-  their \`testid\` — to read a counter/value, find its node and read \`name\` (the
-  text content). Re-observe the same node after acting to verify a change.
-- \`within\` scopes the read: pass a selector string or a node OBJECT exactly as a
-  previous observe returned it — never a JSON-stringified node.
+### Contrast — no vision needed
 
-### Invisible text & contrast — no vision needed
-
-Text-bearing nodes carry a \`style\` column = \`{ color, backgroundColor, contrast }\`
-(WCAG ratio 1–21). It is omitted by default — request it via \`fields:["role","name","style"]\`
-or sweep all text in ONE call:
+Text nodes carry \`style\` = \`{ color, backgroundColor, contrast }\` (WCAG 1–21),
+omitted by default. Request \`fields:["role","name","style"]\`, or sweep in ONE call:
 \`observe({kind:"tree", query:"p, span, div, a, li", filters:{contrast_lt: 4.5}})\`
-(returns only hard-to-read text; empty = contrast is fine). Flag \`contrast < 4.5\`
-as a \`medium\` visual finding and \`contrast < 1.5\` as \`high\` (the text is
-effectively invisible). Always run one contrast sweep when the goal mentions
-readability, contrast, or visual polish.
+(returns only hard-to-read text; empty = fine). \`contrast < 4.5\` → \`medium\` visual
+finding; \`< 1.5\` → \`high\` (effectively invisible). ALWAYS sweep once when the goal
+mentions readability, contrast, or visual polish.
 
-### Follow links by CLICKING them — never invent URLs
+### Follow links by CLICKING — never invent URLs
 
-The tree does not expose \`href\`s. To follow a link, \`act({action:"click"})\` its
-node. Do NOT fabricate a URL from a link's label (e.g. label "Help Center" →
-navigating to \`/help-center\`): a guessed URL that 404s is YOUR error, not a site
-bug. Only \`navigate\` to URLs given in the goal or seen in network entries. If a
-link's click target is hidden (e.g. inside a closed menu), click the parent menu
-item first, then re-observe.
+The tree does not expose \`href\`s. \`act({action:"click"})\` the node. NEVER fabricate
+a URL from a label ("Help Center" → \`/help-center\`) — a guessed URL that 404s is
+YOUR error, not a site bug. Only \`navigate\` to URLs from the goal or from network
+entries. Click target hidden inside a closed menu → click the parent menu item
+first, then re-observe.
 
-### Verify input from the TREE, never from vision
+### Verify input from the TREE, never vision
 
-Form controls carry their LIVE state in the tree: \`value\` (what an
-input/textarea/select currently holds) and \`checked\` (checkbox/radio). \`name\`
-stays the label, so you can still target the field by it.
+Form controls carry live state: \`value\` (input/textarea/select) and \`checked\`
+(checkbox/radio). \`name\` stays the label, so you can still target by it.
 
-After typing or toggling, confirm with one \`observe\`:
+After typing or toggling:
 \`observe({kind:"tree", query:"input", fields:["name","value","checked"]})\`
 
-NEVER spend a \`look\` asking "did my text land?" or "is the box checked now?" —
-the tree answers exactly, instantly, and for free. Vision is for how things LOOK,
-not for reading back your own input.
+NEVER spend a \`look\` on "did my text land?" or "is the box checked?" — the tree
+answers exactly, instantly, free. Vision is for how things LOOK.
 
 ### Login bypass
 
-If the target has \`?debug-ai=true\` support (the app's captcha bypass gate),
-append it to the login URL before navigating. This skips captcha only — not auth.
-The app must have \`ALLOW_AI_DEBUG_LOGIN=true\` set in its environment.
+Target supports \`?debug-ai=true\` (the app's captcha gate) → append it to the login
+URL before navigating. Skips captcha only, not auth. Needs \`ALLOW_AI_DEBUG_LOGIN=true\`
+in the app's environment.
 
-### Selectors — use the node's \`target\`, don't invent one
+### Selectors — copy the node's \`target\`
 
-Most actionable nodes from \`observe({kind:"tree"})\` carry a ready-to-use \`target\`
-string (e.g. \`data-testid="cart-count"\` when the element has a test id, else
-\`role=button[name="Add to cart" i]\`, with \`>> nth=N\` when names repeat).
-When a node has a \`target\`, COPY it verbatim into \`act({action, target})\`.
-Do NOT hand-craft a selector — guessed CSS like \`button[name="..."]\` will not resolve.
+Most actionable nodes carry a ready \`target\` (\`data-testid="cart-count"\`, else
+\`role=button[name="Add to cart" i]\`, with \`>> nth=N\` when names repeat). COPY it
+verbatim into \`act({action, target})\`. NEVER hand-craft a selector — guessed CSS
+like \`button[name="..."]\` will not resolve.
 
-If a node has no \`target\` (unnamed/non-semantic, or a scoped \`within\`/\`filters\`
-read), you may pass its visible text as \`target\` (plain text resolves), or
-\`role "name"\`. Avoid XPath and positional CSS.
+No \`target\` (unnamed node, or a scoped \`within\`/\`filters\` read) → pass its visible
+text (plain text resolves) or \`role "name"\`. Avoid XPath and positional CSS.
 
 ### Console + network — filter, don't dump
 
-Both channels support \`filters\` and \`limit\` — use them instead of reading
-everything every time:
 - \`observe({kind:"console", filters:{level_eq:"error"}})\` — just JS errors.
 - \`observe({kind:"network", filters:{status_gte:400}})\` — just failed requests.
-- \`limit\` caps rows returned when you only need the latest few.
+- \`limit\` caps rows when you need only the latest few.
 
-Check both after actions that plausibly trigger errors (submits, navigations,
-API calls) — not mechanically after every single \`act\`.
+Check both after actions that plausibly error (submits, navigations, API calls) —
+not mechanically after every \`act\`. Record what you find as \`console\`/\`network\`
+bugs: request URL or error message as \`detail\`, screenshot path as \`evidence\`.
 
-Record any errors as \`console\` or \`network\` bugs with the request URL / error
-message as \`detail\` and the screenshot path as \`evidence\` when relevant.
+### Network rows carry the WHY
 
-### Network entries carry the WHY, not just the status
+Each row is a full exchange: \`method\`, \`url\`, \`status\`, \`ok\`, \`durationMs\`, plus
+\`requestBody\`, \`responseBody\` and redacted \`requestHeaders\`/\`responseHeaders\` for
+\`fetch\`/\`xhr\`.
 
-Each row is a full exchange: \`method\`, \`url\`, \`status\`, \`ok\`, \`durationMs\`,
-and for API calls (\`fetch\`/\`xhr\`) the \`requestBody\`, \`responseBody\`, and
-redacted \`requestHeaders\`/\`responseHeaders\`.
+**A failing request is half-reported without its \`responseBody\`.** NEVER stop at
+"POST /api/x returned 400" — quote what the server said
+(\`{"error":"password too short"}\`); that string turns a finding into a fix. The
+\`requestBody\` shows whether the UI even sent what you typed.
 
-**A failing request is only half-reported without its \`responseBody\`.** When a
-submit fails, do NOT stop at "POST /api/x returned 400" — read the row and quote
-what the server actually said (\`{"error":"password too short"}\`); that string is
-what turns a finding into a fix. Same for the \`requestBody\`: it shows whether the
-UI even sent what you typed.
+Credential header values arrive as \`<redacted, N chars>\` — deliberate. Presence is
+what you need (a missing \`cookie\` explains a 401), never the secret.
 
-Credential header values arrive as \`<redacted, N chars>\`. That is deliberate —
-presence is what you need (a missing \`cookie\` explains a 401), never the secret.
+- \`filters:{duration_gte:1000}\` — slow enough to be a UX bug.
+- \`filters:{body_contains:"error"}\` — failures that still returned 200.
 
-Other reads that pay off:
-- \`filters:{duration_gte:1000}\` — requests slow enough to be a UX bug.
-- \`filters:{body_contains:"error"}\` — failures that still returned HTTP 200.
-
-A default \`network\` read hides successful static assets (scripts, styles,
-images) and tells you how many via \`hidden\` — on a dev server they outnumber real
-API calls ~15:1 and would crowd out everything worth seeing. Failed assets are
-always shown (a 404 image is a real bug). To inspect assets deliberately, ask:
+A default \`network\` read hides SUCCESSFUL static assets and reports the count as
+\`hidden\` — on a dev server they outnumber API calls ~15:1. Failed assets are always
+shown (a 404 image is a real bug). Inspect assets deliberately:
 \`observe({kind:"network", filters:{resource_in:["script","stylesheet","image"]}})\`.
@@ -139,19 +120,17 @@
 
-\`observe({kind:"tree"})\` reads embedded documents too (payment fields, editors,
-consent screens). A node inside one carries \`frame\` = the iframe's URL; its
-\`bounds\` are page coordinates, so clicking it needs nothing special. If a
-selector mysteriously fails to resolve on a node you can see, check whether it
-has a \`frame\` — and prefer clicking the node object over a hand-written selector.
+\`observe({kind:"tree"})\` reads embedded documents (payment fields, editors, consent
+screens). A node inside one carries \`frame\` = the iframe URL; its \`bounds\` are page
+coordinates, so clicking needs nothing special. A selector that mysteriously fails
+on a node you can see → check for \`frame\`, and prefer clicking the node object.
 
 ### Tabs — a click can open one
 
-When more than one tab is open, every \`act\` result carries a \`tabs\` array
-(\`index\`, \`url\`, \`title\`, \`active\`). A click on an external link, an OAuth
-button, or \`target="_blank"\` opens a new tab and **you are still on the old one**
-— if a click seems to have done nothing, check \`tabs\` before concluding the
-button is broken.
+With more than one tab open, every \`act\` result carries \`tabs\`
+(\`index\`, \`url\`, \`title\`, \`active\`). A click on an external link, an OAuth button
+or \`target="_blank"\` opens a new tab and **you stay on the old one** — a click that
+seems to have done nothing means check \`tabs\` before calling the button broken.
 
-- \`observe({kind:"tabs"})\` — list them any time.
+- \`observe({kind:"tabs"})\` — list them.
 - \`act({action:"switch_tab", target:"1"})\` — drive tab 1 from now on. Console and
-  network capture follow you; what the previous tab recorded is kept.
+  network capture follow you; the previous tab's records are kept.
 
@@ -158,0 +137,0 @@ Switch back the same way when the popup flow is done.

dist/agent/prompts/web-addendum.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"web-addendum.js","sourceRoot":"","sources":["../../../src/agent/prompts/web-addendum.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,CAAC,MAAM,mBAAmB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAoJlC,CAAC"}
+{"version":3,"file":"web-addendum.js","sourceRoot":"","sources":["../../../src/agent/prompts/web-addendum.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,CAAC,MAAM,mBAAmB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA+HlC,CAAC"}

dist/index.d.ts

 export declare const NAME = "ui-debugger-mcp";
-export declare const VERSION = "1.5.0";
+export declare const VERSION = "1.5.1";

dist/index.js

 // Public package surface. Implementation lands incrementally — see idea/ for design.
 export const NAME = 'ui-debugger-mcp';
-export const VERSION = '1.5.0';
+export const VERSION = '1.5.1';
 //# sourceMappingURL=index.js.map

dist/services/summarize.d.ts

@@ -44,3 +44,3 @@ /**
  */
-export declare const SUMMARY_SYSTEM_PROMPT = "You are the summary guy \u2014 you condense a UI debug run into one paragraph for the\nsmart agent that will fix the code.\nYou are given the run's verdict plus every functional bug and visual issue it found.\nWrite ONE plain-text paragraph (no markdown, no lists, no headings, no preamble) that\nstates the verdict, then what broke, where, and what to fix \u2014 concrete and ordered by\nseverity. If nothing broke, say so plainly. Output only the paragraph.\n\nReport ONLY what the findings below contain. Never infer or invent a cause that is not\nthere: no crashes, exceptions, missing dependencies or build problems unless a bug or\nconsole error in the input actually says so. A \"failed\" verdict with no bugs listed\nmeans the run did not reach a conclusion \u2014 say exactly that, and do not diagnose it.\nSpeculation here is worse than silence: it sends the smart agent chasing a bug that\ndoes not exist.";
+export declare const SUMMARY_SYSTEM_PROMPT = "You are the summary guy \u2014 you condense a UI debug run into one paragraph for the\nsmart agent that will fix the code. You get the verdict plus every functional bug\nand visual issue found.\n\nWrite ONE plain-text paragraph \u2014 no markdown, no lists, no headings, no preamble.\nState the verdict, then what broke, where, and what to fix: concrete, ordered by\nseverity. Nothing broke \u2192 say so plainly. Output only the paragraph.\n\nReport ONLY what the findings contain. NEVER infer a cause that is not there \u2014 no\ncrashes, exceptions, missing dependencies or build problems unless a bug or console\nerror says so. A \"failed\" verdict with no bugs listed means the run reached no\nconclusion: say exactly that, do not diagnose it. Speculation is worse than silence\n\u2014 it sends the smart agent chasing a bug that does not exist.";
 /**
@@ -47,0 +47,0 @@ * Render the findings into the compact, plain-text digest the summary model reasons

dist/services/summarize.js

@@ -31,14 +31,14 @@ /**
 You are the summary guy — you condense a UI debug run into one paragraph for the
-smart agent that will fix the code.
-You are given the run's verdict plus every functional bug and visual issue it found.
-Write ONE plain-text paragraph (no markdown, no lists, no headings, no preamble) that
-states the verdict, then what broke, where, and what to fix — concrete and ordered by
-severity. If nothing broke, say so plainly. Output only the paragraph.
+smart agent that will fix the code. You get the verdict plus every functional bug
+and visual issue found.
 
-Report ONLY what the findings below contain. Never infer or invent a cause that is not
-there: no crashes, exceptions, missing dependencies or build problems unless a bug or
-console error in the input actually says so. A "failed" verdict with no bugs listed
-means the run did not reach a conclusion — say exactly that, and do not diagnose it.
-Speculation here is worse than silence: it sends the smart agent chasing a bug that
-does not exist.`;
+Write ONE plain-text paragraph — no markdown, no lists, no headings, no preamble.
+State the verdict, then what broke, where, and what to fix: concrete, ordered by
+severity. Nothing broke → say so plainly. Output only the paragraph.
+
+Report ONLY what the findings contain. NEVER infer a cause that is not there — no
+crashes, exceptions, missing dependencies or build problems unless a bug or console
+error says so. A "failed" verdict with no bugs listed means the run reached no
+conclusion: say exactly that, do not diagnose it. Speculation is worse than silence
+— it sends the smart agent chasing a bug that does not exist.`;
 /** One functional bug, rendered for the digest. */
@@ -45,0 +45,0 @@ function bugLine(bug) {

dist/services/summarize.js.map

@@ -1,1 +0,1 @@
-{"version":3,"file":"summarize.js","sourceRoot":"","sources":["../../src/services/summarize.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,OAAO,EAAE,YAAY,EAAE,MAAM,IAAI,CAAC;AAClC,OAAO,EAAE,UAAU,EAAE,MAAM,cAAc,CAAC;AAkB1C;;;;GAIG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG;;;;;;;;;;;;;gBAarB,CAAC;AAEjB,mDAAmD;AACnD,SAAS,OAAO,CAAC,GAA6B;IAC5C,MAAM,QAAQ,GAAG,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,eAAe,GAAG,CAAC,QAAQ,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;IACpE,OAAO,MAAM,GAAG,CAAC,IAAI,KAAK,GAAG,CAAC,MAAM,GAAG,QAAQ,EAAE,CAAC;AACpD,CAAC;AAED,iDAAiD;AACjD,SAAS,UAAU,CAAC,KAAiC;IACnD,OAAO,MAAM,KAAK,CAAC,QAAQ,KAAK,KAAK,CAAC,KAAK,MAAM,KAAK,CAAC,KAAK,EAAE,CAAC;AACjE,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,cAAc,CAAC,QAAkB;IAC/C,MAAM,KAAK,GAAa,CAAC,YAAY,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC;IAEzD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,oBAAoB,QAAQ,CAAC,IAAI,CAAC,MAAM,IAAI,EAAE,GAAG,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC;IAC1F,CAAC;SAAM,CAAC;QACN,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;IACvC,CAAC;IAED,IAAI,QAAQ,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC/B,KAAK,CAAC,IAAI,CAAC,kBAAkB,QAAQ,CAAC,MAAM,CAAC,MAAM,IAAI,EAAE,GAAG,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC;IAC/F,CAAC;SAAM,CAAC;QACN,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;IACrC,CAAC;IAED,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,QAAyB,EACzB,QAAkB,EAClB,MAAoB;IAEpB,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,QAAQ,CAAC;QAC9B,MAAM,EAAE,qBAAqB;QAC7B,MAAM,EAAE,cAAc,CAAC,QAAQ,CAAC;QAChC,WAAW,EAAE,MAAM;KACpB,CAAC,CAAC;IACH,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC9B,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC3B,MAAM,IAAI,UAAU,CAAC,kDAAkD,CAAC,CAAC;IAC3E,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,eAAe,CAAC,KAAoB;IAClD,kFAAkF;IAClF,oEAAoE;IACpE,MAAM,QAAQ,GAAoB,KAAK,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,EAAE;QAC1E,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,YAAY,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,CAAC;QAC5E,OAAO,EAAE,IAAI,EAAE,CAAC;IAClB,CAAC,CAAC;IACF,OAAO,CAAC,QAAQ,EAAE,MAAM,EAAE,EAAE,CAAC,SAAS,CAAC,QAAQ,EAAE,QAAQ,EAAE,MAAM,CAAC,CAAC;AACrE,CAAC"}
+{"version":3,"file":"summarize.js","sourceRoot":"","sources":["../../src/services/summarize.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,OAAO,EAAE,YAAY,EAAE,MAAM,IAAI,CAAC;AAClC,OAAO,EAAE,UAAU,EAAE,MAAM,cAAc,CAAC;AAkB1C;;;;GAIG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG;;;;;;;;;;;;;8DAayB,CAAC;AAE/D,mDAAmD;AACnD,SAAS,OAAO,CAAC,GAA6B;IAC5C,MAAM,QAAQ,GAAG,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,eAAe,GAAG,CAAC,QAAQ,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;IACpE,OAAO,MAAM,GAAG,CAAC,IAAI,KAAK,GAAG,CAAC,MAAM,GAAG,QAAQ,EAAE,CAAC;AACpD,CAAC;AAED,iDAAiD;AACjD,SAAS,UAAU,CAAC,KAAiC;IACnD,OAAO,MAAM,KAAK,CAAC,QAAQ,KAAK,KAAK,CAAC,KAAK,MAAM,KAAK,CAAC,KAAK,EAAE,CAAC;AACjE,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,cAAc,CAAC,QAAkB;IAC/C,MAAM,KAAK,GAAa,CAAC,YAAY,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC;IAEzD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,oBAAoB,QAAQ,CAAC,IAAI,CAAC,MAAM,IAAI,EAAE,GAAG,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC;IAC1F,CAAC;SAAM,CAAC;QACN,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;IACvC,CAAC;IAED,IAAI,QAAQ,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC/B,KAAK,CAAC,IAAI,CAAC,kBAAkB,QAAQ,CAAC,MAAM,CAAC,MAAM,IAAI,EAAE,GAAG,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC;IAC/F,CAAC;SAAM,CAAC;QACN,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;IACrC,CAAC;IAED,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,QAAyB,EACzB,QAAkB,EAClB,MAAoB;IAEpB,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,QAAQ,CAAC;QAC9B,MAAM,EAAE,qBAAqB;QAC7B,MAAM,EAAE,cAAc,CAAC,QAAQ,CAAC;QAChC,WAAW,EAAE,MAAM;KACpB,CAAC,CAAC;IACH,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC9B,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC3B,MAAM,IAAI,UAAU,CAAC,kDAAkD,CAAC,CAAC;IAC3E,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,eAAe,CAAC,KAAoB;IAClD,kFAAkF;IAClF,oEAAoE;IACpE,MAAM,QAAQ,GAAoB,KAAK,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,EAAE;QAC1E,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,YAAY,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,CAAC;QAC5E,OAAO,EAAE,IAAI,EAAE,CAAC;IAClB,CAAC,CAAC;IACF,OAAO,CAAC,QAAQ,EAAE,MAAM,EAAE,EAAE,CAAC,SAAS,CAAC,QAAQ,EAAE,QAAQ,EAAE,MAAM,CAAC,CAAC;AACrE,CAAC"}

package.json

 {
   "name": "@developerz.ai/ui-debugger-mcp",
-  "version": "1.5.0",
+  "version": "1.5.1",
   "description": "Autonomous UI debugging MCP server. Give a goal; a fast agent drives the browser/desktop, finds bugs + visual issues, and reports back.",
@@ -5,0 +5,0 @@ "mcpName": "io.github.developerz-ai/ui-debugger-mcp",

New alerts

Long strings

Supply chain risk

Contains long string literals, which may be a sign of obfuscated or packed code.

URL strings

Supply chain risk

Package contains fragments of external URLs or IP addresses, which the package may be accessing at runtime.

Fixed alerts

Long strings

Supply chain risk

Contains long string literals, which may be a sign of obfuscated or packed code.

URL strings

Supply chain risk

Package contains fragments of external URLs or IP addresses, which the package may be accessing at runtime.

Improved metrics

Number of low supply chain risk alerts

35

-2.78%

Worsened metrics

Total package byte prevSize

998954

-0.86%

Lines of code

14851

-0.46%

No dependency changes