pi-ui-extend 0.1.38 → 0.1.41

package/dist/markdown-format.js

@@ -1,4 +1,4 @@
-import { expandTabs, stringDisplayWidth } from "./terminal-width.js";
+import { displayGraphemes, expandTabs, stringDisplayWidth } from "./terminal-width.js";
 import { syntaxHighlightLanguageForMarkdownFence, } from "./syntax-highlight.js";
 const MIN_TRAILING_WORD_WIDTH_TO_REBALANCE = 5;
 export function formatMarkdownTables(text, maxWidth) {
@@ -211,20 +211,16 @@ function displayTokensWithRanges(text) {
     let current = "";
     let currentStart = 0;
     let currentWhitespace;
-    for (let index = 0; index < text.length;) {
-        const codePoint = text.codePointAt(index) ?? 0;
-        const char = String.fromCodePoint(codePoint);
+    for (const { text: char, start } of displayGraphemes(text)) {
         const whitespace = /\s/u.test(char);
         if (current && currentWhitespace !== whitespace) {
-            tokens.push({ text: current, start: currentStart, end: index, whitespace: currentWhitespace ?? false });
+            tokens.push({ text: current, start: currentStart, end: start, whitespace: currentWhitespace ?? false });
             current = "";
-            currentStart = index;
         }
         if (!current)
-            currentStart = index;
+            currentStart = start;
         current += char;
         currentWhitespace = whitespace;
-        index += char.length;
     }
     if (current)
         tokens.push({ text: current, start: currentStart, end: text.length, whitespace: currentWhitespace ?? false });
@@ -235,19 +231,16 @@ function wrapDisplayTokenByWidth(token, width) {
     let chunkText = "";
     let chunkStart = token.start;
     let chunkWidth = 0;
-    for (let index = token.start; index < token.end;) {
-        const codePoint = token.text.codePointAt(index - token.start) ?? 0;
-        const char = String.fromCodePoint(codePoint);
-        const charWidth = stringDisplayWidth(char);
+    for (const { text: char, width: charWidth, start } of displayGraphemes(token.text)) {
+        const absoluteStart = token.start + start;
         if (chunkText && chunkWidth + charWidth > width) {
-            chunks.push({ text: chunkText, start: chunkStart, end: index });
+            chunks.push({ text: chunkText, start: chunkStart, end: absoluteStart });
             chunkText = "";
-            chunkStart = index;
+            chunkStart = absoluteStart;
             chunkWidth = 0;
         }
         chunkText += char;
         chunkWidth += charWidth;
-        index += char.length;
     }
     chunks.push({ text: chunkText, start: chunkStart, end: token.end });
     return chunks;
@@ -506,20 +499,16 @@ function smartDisplayBreakIndex(text, width) {
     let used = 0;
     let fallbackIndex = 0;
     let breakIndex = 0;
-    for (let index = 0; index < text.length;) {
-        const codePoint = text.codePointAt(index) ?? 0;
-        const char = String.fromCodePoint(codePoint);
-        const nextIndex = index + char.length;
-        const nextUsed = used + stringDisplayWidth(char);
+    for (const { text: char, width: charWidth, start, end } of displayGraphemes(text)) {
+        const nextUsed = used + charWidth;
         if (nextUsed > width)
             break;
         used = nextUsed;
-        fallbackIndex = nextIndex;
-        if (char === "/" && index > 0)
-            breakIndex = index;
+        fallbackIndex = end;
+        if (char === "/" && start > 0)
+            breakIndex = start;
         else if (/[._:-]/u.test(char))
-            breakIndex = nextIndex;
-        index = nextIndex;
+            breakIndex = end;
     }
     return breakIndex > 0 ? breakIndex : fallbackIndex;
 }

package/dist/terminal-width.d.ts

@@ -3,6 +3,20 @@ export declare function stringDisplayWidth(text: string): number;
 export declare function sliceByDisplayWidth(text: string, width: number): string;
 export declare function displayIndexForColumn(text: string, column: number): number;
 export declare function sliceByDisplayColumns(text: string, startColumn: number, endColumn: number): string;
+export type DisplayGrapheme = {
+    text: string;
+    width: number;
+    start: number;
+    end: number;
+};
+/**
+ * Grapheme clusters of `text` with their display width and absolute string
+ * indices. Iterating graphemes (instead of code points) is required for correct
+ * width accounting: multi-codepoint emoji such as `⚠️` (U+26A0 U+FE0F), keycaps,
+ * skin-tone modifiers and regional-indicator flags are one width-2 cluster even
+ * though several of their code points have zero width.
+ */
+export declare function displayGraphemes(text: string): DisplayGrapheme[];
 export declare function padOrTrimDisplay(text: string, width: number): string;
 export declare function wrapDisplayLine(text: string, width: number): string[];
 export declare function wrapDisplayLineByWords(text: string, width: number): string[];

package/dist/terminal-width.js

@@ -1,7 +1,7 @@
 const TAB_WIDTH = 4;
 const ANSI_RESET = "\x1b[0m";
 const EMOJI_PRESENTATION_REGEX = /\p{Emoji_Presentation}/u;
-const EMOJI_REGEX = /\p{Emoji}/u;
+const REGIONAL_INDICATOR_REGEX = /[\u{1F1E6}-\u{1F1FF}]/u;
 const GRAPHEME_SEGMENTER = typeof Intl.Segmenter === "function" ? new Intl.Segmenter(undefined, { granularity: "grapheme" }) : undefined;
 export function expandTabs(text, tabWidth = TAB_WIDTH) {
     if (!text.includes("\t"))
@@ -86,6 +86,20 @@ export function sliceByDisplayColumns(text, startColumn, endColumn) {
     const endIndex = Math.max(startIndex, displayIndexForColumn(text, endColumn));
     return text.slice(startIndex, endIndex);
 }
+/**
+ * Grapheme clusters of `text` with their display width and absolute string
+ * indices. Iterating graphemes (instead of code points) is required for correct
+ * width accounting: multi-codepoint emoji such as `⚠️` (U+26A0 U+FE0F), keycaps,
+ * skin-tone modifiers and regional-indicator flags are one width-2 cluster even
+ * though several of their code points have zero width.
+ */
+export function displayGraphemes(text) {
+    const graphemes = [];
+    for (const cluster of indexedDisplayClusters(text)) {
+        graphemes.push({ text: cluster.text, width: cluster.width, start: cluster.start, end: cluster.end });
+    }
+    return graphemes;
+}
 export function padOrTrimDisplay(text, width) {
     const safeWidth = Math.max(0, width);
     if (isPrintableAscii(text)) {
@@ -280,7 +294,22 @@ function graphemeDisplayWidth(text) {
     return width;
 }
 function isEmojiGrapheme(text) {
-    return EMOJI_PRESENTATION_REGEX.test(text) || (EMOJI_REGEX.test(text) && (text.includes("\ufe0f") || text.includes("\u20e3")));
+    // Default-presentation emoji (⛔ ✅ 🚀 ❌), supplementary pictographs, and
+    // regional-indicator flags render two cells wide in conforming terminals
+    // including iTerm2 and Zed, so they are measured at width 2.
+    if (EMOJI_PRESENTATION_REGEX.test(text))
+        return true;
+    // Keycap sequences (base + U+FE0F + U+20E3, e.g. 1️⃣) and regional-indicator
+    // pairs (🇷🇺) also occupy two cells.
+    if (text.includes("\u20e3"))
+        return true;
+    if (REGIONAL_INDICATOR_REGEX.test(text) && /[\u{1F1E6}-\u{1F1FF}]{2}/u.test(text))
+        return true;
+    // Symbols promoted to an emoji glyph only by a variation selector (⚠️ ✔️ ©️
+    // ☀️) keep their base width of 1. Their base code point is East-Asian-Width
+    // Ambiguous, and iTerm2/Zed/wcwidth render them one cell wide; counting them
+    // as 2 would misalign table columns and shorten rendered rows.
+    return false;
 }
 function codePointLength(codePoint) {
     return codePoint > 0xffff ? 2 : 1;

package/dist/theme.js

@@ -9,7 +9,7 @@ export const THEMES = {
             muted: "#7d8590",
             headerForeground: "#c9d1d9",
             headerBackground: "#161b22",
-            statusForeground: "#8b949e",
+            statusForeground: "#9ba5af",
             statusBackground: "#0f1520",
             inputForeground: "#f0f6fc",
             inputBackground: "#090d13",
@@ -64,7 +64,7 @@ export const THEMES = {
             muted: "#64748b",
             headerForeground: "#0f172a",
             headerBackground: "#e2e8f0",
-            statusForeground: "#475569",
+            statusForeground: "#566578",
             statusBackground: "#edf0f4",
             inputForeground: "#0f172a",
             inputBackground: "#f8fafc",

package/external/pi-tools-suite/README.md

@@ -4,11 +4,12 @@ Local all-in-one Pi extension package.
 
 This package keeps shared Pi tools as ordinary source folders under `src/` and registers them through one entrypoint.
 
-- `src/glm-coding-discipline` — injects a deduplicated silent-mode and quality-discipline block at the very top of the main-session per-turn system prompt for GLM-family models immediately before the LLM request; disabled for async sub-agents
+- `src/coding-discipline` — injects a deduplicated silent-mode and quality-discipline block at the very top of the main-session per-turn system prompt for all main-session models immediately before the LLM request; disabled for async sub-agents
 - `src/ast-grep` — `ast_grep` / `ast_apply`
-- `src/async-subagents` — `subagents` tool and sub-agent slash commands, including oh-my-openagent-style `/ultrawork` (`/ulw`) and `/hyperplan` orchestration prompts, plus config-defined sub-agent model/thinking/args presets selected via `/subagent-preset` from `asyncSubagents` in `~/.config/pi/pi-tools-suite.jsonc`; includes the `frontend` profile for Gemini-friendly UI/UX and visual frontend work plus the `vision` profile for screenshot/image description via `openai-codex/gpt-5.4-mini`; enforces a 30-minute per-agent execution timeout, project-wide `maxConcurrent` queueing, optional retry/backoff, and `result.json` structured metadata/chaining fields next to raw `result.md`; stores project-local run files and a registry under `.pi/subagents/` so result/status collection can recover after compaction or reload while the main session remains alive
+- `src/async-subagents` — `subagents` tool and sub-agent slash commands, including oh-my-openagent-style `/ultrawork` (`/ulw`) and `/hyperplan` orchestration prompts, plus config-defined sub-agent model/thinking/args presets selected via `/subagent-preset` from `asyncSubagents` in `~/.config/pi/pi-tools-suite.jsonc`; includes the `frontend` profile for Gemini-friendly UI/UX and visual frontend work and the `oracle` profile for cross-provider second opinions; enforces a 30-minute per-agent execution timeout, project-wide `maxConcurrent` queueing, optional retry/backoff, and `result.json` structured metadata/chaining fields next to raw `result.md`; stores project-local run files and a registry under `.pi/subagents/` so result/status collection can recover after compaction or reload while the main session remains alive
 - `src/lsp` — shared LSP diagnostics hook/library that enriches mutating tool results with diagnostics and shuts down language servers on session shutdown
 - `src/comment-checker` — AI-slop comment guard that listens to the `tool_result` event for `write` / `edit` / `apply_patch` mutations, extracts net-new code comment lines, classifies them (filler phrasing, restating code, decorative separators, generic paraphrasing, or — under aggressive strictness — any non-valuable comment), and appends a short nudge to the tool result so the agent removes unnecessary comments on its next turn; TODO/FIXME, license headers, docstrings, pragmas, linter directives, shebangs, and decorators are never flagged; language-agnostic across `//` / `/* */` / `#` / `--` / `<!-- -->` / triple-quote comment styles; per-session deduplication (at most one nudge per 30 s) prevents fix/remark loops; configured via the `commentChecker` section (`enabled`, `strictness`: `conservative` | `balanced` | `aggressive`, default `balanced`) or `PI_COMMENT_CHECKER_ENABLED` / `PI_COMMENT_CHECKER_STRICTNESS`
+- `src/session-name` — `session_name` tool for reading or setting the current session title directly from tool calls, without relying on slash-command parsing
 - `src/repo-discovery` — `/idx-init`, `/idx-update`, and indexed-only `repo_architecture` / `repo_structure` / `repo_ast` / `repo_search` / `repo_explain` / `repo_deps`; tools register only when the launch project has `.indexer-cli`
 - `src/antigravity-auth` — `antigravity` custom provider with Google Antigravity OAuth login, startup account list, auth.json-only runtime account loading, `/antigravity-add-account` OAuth append into rotation, `/antigravity-account` status display, account rotation/failover, Antigravity plus Gemini CLI model registration, and streaming through the Cloud Code Assist unified gateway
 - `src/todo` — `todo` tool, `/todos`, `/todos-persist`, and `/todos-scope`; supports parent/subtask hierarchy, blockers, ready-task filtering, deferred out-of-scope items, batch operations, JSON/Markdown import/export, automatic clearing when all visible todos are completed, and optional project persistence via `/todos persist on` or `/todos-persist on`; localization/i18n has been removed
@@ -20,7 +21,7 @@ This package keeps shared Pi tools as ordinary source folders under `src/` and r
 
 `index.ts` is intentionally only a thin auto-discovery shim that re-exports `src/index.ts`. There is no `pi.extensions` manifest here, so local Pi auto-discovery loads the suite once via `~/.pi/agent/extensions/pi-tools-suite/index.ts` and does not double-register tools.
 
-Registration order is preserved in `src/index.ts`: glm-coding-discipline, ast-grep, async-subagents, lsp, comment-checker, repo-discovery command/tool gate, antigravity-auth provider, todo, model-tools, usage, web-search, dcp, then prompt-commands. Tool metadata and active model-specific tool sets have two modes: standard and repo-aware. When `.indexer-cli` enables `repo_*`, those tools stay active ahead of overlapping lower-level aliases so the indexed discovery surface has priority.
+Registration order is preserved in `src/index.ts`: coding-discipline, ast-grep, async-subagents, lsp, comment-checker, session-name, repo-discovery command/tool gate, antigravity-auth provider, todo, model-tools, usage, web-search, dcp, then prompt-commands. Tool metadata and active model-specific tool sets have two modes: standard and repo-aware. When `.indexer-cli` enables `repo_*`, those tools stay active ahead of overlapping lower-level aliases so the indexed discovery surface has priority.
 
 ## Disabling modules
 
@@ -68,13 +69,14 @@ DCP settings are stored only under `dcp` in the user shared config file `~/.conf
       "maxContextLimit": 160000,
       "nudgeFrequency": 1,
       "iterationNudgeThreshold": 6,
+      "nudgeForce": "strong",
       "protectedTools": ["compress", "write", "edit", "subagents"]
     },
     "modelOverrides": {
-      "openai-codex/gpt-5.5": {
+      "openai-codex/gpt-5*": {
         "compress": {
-          "minContextPercent": "28%",
-          "maxContextPercent": "48%"
+          "minContextPercent": "26%",
+          "maxContextPercent": "46%"
         }
       },
       "openai-codex/gpt-5.4-mini": {
@@ -208,11 +210,11 @@ Project-local overrides can be added in `.pi/pi-tools-suite.jsonc`; pi-tools-sui
 
 Sub-agent model routing normally follows task overrides, subagent type config, then `ASYNC_SUBAGENTS_MODEL` / `PI_SUBAGENTS_MODEL` fallbacks. Set `ASYNC_SUBAGENTS_FORCE_CURRENT_MODEL=1` (or `PI_SUBAGENTS_FORCE_CURRENT_MODEL=1`) to ignore task/config/env model choices and launch every sub-agent with the current parent session model. When this flag is enabled, any `--model` entries in sub-agent extra args are stripped so they cannot override the current model.
 
-For an oh-my-openagent-style workflow, run `/ultrawork` or `/ulw` to ask the parent agent to split broad work into configured async-subagents roles (`quick`, `scan`, `research`, `docs`, `frontend`, `implement`, `tests`, `review`, `deep`, `vision`). Set `ULTRAWORK=1` before launching Pi to apply that compact routing prompt to normal non-slash user inputs automatically. Set `ULTRAWORK_AUTO=1` to ask the lightweight router model to classify only the first normal user input on non-GPT parent models: clear broad/parallel work is transformed into ultrawork, vague potentially-complex work gets a soft delegation hint, and narrow work is left unchanged. GPT-like parent models skip only this automatic transform; they can still use `/ultrawork` and `subagents` normally. `frontend` is for UI/UX, styling, layout, responsive behavior, and visual component polish; `review` covers security/performance/audit tracks; `implement` covers refactors; `deep` covers debugging/root-cause; `vision` is only for screenshots/images when the parent model is a non-vision GLM-series model. Run `/hyperplan` to pressure-test a plan before implementation.
+For an oh-my-openagent-style workflow, run `/ultrawork` or `/ulw` to ask the parent agent to split broad work into configured async-subagents roles (`quick`, `scan`, `research`, `docs`, `frontend`, `implement`, `tests`, `review`, `deep`, `oracle`). Set `ULTRAWORK=1` before launching Pi to apply that compact routing prompt to normal non-slash user inputs automatically. Set `ULTRAWORK_AUTO=1` to ask the lightweight router model to classify only the first normal user input on non-GPT parent models: clear broad/parallel work is transformed into ultrawork, vague potentially-complex work gets a soft delegation hint, and narrow work is left unchanged. GPT-like parent models skip only this automatic transform; they can still use `/ultrawork` and `subagents` normally. `frontend` is for UI/UX, styling, layout, responsive behavior, and visual component polish; `review` covers security/performance/audit tracks; `implement` covers refactors; `deep` covers debugging/root-cause; `oracle` is for sparse cross-provider second opinions on high-stakes uncertainty. Run `/hyperplan` to pressure-test a plan before implementation.
 
 Async-subagents also injects a lightweight oh-my-openagent-style system-prompt strategy by model: non-GPT parents get `parallel-first`, an orchestration-first hint that favors ultrawork/subagents for broad work, while GPT-like parents get `deep-work`, a direct deep-worker hint that uses subagents only when clearly useful. Explicit custom system prompts (`--system-prompt`, `SYSTEM.md`, custom templates) are respected and skip this injection by default. Disable it with `PI_AGENT_STRATEGY=off`; force a strategy with `PI_AGENT_STRATEGY=parallel-first` or `PI_AGENT_STRATEGY=deep-work`; set `PI_AGENT_STRATEGY_WITH_CUSTOM_PROMPT=1` to append it even when a custom prompt is present.
 
-When the parent model cannot inspect images, async-subagents adds vision-delegation guidance and can save current-turn image attachments under `.pi/subagents/attachments/` so a `vision` sub-agent can receive them as `imagePaths`. Dynamic provider capabilities can be missing or stale after switching models, so blind parent models can still be configured explicitly with case-insensitive `*` masks under `asyncSubagents.vision.blindModelPatterns` in `~/.config/pi/pi-tools-suite.jsonc`. GLM is no longer treated as blind by async-subagents by default; the main-session `glm-coding-discipline` lookup tool is the preferred path for GLM visual lookups.
+For blind-model screenshot/image inspection, use the main-session `coding-discipline` lookup tool. Async-subagents still supports `imagePaths` on tasks when a broader delegated track genuinely needs images, but it no longer ships a dedicated `vision` role. Dynamic provider capabilities can be missing or stale after switching models, so blind parent models can still be configured explicitly with case-insensitive `*` masks under `asyncSubagents.vision.blindModelPatterns` in `~/.config/pi/pi-tools-suite.jsonc`; this keeps guidance honest, not a sub-agent role.
 
 When a task omits `subagentType`, async-subagents asks a lightweight router model to choose one configured type for each task from the task text/scope and the `types.<name>.description` metadata. Explicit task `subagentType` still wins. Keep type descriptions short, literal, and distinct because they are inserted into the router prompt for a small model. Router settings live under `asyncSubagents.routing` (`enabled`, `model`, `maxTaskChars`, `maxTokens`, `maxRetries`, `timeoutMs`, `debug`); the default router model is `zai/glm-4.5-air`. If the router is disabled, unavailable, aborted, or returns invalid JSON, omitted types fall back to `defaultType`.
 
@@ -231,7 +233,7 @@ Example shared async-subagents config section:
     },
     "presets": {
       "cheap": {
-        "description": "Use GLM for text/code roles; keep vision on its dedicated model.",
+        "description": "Use GLM for text/code roles.",
         "types": {
           "quick": { "model": "zai/glm-5.2", "thinking": "off" },
           "frontend": { "model": "antigravity/gemini-3-flash-preview", "fallbackModels": ["zai/glm-5.2"], "thinking": "medium" },
@@ -253,6 +255,29 @@ Example shared async-subagents config section:
 }
 ```
 
+### Parent-model-aware model selection (`modelByParent`)
+
+Any type profile can carry `modelByParent`: a map from glob model refs (matched against the **current parent model**, e.g. `"zai/*"`) to a model for that role. The first matching key wins. Values may be a model string or `{ "model": "...", "fallbackModels": [...] }`. It is resolved after an explicit task `model` / `forcedModel`, but **before** the preset/static profile `model`, so a role can always pick a model based on who the parent is — independent of the active preset.
+
+The canonical use case is an **`oracle`** role that consults a flagship model from a *different* provider than the parent for a second opinion:
+
+```jsonc
+"oracle": {
+  "description": "Cross-provider second opinion: consult a flagship from a different provider than the parent to pressure-test a hard decision. Read-only; advise, do not edit.",
+  "model": "openai-codex/gpt-5.5",
+  "fallbackModels": ["zai/glm-5.2"],
+  "thinking": "xhigh",
+  "modelByParent": {
+    "zai/*":         { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"] },
+    "openai-codex/*": "zai/glm-5.2",
+    "antigravity/*": { "model": "zai/glm-5.2", "fallbackModels": ["openai-codex/gpt-5.5"] },
+    "anthropic/*":   { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"] }
+  }
+}
+```
+
+With this config a GLM parent (`zai/*`) spawns the oracle on `gpt-5.5`, a GPT parent (`openai-codex/*`) spawns it on `glm-5.2`, and so on — automatically, at spawn time, with no `task.model` needed. The parent model ref is read from the spawn context (`ctx.model`) and passed into resolution. Pattern matching is case-insensitive `*` glob (same engine as `vision.blindModelPatterns`). When no key matches (or no parent model is known), the role falls back to its static `model` + `fallbackModels`. An explicit `task.model` or `ASYNC_SUBAGENTS_FORCE_CURRENT_MODEL=1` still overrides the match.
+
 Sub-agents run with `--no-session` by default to avoid writing duplicate Pi session JSONL files for fire-and-forget background work. Set `ASYNC_SUBAGENTS_ENABLE_SESSIONS=1` to restore persisted per-agent sessions under each agent's `sessions/` directory; this also registers the session-navigation slash commands (`/sub-open`, `/sub-back`, `/sub-where`) needed for switching and deeper post-mortem navigation.
 
 Sub-agent runs are stored in the current project's `.pi/subagents/` directory while the main session is alive. Each spawn updates `.pi/subagents/registry.json` with the latest run and `agentId -> runDir` mappings. Because of that, `subagents({ action: "status" })`, `wait`, and `stop` can omit `runDir` to target the latest run, and `subagents({ action: "result", agentId: "..." })` can resolve the run from the registry even if the exact `runDir` was lost during compaction. Result reads always return a summary-first response with artifact paths; raw `result.md` and `stderr.log` are not inlined, which avoids IPC/socket buffer overflows. Include `runDir` when you need an older or non-latest run, and use `cleanup` with `delete=true` to remove collected old runs before the session ends. On normal main-session shutdown, Pi stops sub-agents and removes the project-local run files/registry to avoid leaving `.pi/subagents/` clutter behind; reload and fork shutdowns preserve them so in-process recovery still works.

package/external/pi-tools-suite/package.json

@@ -38,9 +38,9 @@
     "vscode-languageserver-protocol": "^3.17.5"
   },
   "peerDependencies": {
-    "@earendil-works/pi-ai": "0.79.6",
-    "@earendil-works/pi-coding-agent": "0.79.6",
-    "@earendil-works/pi-tui": "0.79.6",
+    "@earendil-works/pi-ai": "0.79.7",
+    "@earendil-works/pi-coding-agent": "0.79.7",
+    "@earendil-works/pi-tui": "0.79.7",
     "typebox": "*"
   },
   "devDependencies": {

package/external/pi-tools-suite/src/async-subagents/async-subagents.sample.jsonc

@@ -28,10 +28,11 @@
   // for short synthetic tests.
   // "timeoutMs": 1800000,
 
-  // Parent-model vision overrides. Pi provider metadata can be missing or stale
-  // after switching models, so async-subagents treats these glob-like model refs
-  // as text-only even if dynamic capability detection is inconclusive. Matching
-  // is case-insensitive and supports `*` wildcards. Set [] to disable built-ins.
+  // Parent-model image-capability overrides. Pi provider metadata can be missing
+  // or stale after switching models, so async-subagents treats these glob-like
+  // model refs as text-only when dynamic detection is inconclusive. Blind-model
+  // screenshot inspection should use lookup; this only keeps guidance honest.
+  // Matching is case-insensitive and supports `*` wildcards.
   "vision": {
     "blindModelPatterns": ["zai/glm*", "glm*", "*/glm*"]
   },
@@ -47,6 +48,19 @@
   // ],
   // "promptOverride": "You are a specialized {subagentType} sub-agent.\nParent: {parentObjective}\nTask: {task}\nScope: {scope}",
 
+  // Parent-model-aware model selection (optional, any type profile).
+  // modelByParent maps glob model refs (matched against the current parent
+  // model, e.g. "zai/*") to a model for this role. The first matching key wins.
+  // Values may be a model string or { model, fallbackModels? }. It is resolved
+  // after an explicit task.model / forcedModel, but before the preset/static
+  // profile model, so a role like an "oracle" can always pick a flagship from a
+  // DIFFERENT provider than the parent. Example:
+  // "modelByParent": {
+  //   "zai/*": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"] },
+  //   "openai-codex/*": "zai/glm-5.2",
+  //   "antigravity/*": { "model": "zai/glm-5.2", "fallbackModels": ["openai-codex/gpt-5.5"] }
+  // }
+
   // Role auto-selection when subagentType is omitted is handled by a small LLM
   // router request. The router receives each task plus the configured type
   // descriptions, then returns one allowed subagentType per task. Explicit
@@ -95,7 +109,7 @@
 
   "presets": {
     "cheap": {
-      "description": "Use cheap GLM/Gemini Flash models for text/code roles; keep vision on the enabled GPT vision model.",
+      "description": "Use cheap GLM/Gemini Flash models for text/code roles.",
       "types": {
         "quick": { "model": "zai/glm-4.5-air", "thinking": "off" },
         "scan": { "model": "zai/glm-4.5-air", "thinking": "off" },
@@ -105,8 +119,7 @@
         "tests": { "model": "zai/glm-5-turbo", "thinking": "medium" },
         "review": { "model": "zai/glm-5.2", "thinking": "high" },
         "implement": { "model": "zai/glm-5.2", "thinking": "high" },
-        "deep": { "model": "zai/glm-5.2", "thinking": "high" },
-        "vision": { "model": "openai-codex/gpt-5.4-mini", "thinking": "off" }
+        "deep": { "model": "zai/glm-5.2", "thinking": "high" }
       }
     },
 
@@ -121,8 +134,7 @@
         "tests": { "model": "openai-codex/gpt-5.4-mini", "fallbackModels": ["zai/glm-5-turbo"], "thinking": "medium" },
         "review": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"], "thinking": "high" },
         "implement": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"], "thinking": "high" },
-        "deep": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"], "thinking": "high" },
-        "vision": { "model": "openai-codex/gpt-5.4-mini", "thinking": "off" }
+        "deep": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"], "thinking": "high" }
       }
     },
 
@@ -137,8 +149,7 @@
         "tests": { "model": "antigravity/antigravity-claude-sonnet-4-6", "fallbackModels": ["openai-codex/gpt-5.4-mini", "zai/glm-5-turbo"], "thinking": "high" },
         "review": { "model": "antigravity/antigravity-claude-sonnet-4-6", "fallbackModels": ["openai-codex/gpt-5.5", "zai/glm-5.2"], "thinking": "high" },
         "implement": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"], "thinking": "high" },
-        "deep": { "model": "antigravity/antigravity-claude-opus-4-6-thinking", "fallbackModels": ["openai-codex/gpt-5.5", "zai/glm-5.2"], "thinking": "high" },
-        "vision": { "model": "openai-codex/gpt-5.4-mini", "thinking": "off" }
+        "deep": { "model": "antigravity/antigravity-claude-opus-4-6-thinking", "fallbackModels": ["openai-codex/gpt-5.5", "zai/glm-5.2"], "thinking": "high" }
       }
     }
   },
@@ -212,16 +223,19 @@
       "thinking": "high"
     },
 
-    "vision": {
-      "description": "Use only when task has imagePaths, screenshots, or asks to inspect visible UI/image content for a text-only parent.",
-      "model": "openai-codex/gpt-5.4-mini",
-      "thinking": "off",
-      "promptAppend": [
-        "You are a vision helper for a parent model that may not be able to see images.",
-        "Inspect any attached images and any image paths mentioned in the task/scope. Describe concrete visible details, UI state, text, layout, errors, and uncertainties.",
-        "If focus instructions are provided, prioritize them, but still mention other important visible findings.",
-        "Do not make code changes. Return a compact visual description that the parent agent can rely on."
-      ]
+    "oracle": {
+      "description": "Oracle: cross-provider flagship second opinion for hard or high-stakes uncertainty. Use sparingly to pressure-test architecture, plans, root-cause hypotheses, risk/security calls, or final recommendations when independent disagreement is valuable. Read-only; advise, do not edit.",
+      "model": "openai-codex/gpt-5.5",
+      "fallbackModels": ["zai/glm-5.2"],
+      "thinking": "xhigh",
+      "tools": ["read", "grep", "bash"],
+      "modelByParent": {
+        "zai/*": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"] },
+        "openai-codex/*": { "model": "zai/glm-5.2", "fallbackModels": ["openai-codex/gpt-5.5"] },
+        "antigravity/*": { "model": "zai/glm-5.2", "fallbackModels": ["openai-codex/gpt-5.5"] },
+        "anthropic/*": { "model": "openai-codex/gpt-5.5", "fallbackModels": ["zai/glm-5.2"] }
+      },
+      "promptAppend": "You are an oracle: a flagship model from a different provider giving a second opinion to the parent agent. Give a concise, decisive recommendation with key tradeoffs and risks. Disagree when warranted; do not rubber-stamp. Do not edit unless explicitly asked."
     }
   }
 }

package/external/pi-tools-suite/src/async-subagents/commands.ts

@@ -53,7 +53,7 @@ const COPY_SAMPLE_CONFIG_LABEL = "Copy sample asyncSubagents config";
 
 export const ULTRAWORK_PROMPT = `Run ultrawork mode for the current objective.
 
-Use subagents when independent parallel tracks help. Pick subagentType from configured roles: quick, scan, research, docs, frontend, implement, tests, review, deep, vision. Use review for security/performance/audit tracks, implement for refactors, deep for debugging/root-cause. Use frontend for UI/UX and visual frontend implementation; use vision only for screenshots/images when the parent model is a non-vision GLM-series model.
+Use subagents when independent parallel tracks help. Pick subagentType from configured roles: quick, scan, research, docs, frontend, implement, tests, review, deep, oracle. Use review for security/performance/audit tracks, implement for refactors, deep for debugging/root-cause. Use frontend for UI/UX and visual frontend implementation; use oracle sparingly for cross-provider second opinions on high-stakes uncertainty.
 
 Keep parent context lean: spawn for broad parallel work, read results only when needed, and finish unless genuinely blocked.`;
 

package/external/pi-tools-suite/src/async-subagents/core/agent-strategy.ts

@@ -14,7 +14,7 @@ const TRUE_ENV_PATTERN = /^(1|true|on|yes|auto)$/i;
 const PARALLEL_FIRST_STRATEGY_PROMPT = `<agent_strategy name="parallel-first">
 This is an orchestration hint for Pi, not a replacement for the user's instructions.
 
-Default posture: orchestration-first for non-GPT models. For broad, multi-file, ambiguous, review/audit, frontend, test-strategy, architecture, or root-cause work, prefer ultrawork mode: split independent tracks and spawn focused async subagents with the configured roles. Keep the parent context lean, collect compact results only when needed, synthesize the findings, then verify before finishing.
+Default posture: orchestration-first for non-GPT models. For broad, multi-file, ambiguous, review/audit, frontend, test-strategy, architecture, or root-cause work, prefer ultrawork mode: split independent tracks and spawn focused async subagents with the configured roles. For high-stakes uncertainty, add one oracle track for a cross-provider second opinion. Keep the parent context lean, collect compact results only when needed, synthesize the findings, then verify before finishing.
 
 Before DCP/compress while work is unfinished, keep one in_progress todo with objective + next step; compression summaries must preserve Active objective and Next step.
 
@@ -28,7 +28,7 @@ Default posture: autonomous deep worker. Build context directly, make concrete p
 
 Before DCP/compress while work is unfinished, keep one in_progress todo with objective + next step; compression summaries must preserve Active objective and Next step.
 
-For broad work, keep delegation explicit and bounded: spawn focused review/research/tests/frontend/deep tracks, read compact results, make the final decisions in the parent session, and report only what matters.
+For broad work, keep delegation explicit and bounded: spawn focused review/research/tests/frontend/deep tracks, plus one oracle track only for high-stakes uncertainty or final plan checks. Read compact results, make the final decisions in the parent session, and report only what matters.
 </agent_strategy>`;
 
 export function agentStrategyPrompt(options: AgentStrategyOptions = {}): string | undefined {

package/external/pi-tools-suite/src/async-subagents/core/config.ts

@@ -6,11 +6,25 @@ import { applyEdits, modify, parse as parseJsonc } from "jsonc-parser";
 import { ensurePiToolsSuiteUserConfig, getPiToolsSuiteUserConfigPath } from "../../config.js";
 import type { AgentTask, RetryConfig } from "./types.js";
 
+export interface ModelByParentEntry {
+	/** Model ref to use when the parent model matches the entry's pattern. */
+	model: string;
+	/** Ordered fallbacks used when this entry's model hits quota/rate limits; replaces the normal fallback chain. */
+	fallbackModels?: string[];
+}
+
 export interface SubagentTypeConfig {
 	description?: string;
 	model?: string;
 	/** Ordered model fallbacks used when the selected model hits quota/rate limits. */
 	fallbackModels?: string[];
+	/**
+	 * Parent-model-aware model selection. Keys are glob model refs (e.g. "zai/*")
+	 * matched against the current parent model; the first matching key wins.
+	 * Values may be a model ref string or { model, fallbackModels? }.
+	 * Resolved after explicit task.model / forcedModel, before preset/static model.
+	 */
+	modelByParent?: Record<string, ModelByParentEntry>;
 	thinking?: string;
 	tools?: string[];
 	extraArgs?: string[];
@@ -127,6 +141,8 @@ export interface ResolveAgentTaskOptions {
 	extraArgs?: string[];
 	/** Force every sub-agent to use this model, ignoring task/profile/env model selection. */
 	forcedModel?: string;
+	/** Current parent model ref (provider/model). Enables type.modelByParent matching. */
+	parentModel?: string;
 	/** Force a wall-clock timeout for every sub-agent spawned by this call. */
 	timeoutMs?: number;
 }
@@ -201,16 +217,19 @@ const BUILTIN_CONFIG: SubagentConfig = {
 			description: "Use for broad hard reasoning: architecture, root-cause analysis, cross-module impact, complex debugging or tradeoffs.",
 			thinking: "high",
 		},
-		vision: {
-			description: "Use only when task has imagePaths, screenshots, or asks to inspect visible UI/image content for a text-only parent.",
-			model: "openai-codex/gpt-5.4-mini",
-			thinking: "off",
-			promptAppend: [
-				"You are a vision helper for a parent model that may not be able to see images.",
-				"Inspect any attached images and any image paths mentioned in the task/scope. Describe concrete visible details, UI state, text, layout, errors, and uncertainties.",
-				"If focus instructions are provided, prioritize them, but still mention other important visible findings.",
-				"Do not make code changes. Return a compact visual description that the parent agent can rely on.",
-			].join("\n"),
+		oracle: {
+			description: "Oracle: cross-provider flagship second opinion for hard or high-stakes uncertainty. Use sparingly to pressure-test architecture, plans, root-cause hypotheses, risk/security calls, or final recommendations when independent disagreement is valuable. Read-only; advise, do not edit.",
+			model: "openai-codex/gpt-5.5",
+			fallbackModels: ["zai/glm-5.2"],
+			modelByParent: {
+				"zai/*": { model: "openai-codex/gpt-5.5", fallbackModels: ["zai/glm-5.2"] },
+				"openai-codex/*": { model: "zai/glm-5.2", fallbackModels: ["openai-codex/gpt-5.5"] },
+				"antigravity/*": { model: "zai/glm-5.2", fallbackModels: ["openai-codex/gpt-5.5"] },
+				"anthropic/*": { model: "openai-codex/gpt-5.5", fallbackModels: ["zai/glm-5.2"] },
+			},
+			thinking: "xhigh",
+			tools: ["read", "grep", "bash"],
+			promptAppend: "You are an oracle: a flagship model from a different provider giving a second opinion to the parent agent. Give a concise, decisive recommendation with key tradeoffs and risks. Disagree when warranted; do not rubber-stamp. Do not edit unless explicitly asked.",
 		},
 	},
 };
@@ -325,14 +344,19 @@ export function resolveAgentTaskConfig(
 	const promptAppend = joinTextBlocks(profile?.promptAppend, task.promptAppend);
 	const forcedModel = trimString(globalOptions.forcedModel);
 	const taskModel = trimString(task.model);
+	const parentMatch = resolveModelByParent(profile, trimString(globalOptions.parentModel));
+	const parentMatchModel = trimString(parentMatch?.model);
 	const presetTypeModel = trimString(presetType?.model);
 	const globalModel = trimString(globalOptions.model);
 	const presetModel = trimString(preset?.model);
 	const profileModel = trimString(profile?.model);
-	const model = forcedModel || taskModel || presetTypeModel || globalModel || presetModel || profileModel;
+	const model = forcedModel || taskModel || parentMatchModel || presetTypeModel || globalModel || presetModel || profileModel;
+	const usedParentMatch = Boolean(parentMatchModel) && model === parentMatchModel;
 	const fallbackModels = forcedModel || taskModel
 		? []
-		: resolveFallbackModels({ model, presetType, preset, profile });
+		: usedParentMatch && parentMatch?.fallbackModels && parentMatch.fallbackModels.length > 0
+			? parentMatch.fallbackModels
+			: resolveFallbackModels({ model, presetType, preset, profile });
 	const extraArgs = forcedModel
 		? stripModelArgs([...profileExtraArgs, ...presetTypeExtraArgs, ...taskExtraArgs, ...presetExtraArgs, ...globalExtraArgs])
 		: [...profileExtraArgs, ...presetTypeExtraArgs, ...taskExtraArgs, ...presetExtraArgs, ...globalExtraArgs];
@@ -480,6 +504,7 @@ function normalizeConfig(value: Record<string, unknown>, file: string): Partial<
 			description: trimString(rawProfile.description),
 			model: trimString(rawProfile.model),
 			fallbackModels: modelList(rawProfile.fallbackModels, rawProfile.fallbackModel),
+			modelByParent: normalizeModelByParent(rawProfile.modelByParent, name, file),
 			thinking: trimString(rawProfile.thinking),
 			tools: arrayOfStrings(rawProfile.tools),
 			extraArgs: arrayOfStrings(rawProfile.extraArgs),
@@ -555,6 +580,7 @@ function compactProfile(profile: SubagentTypeConfig): SubagentTypeConfig {
 	if (profile.description) compact.description = profile.description;
 	if (profile.model) compact.model = profile.model;
 	if (profile.fallbackModels && profile.fallbackModels.length > 0) compact.fallbackModels = profile.fallbackModels;
+	if (profile.modelByParent) compact.modelByParent = profile.modelByParent;
 	if (profile.thinking) compact.thinking = profile.thinking;
 	if (profile.tools && profile.tools.length > 0) compact.tools = profile.tools;
 	if (profile.extraArgs && profile.extraArgs.length > 0) compact.extraArgs = profile.extraArgs;
@@ -606,6 +632,38 @@ function normalizePresetTypeOverrides(value: unknown, file: string, presetName:
 	return Object.keys(types).length > 0 ? types : undefined;
 }
 
+function normalizeModelByParent(value: unknown, typeName: string, file: string): Record<string, ModelByParentEntry> | undefined {
+	if (value === undefined || value === null) return undefined;
+	if (!isRecord(value)) throw new Error(`Subagent type "${typeName}" modelByParent must be an object: ${file}`);
+	const out: Record<string, ModelByParentEntry> = {};
+	for (const [pattern, raw] of Object.entries(value)) {
+		const pat = trimString(pattern);
+		if (!pat) continue;
+		if (typeof raw === "string") {
+			const model = trimString(raw);
+			if (model) out[pat] = { model };
+			continue;
+		}
+		if (isRecord(raw)) {
+			const model = trimString(raw.model);
+			if (!model) throw new Error(`Subagent type "${typeName}" modelByParent["${pat}"].model must be a non-empty string: ${file}`);
+			out[pat] = { model, fallbackModels: modelList(raw.fallbackModels, raw.fallbackModel) };
+			continue;
+		}
+		throw new Error(`Subagent type "${typeName}" modelByParent["${pat}"] must be a string or object: ${file}`);
+	}
+	return Object.keys(out).length > 0 ? out : undefined;
+}
+
+function resolveModelByParent(profile: SubagentTypeConfig | undefined, parentModelRef: string | undefined): ModelByParentEntry | undefined {
+	const entries = profile?.modelByParent;
+	if (!entries || !parentModelRef) return undefined;
+	for (const [pattern, entry] of Object.entries(entries)) {
+		if (modelPatternRegExp(pattern).test(parentModelRef)) return entry;
+	}
+	return undefined;
+}
+
 function resolveFallbackModels(options: {
 	model?: string;
 	presetType?: SubagentPresetTypeOverride;

package/external/pi-tools-suite/src/async-subagents/core/routing.ts

@@ -32,7 +32,7 @@ export interface RoutedSubagentTasks {
 const ROUTER_SYSTEM_PROMPT = [
 	"You route Pi async sub-agent tasks to the best configured subagentType.",
 	"Choose exactly one allowed type for each task. Use the allowed type descriptions as the source of truth.",
-	"Prefer the most specific matching type over generic quick/deep. Use frontend for UI/UX implementation or visual frontend polish; use vision only for attached images/screenshots or image inspection.",
+	"Prefer the most specific matching type over generic quick/deep. Use frontend for UI/UX implementation or visual frontend polish. For pure image inspection, use the lookup tool rather than subagents.",
 	"Return only strict JSON with this shape: {\"routes\":[{\"id\":\"task-id\",\"subagentType\":\"type\"}]}",
 	"Do not include markdown, comments, explanations, or unknown types.",
 ].join("\n");

package/external/pi-tools-suite/src/async-subagents/core/spawn.ts

@@ -677,7 +677,7 @@ function subagentEnvironment(env: NodeJS.ProcessEnv): NodeJS.ProcessEnv {
 		PI_TERMINAL_BELL_DISABLED: "1",
 		PI_TOOLS_SUITE_DISABLED_MODULES: appendEnvList(env.PI_TOOLS_SUITE_DISABLED_MODULES, [
 			"async-subagents",
-			"glm-coding-discipline",
+			"coding-discipline",
 			"question",
 		]),
 	};

package/external/pi-tools-suite/src/async-subagents/core/types.ts

@@ -14,7 +14,7 @@ export interface AgentTask {
 	promptAppend?: string;
 	/** Full prompt replacement for this task. Supports prompt template variables. */
 	promptOverride?: string;
-	/** Optional visual attention instructions for vision-capable sub-agents. */
+	/** Optional visual attention instructions for agents that receive imagePaths. */
 	focus?: string;
 	/** Local image files to attach to the sub-agent RPC prompt. */
 	imagePaths?: string[];